MASTER THESIS

Jonáš Dujava

Strongly Coupled
Quantum Field Theory
in Anti–de Sitter Spacetime

Institute of Particle and Nuclear Physics

Supervisor of the Thesis Mgr. Petr Vaško, Ph.D.

Study Programme Theoretical Physics

Prague 2025
Revised version 2025
Dedication ii

I would like to express my deep gratitude to my advisor, Petr Vaško, for his
guidance, patience, and invaluable advice during our collaboration.
Throughout my studies at Charles University I had the pleasure to interact with
many great lecturers. In particular, I would like to thank Pavel Krtouš, whose
style and dedication to clarity left an unmistakable imprint on my work.
These years would not have been the same without my friends — Matej, Marek,
Tomáš (and their amazing girlfriends), Kubo, Jožko, Tomáš, Emil, Mišo, Adam,
Tony, Jano, Robo, among others — whose companionship, endless conversations,
and shared laughter filled them with joy and made them truly unforgettable.
I am especially grateful to my parents and family for their unwavering support,
love, and encouragement throughout my academic journey and my whole life.
Miška, thank you for standing by my side and for bringing light and encouragement
even in the most challenging moments.

Jonáš Dujava

©2025 Jonas Dujava

Information & Abstract iii

Information
Title Strongly Coupled Quantum Field Theory in Anti–de Sitter Spacetime
Author Jonáš Dujava
Institute Institute of Particle and Nuclear Physics
Supervisor Mgr. Petr Vaško, Ph.D. — Institute of Particle and Nuclear Physics
Keywords QFT in Curved Spacetime, 1/N Expansion, Conformal Field Theory,
Anti–de Sitter Spacetime, AdS/CFT Correspondence, O(N ) Model

Abstract
We introduce the framework of Quantum Field Theories in general backgrounds
through the lens of the path integral, in the formulation known as the Functorial
QFT. With the aim of studying properties of strongly coupled QFTs, we present
key concepts and techniques such as the effective action, renormalization group
flow, and the 1/N expansion.
At the fixed points of the RG flow we typically find Conformal Field Theories,
which are symmetric under local rescaling of the metric. Among other things,
CFTs play a central role in the context of QFT in Anti–de Sitter spacetime, through
what is known as the AdS/CFT correspondence.
Finally, the developed formalism is applied to analyze the O(N ) model in AdS at
finite coupling. In particular, we focus on the non-singlet sector, which requires an
understanding of crossed-channel diagram contributions to the s-channel conformal
block decomposition.

Typeset with LATEX, utilizing the TEXtured template [1]

Copyright © 2025 Jonáš Dujava
This work is licensed under CC BY 4.0. To view a copy of this license,
visit https://creativecommons.org/licenses/by/4.0/ .

Original version submitted to the Charles University — April 2025

Revised version (contains fixed typos) — June 2025
 iv

C
Contents

Information & Abstract iii

Contents iv

Introduction vi

Quick Summary viii

1 Quantum Field Theory Formalism 1

1.1 Path Integral Approach to QFT . . . . . . . . . . . . . . . . . . . . 1
Classical Field Theory 2 • Path Integral Measure 5 • Correlators 6 •
Boundary States 9 • Perturbation Theory 10 • Wick Rotation 12
1.2 Functorial Quantum Field Theory . . . . . . . . . . . . . . . . . . 13
Motivation 14 • Cobordism Category 14 • QFT as a Functor 16 • Quantum
Observables 19
1.3 Symmetries in Quantum Field Theories . . . . . . . . . . . . . . . 22
Finite Symmetry Transformations 23 • Energy–Momentum Tensor 25 •
Ward–Takahashi Identities 27 • Symmetries as Topological Operators 28
1.4 Operator Formalism . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Foliation 29 • Vacuum and Particle Representation 30 • Conjugation 32 •
Unitarity and Reflection Positivity 35

2 Beyond Basic Perturbation Theory 37

2.1 Effective Action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Different Types of Effective Actions 37 • Quantum Effective Action 38 •
1PI Effective Action 39
2.2 Renormalization Flow . . . . . . . . . . . . . . . . . . . . . . . . . 41
Effective Field Theory Picture 41 • Wilsonian Effective Action 43 • Flow
in the Space of Couplings 44 • Scale-Dependence of Theories 46 • Fixed
Points 47 • Universality 48

3 Conformal Field Theory 51

3.1 Conformal Structure Generalities . . . . . . . . . . . . . . . . . . . 52
Conformal Transformations 53 • Conformal Group 57 • Conformal Killing
Vector Fields 58 • Conformal Compactification 60
3.2 Flat-Space Conformal Structure. . . . . . . . . . . . . . . . . . . . 61
Conformal Transformations in Flat Space 61 • Conformal Algebra of Flat
Space 64 • Embedding Space 66 • Conformal Cross-Ratios 67 • Conformal
Frame 68
3.3 Primary Operators and Descendants . . . . . . . . . . . . . . . . . 69
Quantum Conformal Algebra 70 • Conformal Representations of Local
Operators 71 • Anomalous Dimensions 74 • Correlation Functions of
Primary Operators 75
Contents v

3.4 Radial Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . 78

Radial Foliation and Cylinder Picture 79 • State–Operator Correspon-
dence 82 • Implications of Unitarity 84
3.5 Operator Product Expansion . . . . . . . . . . . . . . . . . . . . . 86
Operator Algebra 86 • OPE Coefficients 87
3.6 Conformal Blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Conformal Block Decomposition 88 • Crossing Symmetry 90 • Conformal
Bootstrap 90

4 QFT in Anti–de Sitter Spacetime 91

4.1 Anti–de Sitter Spacetime. . . . . . . . . . . . . . . . . . . . . . . . 91
Maximally Symmetric Space and Solution of Einstein’s Equations 91 •
Hyperboloid/Quadric/Quasi-Sphere 92 • Coordinate Systems 94
4.2 QFT in AdS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Embedding Formalism 95 • Bulk-to-Bulk Two-Point Function/Propagator 97
• Boundary Limit 98

5 O(N ) Model in Anti–de Sitter 100

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.2 Review of O(N ) Model in AdS. . . . . . . . . . . . . . . . . . . . . 103
Generalities of O(N ) Model 104 • CFT on Boundary of AdS 108 • Utilizing
Spectral Representation 112
5.3 CFT Generalities — 4-Point Correlators . . . . . . . . . . . . . . . 113
Conformal Block and Conformal Partial Wave Decompositions 115 • CPW
Orthogonality and Completeness, 6j–Symbol 116 • Contribution of t-Channel
CBs to Anomalous Dimensions 118
5.4 Spectrum of O(N ) Model in AdS . . . . . . . . . . . . . . . . . . . 123
Decomposition into O(N ) Irreducible Representations 123 • Singlet Spec-
trum 125 • Analysis of the Singlet Sector 129 • Criticality in the Bulk 132
• Non-Singlet Spectrum 137
5.5 Analysis of the Non-Singlet Sector . . . . . . . . . . . . . . . . . . 138
Numerical Calculation of Anomalous Dimensions 139 • Twists in the Non-
Singlet Sector 140 • Dependence of Anomalous Dimensions on Coupling 141
• Dependence of Anomalous Dimensions on Trajectory Label n 143 • Large
Spin Asymptotics 144
5.6 Summary and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . 148

Conclusion 152

N Notation & Conventions 153

Miscellaneous 153 • Differential Geometry 153

References 155
 vi

I
Introduction
Quantum Field Theory (QFT) stands as one of the cornerstones of modern
theoretical physics. Its reach extends across a broad spectrum of disciplines,
including particle physics, statistical mechanics, condensed matter physics, and
even gravitational physics.
While its principles and methods are well-established and widely used, QFT still
lacks a fully rigorous and comprehensive mathematical foundation. This becomes
particularly evident when one attempts to formulate QFT in general curved
spacetimes — an essential step for connecting it with cosmology and eventually
Quantum Gravity.
Another major challenge lies in understanding strongly coupled regimes of QFTs,
where standard perturbative techniques fail. In these cases, exact results are
rare, and progress relies heavily on developing and studying special models.
Such simplified yet nontrivial examples provide vital intuition and often serve as
laboratories for testing and refining general ideas.
Throughout all of this, symmetry plays a central role. It not only guides the
formulation of physical theories, but also provides strong constraints that make
otherwise intractable problems manageable. Some symmetries can be emergent,
as is the case for second order phase transitions, where the system becomes scale
invariant. Moreover, they usually enjoy a further symmetry enhancement — they
become conformally invariant, which means we can make even position-dependent
scale transformations. Such critical point are described by Conformal Field
Theories (CFTs), which due to their enlarged symmetry group admit powerful,
non-perturbative techniques.
One of the most interesting features a physical/mathematical model or theory
can have is the presence of a duality — see Figure 1 → p.vii . Different properties of
the theory are often more naturally seen from different perspectives, which can
offer us new intuitions, insights, or computational tools. One of the most famous
examples is the AdS/CFT correspondence, which attempts to relate a Quantum
Gravity in asymptotically AdS spacetime to a CFT on its boundary.
The final part of this thesis focuses on the study of the O(N ) model in anti–
de Sitter (AdS) spacetime. As it admits a large N expansion, allowing us to
compute at finite coupling, the O(N ) model provides a fertile ground for exploring
fundamental questions in QFT. Considering it in AdS spacetime, we can also
leverage the AdS/CFT correspondence — a QFT in a fixed AdS background admits
asymptotic observables, in particular boundary correlation functions, which satisfy
essentially all CFT axioms. This leads to a rich interplay between standard QFT
and CFT techniques, and allows us to study finite spectrum deformations of
interacting CFTs.
Introduction vii

Theory 1

Underlying ?
duality Model
Structure?

Theory 2
©2024 Jonas Dujava

Figure 1 / A model (or its part) can sometimes be described by two distinct theories,
with different starting principles and associated methods/techniques. Such duality
hints at a possible underlying structure.

Remark 2 (On Exposition). The subject of QFTs can be hard to approach, not
least due to the plethora of concepts and techniques one needs to know intimately
before conceptual clarity can eventually be achieved. One of the objectives of this
work is to provide — as much as possible — a self-contained account of various
important topics that lie slightly beyond the standard curriculum of QFT.
While not all of the discussed aspects are used directly in the “application part”
of this work — namely Chapter 5 → p.100 — they serve to highlight and appreciate
the interconnections between various notions and methods, which might otherwise
seem unmotivated or ad hoc. ⌟
Remark 3 (On Rigor). The author’s intention was to produce a clearly structured
text, making it easy to locate and refer to almost any introduced concept. To this
end, a “definition–theorem–proof–remark” structure was adopted, even though
the work should not be considered fully mathematically rigorous.
As a byproduct, the structure makes it rather transparent which parts are more
established and which remain more heuristic — whether due to gaps in the author’s
understanding, the absence of a rigorous mathematical underpinning, or simply
because the necessary background would be too extensive to include.
This is considered a feature, since it avoids (at least partially) the pitfall of giving
a false impression of understanding. A related and interesting read about the
interaction between theoretical physics and mathematics can be found in [2]. ⌟
Remark 4 (On Numbering). To simplify the navigation, all types of document
constructs — for example Definitions, Remarks, Figures, Theorems, and so on —
are numbered within the same counter (inside the given chapter). To give you an
idea, Figure 3.18 → p.57 is shortly after Example 3.15 → p.56 . ⌟
 viii

Q
Quick Summary

In Chapter 1 we briefly introduce the QFT framework, mainly through the lens
of path integral. Its formalization is dubbed as the Functorial QFT (FQFT).
In Chapter 2 we look at methods important to understand QFT at large — the
notion of effective actions, and the renormalization group (RG) flow.
In Chapter 3 we dive deep into the realm of Conformal Field Theories (CFTs),
which usually appear at the fixed points of the RG flow. Apart from the scaling
symmetry, they enjoy further symmetry enhancement to the conformal group.
In Chapter 4 we discuss QFT on Anti–de Sitter spacetime (AdS), and the
connection to CFTs through the AdS/CFT correspondence.
In Chapter 5 we apply previously developed formalism to study the O(N ) model
in AdS at finite coupling, in particular its non-singlet spectrum.
We summarize the overarching narrative in Conclusion .
In Appendix N we collect some of the conventions and notation used throughout
the work, mainly of various differential geometry objects.
 1

1
Quantum Field Theory Formalism
We start by introducing some general aspects of QFT formalism, and in particular
its formulation in general backgrounds. The most straightforward motivation is
the fact that our Universe is not flat, so understanding effects of the spacetime
curvature is required to study physics at cosmological scales, or in the presence of
strong gravitational fields, for example near black holes.

Treating the spacetime as a classical static background may be viewed as an

intermediate step towards Quantum Gravity. A proper understanding of all
implications in such semi-classical approximation may contain valuable hints for
the full theory.

Understanding QFT in general backgrounds can also bring other benefits, both of
conceptual and computational nature. It allows us to consider models on manifolds
with boundary or non-trivial topology, and enlightens the fundamental role of
spacetime symmetries. Having such general formulation we can make statements
about (and relate) theories in different backgrounds. In particular, we can get a
better feeling for the connection between Lorentzian and Euclidean QFTs, the
former describing relativistic particle or condensed matter systems, and the latter
describing statistical systems (at large scales).

We will take the (heuristic) path integral approach as our starting point, and
briefly introduce it in Section 1.1 → p.5 . Then we formalize the properties we usually
require from the path integral into an axiomatic framework in Section 1.2 → p.13 .
Symmetries in context of QFTs are treated in Section 1.3 → p.22 . As discussed in
Section 1.4 → p.29 , they turn out to be central for building the usual “operator
formalism” obtained by appropriately foliating the spacetime.

1.1 Path Integral Approach to QFT

A classical field theory describes a classical theory with infinite degrees of freedom.
It is usually defined by a local action functional, which uniquely determines the
evolution of the fields given the initial conditions.

Going to the quantum realm, that is considering QFT, the possibilities are much
more diverse. Intuitively, the Feynman’s “sum over histories” motto [3] says that
we should integrate over all possible (spacetime) configurations of the field, with
an appropriate measure depending on the action functional.

Performing such “path integral” exactly in a nontrivial interacting theory is

unfeasible, so we often fall back to the perturbation theory. Various features
are sometimes clearer by doing calculations in the Euclidean signature, and then
analytically continuing the results back to the original Lorentzian case.
Chapter 1. Quantum Field Theory Formalism 2

This section is inspired by [4] [5] [6] [7] [8], where you can find more details.
Remark 1.1 (Attitude Towards the Path Integral). We view path integral as a
powerful heuristic tool, which can be made mathematically rigorous only in some
cases — in d = 0 it is just an ordinary (multidimensional) integral, in higher
dimensions we obtain infinite-dimensional integration. We will not delve deeper
concerning (possibly) rigorous definitions of path integral, but rather focus on the
physical intuition behind the path integral. ⌟

Classical Field Theory

Classical field theory is usually formulated by dynamical fields living on the
spacetime manifold, whose dynamics is prescribed by the action functional. We
will review the usual definitions to prepare for following sections, mainly to
fix notations and consolidate understanding of the initial setup. For a more
mathematically thorough account, see for example [6].

Definition 1.2 (Spacetime Manifold, Background Fields). The field theory is

formulated on a smooth manifold M, which we call the spacetime. We allow
a possibility of M having a boundary, or being endowed with an additional
structure such as orientation, metric, or spinor structure.
Such extra structure is often understood as being part of the non-dynamical
background fields, which also includes other background sources.

Remark 1.3. We usually assume orientability, and a metric usually of Lorentzian

or Euclidean signature. Sometimes we explicitly write such structure as (M, g)
for a pseudo-Riemannian manifold, or we implicitly include it in M. ⌟
Example 1.4. For simplicity, most of the time we work with spacetimes enjoying
maximal symmetry, examples being (with dimension D ≡ dim M):
• Lorentzian signature: Minkowski flat space MinkD ≡ R1,D−1 , de Sitter dSD ,
Anti–de Sitter AdSD ,
• Euclidean signature: Euclidean flat space RD , sphere SD , hyperboloid HD .
For Classical Mechanics (CM) or Quantum Mechanics (QM) we have D = 1, that
is Mink1 ≃ R1 — there is just the (affine) time. ⌟

Definition 1.5 (Dynamical Fields, Field Configurations). The space of dynamical

fields (field configurations) on a manifold M is denoted by Fields M . They
are usually (smooth) sections of a given field fiber bundle E → M. We will
collectively denote them by ϕ ∈ Fields M .

Example 1.6 (σ-Model, Real Scalar Field). A so-called σ-model field is given by
a map ϕ : M → X for some auxiliary target manifold X. For a real scalar field
we simply have E = M × X with X = R, whose sections are indeed functions
Chapter 1. Quantum Field Theory Formalism 3

C ∞ (M). An N -tuple of real scalar fields is obtained by taking X = RN . ⌟

Remark 1.7 (Type of Field). Typically, a field theory contains several fields
×
ϕ ≡ {ϕi } with the corresponding field bundle E = Ei being a fiber product.
Roughly speaking each Ei decomposes into an intrinsic part times and extrinsic
part. The extrinsic part is associated to the target values of the field, and the
intrinsic part is associated to the principal frame bundle of M via a representation
of GL (or Spin). This representation determines the type of the field, and in partic-
ular the transformation of its components under the change of frame/coordinates.
For example, a trivial representation gives us a scalar field, a vector/spinor
representation gives us a vector/spinor field, and so on. ⌟

Example 1.8 (Gauge Theories). In gauge theories, the basic field we consider is a
connection on a principal G-bundle P → M for some specified Lie group G. The
matter fields are then additionally specified by a choice of representation under
the gauge group G — they are sections of some associated bundle. ⌟

Definition 1.9 (Action, Lagrangian Density). An action SM is a real functional

of field configurations, that is SM : Fields M → R. The locality of the theory
is expressed by the form of the action
∫︂ ∫︂
SM [ϕ] = L[ϕ] = L(ϕ, ∇ϕ, . . .)
M M

as a spacetime integral of a local functional L called Lagrangian density,

which at a given point m ∈ M depends only on a finite-order jet of ϕ at m
(usually 1-jet) — just the values and derivatives of ϕ at m. The dependence
on background fields (such as the metric) is kept implicit.

Remark 1.10 (Lagrangian). We often factor out the metric density as L ≡ g /2 L,

1

and work with the “Lagrangian” L. ⌟

Remark 1.11. Strictly speaking, when M is not compact, SM will usually diverge.
We should then work on some domain Ω ⊂ M. We will ignore this subtlety and
often just write S instead of SM , keeping the domain/manifold implicit. ⌟

Remark 1.12 (Covariant Lagrangian Description). Not all field theories admit a
Lagrangian description, but we will assume it for the time being. Moreover, we
assume a possibility of covariant formulation in a general background, that is for
a diffeomorphism f : M → M′ we have

S(M,g) [ϕ] = S(M′ ,f∗ g) [f∗ ϕ] ,

where g labels all background fields, and f∗ ≡ (f −1 )∗ is the pushforward by f . ⌟

Example 1.13 (Action of Free Scalar Field). A free real scalar field ϕ has a quadratic
Chapter 1. Quantum Field Theory Formalism 4

action (assuming Lorentzian signature)

∫︂ (︂ )︂
SM [ϕ] =
1
g /2 − 12 ∇a ϕ∇a ϕ − 21 m2 ϕ2 − 12 ξ Rϕ2 ,
M

where the m ≥ 0 is the “mass”, and ξ controls the coupling to the spacetime
curvature through the Ricci scalar R. Everything is written in terms of “geometric”
objects, so the covariance Remark 1.12 → p.3 is manifest. Since ϕ is a scalar, we
could also write ∇ϕ = ∂ϕ ≡ dϕ, as it does not depend on the choice of metric
connection. ⌟

Remark 1.14 (Condensed DeWitt Notation). To simplify the expressions, we will

sometimes use the condensed “DeWitt notation”, For example, the action of a
free scalar field in Example 1.13 → p.3 generalized to multiple flavors ϕ = {ϕ• } can
be written as (after integrating by parts once and neglecting the boundary term)

SM [ϕ] = − 12 ϕ • F • ϕ , with F ≡ (− + m2 + ξ R)1 ,

(︂ )︂
where ϕ ≡ ∇2 ϕ ≡ g ab ∇a∇b ϕ = g− /2 ∂a g /2 g ab ∂b ϕ is the d’Alembertian.
1 1

The dot “ • ” denotes the contraction of both the “spacetime index” — representing
integration over the manifold (when needed we include g /2 ) — and also possible
1

flavor indices and/or field components. The differential operator F is understood

to act on functions through the convolution with the corresponding kernel
[︂ ]︂
F(x, y) ≡ (− + m2 + ξ R)1 (x, y) ≡ (− x + m2 + ξ R)δ(x, y) ,

where the δ-distribution is appropriately normalized such that the identity operator
1 really acts on functions as an identity. ⌟

Remark 1.15 (Conformal Coupling). If we have a massless scalar field m2 = 0, the

choice of ξ = 14 D−2
D−1
(where dim M ≡ D) is called the conformal coupling — we
say that we have a theory of a conformally coupled scalar field.
This refers to the fact that such theory is invariant with respect to Weyl transfor-
mations (see Definition 3.3 → p.53 ) of the form
D−2
g ↦→ Ω2 g , ϕ ↦→ Ω− 2 ϕ.

Invariance of SM under such transformation can be verified by utilizing an

appropriate formula for the transformation of R, or by computing the energy–
momentum tensor Definition 1.74 → p.25

Tab = ∇a ϕ∇b ϕ − 12 gab (∇• ϕ)2 + ξ(Gab + gab − ∇a∇b )ϕ2 ,

which can be checked to satisfy Taa ∼ 0 for the chosen value of ξ.

EL
⌟
Chapter 1. Quantum Field Theory Formalism 5

Definition 1.16 (Dynamics, Euler–Lagrange Equations Of Motion, On-Shell).

Classical fields are governed by the Euler–Lagrange equations of motion given
by the variational principle, which requires the stationarity of the action —
!
δSM = 0 — with some appropriate boundary conditions on ∂M.
When we require fields to satisfy the Euler–Lagrange equations, we say that
we are on-shell. We write A ∼ 0 for a quantity A which vanishes on-shell.
EL

Remark 1.17 (Variational Bicomplex). The formalism is naturally expressed using

the variational bicomplex Ω•,•
loc (M × Fields M ). More details can be found in [6].⌟

Example 1.18 (Euler–Lagrange EOM of Free Scalar Field). Using the symmetry of
F in Remark 1.14 → p.4 (ignoring boundary terms), we have

! δϕ arbitrary
(︂ )︂
δSM = −δϕ • F • ϕ = 0 =======⇒ F •ϕ≡ − + m2 + ξR ϕ = 0 .
⌟

Path Integral Measure

In path integral approach to quantization, we imagine all possible (off-shell)
evolutions of fields (with given boundary conditions), and sum/integrate over
them with a complex functional “measure” written formally as
∫︂
ı̊
Dϕ e ℏ S[ϕ] .
Fields

The classical field theory is then recovered in the “classical limit” ℏ → 0, since
by a stationary-phase/saddle-point argument, the path-integral contributions
concentrate around the classical solution ϕcl of the Euler–Lagrange equations
Definition 1.16 → p.5 , which are precisely the stationary condition of S.
Before introducing some fundamentals of path integral in QFT, let us make some
remark about the “path-integral measure” itself.
Remark 1.19 (Path Integral in QM). In Quantum Mechanics (QM), the path
integral can be made rather rigorous by using the theory of Gaussian measures
on Abstract Wiener spaces. The idea is to work in Euclidean signature — to
be discussed in Wick Rotation → p.12 — and discretize the time interval. We then
define cylindrical functions depending only on finite number N of points/times,
and measure on them (just a finite-dimensional Lebesgue measure). Finally, we
take the continuum limit N → ∞, which however is less straightforward than one
might initially think.
1
The reason is that neither Dϕ , nor e− ℏ S[ϕ] , exist in the continuum limit. Concern-
ing the measure, see next Remark 1.20 → p.6 . Concerning the action, we can clearly
see that continuum limit of cylindrical function need not be smooth/differentiable,
and thus the action is not well-defined.
It actually turns out that differentiable functions form a measure-zero set, and
Chapter 1. Quantum Field Theory Formalism 6

most important are nowhere differentiable continuous functions. This is closely

tied to the non-commutativity of QM, since otherwise commutators would be
trivial. So in the end we have some measure on continuous functions, even though
we naively started with differentiable Fields. ⌟

Remark 1.20 (No Lebesgue Measure in Infinite-Dimensional Spaces). There is no

“Lebesgue measure” in infinite-dimensional Banach spaces, by which we mean it
has the following properties — every point has a neighborhood of positive measure,
every non-empty open set has positive finite measure, and it is translationally
invariant. This is because we can fit infinitely many disjoint balls of some
fixed radius around any point — each having the same positive measure by
translation invariance — which is in contradiction with finite measure of the point
neighborhood. ⌟

Example 1.21 (Gaussian Measures). One can however define Gaussian measures
on infinite-dimensional spaces, which avoids the problem in Remark 1.20 → p.6 by
throwing away the translation invariance requirement.
These correspond to free theories, and we can formally work with them as if
1
we could write Dµ(ϕ) ≃ Dϕ e− 2 ϕ•F • ϕ , where we take Dϕ to be translationally
invariant and satisfy the usual substitution rules, just written in the functional
language. ⌟

Remark 1.22 (Interacting Measures). In the case of interacting theories, the precise
mathematical sense of the path integral is less clear. Two immediate approaches
are:
• Lattice Approach — We discretize the spacetime manifold by replacing it
with a lattice. Fields are then defined by values on the nodes of the lattice,
and the action by some discretized version. If the manifold is compact, the
discretized version has a finite number of nodes, and path integral becomes
a finite-dimensional integral (but we manifestly violate Lorentz symmetry).
Everything is then well-defined, and the only problem is whether appropriate
continuum limit exists (in which we hope to recover Lorentz symmetry).
• Perturbative Approach — For interacting theory “close” to a free theory, we
can view the measure as a perturbative deformation of the Gaussian measure.
We will briefly discuss such approach below in Perturbation Theory → p.10 .
Let us quickly note that it requires some regularization and renormalization
scheme, which is usually chosen to respect as many symmetries of the theory
as possible, but generally not all classical symmetries can be preserved.
⌟

Correlators
Now we will present basics of the path integral formalism, with examples being
provided by the free scalar theory.
Chapter 1. Quantum Field Theory Formalism 7

Definition 1.23 (Field, Local Operator, Observable). A (complex topological)

vector space Ox M of observables at x ∈ M is the space of smooth functions
f : Fields M → C such that f [ϕ] depends only on the restriction of ϕ ∈ Fields M
to an arbitrarily small neighborhood of x.
Spaces Ox M ≃ V are isomorphic to some standard fiber of local/point
observables V , and fit together to form a vector bundle OM over M, which
is the space of local operators/fields of the theory.

Example 1.24 (Fundamental, Elementary, Composite Operator). A field/operator O

directly integrated over in the path integral is called fundamental or elementary,
example being O = ϕ for a scalar field ϕ. Local functions of (multiple) fundamental
fields are called composite, example being O ∼ ϕ p with p ≥ 2, where p = 2
corresponds to the mass term. ⌟
Remark 1.25 (Nonlocal, Extended Operator). There are also important nonlocal
operators, which are not localized at a single point, but for example on some
surface — symmetry charges as spatial integrals of conserved currents — or a
loop — Wilson loop as a holonomy of the gauge field around a closed loop. ⌟
In path integral approach, correlators are the fundamental objects in QFT.

Definition 1.26 (Correlation Function, Correlator). The correlation function or

simply correlator of an operator O : Fields → C is defined as the functional
average over the space of field configurations
1 ∫︂ ı̊
⟨O⟩ ≡ Dϕ e ℏ S[ϕ] O[ϕ] ,
Z
Fields

where Z ≡ ⟨1⟩ is the normalization. We often suppress writing Fields.

Remark 1.27. A lot of information about a QFT can be extracted from such
correlators, for example flat-space scattering amplitudes are obtained via LSZ
reduction formulas by looking at their poles in momentum representation.
Alternatively, in Euclidean QFTs describing continuum limit of statistical systems,
one can read out correlation lengths or other measurable coefficients from their
(asymptotic) behavior. ⌟
All correlators are compactly encoded in the following object of central importance.

Definition 1.28 (Partition Function, Background Sources). The partition func-

tion depending on background sources J is defined as
∫︂
ı̊ ı̊
⟨︂ ⟩︂
Z[J ] ≡ e ℏ J • ϕ ≡ Dϕ e ℏ (S[ϕ]+J • ϕ) .

Sometimes we implicitly include the source term J • ϕ in the action, and we

can then write Z[J ] = ⟨1⟩J .
Chapter 1. Quantum Field Theory Formalism 8

Remark 1.29 (Normalization of Partition Function). When working in a fixed

background, we usually normalize the partition function such that Z[0] = 1. ⌟

Example 1.30 (Partition Function for Free Scalar Field). Completing the square in
the exponent and assuming normalization as described above, we have

Dϕ e ℏ (− 2 ϕ F )= Dϕ e ℏ (− 2 (ϕ−F )
∫︂ ∫︂
ı̊ 1 ı̊ 1 −1 •J) • F • (ϕ−F −1 •J)+ 1 J •F −1 •J
Z[J] =
• • ϕ+J • ϕ
2

Dϕ e (− 2 ϕ F ϕ ) e ℏ 2 J •F •J ≡ e( ℏ ) 2 J •G•J
(︃∫︂ )︃
1 ′• • ′ ı̊ ı̊ 1 −1 ı̊ 2 1
= ′ ℏ

We used the translation invariance of the measure Dϕ to shift the integration

variable ϕ ↦→ ϕ′ ≡ ϕ − F −1 • J, after which the J-independent functional determi-
nant obtained by Gaussian/Fresnel integration cancels with the normalization of
the measure.
As we can see, such Gaussian integration is equivalent to evaluating the exponent
at the solution of classical Euler–Lagrange EOM

−F • ϕ + J = 0 ⇐⇒ ϕJ = F −1 • J ≡ ℏı̊ G • J .
⌟

Remark 1.31 (Green’s Functions). The Green’s function satisfying the equation

!
−F • G =ı̊ℏ1

is not uniquely defined in the Lorentzian signature, and one needs to provide
appropriate “ı̊ε prescription”. We will discuss this later in Wick Rotation → p.12 .⌟

The partition function Z[J ] is a generating functional for all correlators of local
operators, in the following sense.

Proposition 1.32 (Partition Function as Generating Functional). Correlators of

local functionals, which are monomials of fundamental fields ϕ, are given by
appropriate functional derivatives of the partition function with respect to
the sources J. In particular, we have
(︄ )︄n
1 δ δ
⃓
⟨︂ ⟩︂ ℏ
ϕ(x1 ) . . . ϕ(xn ) = Z[J ] ⃓⃓
⃓
··· .
ı̊ Z[J ] δJ (x1 ) δJ (xn ) J =0

Proof. Directly follows from δ

δJ (x)
(J • ϕ) = δx • ϕ = ϕ(x) and

ℏ δ ℏ δ ⟨︂ ı̊ J • ϕ ⟩︂ ⟨︂ ı̊
⟩︂
Z[J ] ≡ eℏ = ϕ(x) e ℏ J • ϕ .
ı̊ δJ (x) ı̊ δJ (x)
Chapter 1. Quantum Field Theory Formalism 9

Example 1.33 (Wick’s theorem). Using Proposition 1.32 → p.8 in the case of free
scalar field Example 1.30 → p.8 , we obtain

0 for n odd ,
⎧
⎪
⎪
⎪
G(x1 , x2 ) for n = 2 ,
⎪
⎪
⟨︂ ⟩︂ ⎪
⎨
ϕ(x1 ) . . . ϕ(xn ) = ⎪ ∑︂
G(xi+ , xi− ) for n even ,
∏︂
⎪⎪
⎪
⎩ pairings (i+ ,i− )
⎪
⎪
{(i+ ,i− )}

since only when all δJ(x

δ δ
• ) δJ(x• )
act in pairs on 12 J • G • J, we can obtain nonzero
contribution after finally evaluating at J = 0. ⌟

Boundary States
Up to now we always imagined performing the path integral over all field con-
figurations in the whole spacetime, and did not consider any boundaries. In
their presence, we need to specify the boundary conditions on the fields, which
usually means fixing values of the fields on the boundary, see Figure 1.34 → p.9 for
visualization.

Such setting is well known from QM, where path integral with constrained positions
at the ends gives us the kernel of the evolution operator in the position basis.

Similarly in QFT, constraining the field values on the boundary of the region, the
path integral gives us the kernel of the evolution operator in the field bases of the
Hilbert spaces of states associated to the boundaries. We will not delve into details
of this here, since everything will be properly formalized in Section 1.2 → p.13 .

ϕ|Σ

ϕb

∫︂
ı̊
Dϕ e ℏ S[ϕ]
Σ
ϕ|Σ =ϕb ©2025 Jonas Dujava

Figure 1.34 / Visualization of path integral with fixed boundary

values ϕb on the boundary Σ. Only the field fluctuations are
integrated over, the underlying structure of manifold is fixed.
Chapter 1. Quantum Field Theory Formalism 10

Perturbation Theory
Exact evaluation of the path integral is usually not feasible. Among few exceptions
a special role is played by free theories, which have quadratic actions, leading to
a Gaussian integration. In general, including any non-trivial interactions makes
the explicit evaluation intractable. The most common way to proceed, applicable
for weak couplings, is to employ perturbation theory around a free theory.
This amounts to splitting the action into free (quadratic) and interaction part,
and then expanding the interaction part into series in the coupling.
Example 1.35 (Interacting Scalar Field, Feynman Diagrams). For simplicity, consider
an action symbolically of the form (integrations are done over x ∈ M with
appropriate metric density)

S[ϕ] = Sfree [ϕ] + Sint [ϕ]

∫︂
≡ − 21 ϕ • F ϕ−λ
• V (x) ,
x

where λ is the coupling assumed to be small, and V is some simple polynomial in

ϕ, for example V (x) ∼ ϕ4 (x). Then the partition function or generating functional
can be formally written as

Dϕ e ℏ (− 2 ϕ F ϕ−λ V (x)+J • ϕ)
∫︂ ∫︁
ı̊ 1 • •
Z[J] ≃ x

⟨︂ ı̊λ
∫︁ ı̊
⟩︂ ∞
1 (︂ −ı̊λ )︂n ∫︂ ⟨︂ ı̊ •
⟩︂
= e− ℏ V (x)
e ℏ J •ϕ = (x ) (x ) e J ϕ
∑︂
x V 1 · · · V n ℏ .
n=0 n!
free ℏ free
x1 ,...,xn

To compute arbitrary correlator, we take functional derivatives with respect to

the source J as in Proposition 1.32 → p.8 , which brings down additional factors ϕ
into the free correlator. These are then evaluated using Example 1.33 → p.9 , just
by pairing up all the ϕs with the free propagators G.
There is a well known diagrammatical technique — called Feynman diagrams —
for organizing such an expansion. While more detail can be found in any QFT
textbook, let us briefly summarize the main points.
To calculate a correlator ⟨ϕ(y1 ) · · · ϕ(yn )⟩, we draw external (degree one) vertices
for each ϕ(yi ), and then construct all diagrams by additionally inserting any number
of internal vertices (with degree corresponding to the power of the polynomial V )
and connecting them with simple edges/lines.
All such diagrams Γ are summed up with the corresponding weights w[Γ] — each
insertion of internal vertex carries a factor of −ı̊λ
ℏ
, each edge between vertices at
x+ and x− corresponds to the free propagator G(x+ , x− ), and all internal vertices
are integrated over the spacetime manifold. One just needs to take care about
combinatorial factors, which results in
⟨︂ 1 ∫︂ ⟩︂ ı̊ 1
ϕ(y1 ) · · · ϕ(yn ) ≃ Dϕ e ℏ S[ϕ] ϕ(y1 ) · · · ϕ(yn ) ≡
∑︂
w[Γ] ,
Z pert Γ |Aut(Γ)|
Chapter 1. Quantum Field Theory Formalism 11

where |Aut(Γ)| is the symmetry factor of the diagram Γ. Here Γ is any diagram
as described above, but we exclude vacuum bubbles — all interaction vertices are
just connected to each other, and not to any external vertex — since they factor
out and are cancelled by the normalization Z. ⌟

Remark 1.36 (Asymptotic Series). Such perturbation expansions are typically

not convergent for any finite value of the coupling constant λ, so should not be
viewed as Taylor series. However, the formal manipulations are still justified in
the sense of asymptotic series, which means that truncating it at any finite order
gives increasingly good approximations when taking the limit λ → 0. ⌟

Remark 1.37 (Regularization, Counterterms, Renormalization). In fact, the previous

discussion swept under the rug the issue of divergences, which come from the
coincidence limit of points during the integration over the internal vertices.
We should somehow regularize the integrals, and add appropriate counterterms in
the interaction part such that in the end (after removing the regularization) we
obtain finite results. This goes under the name of renormalization, and we will
discuss it from slightly different perspective in Section 2.2 → p.41 . ⌟

To calculate the partition function Z[J ] in the perturbation theory, we draw all
possible diagrams discussed above, but now external vertices come from expanding
ı̊
the exponential e ℏ J • ϕ , so they have an extra factor of ℏı̊ J associated with them
(and we integrate over them). We can simplify the diagrammatic expansion only
to the connected diagrams — the motto is that exponential map generates all
graphs/contributions from just connected graphs/contributions.

Proposition 1.38 (Exponential and Connected Graphs). Perturbative expansion

of the partition function Z[J ] can be expressed as
⎛ ⎞
1 1
w[γ]⎠ ∼
ı̊
Z[J ] ≃ w[Γ] = exp⎝ = e ℏ W[J ] ,
∑︂ ∑︂

Γ |Aut(Γ)| γ connected |Aut(γ)|

where W[J ] is the generating functional for the connected correlators.

Proof. We can decompose any graph Γ = γ1⊔r1 ⊔ · · · ⊔ γk⊔rk into its connected
components, γ• being pairwise non-isomorphic connected graphs. Then we have
for the symmetry factor |Aut(Γ)| = i ri !|Aut(γi )|ri . Using the multiplicativity of
∏︁

the weights w[Γ1 ⊔ Γ2 ] = w[Γ1 ]w[Γ2 ], we can write

⎛ ⎞
1 ∞
1
exp⎝ w[γ]⎠ = r
∑︂ ∏︂ ∑︂
r w[γ]
γ connected |Aut(γ)| γ connected r=0 r!|Aut(γ)|
∞
1 1
= w[γi ]ri =
∑︂ ∏︂ ∑︂
w[Γ] .
i ri !|Aut(γi )|
ri
Γ |Aut(Γ)|
∏︁
r=0 i
Chapter 1. Quantum Field Theory Formalism 12

Wick Rotation
As already noted in Remark 1.31 → p.8 , the inverse of the differential operator F
necessary for computing the functional Gaussian/Fresnel integrals is not unique
in the Lorentzian signature.
Remark 1.39 (Fresnel and Gauss Integrals). What the appropriate choice should
be can be understood by considering one-dimensional Fresnel integral of the form
∫︂
2 √ ı̊π
dx eı̊x = πe4 .
R

The integrand is highly oscillatory, and the integral is not absolutely convergent.
Still, it can be understood as a boundary value of the absolutely convergent Gauss
integral (taking α ∈ C with Re α > 0)
π
∫︂ √︃
−αx2
dx e = ,
R α

in particular taking the limit α → −ı̊ gives us the above result. ⌟

We thus infer that the path integral in Lorentzian signature should be understood
as a boundary value of an appropriate analytical continuation from the Euclidean
signature, where the propagator is unique. In this way we recover the usual
Feynman propagator corresponding to correlators understood as time-ordered
vacuum expectation values. We will see this explicitly in Section 1.2 → p.13 and
Section 1.4 → p.29 .
Example 1.40 (Wick Rotation in Flat Spacetime, Euclidean QFT). In flat spacetime,
the so-called Wick rotation to the Euclidean signature is obtained by setting the
time coordinate to be imaginary
t ≡ −ı̊τ , τ ∈ R.
where τ is the Euclidean time. This can be seen as taking t = e−ı̊δ τ for δ ∈ [0, π2 ]
and τ ∈ R, where δ = 0 corresponds to the Lorentzian signature, and δ = π2 to
the Euclidean signature. For the action we have
∫︂ ∫︂ ⃓ ∫︂
S= dt d
d−1
x L = −ı̊ dτ d d−1
dτ dd−1 x EL ≡ı̊ ES ,
L L L
⃓
x L ⃓⃓ t=−ı̊τ ≡ı̊
∂t =ı̊∂τ

so under the path integral we obtain

ı̊ L 1E
Dϕ e ℏ S ↦−→ Dϕ e− ℏ S .
For example, the Euclidean action of a free scalar field is
∫︂ (︂ )︂
E
S[ϕ] = dτ dd−1 x 1
2
(∂ϕ)2 + 12 m2 ϕ2 ,

where contractions are done using Eη ≡ δ, so ES is positive-definite functional,

and actually it is the energy of the field configuration. We can thus conclude that
a flat-space Lorentzian QFT is the Wick rotation of a Euclidean QFT describing a
(classical) statistical system at “inverse temperature” β = ℏ1 . ⌟
Chapter 1. Quantum Field Theory Formalism 13

We can slightly generalize this for arbitrary static spacetimes.

Example 1.41 (Wick Rotation in Static Spacetime). Take a static Lorentzian
spacetime of the form LM = R × Σ with metric
L
g = − dt2 + gΣ .

We can consider its complexification MC = C × Σ with the metric as above, but

now we take t ∈ C. The real slice t ∈ R ⊂ C of MC is the original Lorentzian LM,
and taking the imaginary slice t ∈ı̊R ⊂ C ⇔ τ ≡ı̊t ∈ R we obtain EM which is
topologically the same as LM, but with the Euclidean metric
E
g = dτ 2 + gΣ .
⌟
Remark 1.42. In the above example we complexified the manifold and identified
its slices as our Lorentzian and Euclidean manifolds. One drawback of such
approach is that such embedding into a complexified manifold is not generally
available — however, when possible, it is a crucial tool. Alternative approach is
to leave the spacetime fixed, and complexify the metric. See [9] for an attempt
in the FQFT setting. ⌟

1.2 Functorial Quantum Field Theory

We will now formalize some of the ideas concerning the path integral we have
presented and heuristically motivated in the previous sections. In the context
of Functorial Quantum Field Theory (FQFT), we will take these properties as
axioms of our QFT.
This will elucidate the structure behind QFTs, and provide us a more general
framework to work with. Functorial viewpoint is especially useful for formulating
different types of QFTs on manifolds with various structures, and for conceptual
understanding of various “operator formalisms”/“operator quantizations” to be
discussed in Section 1.4 → p.29 .
Exposition of this chapter is mainly based on [4] [10] [11] [12] [9] [13].
Remark 1.43. FQFT has beginnings in the attempts of Greame Segal and Michael
Atiyah to axiomatize CFTs [14] and TQFTs [15], and is most natural in that
setting. Following should not be considered as a rigorous definition, but rather as
a (sub)set of guiding principles and properties we consider when we talk about
QFTs and work with ideas coming from the path integral viewpoint. ⌟
Remark 1.44 (Other Axiomatic Frameworks). Apart from the FQFT approach,
there are mainly two other types of axiomatic frameworks:
• Correlator-focused approaches (Wightman axioms, Osterwalder–Schrader)
— various (related) axiomatic systems specifying properties of correlation
functions of local operators acting on a Hilbert space,
Chapter 1. Quantum Field Theory Formalism 14

• Algebraic Quantum Field Theory — emphasizes algebras of observables

associated to local spacetime regions, with Hilbert spaces of states being a
derived notion.
For more details, see the reviews [16] [17]. ⌟

Motivation
One key notion we would like to formalize is the locality of the path integral.
Idea 1.45 (Cutting and Gluing). Assuming a local Lagrangian theory, we can cut
the manifold into pieces, compute the path integral individually on each piece
(holding fixed boundary conditions), and then glue/sew them back together by
finally integrating over the boundary configurations. In formula, this reads for
M = M1 ∪Σ M2 as (ϕ1,2 are restrictions of ϕ to M1,2 , respectively)
∫︂ ∫︂ ∫︂ ∫︂
ı̊ ı̊ ı̊
Dϕ e ℏ SM [ϕ] = Dϕb Dϕ1 e ℏ SM1 [ϕ1 ] Dϕ2 e ℏ SM2 [ϕ2 ] ,
Fields M Fields Σ Fields M1 ,ϕ1 |Σ =ϕb Fields M2 ,ϕ2 |Σ =ϕb

where we used additivity of the action SM [ϕ] = SM1 [ϕ1 ] + SM2 [ϕ2 ].
In the well understood (1 + 0)–dimensional case — just QM — the locality of the
time evolution is represented by the functoriality of going from time intervals to
the corresponding evolution operators
QM
[t1 , t3 ] = [t2 , t3 ] ∪ [t1 , t2 ] U (t3 , t1 ) = U (t3 , t2 ) ◦ U (t2 , t1 ) .

Note that convention for writing intervals with t3 > t2 > t1 is opposite to how we
write and compose evolution operators, so we will sometimes use [t2 |t1 ] ≡ [t1 , t2 ]⌟
We will generalize this idea to more general setting in higher dimensions.

Cobordism Category
Oftentimes we imagine cutting/foliating the spacetime into pieces, and then study
the dynamics in-between the cuts/leaves, where we appropriately specify some
boundary conditions. Prototypical example is the evolution between chosen time-
slices in a given coordinate system. To allow for more complicated situations, let
us slightly formalize this idea.

Definition 1.46 (Cobordism). A cobordism Σin M Σout is an oriented compact

(d + 1)–dimensional manifold M with boundary, together with smooth maps
ιin : Σin ↪→ M ←↩ Σout : ιout with image in ∂M such that
∼
(−ιin ) ⊔ ιout : (−Σin ) ⊔ Σout −
→ ∂M
is an orientation-preserving diffeomorphism. The −Σin denotes Σin with the
opposite orientation, and the above means we have the decomposition of the
boundary ∂M = (−Σin ) ⊔ Σout in disjoint (out
in )–components.
Chapter 1. Quantum Field Theory Formalism 15

Remark 1.47 (Direction). Depending on the context, it may be more natural to

write Σout M Σin instead of Σin M Σout . ⌟

Remark 1.48 (Additional Geometric Structure). We will often consider cobordisms

with additional geometric structure, such as a metric or spin structure. Then
both the cobordism and boundaries should be equipped with the same geometric
structure — more precisely, a boundary Σ• should be a germ Σ• × (−ε, ε) of
a (d + 1)-dimensional manifold with the given geometric structure — and the
inclusion maps ι• should be structure-preserving maps. ⌟

Remark 1.49 (Gluing/Sewing of Cobordisms, Cobordism Category). When out-

boundary of one cobordism matches the in-boundary of another cobordism, we
can glue/sew them together to obtain a new cobordism.
In equations, for cobordisms Σ2 M2 Σg and Σg M1 Σ1 , their gluing along Σg
M ≡ M2 ∪Σg M1
is given by the cobordism Σ2 Σ1 ≡ Σ2 M2 Σg M1 Σ1 . This is
illustrated in Figure 1.50 → p.15 .
We have thus the geometric cobordism category Cob, whose objects are closed
oriented d–manifolds and morphisms are (d + 1)–dimensional cobordisms, where
everything can be enriched with additional “geometric” structure. For nontrivial
geometric structure, such as metric, Cob is actually a non-unital category, since
there is no identity morphism. Instead, it has “almost identity” morphisms — the
infinitesimally short cylinders. ⌟

cutting gluing

©2025 Jonas Dujava ©2025 Jonas Dujava

Figure 1.50 / Gluing and cutting of cobordisms.

Chapter 1. Quantum Field Theory Formalism 16

QFT as a Functor

Definition 1.51 (Functorial QFT à la Atiyah–Segal). A (d + 1)–dimensional

QFTd+1 = (H , Z ) is defined by the following assignments:
• H : Boundary ↦→ Space of Quantum States — to each closed d-
dimensional (codimension 1) manifold Σ we assign a complex vector space
HΣ — the space of quantum states,
• Z : Cobordism ↦→ Evolution Operator — a (d + 1)–dimensional
cobordism Σin M Σout is assigned a C-linear map ZM : HΣin → HΣout
called an evolution operator (or sometimes partition function).
These furthermore satisfy following axioms:
(a) Multiplicativity (⊔ ↦→ ⊗): Disjoint unions are mapped to tensor
products, that is

HΣ⊔Σ′ = HΣ ⊗ HΣ′ , ZM⊔M′ = ZM ⊗ ZM′ ,

where for consistency an empty d-dimensional manifold ∅ is assigned a

one-dimensional space of states H∅ = C.
(b) Gluing/Sewing (∪ ↦→ ◦): Gluing of two cobordisms Σ2 M2 Σg
and Σg M1 Σ1 along Σg is mapped to a composition of the evolution
operators ZM2 ∪M1 = ZM2 ◦ ZM1 : HΣ2 ← HΣ1 .
(c) Normalization: We assume a “very short” (notion depending on the
type of geometric structure on cobordisms) cylinder Σ Σ × [0, ε] Σ over
a closed d-manifold Σ tends in the limit ε → 0 to the identity map
limε→0 ZΣ×[0,ε] = 1HΣ : HΣ → HΣ .
(d) Naturality/Equivariance under Diffeomorphisms: For a diffeo-
morphism of closed d–manifolds φ : Σ → Σ′ we have an isomorphism
ρφ : HΣ → HΣ′ between the corresponding state-spaces. It is C-linear if
φ is orientation-preserving, and C-antilinear if φ is orientation-reversing.
Moreover, it is an action, that is ρφ′ ◦φ = ρφ′ ◦ ρφ .
Given a diffeomorphism between cobordisms ϕ : M → M′ , one has a
commutative diagram (we denote ϕ|in/out the restrictions of ϕ to Σin/out )
ZM,g
HΣin HΣout
⌜
ρϕ|in ρϕ|out
,
HΣ′in ZM′ ,g′ ≡ϕ∗ g
HΣ′out

where we explicitly indicated how the background structure g on M is

push-forwarded to g ′ ≡ ϕ∗ g on M′ .
Chapter 1. Quantum Field Theory Formalism 17

Remark 1.52 (QFT as a Functor). The cobordism category Cob was already
introduced in Remark 1.49 → p.15 , and the second category is the vector space
category Vect C , whose objects are complex vector spaces and morphisms are linear
maps. They are even symmetric monoidal categories, monoidal product being
disjoint union for Cob (with empty manifold being the unit) and tensor product
for Vect C (with C being the unit).
Then a QFT can be understood as a functor of symmetric monoidal categories
QFT ≡ (H , Z )
Cob Vect C

natural under diffeomorphisms. Such a description neatly encompasses axioms

listed in Definition 1.51 → p.16 . For additional category-theoretical details, refer to
[18] [19]. ⌟
Remark 1.53 (Closed Manifold, Partition Function). Since closed (d +1)–dimensional
manifold M has no boundary, the evolution operator for cobordism ∅ M ∅ is
simply a map ZM : C → C. This is just a multiplication by a complex number,
which we call the partition function (and use the same symbol ZM ). Oftentimes
we omit M when it is clear from the context. ⌟
Remark 1.54 (Topological QFT). For a topological QFT (TQFT) there is no
geometric background structure on the cobordisms. Cylinders are thus mapped
to identity operators, since any cylinder is “very short”. Making a torus calculates
the trace of the identity operator, which is the dimension of the Hilbert space. ⌟
Remark 1.55 (Reversed Structure and Dual Space, − ↦→ ∗). Considering a “very
short” cylinder Σ Σ × [0, ε] Σ viewed as a cobordism Σ ⊔ (−Σ) Σ × [0, ε] ∅ yields
in the ε → 0 limit a bilinear pairing (for convenience we put H−Σ first)

( , )Σ : H−Σ ⊗ HΣ → C ,

which can be seen to be nondegenerate, and thus provides the identification

H−Σ ∼ = HΣ∗
by H−Σ ∋ s ↦→ (s, ) ∈ HΣ
∗
, where HΣ
∗
is the linear dual space to HΣ
(linear maps HΣ → C).

Nondegeneracy can be seen by taking a second short cylinder ∅ Σ × [0, ε] (−Σ) ⊔ Σ,

and gluing its out-boundary (−Σ) to the in-boundary (−Σ) of the first cylinder,
obtaining thus a cylinder Σ Σ × [0, 2ε] Σ giving the identity in the limit ε → 0 by the
normalization axiom. The naturality axiom is tied to the assumption our functorial
description describes already the physical states without any degeneracy. ⌟
Remark 1.56 (Unitarity). Using the tautological orientation-reversing mapping
r : Σ → −Σ mapping each point to itself and just reversing the geometrical data,
we can define a nondegenerate sesquilinear form
ρ(r) ⊗ 1
⟨ , ⟩Σ : H Σ ⊗ H Σ H−Σ ⊗ HΣ ( , )Σ
C

Unitarity (in the Lorentzian signature) is then an additional (optional) collection

Chapter 1. Quantum Field Theory Formalism 18

of assumptions on a QFT:
(1) For each Σ, the pair (HΣ , ⟨ , ⟩Σ ) forms a Hilbert space, possibly after an
appropriate completion of HΣ . In particular, the corresponding sesquilinear
form is positive definite.
(2) For any cylinder Σ Σ × [0, t] ⟨︂Σ the evolution operator
⟩︂
ZΣ×[0,t] : HΣ → HΣ is
a unitary operator, that is ZΣ×[0,t] , ZΣ×[0,t] = ⟨ , ⟩Σ , or equivalently
Σ
†
ZΣ×[0,t] = ZΣ×[0,t]
−1
.
(3) The representation ρ of diffeomorphisms on the Hilbert spaces of states
is unitary. More precisely, for any diffeomorphism ϕ : Σ → Σ′ the corre-
sponding isomorphism ρϕ : (HΣ , ⟨ , ⟩Σ ) → (HΣ′ , ⟨ , ⟩Σ′ ) is unitary, that is
⟨ρϕ , ρϕ ⟩Σ′ = ⟨ , ⟩Σ .
The Euclidean counterpart of the unitarity is the reflection positivity to be
discussed at the end of Section 1.4 → p.29 . For more FQFT details see [9]. ⌟

As a simple example, lets look at Quantum Mechanics.

Example 1.57 (Quantum Mechanics as FQFT). Quantum Mechanics can be un-

derstood as a (1 + 0)–dimensional Riemannian Atiyah–Segal QFT. We will clearly
see how the ordinary Schrödinger picture emerges.
Since d = 0, the objects from the cobordism/spacetime category Cob are oriented
0-dimensional manifolds, that is collections of points {pt± } with orientation ±.
If we denote the space of states for pt+ by H ≡ Hpt+ , we automatically have
Hpt− = H−pt+ = H∗ using Remark 1.55 → p.17 .
Morphisms (1-cobordisms) are then collections of oriented intervals and circles
equipped with a Riemannian metric. By naturality (see Definition 1.51 → p.16 ), the
partition function depends only on the metric structure modulo diffeomorphisms,
that is only on the lengths of the connected components.
We denote the evolution operator for an interval of length t — viewed as a
cobordism pt+ [0, t] pt+ — by Zt ≡ Z[0,t] ≡ Ut : H → H. Functoriality/gluing
implies (semi)group law for evolution operators, that is
!
Zt2 +t1 ≡ Z[0,t2 +t1 ] = Z[t1 ,t1 +t2 ]∪[0,t1 ] = Z[t1 ,t1 +t2 ] ◦ Z[0,t1 ] ≡ Zt2 ◦ Zt1 .

The normalization axiom implies that the evolution operator in the limit of zero
length interval is the identity operator, that is limε→0 Zε = 1H . Expanding around
zero length (assuming analyticity), we obtain

ı̊ (︂ )︂
Zε ∼ 1H − Hε + O ε2 ,
ℏ
where H ∈ End(H) is the Hamiltonian operator — time translation generator. In
Example 1.65 → p.20 , we will discuss QM operators in more detail.
 Chapter 1. Quantum Field Theory Formalism 19

Either by forming a Schrödinger differential equation for Zt ≡ Ut in the form

∂ ∂ ! ∂ ı̊ ∂
⃓ ⃓
= = − HUt =⇒ U = HUt ,
⃓ ⃓
Ut ≡ Uε+t ⃓⃓ Uε ◦ Ut ⃓⃓ ı̊ℏ
∂t ∂ε ε=0 ∂ε ε=0 ℏ ∂t t
(︂ )︂N
or by using semigroup law iteratively as Zt = Z Nt and taking the limit N → ∞,
we obtain the solution
ı̊
Zt ≡ Ut = e− ℏ Ht .
If we take our QM to be unitary, see Remark 1.56 → p.17 , we require H ≡ Hpt+ to
be a Hilbert space with the natural inner product ⟨ , ⟩ ≡ ⟨ , ⟩pt+ , and the
Hamiltonian H to be Hermitian/self-adjoint, such that Zt is unitary.
Considering a circle S1t of circumference t — interval with the ends glued together
— corresponds to making a trace in the end
ı̊
ZS1t = TrH Zt = TrH e− ℏ Ht .
⌟
Remark 1.58. To describe Hamiltonians depending on time, we need to include
extra structure to know where we are located with respect to this “global” time,
on which the Hamiltonian parametrically depends. ⌟
Remark 1.59 (Finite Temperature). In the Euclidean signature we set t = −ı̊τ , so
for a Euclidean circle S1β of circumference τ ≡ β > 0 we have (ℏ ≡ 1)

Z E
S1β = TrH EZβ = TrH e−βH .

which is the partition function for QM system at inverse temperature T1 ≡ β.

Considering a QFT at finite temperature thus geometrically corresponds to curling
up the Euclidean time direction into a circle of circumference β. ⌟

Quantum Observables
We will now show how to understand quantum operators/observables as insertions
of states on the boundaries of infinitesimally small cuts of spacetime. First, we
state the rather general definition, and then mention some important examples.
Idea 1.60 (Operator as State Insertion on Cut). Let Σin M Σout be a cobordism,
and let Γ ⊂ M be some CW subcomplex — collection of points, lines, surfaces,
and so on in M — disjoint from ∂M.
Consider a family of ε-thickenings Uε (Γ) of Γ with ε ∈ (0, ε0 ), for example by
equipping M with a metric, and taking all points with distance ≤ ε from Γ. Then
M\Uε (Γ) is a cobordism, where we additionally made cuts around Γ with some
small padding parametrized by ε, so it has a boundary
(︂ )︂
∂ M\Uε (Γ) = (−∂Uε (Γ)) ⊔ (−Σin ) ⊔ Σout .

We can insert states on the boundaries of these cuts, which are viewed as extra
Chapter 1. Quantum Field Theory Formalism 20

pieces of the initial boundary, and in the limit ε → 0 they will correspond to some
local operators. ⌟

Definition 1.61 (Insertion, Quantum Observable, Correlator, VEV). A quantum

observable supported on Γ is a family of elements OΓ,ε ∈ H∂Uε (Γ) — states on
the boundary of the ε-tube around Γ — such that the insertion/correlator
defined by the limit
⟨︂ ⟩︂ [︂ ]︂
OΓ ≡ lim ZM\Uε (Γ) OΓ,ε : HΣin → HΣout
M ε→0

is well-defined. Here ZM\Uε (Γ) is a map H∂Uε (Γ) ⊗ HΣin → HΣout , and we
substitute the states defining the operator into the first factor. An important
case is when M is closed (Σin = Σout = ∅ = ∂M), in which case the correlator
is just a complex number called Vacuum Expectation Value (VEV).

Remark 1.62. Usually we require for ε0 > ε′ > ε > 0 that

!
[︂ ]︂
OΓ,ε′ = ZUε′ (Γ)\Uε (Γ) OΓ,ε ,

so the states on a (bigger) ε′ -thickenings are obtained by the “free” evolution

(without additional insertions) from a (smaller) ε-thickening. In such a case, the
expression under the limit in Definition 1.61 → p.20 actually does not depend on
ε ∈ (0, ε0 ) due to the gluing axiom in Definition 1.51 → p.16 . ⌟
Remark 1.63 (Point Observables, Local Operators). Now we will compare the
FQFT definition with Definition 1.23 → p.7 and Definition 1.26 → p.7 . Consider a
space with n punctures {xi } understood as “infinitesimally” small balls {Bi }
centered on punctures with boundaries {Si }. The corresponding local operators
are Oxi M ∼
= HSi . If there is no other boundary (or it is infinitely far away), we
can consider insertions of {Oi } as a vacuum n-point correlation function
⟨︂ ⟩︂
O1 (x1 ) · · · On (xn ) .
⌟
Example 1.64 (Identity Operator). The identity operator/field 1 ≡ 1 corresponds
to a point observable satisfying

⟨1(x) OΓ ⟩ ≡ ⟨OΓ ⟩

for any operator OΓ . It essentially means that putting the operator 1 on a puncture
just fills the puncture as if it was never there. Alternatively, it corresponds to
the vacuum vector |vac⟩ ≡ ZB ∈ HS for an infinitesimally small ball B with a
boundary S — it is the state obtained by path integrating over the ball B inside
S without any additional operator insertions. ⌟
Example 1.65 (Point Observables in QM — Operators). For an elementary but
enlightening example, we will look at point observables in the context of Quantum
Mechanics, see Example 1.57 → p.18 . To understand an observable supported on
Chapter 1. Quantum Field Theory Formalism 21

a single point Γ ≡ t inside a cobordism (time interval) M = [tin , tout ] ≡ [tout |tin ],
we consider the corresponding ε-thickenings

Uε (Γ) ≡ Uε (t) = (t + ε|t − ε) ≡ (t − ε, t + ε)

with boundaries
∂Uε (t) = pt+ −
t+ε ⊔ ptt−ε .

A quantum observable is then — in a limit ε → 0, but we can take it to be

constant even for nonzero ε — an element

O ∈ H∂Uε (t) = Hpt+ ⊔pt− = Hpt+ ⊗ Hpt− ≡ H ⊗ H∗ ∼

= End(H) ,

so it is a linear operator acting on the space of states H.

If we consider several point observables supported on Γ = {tn , . . . , t1 } with
tout > tn > · · · > t1 > tin , it is the same as choosing a collection of operators
{On , . . . , O1 } ⊂ End(H). The “correlator” (or insertion into evolution operator)
is then obtained by inserting these operators at the corresponding points (times),
leading to
⟨︂ ⟩︂
On (tn ) · · · O1 (t1 ) = Z[tout |tn ] ◦ On ◦ · · · ◦ Z[t2 |t1 ] ◦ O1 ◦ Z[t1 |tin ]
M
ı̊ ı̊ ı̊
= e− ℏ H(tout −tn ) ◦ On ◦ · · · ◦ e− ℏ H(t2 −t1 ) ◦ O1 ◦ e− ℏ H(t1 −tin ) .

The order of operators on the left-hand side is not important, since they are by
definition inserted at the appropriate times, automatically time-ordered. ⌟

Example 1.66 (Topological Surface Operators). Very important class of nonlocal

operators are the so-called conserved charges. Usually, they are integrals of
conserved currents (coming from Noether’s theorem) over some codimension-1
submanifolds (spatial slices). Due to the conservation of the integrated current,
these operators do not depend on the particular choice of the integration surface
(as long as we do not cross other operator insertions), and are therefore called
topological surface operators. We will discuss them in Section 1.3 → p.28 . ⌟
As we have seen, the Atiyah–Segal formulation of QFTs naturally generalizes
the Schrödinger picture. In the following remark we briefly touch on the various
“quantum pictures” in QM.
Remark 1.67 (Various Quantum Pictures). Special things happen when the space-
time has (continuous) symmetry (generating a family F of (space)slices foliating
the spacetime) relating vector spaces HΣ at different (space)slices Σ ∈ F. Such
isomorphic vector spaces can be then identified into one quantum state space H,
and similarly all operators can be viewed as acting on this one space. We will
further discuss the notion of Operator Formalism in Section 1.4 → p.29 .
Such simplification precisely happens in QM, where the family of “(space)slices”
is just collection of points along an interval F = {pt+
t }t∈I⊂R , and corresponding
Hilbert spaces (of quantum states) at different times are isomorphic. Put another
Chapter 1. Quantum Field Theory Formalism 22
′
way, they form a family HF ≡ {Ht }t∈I⊂R linked by isomorphisms Ht′ U (t , t) Ht
generated by the time translation evolution operators U (t′ , t). To describe the
dynamics in such a setting, we can use various Quantum Pictures:
• Schrödinger Picture. Quantum states |ϕ(t)⟩S ∈ H evolve in time, operators
OS ∈ End(H) are fixed (their “time of insertion” is determined by the time
parameter of the state) — closest to the FQFT picture, we just represent
evolution on one fixed Hilbert space (because we can). Time evolution is
given by the Schrödinger equation
d
ı̊ℏ |ϕ(t)⟩S = H|ϕ(t)⟩S ,
dt
where H is the Hamiltonian operator.
• Heisenberg Picture. Quantum states |ϕ⟩H ≡ |ϕ(t0 )⟩S ∈ H are fixed and
describe initial conditions at a reference time t0 , while operators OH (t) ∈
End(H) evolve in time (time evolution pullback of some fixed “Schrödinger”
operator acting on Ht , so that it acts on reference Hilbert space Ht0 )

U (t,t0 )
Ht Ht0
OH (t) ≡ U (t, t0 )−1 OS U (t, t0 )
⌜

OS OH (t) .
OH (t0 ) ≡ 1 OS 1
Ht Ht0
U (t,t0 )−1

Infinitesimally, we obtain the evolution equation for Heisenberg operators

d
ı̊ℏ O (t) = [OH (t), H] .
dt H
Comparing with Example 1.65 → p.20 , we see that matrix elements (after in-
serting states at the boundary of correlator/evolution function) calculated in
either picture lead to the same result.
• Interaction/Dirac Picture. Sometimes it is useful to split the evolution
operator into free (which we can explicitly solve) and interaction part. We
evolve the operators with the “uninteresting” free part, and states with
the interaction part. If the interaction is weak, we can conveniently do
perturbation theory.
⌟

1.3 Symmetries in Quantum Field Theories

When we change the system (transform the fields), but the system nevertheless
looks the same, we say that the given transformation is a symmetry. We will show
how these often correspond to symmetries of the background structure.
Symmetries of a theory constrain the form of correlation functions. Infinitesimal
form of these constraints are encoded in the so-called Ward–Takahashi identities.
Chapter 1. Quantum Field Theory Formalism 23

Generators of symmetries are given by conserved currents, which are represented

as topological (surface) operators in the theory. Algebra of generators encodes,
how we should “commute” surfaces of these operators.
Remark 1.68 (Types of Symmetries). There are various types of symmetries —
discrete or continuous, spacetime or internal, global or local, and so on. We will
mainly focus on continuous spacetime symmetries, or their slight generalizations.
Internal symmetries correspond to the target values of the fields, that is to the
extrinsic part of field bundle, as discussed in Remark 1.7 → p.3 . Global internal
symmetries are conceptually pretty simple, especially when they are linear, see
Remark 2.9 → p.40 .
Local internal symmetries are topic of gauge theories, and should be viewed more
as a redundancy of description, similarly to diffeomorphism invariance. ⌟

Finite Symmetry Transformations

Invariance of the functional measure + action under some transformation implies
relations between correlation functions.
Remark 1.69 (Correlation Relations from General Measure + Action Invariance).
Suppose a transformation of fundamental fields ϕ ↦→ ϕ′ together with background
fields/structure g ↦→ g ′ (for example metric) leaves the measure + action invariant,
that is
ı̊ ı̊ ′
[Dϕ]g e ℏ Sg [ϕ] = [Dϕ′ ]g′ e ℏ Sg′ [ϕ ] .
independent of other terms under the path integral. Then we have a relation for
correlators (X is an arbitrary operator, X ↦→ X ′ is induced by ϕ ↦→ ϕ′ )
∫︂
ı̊
⟨︂ ⟩︂
X′ ≡ [Dϕ]g e ℏ Sg [ϕ] X ′
g
∫︂
ı̊ ′
= [Dϕ′ ]g′ e ℏ Sg′ [ϕ ] X ′
∫︂
ı̊
⟨︂ ⟩︂
= [Dϕ]g′ e ℏ Sg′ [ϕ] X ≡ X ,
g′

where we first used the assumed invariance, and then we just relabeled the
integration variables. We did not need to care of the normalizations Zg or Zg′ ,
since the invariance implies they are the same. ⌟
Remark 1.70 (Diffeomorphism Relations, Spacetime Symmetries). Assuming the
theory is diffeomorphism-invariant, we can consider transformations induced by
a diffeomorphism f : M → M, that is O ↦→ f ∗ O for any operator, and g ↦→ f∗ g.
The previous remark then gives us the following relation for any diffeomorphism
f in the form (to have both transformations on the same side, we pushforward g)
⟨︂ ⟩︂ ⟨︂ ⟩︂
O1 (x1 ) · · · On (xn ) = [f ∗ O1 ](x1 ) · · · [f ∗ On ](xn ) .
g f∗ g

This does not mean that we have an infinite number of symmetries, because a
general f relates correlators in different backgrounds — it should be understood
Chapter 1. Quantum Field Theory Formalism 24

as the usual redundancy in the choice of coordinates.

But when the chosen background g has some spacetime symmetry, that is there
!
exists a (nontrivial) diffeomorphism f preserving g as f∗ g = g, we obtain
⟨︂ ⟩︂ ⟨︂ ⟩︂
O1 (x1 ) · · · On (xn ) = [f ∗ O1 ](x1 ) · · · [f ∗ On ](xn )
g f∗ g
!
⟨︂ ⟩︂
= [f O1 ](x1 ) · · · [f On ](xn )
∗ ∗
,
g

which is a nontrivial relation between correlators in the same background. ⌟

Example 1.71 (Symmetries from Flat Space Isometries). As a simple example we

consider flat space M = Rd with the usual flat metric, which we will suppress in
the notation. The symmetries of the metric structure are called isometries, and in
Rd we have the usual translations and rotations. For a translation f : x ↦→ x + a
we have the relation
⟨︂ ⟩︂ ⟨︂ ⟩︂
O1 (x1 ) · · · On (xn ) = O1 (x1 + a) · · · On (xn + a) ,

and similarly for rotations (where we also act by an appropriate representation of

the rotation, recall Remark 1.7 → p.3 ). Translation symmetry implies correlators
can depend only on the relative positions of the operators. Rotation symmetry is
particularly simple for scalar operators, where we can conclude they depend only
on the relative distances. ⌟

Remark 1.72 (Symmetries of FQFT). From the FQFT point of view, this is just
the naturality axiom (see (d) of Definition 1.51 → p.16 ) restricted to the subgroup
Sym(M, g) ⊂ Diff(M) of diffeomorphisms f : M → M preserving the background
!
structure g, that is f∗ g = g. For a cobordism Σout M Σin , such a symmetry f
implies
ZM,g = ρf |out ◦ ZM,g ◦ ρ−1
f |in .

One can try to think through how this works out when we include operator
insertions along the lines of Definition 1.61 → p.20 . ⌟
Now consider a slightly different situation. Suppose the theory is invariant under
simultaneous active transformation of the operators and the background structure
(now we allow different local operators O• to transform in different ways, not
necessarily induced by the transformation of the fundamental fields)

O• ↦→ O
˜︁ ,
• g ↦→ g̃ ,

that is, we have a relation

⟨︂ ⟩︂ ⟨︂ ⟩︂
O1 (x1 ) · · · On (xn ) = O
˜︁ (x ) · · · O
1 1
˜︁ (x )
n n .
g g̃

When does such a relation imply a symmetry in a given background g? The

answer is that we must find a diffeomorphism f , with which we can pushforward
Chapter 1. Quantum Field Theory Formalism 25
!
g̃ back to the original background g = f∗ g̃. Then we obtain symmetry relation
⟨︂ ⟩︂ ⟨︂ ⟩︂
O1 (x1 ) · · · On (xn ) = [f ∗ O
˜︁ ](x ) · · · [f ∗ O
1 1
˜︁ ](x )
n n
g f∗ g̃
!
⟨︂ ⟩︂
= [f O
˜︁ ](x ) · · · [f O
1
∗
1
˜︁ ](x )
∗
n n .
g

Example 1.73 (Weyl Transformation). Now we will consider a Weyl-invariant

theory, so the correlators are invariant under an active Weyl transformation
g ↦→ g̃ ≡ Ω2 g together with O• ↦→ O ˜︁ ≡ Ω∆• O (we consider O to be primary
• • •
scalar operators with scaling dimension ∆• , more details will be in Chapter 3 → p.51 ).
For example, we have the following relation for 2-point functions
⟨︂ ⟩︂ ⟨︂ ⟩︂
O1 (x)O2 (y) = Ω(x)∆1 Ω(y)∆2 O1 (x)O2 (y) .
g Ω2 g

When the same Weyl rescaled metric can be obtained also by diffeomorphism
(which we then call conformal transformation, see Definition 3.6 → p.53 ), this induces
relation between correlators on the same background.
Considering flat space Rd and scaling/dilatation f : x ↦→ x′ ≡ λx, simple calcula-
!
tion gives f ∗ g = λ2 g = Ω2 g ≡ g̃, so we obtain
⟨︂ ⟩︂ ⟨︂ ⟩︂
O1 (x)O2 (y) = λ∆1 +∆2 O1 (λx)O2 (λy) ,

and thus the possibly consistent behavior of the 2-point function of a CFT is
given by a power law (where we implicitly used the translational and rotational
invariance)
⟨︂ ⟩︂ 1
O1 (x)O2 (y) ∝ .
|x − y|∆1 +∆2
In Corollary 3.80 → p.77 , by further considering “special conformal transformations”,
we will see that the correlator can be nontrivial only when ∆1 = ∆2 . ⌟

Energy–Momentum Tensor
In the following we will assume a classical theory with Lagrangian description on a
manifold with a metric g. Since spacetime symmetries are essentially symmetries
of g, they are intimately connected with the energy–momentum tensor.

Definition 1.74 (Energy–Momentum Tensor). The energy–momentum tensor

T is defined as the functional derivative of the action with respect to the
metric
δSg 1 ∫︂ 1/2 ab
T ab ≡ −2 ⇐⇒ δg Sg ≡ − g T δg ab .
δgab 2 M
It is also often called the stress tensor or stress–energy tensor.

Remark 1.75. We will always work with such “Hilbert” energy–momentum tensor,
as opposed to the “canonical” one, which is non-uniquely defined by the Noether’s
prescription. ⌟
Chapter 1. Quantum Field Theory Formalism 26

Proposition 1.76 (Properties of Hilbert Energy–Momentum Tensor). We have

(1) T •• = T (••) ⇐⇒ T ab = T ba — T •• is a symmetric tensor.
(2) ∇• T ∼ 0 ⇐⇒ ∇a T ab ∼ 0 — T •• is conserved on-shell.
EL EL

(3) Tr T ≡ Taa ∼ 0 — T •• is traceless on-shell for Weyl-invariant theories.

EL

Proof. The symmetry (1) follows from the definition and symmetry of g.
The conservation property (2) follows from the fact that the action is invariant
under diffeomorphisms. Indeed, using the covariance Remark 1.12 → p.3 for auto-
morphism generated by a vector field ξ ∈ T M vanishing in a neighborhood of
the boundary ∂M, we have

! 1 ∫︂ 1/2 ab ∫︂
δSg
0 = δξ Sg [ϕ] ≡ − g T δξ gab +
1
g /2 δξ ϕ
2 M M δϕ
∫︂
g /2 T ab ∇(a ξb) + 0
EL 1
∼−
∫︂M
ξ arbitrary
g /2 (∇a T ab )ξb + 0 ======⇒ ∇• T ∼ 0 .
EL 1 EL
∼
M

We used definition of T and δξ gab ≡ Lξ gab = ∇(a ξb) . In the last step we used
symmetry of T and integrated by parts — the boundary term vanishes due to
the assumption on ξ. Since ξ is otherwise arbitrary, we obtain the conservation
property (2).
For Weyl transformation Example 1.73 → p.25 we have the infinitesimal form δω g =
2ωg, so we calculate (similarly as previously for diffeomorphism)

! 1 ∫︂ 1/2 ab ∫︂
δSg
0 = δω Sg [ϕ] ≡ − g T (2ωgab ) + δω ϕ
2 M M δϕ
∫︂
ω arbitrary
g /2 Taa ω + 0 ======⇒ Tr T ∼ 0 .
EL 1 EL
∼−
M

Corollary 1.77 (Conserved Currents from Isometries). Let f : M → M be

an isometry generated by a (Killing) vector field ξ ∈ T M. Then T • ξ is a
conserved current, that is ∇• (T • ξ) ≡ ∇a (T ab ξb ) = 0.

Proof. By direct calculation we have

!
∇• (T • ξ) = (∇a T ab )ξb + T ab ∇a ξb = 0 + T ab ∇(a ξb) = 0 ,

where we used both the conservation and symmetry of T , and finally the infinites-
!
imal form of the isometry Lξ gab = 2∇(a ξb) = 0.

Remark 1.78 (Conserved Currents from Conformal Transformations). Refer to

Section 3.1 → p.52 for the definition of conformal Killing vector fields (CKV), where
we discuss in detail the conformal structure of manifolds. The infinitesimal
Chapter 1. Quantum Field Theory Formalism 27

condition for CKV is given by Lξ g = ∇(a ξb) ∝ gab , so assuming a traceless

energy–momentum tensor, we again obtain a conserved current T • ξ. Thus, a
Weyl invariant theory has a conserved current for every CKV. ⌟

Ward–Takahashi Identities
Now we will look at the infinitesimal forms of the continuous spacetime symmetries.
Remark 1.79. Even though it is not strictly necessary for our purposes, the
following will be formulated in the “tetrad” formulation — see Appendix N → p.153 .
This would allow us to consider also spinors, since just the metric is not sufficient.
For example, the energy–momentum tensor is classically given by
δS k δS
T ab ≡ Tka ekb ≡ − k
eb = −2 gnb ,
δea δgan

since gab = ηkl eka elb . ⌟

In the quantum setting, we also include possible contribution from the variation
of the (renormalized) path integral measure. For simplicity (to avoid factors of ı̊)
we will consider Euclidean signature and set ℏ ≡ 1.

Definition 1.80 (Quantum Energy–Momentum Tensor). We define the quantum

energy–momentum tensor operator as the response of functional measure +
action to an infinitesimal metric/tetrad variation
(︂ )︂ ∫︂ (︂ )︂
δe [Dϕ]e e−Se [ϕ] ≡ [Dϕ]e e−Se [ϕ] e δeka Tka + O (δe)2 ,
M

by which we mean
⟨︂ ⟩︂ ⟨︂ ⟩︂ ⟨︂ ⟩︂ δ ⟨︂ ⟩︂
Tka (x)X − Tka (x) X ≡ X ,
e e e δeka (x) e

where the second term comes from the variation of normalization Ze in the
definition of the correlator, because (Ze ≡ e−We )

δZe ∫︂ ⟨︂ ⟩︂ ⟨︂ ⟩︂ δWe
≡ [Dϕ] e −Se [ϕ]
T a
≡ Ze T a
(x) ⇐⇒ T a
(x) =− k .
δea (x)
k e k k e k e δea (x)

Calculation 1.81 (Ward–Takahashi Identities). Recalling Remark 1.69 → p.23 and

Remark 1.70 → p.23 , for an infinitesimal diffeomorphism generated by a vector field
ξ we have
d ⟨︂ ⟩︂ ⃓⃓
⟨︂ ⟩︂ ⃓ ∫︂ (︃⟨︂ ⟩︂ ⟨︂ ⟩︂ ⟨︂ ⟩︂ )︃
= =
1/ 1 ab ab
Lξ X X ∗ ⃓ g 2 Lξ gab T X − T
2
X
g dε fε g ε=0 M
∫︂
g g g
⟨︂ ⟩︂
=− + 0,
1
g /2 ξb ∇a T ab X
M g

where we integrated by parts and used Zg = Zg+Lξ g =⇒ ⟨∇• T ⟩g = 0.

Chapter 1. Quantum Field Theory Formalism 28

Thus, ∇• T = 0 up to contact terms. For example, in flat spacetime, and for

scalar operators, we have
⟨︂ ⟩︂ n ⟨︂ ⟩︂
∂a T ab (x)O1 (x1 ) · · · On (xn ) = − δ(x, xi )∂(i)
b
O1 (x1 ) · · · On (xn ) .
∑︂

i=1

These are called the Ward–Takahashi identities. ⌟

Symmetries as Topological Operators

As foreshadowed in Example 1.66 → p.21 , we can form symmetry generators (charges)
as space-slice (surface) integrals of conserved currents, that is
∫︂
Qξ (Σ) ≡ − dSa T ab ξb ,
Σ

where now ξ is a Killing vector (or conformal Killing vector if T is traceless).

The Gauss law, Ward–Takahashi identities Calculation 1.81 → p.27 , and the con-
servation of currents Corollary 1.77 → p.26 imply that the correlation function of
Qξ (Σ) do not change as we deform the surface Σ, as long as we do not cross the
support of other operators (see Figure 1.82 → p.28 ).

Q(Σ2 ) Q(Σ1 )
Q(Σ3 )
O1 O2

©2025

Figure 1.82 / Topological operator can be deformed to small surfaces around the
operator insertions, so even though the operator is not local, the correlation functions
depend locally on the points of insertions.

When we move the surface over the operator insertions, we obtain a contact term,
which is the infinitesimal transformation of the operator under the symmetry
generated by Qξ . In particular, taking flat space and ξ ≡ ∂a , we obtain momentum
generators ∫︂
Pa (Σ) ≡ Q∂a (Σ) ≡ − dSn T na ,
Σ
which by Calculation 1.81 → p.27
act as derivatives
⟨︂ ⟩︂ ⟨︂ ⟩︂
Pa (Σ)O(x) · · · = ∂a O(x) · · ·

where we took Σ ≡ ∂B to be the boundary of a ball B surrounding x and no

other operator insertions.
If we choose some particular foliation of spacetimes — as we will discuss below
in Section 1.4 → p.29 — and define the symmetry charges as surface integrals over
Chapter 1. Quantum Field Theory Formalism 29

these slices, the above statement would be interpreted in an operator formalism

as a commutation relation
[Pa , O(x)] = ∂a O(x) .
So even though the symmetry generators are not local operators, their commutators
with local operators (or surrounding of local operators) are again local.
The algebra (commutation relations) of symmetry generators corresponds to
how should one cross the surfaces of two topological operators. The important
point of this discussion is, that this is a property independent of the particular
foliation/operator formalism we choose.

1.4 Operator Formalism

Virtue of the functorial (path integral) formalism is that symmetries — both
spacetime and internal ones — are usually manifest. Such a formulation is natural
for conjuring up, comparing, and thinking about theories of various types, placed
on backgrounds with diverse structures. Furthermore, calculations carried out via
(formal) path-integral manipulations — oftentimes inspired by finite-dimensional
analogs — can be especially convenient.
However, much less transparent are questions about the number of physical degrees
of freedom, and the unitarity of the theory. It is necessary to address these issues
on the level of Hilbert spaces and corresponding states/operators.
We will only briefly discuss such “operator formalisms”, and by no means exhaus-
tively treat the complicated subject of QFT in curved spacetimes.

Foliation
Different foliations of spacetime in general give rise to different representations of
the theory, since we obtain in general different Hilbert spaces of quantum states
with different notions of “time evolution” and “time ordering” of operators. These
can be sometimes related by symmetry transformations, but as we will briefly
discuss below, in general they can be rather different.
By picking a specific foliation we always lose manifest covariance of the theory,
but the spacetime symmetries naturally provide preferred foliations respecting
given background structure.
Example 1.83 (Static/Stationary Spacetimes). If the (Lorentzian) spacetime is
static/stationary — we have a timelike Killing vector, which in static spaces is
even surface orthogonal — then the preferred foliations (choice of time coordinate)
respecting the time-translation symmetry share the same spatial geometry at each
time slice. Corresponding Hilbert spaces are then naturally isomorphic as was
discussed in Remark 1.67 → p.21 , and we can conveniently describe the system on
one Hilbert space. ⌟
Chapter 1. Quantum Field Theory Formalism 30

Remark 1.84. Without symmetries, we are quite lost, as computations become

very complicated. This is why we usually work with spacetimes with maximum
available symmetry — maximally symmetric spacetimes. One example is of course
the Minkowski/Euclidean flat spacetime, and another one which we will cover in
Chapter 4 → p.91 is the anti–de Sitter spacetime. ⌟

Vacuum and Particle Representation

In general, there is no preferred/distinguished ground state (vacuum vector) in
Hilbert space H, since there is no “natural” way to foliate spacetime — no preferred
notion of time or energy (Hamiltonian). Usefulness of various descriptions depends
on what questions we want to answer, which in turn depends on the observers
asking those questions.
Remark 1.85 (Particle Representations). The notion of particles is tied to the
choice of the vacuum state |vac⟩, since particles are then excitations of this vacuum
state.
In static/stationary spacetime (and for simplicity free theory), this is tied to the
choice of positive-frequency solutions of Euler–Lagrange EOM — we pick out a
subspace invariant under time-translation symmetry in space of complex solutions,
to which we then assign annihilation/creation operators. This way we build up
the Fock space of states over the vacuum state |vac⟩.
In general, for different observers their natural particle representations are differ-
ent, since foliations adapted to their worldlines are different, and their particle
interpretations are different. In particular, a no-particle (vacuum) state in one
representation can be viewed by another observer as a state with nonzero (or even
infinite) number of particles. Such transformations between different representa-
tions are called Bogoliubov transformations.
So without enough symmetry, there is no natural way how to define what is a
vacuum, or what is a particle. ⌟
Remark 1.86 (Correlators as Vacuum Expectation Values). Correlators in the sense
of Definition 1.26 → p.7 or Definition 1.61 → p.20 are interpreted in the operator
formalism as vacuum expectation values of time-ordered — perhaps better to say
foliation ordered — products of operators, that is
⟨︂ ⟩︂ (︂ )︂
O1 (x1 ) · · · On (xn ) = ⟨vac| T O1 (x1 ) · · · On (xn ) |vac⟩ ,

where O• are now represented on given Hilbert space H as elements of End(H),

and T orders them according to the chosen foliation. We have seen this in the
context of QM in Example 1.65 → p.20 .
Let us stress again that different foliations give rise to different vacuum states
|vac⟩ and Hilbert space, different representations of operators on these Hilbert
spaces, and different notions of time ordering. Nevertheless, in the end all of
them calculate the same correlation functions. From this point of view, different
Chapter 1. Quantum Field Theory Formalism 31

operator formalisms are different interpretations of the same underlying physics,

same underlying path integral. ⌟

Remark 1.87 (Symmetry Charges). As mentioned briefly in Section 1.3 → p.28 , the
charges corresponding to the symmetries of the theory are defined as integrals
of the conserved currents over the slices of the foliation. Then the surrounding
of some local operator O(x) by a charge Q is represented as the commutator
[Q, O(x)] — you can imagine how we can deform small surface around the operator
insertion to a one surface above, and one below, with the orientation giving the
relative minus sign. The finite exponentiated action (with some parameter t) is
then given by
et[Q, ] O(x) = etQ O(x) e−tQ .
⌟

Example 1.88 (Minkowski Spacetime — Inertial Foliation). In flat Minkowski

spacetime, we can foliate spacetime by hyperplanes of constant time, which
corresponds to some class of inertial observers. This is adapted to the action of the
Poincaré group (namely the subgroup of translations and spatial rotations). As
usual, the vacuum state is invariant under the whole Poincaré group represented
by unitary transformations. Foliation-ordering of operators is in this case given
by the standard notion of time ordering (in the given time parameter tied to a
chosen class of inertial observers).
Different choice of inertial observer corresponds to the foliations obtained by
boosting the original foliation, see Figure 1.89 → p.31 on the left. Since the vacuum
is invariant, they share the same vacuum state, and thus the corresponding particle
representation. We say that the particle interpretation is “unitarily equivalent” —
descriptions of the same quantum state by different inertial observers are related
by unitary transformation, with basis-independent characteristics (such as number
of particles in a given state) staying the same. ⌟

©2025 Jonas Dujava

Figure 1.89 / Common foliations of Minkowski spacetime (or its part). The left figure
shows the usual inertial foliation, while the right figure shows the Rindler foliation of
the right/left wedges adapted to the trajectory of uniformly accelerated observer.
Chapter 1. Quantum Field Theory Formalism 32

Example 1.90 (Minkowski Spacetime — Rindler Foliation). Suppose we wanted

to describe the system from the point of view of uniformly accelerated observer.
Observer then moves along an orbit of the boost action. A more natural foliation
(adapted to the action of corresponding boost) is then the Rindler foliation (Rindler
coordinates), see Figure 1.89 → p.31 on the right.
Considering the causality structure of Minkowski, we see that there appear horizons
(enclosing for example the right wedge containing the worldline of the observer),
beyond which events cannot reach the observer.
To properly describe quantum field pervading the whole spacetime, we need to
include into description also the left wedge — space of quantum states will be
H ≃ HL ⊗ HR — such that we can form Cauchy slice (with enough initial data
for evolution of quantum field). At the end we can trace out through the left
wedge HL , and obtain density matrix operator carrying the whole information
which concerns the uniformly boosted observer in the right wedge.
From the point of view of the quantum field, the more natural choice is the inertial
foliation discussed in Example 1.88 → p.31 . Most likely it rests in the corresponding
Minkowski vacuum |vac⟩Mink , or some low excitation of it. How is such a state
perceived by the accelerated observer? After partially tracing over the left wedge,
we obtain a density matrix operator ρR ≡ TrL |vac⟩Mink⟨vac|, which actually turns
out to be a thermal state — a uniformly accelerating observer interprets |vac⟩Mink
as a thermal bath — this is called the Unruh effect. ⌟
Example 1.91 (Radial Quantization in CFTs). For Conformal Field Theories (CFTs),
which have in particular the scaling/dilatation symmetry, the common choice is
the so-called radial quantization. Foliation by concentric spheres is adapted to
the actions of rotations and dilatation. More on this in Section 3.4 → p.78 . ⌟

Conjugation
The usual goal is to study unitary QFTs, but as we discussed in Section 1.1 → p.12 ,
when it is possible, it is often convenient to work in the Euclidean signature. In
the following we will look at the Euclidean version of the Hermitian conjugation.
For the purposes of this and following chapter, we will explicitly indicate by a
pre-superscript L/E in what signature we consider the given object. As we will
see shortly, the conjugation in the Euclidean signature slightly differs from the
usual one in Lorentzian signature — a Euclidean version of Hermitian operator is
typically not Hermitian in the usual sense. To avoid confusion, it is helpful to
introduce extra nomenclature.

Definition 1.92 (Real and Complex Operators). Considering a standard inertial

quantization in the Lorentzian signature, we say that an operator is real if it
is Hermitian, that is LO† = LO. Otherwise, we say it is complex and denote
its Lorentzian conjugation by LO∗ ≡ LO† , and its continuation to Euclidean
signature similarly by EO∗ .
Chapter 1. Quantum Field Theory Formalism 33

Remark 1.93. Alternatively, we can think of LO∗ as first performing the complex
conjugation of the underlying classical field, and only then representing it as an
operator on the Hilbert space, giving us precisely LO† . ⌟

Remark 1.94. In a unitary Lorentzian QFT, the symmetry generators are rep-
resented as conserved charges which are (anti)unitary operators — depending
on whether we include ı̊ in the finite exponentiated transformation — such that
their exponentiated actions are unitary transformations of the Hilbert space. In
particular, the energy–momentum tensor is real. ⌟

Example 1.95 (Conjugation of Scalar Operators in Euclidean Signature). Starting

with a Lorentzian theory where time translations are generated by a (Hermitian)
Hamiltonian H as
L
O(t, x) = eı̊Ht LO(0, x) e−ı̊Ht ,
an analytical continuation to the Euclidean time τ ≡ tE ≡ı̊t gives
E
O(τ, x) ≡ LO(t = −ı̊τ, x) = eτ H LO(0, x) e−τ H ,
=⇒ EO(τ, x)† = e−τ H LO∗ (0, x) eτ H = EO∗ (−τ, x) ,

so the conjugation in Euclidean signature contains extra time reflection. ⌟

Example 1.96 (Euclidean Version of Tensor Operators). Before discussing the

conjugation of tensor operators, let us first comment on an additional detail in
their Wick-rotation. Since dt ≡ d(−ı̊τ ) = −ı̊ dτ , ∂t
∂ ∂
≡ ∂(−ı̊τ )
=ı̊ ∂τ
∂
, we typically
add factors of ∓ı̊ to the time components of tensors. In this way Lorentzian
correlators covariant under SO(1, d − 1) become (after analytical continuation)
Euclidean correlators covariant under SO(d) rotations.
As an example, for a vector operator we have
E
O0 (τ, x) ≡ −ı̊ LO0 (−ı̊τ, x) ,
E
Oi (τ, x) ≡ L
Oi (−ı̊τ, x) .
⌟
The following definition slightly generalizes examples provided above.

Definition 1.97 (Euclidean Conjugation, Time Reflection). A local Euclidean

tensor operator with l spacetime indices behaves under conjugation as
[︂ ]︂
E
Oa1 ...al (τ, x)† = Θab11 · · · Θabll EOb1 ...bl ∗ (−τ, x) ≡ Θ EOa1 ...al ∗ (τ, x) ,

where Θab ≡ δba − 2δ0a δb0 , so the time-reflection operation Θ in addition to

performing (τ ↦→ −τ ) includes a factor of (−1)N⊥ with N⊥ being the number
of time components — indices perpendicular to the reflection plane τ = 0.

Remark 1.98. We could freely take all indices to be the upper ones, since we are
in the Euclidean signature. ⌟
Chapter 1. Quantum Field Theory Formalism 34

Remark 1.99. Such conjugation is in particular consistent for operators whose

indices come from derivatives, because
(︂ )︂ (︂ )︂ (︂ )︂†
∂τ E
O(τ, x)† = ∂τ O∗ (−τ, x) = −E∂τ O∗ (−τ, x) =
E E
∂τ O(τ, x) ,
(︂ )︂ (︂ )︂ (︂ )︂†
∂xi EO(τ, x)† = ∂xi EO∗ (−τ, x) = E
∂xi O∗ (−τ, x) = E
∂xi O(τ, x) .
⌟

Remark 1.100 (Conjugation of Conserved Charges). One needs to be slightly careful

when applying Euclidean conjugation to conserved charges, which are nonlocal
operators. However, since they are integrals of local energy–momentum tensor,
we can calculate (for simplicity we consider flat space Rd )
(︃ ∫︂ )︃† ∫︂ (︂ )︂
E
Q†ξ ≡ − dd−1 x ET 0b (τ, x) ξ b (τ, x) =− dd−1 x −ET 0a (−τ, x)Θab ξ b (τ, x)
∫︂ τ =const. (︂ )︂
τ =const.

=− d d−1
x T a (τ, x)
E 0
−Θab ξ b (−τ, x) ≡ EQ−AdΘ ξ ≡ EQ−ΘξΘ .
τ =−const.

where the steps are given as

(1) definition of Euclidean charge EQξ ,
(2) reality of T•• and ξ • , and Definition 1.97 → p.33 which gives a minus sign for
the upper index “0”, and Θ•• for the lower one,
(3) exchange (τ ↦→ −τ ) and simple rearrangement,
(4) conserved charge being topological, it does not depend on the choice of τ
“const.”, and definition of AdΘ ξ ≡ Θ∗ ξ,
(5) understanding both ξ and Θ as operators on functions, and Θ−1 = Θ.
⌟

Example 1.101 (Conjugation of Momentum Operators). Either by using the previ-

ous Remark 1.100 → p.34 for isometry generators — AdΘ ∂τ
∂
= − ∂τ
∂
and AdΘ ∂x
∂
i =
∂
∂xi
— or by writing ∫︂
E
Pa ≡ − dd−1 x ET 0a (τ, x) ,
τ =const.

and applying Definition 1.97 , it follows that EP0 is Hermitian and EPi are
→ p.33

anti-Hermitian. So by identifying
E
P0 ≡ H ≡ LP0 , E
Pi ≡ −ı̊ LPi ,

we recover the standard Lorentzian (Hermitian) momentum operators, together

with agreement of translation formulas (for scalar operator)
E E L L
E
O(x) = ex P EO(0) e−x P ⇐⇒ E
O(τ, x) = eτ H−ı̊x P LO(0, 0) e−τ H+ı̊x P ,
• • • •

the first using directly the Euclidean generators of translations, and the second
obtained by Wick rotation of the standard Lorentzian formula. ⌟
Chapter 1. Quantum Field Theory Formalism 35

Remark 1.102. From the Lorentzian point of view the time-direction is clearly
special, and this is reflected in the way Euclidean conjugation acts. If we take
as a starting point the Euclidean theory, there is initially no preferred direction.
It only appears when we choose a direction we call “time” — we conventionally
label the associated index with “0” — and attempt to quantize it in a way such
that it can be analytically continued to a unitary Lorentzian theory. ⌟

Unitarity and Reflection Positivity

We have seen that different foliations lead to different Hilbert spaces, and to
possibly different notions of conjugation. As shown in previous subsection, when
we Wick-rotate a Lorentzian theory to the Euclidean signature, we obtain a specific
type of conjugation which reflects the time direction.
Can we also turn this around? Given a Euclidean path integral, how do we know
it calculates Wick-rotated correlators of a unitary Lorentzian theory?
The necessary condition for a unitary theory is that the norms ⟨ψ|ψ⟩ for every
ψ ∈ H must be non-negative. It is reflected in the Euclidean theory by the
so-called reflection positivity condition.

Definition 1.103 (Reflection Positivity). Let P be a polynomial of Euclidean

operators EO (possibly smeared) in a half-space with τ < 0. Then we say that
the Euclidean theory is reflection positive if all correlation functions satisfy
⟨︂ ⟩︂
Θ[P ∗ ] P ≥ 0 .

Remark 1.104 (Lorentzian Positivity). The Lorentzian counterpart of positivity is

simply ⟨P ∗ P ⟩ ≥ 0, where P is build out of LO. ⌟

Proposition 1.105 (Unitarity Leads to Reflection Positivity). A Euclidean

theory obtained from a Wick-rotated unitary Lorentzian theory is reflection
positive.

Proof. Consider a state given by a time-ordered product of Euclidean operators

acting on the vacuum
|ψ⟩ ≡ EO1 (x1 ) · · · EOn (xn )|vac⟩ ≡ P |vac⟩
with 0 > τ1 ≥ · · · ≥ τn . Since Definition 1.97 → p.33 gives us
(︂ )︂†
⟨ψ| ≡ E
O1 (x1 ) · · · EOn (xn )|vac⟩
[︂ ]︂ [︂ ]︂
= ⟨vac|Θ EOn∗ (xn ) · · · Θ EO1∗ (x1 ) ≡ ⟨vac|Θ[P ∗ ] ,
we obtain the positivity reflection condition for the correlator as
⟨︂ ⟩︂ !
0 ≤ ⟨ψ|ψ⟩ = ⟨vac| Θ[P ∗ ] P |vac⟩ ≡ Θ[P ∗ ] P ≥ 0 .
Note that inside VEV was a correctly time-ordered product of operators.
Chapter 1. Quantum Field Theory Formalism 36

Remark 1.106 (Reconstruction Theorems). For a unitary QFT in a flat space, the
famous reconstruction theorems assert that the correlation functions — called
Wightman distributions in the Lorentzian Minkd ≡ R1,d−1 and Schwinger functions
in the Euclidean Rd — completely determine the quantum theory, including its
Hilbert space.
The positivity requirements mentioned above are absolutely crucial for such
reconstruction (in addition to other technical assumptions), otherwise one could
not invert the Wick-rotation to obtain a local unitary QFT in Lorentzian signature,
as there would be no probability interpretation for states in H [20] [21] [22].⌟
As unitarity and reflection positivity in the respective signatures are essentially
equivalent, we will often refer to both as “unitarity”, even though most of the
time we will be working in the Euclidean signature. Some of their implications in
the context of CFTs will be discussed in Section 3.4 → p.78 and Section 3.5 → p.86 .
Finally, let us remark that — at least on a formal level — the Euclidean path
integral is reflection-positive.
Remark 1.107 (Reflection Positivity of Euclidean Path Integral). Using the locality as
in Idea 1.45 → p.14 , we can split the path integral into positive (τ > 0) and negative
(τ > 0) Euclidean “time” parts. Assuming that the Euclidean path integral
E

measure Dϕ e− S[ϕ] is invariant under the combined action of the time-reflection

and complex conjugation of all fields ϕ, we can write
∫︂ ∫︂ ⃓ ∫︂ ⃓2
⟨︂ ⟩︂ E
⃓ E
⃓
Θ[P ] P =
∗
Dϕ e − S[ϕ]
P [Θϕ ]P [ϕ] =
∗
Dϕ0 ⃓⃓
⃓
Dϕ− e − Sτ <0 [ϕ− ]
P [ϕ− ]⃓⃓
⃓
≥ 0.
Fields Fields τ =0 Fields τ <0 , ϕ− |τ =0 =ϕ0

Alternatively, since the action is real, we can separate and split the interaction
part of the action as (again the locality is crucial)
E int [ϕ] E int E int E int
e− S = e− Sτ >0 [ϕ] e− Sτ <0 [ϕ] ≡ Θ[F ∗ ] F , with F ≡ e− Sτ <0 [ϕ] ,

giving us (assuming the reflection positivity of the free theory)

⟨︂ ⟩︂ ⟨︃ [︂ ]︂ ⟩︃
Θ[P ] P = Θ (F P ) (F P )
∗ ∗
≥ 0.
free

Of course, while this is formally true for local Euclidean theories, difficulties
can (and do) arise when one tries to turn these expressions into well-defined
quantities. ⌟
 37

2
Beyond Basic Perturbation Theory
It is often the case that a “basic” perturbation theory — Feynman diagrams to
some finite order — is not enough to describe the phenomena we are interested in.
Alternatively, we want to know which properties are present in the full theory,
and perturbation theory is sometimes able to give us some hints, but more often
falls short of giving us an adequate picture.
In this chapter we will only briefly comment on some notions and concepts which
are useful in this context, and for understanding QFT in general. For much more
thorough accounts and deeper insights we refer the reader to the literature, for
example [5] [23] [24] [25] [26] [27] [28].

2.1 Effective Action

In the classical theory, the action S directly governs the dynamics of the system,
in particular the classical fields ϕcl must extremize the action. Things complicate
in the quantum theory, where we need to perform path integrals over all possible
field configurations.
In a free theory, or in the classical limit ℏ → 0, the path integral is dominated by
the classical solutions, and we can directly use the classical action S to analyze
the system. However, for an interacting QFT (outside the classical limit), we
encounter non-classical contributions to Z from “quantum fluctuations”, which
can have huge impact on the physics.
Is there a way to package (at least partially) quantum effects into some kind of
effective action, for which we can again use a classical intuition? Access to such
an object would allow us to properly study various aspects of QFTs — the phase
structure, symmetries, or renormalizability.

Different Types of Effective Actions

Depending on the context, the term “effective action” can mean different things:
(1) Wilsonian low-energy effective action — We integrate out the “heavy” fields
(or high-momentum field modes), leaving us with an effective action for the
remaining degrees of freedom.
Such effective action contains contributions from the high-energy physics,
and defines an Effective Field Theory (EFT) valid only for the description of
low-energy phenomena. We will discuss this more in Section 2.2 → p.41 .
(2) Quantum (1PI) effective action — We “integrate out the loops”, and obtain
effective action including the quantum corrections.
 Chapter 2. Beyond Basic Perturbation Theory 38

By examining its form, we can read out exact/full propagators and 1PI
vertices, reducing the calculation to drawing only “classical” tree graphs. We
focus on this type of effective action in the following.
Remark 2.1. As we can see, effective actions correspond to a certain partial
evaluation of the path integral. For interacting theories, even such partial path-
integration is usually not feasible to perform exactly, and we have to resort to
some perturbation scheme. ⌟

Quantum Effective Action

For discussing the phase structure — and possible spontaneous symmetry breaking
— we ultimately want to know whether field ϕ obtains nonzero VEV, so we want
to study the expectation values of the fields. It turns out to be fruitful to take
into consideration the presence of the sources J .

Definition 2.2 (Expectation Values). The expectation values Φ of the fields ϕ

in the presence of sources J are given by
⟨︂ ⟩︂ ℏ 1 δ δW[J ]
ΦJ (x) ≡ ϕ(x) = Z[J ] ≡ .
J ı̊ Z[J ] δJ (x) δJ (x)

Our goal is to obtain Quantum Effective Action Γ , which would govern Φ similarly
to how S governs ϕcl . We thus promote ΦJ to a primary variable Φ and demote
J to a derived one JΦ , which is (implicitly) defined such that
δW δW ⃓⃓
⃓
ΦJ = Φ=
δJ δJ ⃓J =JΦ
holds.
Inspired by similar scenarios in
Legendre Transform
• Classical Mechanics — Lagrangian L velocities q̇ ←→ p momenta
H Hamiltonian ,

Legendre Transform
• Thermodynamics — Helmholtz Energy F G Gibbs Energy ,
volume V ←→ p pressure

we come to the following definition.

Definition 2.3 (Quantum Effective Action). The Quantum Effective Action Γ

is defined by the following relation
Legendre Transform
Connected Generating Functional W expectation
Γ Quantum Effective Action ,
sources J ←→ Φ values

that is (︂ )︂ ⃓⃓
Γ [Φ] = W[J ] − J •Φ ⃓ ,
J =JΦ : Φ= δW
⃓
δJ

or conversely
(︂ )︂ ⃓⃓
W[J ] = Γ [Φ] + J •Φ ⃓ .
δΓ
⃓
Φ=ΦJ : J =− δΦ
Chapter 2. Beyond Basic Perturbation Theory 39

Remark 2.4. As is usual — both in the context of classical mechanics and

statistical physics/thermodynamics — in the following we will oftentimes omit the
dependence of the variables, since it will be clear from the context if we should
express J as function of Φ, or vice versa. ⌟
Remark 2.5. As is usual for Legendre Transform, we have (using the chain rule)
(︄ )︄
δΓ δW δJ δJ
= • − •Φ −J = 0−J ,
δΦ δJ δΦ δΦ

so in the absence of sources J = 0, we obtain equation for Φ as a stationary point

of Γ . This is one sense in which Γ is seen as the quantum counterpart of S. ⌟

1PI Effective Action

Can we take this analogy even further? It turns out interesting to take Γ as a
full-fledged classical action on its own, and consider the corresponding quantum
theory given by ∫︂
ı̊ ı̊
e k̄
WΓ [J ]
≡ DΦ e k̄ (Γ [Φ]+J • Φ) ,
where we keep the “Planck constant” k̄ distinct from ℏ to avoid mixing them,
since ℏ is already present in Γ .

Proposition 2.6 (Quantum Effective Action as 1PI Effective Action). The

quantum effective action Γ is also the One-Particle-Irreducible (1PI) effective
action, meaning that W is (perturbatively) given by all connected tree graphs,
with vertices and propagators given by Γ . Schematically, we have
∫︂ ∫︂
ı̊
ı̊W[J ] ≃ DΦ e ı̊(Γ [Φ]+J • Φ)
⇐⇒ ı̊
ℏ
Γ [Φ] ≃ Dϕ e ℏ S[ϕ+Φ] .
connected connected
tree graphs 1PI graphs

Idea of the Proof. See any of [26] [27] [5] [10] for the detailed proof. The
idea is that connected graphs contributing to WΓ [J ], see Proposition 1.38 → p.11 ,
have the following dependence on k̄ — propagator is inverse of the quadratic term
in k̄1 Γ and hence proportional to k̄, while each interaction vertex carries a factor
of 1/k̄ from the k̄ı̊ in front of Γ — so a graph with I internal lines and V vertices
is proportional to k̄ I−V = k̄ L−1 , where L = I − V + 1 is the number of loops for a
connected graph.
(0)
The leading contribution WΓ ∝ k̄ 0 for k̄ → 0 is thus given by the connected
tree graphs, but alternatively it is given by the stationary-phase/saddle-point
approximation of the path integral, giving us
(︂ )︂ ⃓⃓
(0)
WΓ [J ] = Γ [Φ] + J •Φ ⃓ ≡ W[J] ,
δΓ
⃓
Φ=ΦJ : J =− δΦ

where we recognized the Legendre transform relation between Γ and W. Since any
connected graph can be understood as a tree graph made with vertices representing
Chapter 2. Beyond Basic Perturbation Theory 40

1PI graphs, the interpretation of Γ as the 1PI effective action follows.

Remark 2.7 (1PI Vertices, Exact Propagator). In the light of Proposition 2.6 → p.39 ,
we can see that Γ is the quantum corrected counterpart of S, having for the
interactions the renormalized 1PI vertices, and for the propagator the exact/full
propagator. Indeed, using Remark 2.5 → p.39 and Definition 2.2 → p.38 we have
)︄−1 )︄−1
δ2Γ δ2W
(︄ (︄
δJ δΦ
=− =− =−
δΦδΦ δΦ δJ δJ δJ

so the quadratic part of Γ is the inverse of the full propagator

)︄−1 ⃓
ℏ δ 2 W ⃓⃓ ı̊ δ 2 Γ
⃓ (︄
⟨︂ ⟩︂
= =
⃓
ϕϕ − ⃓ .
conn ı̊ δJ δJ ⃓J =0 ℏ δΦδΦ ⃓
Φ=0
⌟
Remark 2.8 (Energy Interpretation, Effective Potential). There is also very important
energy interpretation of Γ . For constant Φ = Φ0 we can write

Γ [Φ0 ] = −VM V (Φ0 ) ,

where we factored out the volume VM ≡ vol(M) of the spacetime, and V (Φ0 )
is the effective potential — for constant Φ = Φ0 we are only left with potential
terms in the effective Lagrangian.
It can be then shown that the value of V (Φ0 ) is the minimum of the expectation
value of the energy density among all states satisfying ⟨ϕ⟩Ω ≡ ⟨Ω|ϕ|Ω⟩ = Φ0 . The
ground state (vacuum) is thus found by minimizing the effective potential V .
Say a minimum is at Φg . To avoid working with linear “tadpole” terms in the
effective action Γ (or fields with nonzero VEV), the calculations naturally proceed
in terms of the shifted fields ϕ˜︁ ≡ ϕ − Φg . If the minimum Φg is not invariant
under some symmetry of the original action S, we say that the symmetry is
spontaneously broken. ⌟
Remark 2.9 (Slavnov–Taylor Identities). Consider a continuous symmetry transfor-
mation generated by (F depends locally on ϕ)

ϕ ↦−→ ϕ′ ≡ ϕ + εF [ϕ] ,

which leaves the (path-integral measure×action) invariant, that is

ı̊ ′ ı̊
Dϕ′ e ℏ S[ϕ ] = Dϕ e ℏ S[ϕ] .

Then we have
∫︂ ∫︂
ı̊ ı̊
(S[ϕ′ ]+J • ϕ′ )
Z[J ] = Dϕ e ′ ℏ = Dϕ e ℏ (S[ϕ]+J • ϕ+εJ • F [ϕ])
∫︂
ı̊
(︂ )︂
= Dϕ e ℏ (S[ϕ]+J • ϕ) 1 + ε ℏı̊ J • F [ϕ] + . . .
(︂ ⟨︂ ⟩︂ )︂ ⟨︂ ⟩︂
≡ Z[J ] 1 + ε ℏı̊ J • F [ϕ] + ... =⇒ J • F [ϕ] = 0.
J J
Chapter 2. Beyond Basic Perturbation Theory 41

Setting J = JΦ , and using Remark 2.5 → p.39 , we can write

δΓ ⟨︂ ⟩︂
• F [ϕ] = 0.
δΦ JΦ

In other words, the effective action Γ [Φ] is invariant under the transformation
⟨︂ ⟩︂
Φ ↦−→ Φ′ ≡ Φ + ε F [ϕ] .
JΦ

Such symmetries are known as Slavnov–Taylor identities.

⟨︂ ⟩︂
In the important special case where F is linear, we have F [ϕ] = F [Φ], and
JΦ
the classical symmetry is (at least formally) preserved also at the quantum level.
Quantum corrections in such a case do not generate terms in Γ which would break
the symmetry, so we do not have to worry about renormalizing them. Since we
can remove all ultraviolet divergences by adding appropriate counterterms before
we shift the fields — see Remark 2.8 → p.40 — the possible spontaneous symmetry
breakdown does not affect the divergence structure of the theory.
For non-linear F , the obtained symmetry of Γ is not generally the same as the
original symmetry of S. One important example is the BRST transformation
encountered in gauge theories, which complicates the analysis of their renormal-
ization. ⌟

2.2 Renormalization Flow

Since we do not have access to arbitrary energies to probe the high-energy “ultra-
violet” (UV) physics, our experience mostly lies in the large-distance “infrared”
(IR) domain. In the following we will look on the big picture ideas concerning
effective description of such long-distance physics.
Idea 2.10 (Renormalization Group Framework). A common idea of Renormalization
Group techniques is the following:
(1) have a system, and identify the physical aspects of interest,
(2) re-express the parameters in terms of another set — usually we choose simpler
ones by some kind of coarse-graining, thus decimating certain degrees of
freedom — while keeping the physical aspects of interest unchanged,
(3) study how the description of the system changes under the corresponding
Renormalization Flow, and extract information about the desired physical
phenomena.
⌟

Effective Field Theory Picture

It is unreasonable to expect that any theory at our current disposal is accurately
applicable at arbitrarily high energy/momenta, or equivalently arbitrarily small
distances.
Chapter 2. Beyond Basic Perturbation Theory 42

Unavoidably, every (realistic) model works with particular set of relevant degrees
of freedom (DOF), while the rest can at most induce effective interactions between
our chosen DOF. Thus, it is natural that our effective theory has finite domain of
applicability, beyond which it is not able to encompass all interesting/relevant
phenomena.

Remark 2.11. Why is it only in QFT that such detailed discussion is necessary?
Why did not questions of similar nature appear already in Classical Mechanics,
Relativity, or ordinary Quantum Mechanics? The answer lies in the unique
combination of crucial aspects present in Relativity and Quantum Mechanics:
• Classical Mechanics (CM) — There is not much to discuss, since our system
is described precisely by its location in the phase space, and the evolution is
deterministic — not really space for crazy stuff.
• Relativistic Mechanics (RM) — new aspect enters — Mass is just a form
of energy, allowing us to create and destroy particles. Still, energy being
conserved, we have only so much options, and we are not bothered by any
DOF lying (energetically) far beyond our reach.
• Quantum Mechanics (QM) — new aspect enters — Evolution is not deter-
ministic in the strong sense of CM, since from path integral point of view all
possible evolutions contribute in certain sense. Yes, the physical language and
intuitions drastically change, nonetheless, choosing basically any particular
QM system we obtain a self-consistent description.
• Quantum Field Theory (QFT) — Only when we combine RM and QM we
face a fundamental difficulty. By QM our system goes through all available
processes, which by RM can include producing particles of essentially arbitrary
type (supposing the fields are not in totally decoupled sectors). We thus
necessarily encounter cross-linking and interaction of various DOF, which
would be otherwise (classically) unreachable. ⌟

After reading the last point, it could seem hopeless to obtain any reasonably
predictive theory without “knowing everything”. This is in stark contrast with
our experience, where even idealized physical models prove immensely useful.

This is rooted in the fact that vast amount of high-energy physics effects (say at
energy scales higher than Λ) essentially decouples. Certain aspects of this can be
illuminated in the framework colloquially known as renormalization theory, to be
discussed below.

Remark 2.12 (Lattice Models, Condensed Matter Physics). All of this is conceptu-
ally clear in lattice models or condensed matter physics, where the field theory
description is effective from the start, and valid only on length scales |x| ≫ a
much larger than the lattice spacing a. ⌟
Chapter 2. Beyond Basic Perturbation Theory 43

Wilsonian Effective Action

We will now look in more detail on how the renormalization idea Idea 2.10 → p.41
plays out in (flat-space) QFT. For better contact with statistical physics we work
throughout in Euclidean signature with ℏ ≡ 1, and from start we impose a UV
cut-off Λ0 .
Including appropriate sources J in the action S Λ0 , the partition function
∫︂
Λ0 [ϕ]
Z≡ [Dϕ]k<Λ0 e−S

carries all information about the theory for energies/momenta k ≡ |k| < Λ0 , or
equivalently for distances |x| > L0 ≡ Λ−10 . If we are interested in low-energy
large-distance physics, we can try to simplify our description by integrating out
the high-energy degrees of freedom.

Definition 2.13 (Wilsonian Effective Action). Let Λ < Λ0 be a lower cut-off,

and split the fields into long-wavelength/low-momentum modes ϕ− with
k < Λ, and short-wavelength/high-momentum modes ϕ+ with k > Λ. Since
∫︂ ∫︂ ∫︂
Λ0 [ϕ] Λ0 [ϕ− +ϕ+ ]
Z= [Dϕ]k<Λ0 e−S = [Dϕ− ]k<Λ [Dϕ+ ]Λ<k<Λ0 e−S ,

the Wilsonian effective action defined by

∫︂
Λ [ϕ− ] Λ0 [ϕ− +ϕ+ ]
e −Seff
≡ [Dϕ+ ]Λ<k<Λ0 e−S

describes the same physics of the low-momentum modes ϕ− as the original

action S Λ0 .

Remark 2.14 (Partial Trace Analogy, Real-Space Coarse-Graining). In lattice models

we naturally encounter real-space variant of the above. By grouping multiple
lattice sites into blocks, we coarse-grain by partially tracing over the internal
degrees of freedom inside the block. Only thing left is a block variable describing
the overall/mean/dominant state of the block. ⌟
Remark 2.15 (Locality of Effective Action). Integrating out high-momentum modes
ϕ+ generally introduces non-local interactions for ϕ− . However, these are sup-
pressed by the high-momentum scale Λ — coming from their propagator — so
as long as we want effective low-energy description for E ≪ Λ, we can write the
effective Lagrangian as a series in “E/Λ” of local terms.
The same happens in lattice models. Starting with nearest-neighbor interactions,
coarse-graining into blocks introduces next-to-nearest–neighbor interactions. ⌟
Remark 2.16 (Wilson Renormalization Group). We can iterate this procedure, and
go to a still lower scale Λ′ < Λ — it yields the same effective action as going
directly from Λ0 to Λ′ . For this reason it is known as the renormalization group,
even though “renormalization semigroup” would be more appropriate. ⌟
Chapter 2. Beyond Basic Perturbation Theory 44

Flow in the Space of Couplings

Having fixed the underlying degrees of freedom, even if we start with a particular
set of couplings, the effective theory (after integrating out heavy modes) will have
different couplings. In general, it can generate all possible terms compatible with
symmetries of the theory.

Definition 2.17 (Space of Couplings, Most General Effective Action). Fixing a

set of relevant degrees of freedom ϕ, we consider the most general effective
action (respecting the locality and symmetries) in the form
[︄ ]︄
∫︂
ci
ScΛ0 [ϕ] = d x Lkin [ϕ] +
d
Oi [ϕ] .
∑︂
M i Λd0i −d

As such, we have allowed all possible local (Lorentz-invariant) operators Oi

of classical mass dimensions di . For convenience, we have included explicit
factors of Λ0 such that the couplings c = {ci } are dimensionless.

With such generic initial action (for simplicity we consider a single real scalar ϕ),
the Wilsonian effective action for some cutoff Λ will be of the form
∑︂ ci (Λ) n /2
[︄ ]︄
∫︂
ZΛ
Λ
Seff [φ] = d x (∂φ)2 +
d
ZΛi Oi [φ] ,
M 2 i Λ d i −d

where the field renormalization factor ZΛ accounts for the possible corrections to
n /2
the kinetic term. We also introduced a factor ZΛi before Oi ∝ φni , such that in
terms of the renormalized field
1/2
ϕ ≡ ZΛ φ

we have again a canonically normalized kinetic term, and the whole effective
action of the form Seff
Λ
[φ] = Sc(Λ)
Λ
[ϕ] as defined in Definition 2.17 → p.44 .

By Definition 2.13 → p.43 , changing the cutoff scale together with the appropriate
“running” of the couplings leaves the low-energy physics unchanged.

Definition 2.18 (Renormalization Flow, Beta Functions). The renormalization

flow of the couplings

c(Λ) ↦−→ c(Λ′ ) ≡ RΛ′ ←Λ [c(Λ)] ,

is infinitesimally generated by the so-called β-functions

(︂ )︂ ∂ci (Λ) ∂
βi c(Λ) ≡ Λ ≡ ci (Λ) ,
∂Λ ∂ ln Λ
which can be viewed as a vector field β on the space of couplings.
Chapter 2. Beyond Basic Perturbation Theory 45

Remark 2.19. For the dimensionless couplings {ci } it is convenient to write

(︂ )︂ (︂ )︂
βi c(Λ) = (di − d)ci (Λ) + βiquant c(Λ) ,

where the first term compensates the explicit powers of Λ in Definition 2.17 → p.44 ,
and the second term represents the quantum effect of integrating out the high-
momentum modes.
One can see that we have β quant = 0 in a free (massless) theory with c = 0, since
the high-momentum modes are completely decoupled from the low-momentum
ones. Thus, β = 0, and the couplings are constant. We will discuss such fixed
points more generally in a moment. ⌟
Suppose we want to calculate n-point correlators inserted at well separated points
x1 , . . . , xn ∈ M. Since high-energy modes do not play much of a role, we can use
the effective action at some scale Λ as
⟨︂ ⟩︂ 1 ∫︂ −S Λ [Z
1/2
φ]
φ(x1 ) · · · φ(xn ) ≡ [Dφ]k<Λ e c(Λ) Λ φ(x1 ) · · · φ(xn ) ,
Λ,c(Λ) Z
where we allowed the possibility that we did not yet canonically normalize the
1/2
field. For the renormalized field ϕ ≡ ZΛ φ we define renormalized correlation
functions
⟨︂ ⟩︂ ⟨︂ ⟩︂
(n) n/2
ΓΛ,c(Λ) (x1 , . . . , xn ) ≡ ϕ(x1 ) · · · ϕ(xn ) ≡ ZΛ φ(x1 ) · · · φ(xn ) .
Λ,c(Λ) Λ,c(Λ)

If the field insertions are really far apart — equivalently involve modes with
energies E ≪ Λ — we would obtain the same (nonrenormalized) correlators if we
first integrated out modes in the range (sΛ, Λ] for some s < 1. For an arbitrary
configuration of (distinct) points, we can choose s sufficiently close to 1 such that
effect of integrating out the high-momentum modes is still negligible.

Proposition 2.20 (Callan–Symanzik Equation, Anomalous Dimension). The

renormalized correlation function obey the Callan–Symanzik equation
(︄ )︄
∂ ∂ (n)
Λ +β + nγϕ ΓΛ,c(Λ) (x1 , . . . , xn ) = 0 ,
•

∂Λ ∂c

where we defined the anomalous dimension of the field ϕ as

−1/2
1 ∂ ln ZΛ ∂ ln ZΛ
γϕ (Λ) ≡ − Λ ≡ .
2 ∂Λ ∂ ln Λ

Proof. From the discussion above we have (s infinitesimally small)

−n/2 (n) ! −n/2 (n)

⟨︂ ⟩︂
φ(x1 ) · · · φ(xn ) = ZΛ ΓΛ,c(Λ) (x1 , . . . , xn ) = ZsΛ ΓsΛ,c(sΛ) (x1 , . . . , xn ) ,
Λ,c(Λ)

and the result is obtained by performing d

|
ds s=1
and using the chain rule together
with the definition of β and γϕ .
Chapter 2. Beyond Basic Perturbation Theory 46

Scale-Dependence of Theories
Up to now we discussed effective descriptions (with different UV cutoffs) of the
same low-energy theory. As we will see shortly, this is connected to the behavior
at different length scales.
Remark 2.21 (RG Flow and Scaling). Suppose that after integrating out modes in
the range (sΛ, Λ] as above, we additionally perform a scaling transformation

x ↦−→ x′ ≡ sx , dd x ↦−→ dd x′ = sd dd x ,
Λ ↦−→ Λ′ ≡ 1s Λ , ∂ ↦−→ ∂ ′ = 1s ∂ ,
d−2
ϕ ↦−→ ϕ′ (sx) ≡ s−dϕ ϕ(x) ≡ s− 2 ϕ(x) ,

under which the action in Definition 2.17 → p.44 is invariant. Therefore, we have
[︄ ]︄n/2
x1 xn ZΛ x1 xn
(︃ )︃ (︃ )︃
(n) (n)
ΓΛ,c(Λ) ,..., = ΓsΛ,c(sΛ) ,...,
s s ZsΛ s s
[︄ ]︄n/2
ZΛ d−2 (n)
= s 2
n
ΓΛ,c(sΛ) (x1 , . . . , xn ) ,
ZsΛ

where in the second line we just performed the scaling transformation, which does
not affect the couplings c(sΛ).
In this way we ended up with the same UV cutoff Λ, and related correlation
functions at different scales. When s → 0, the left-hand side probes correlations
at longer and longer distances, while the distances on the right-hand side are kept
constant, but (apart from the prefactor) we let the coupling run to the IR.
This is what we would expect — the IR behavior is governed by the low-energy
effective theory (and corresponding couplings which survive) obtained after we
integrate out all of the high-energy modes. ⌟

Remark 2.22 (Interpretation of Anomalous Dimension, Scaling Dimension). For

each field, the correlator above in Remark 2.21 → p.46 has a scaling factor on the
right-hand side, which for s = 1 + δs with |δs| ≪ 1 is approximately
]︄1/2
d −2
[︄ (︄ )︄
ZΛ d−2
s 2 ≈1+ + γϕ δs + · · · .
ZsΛ 2

The scaling is thus determined by the scaling dimension of the field ϕ defined as
d −2
∆ϕ ≡ + γϕ = dϕ + γϕ ,
2
so in addition to the classical mass dimension dϕ = (d − 2)/2 we also need to
include the anomalous dimension γϕ defined in Proposition 2.20 → p.45 . It generally
depends on the energy scale Λ and the couplings c(Λ), and similarly to β-functions,
it is zero in the free massless theory, see Remark 2.19 → p.45 . ⌟
Chapter 2. Beyond Basic Perturbation Theory 47

Fixed Points
Very special things happen, when the β-functions vanish at some point in the
space of couplings.

Definition 2.23 (Fixed Point). A point c∗ in the space of couplings is called a

fixed point if
β(c∗ ) = 0 =⇒ c∗ (Λ) = const. .
The corresponding theory is then invariant under the renormalization group
flow, or equivalently under the scaling transformations.

Remark 2.24 (Correlation Length, Mass Gap, Critical Point). The (connected)
two-point correlators at large distances typically decay as
⟨︂ ⟩︂ 1 |x−y|
− ξ
ϕ(x)ϕ(y) ∼ ϑ e ,
conn |x − y|

where ξ is the correlation length of the theory, and ϑ is some exponent.

When the theory has a “mass gap” — that is mgap ≡ E1 − E0 > 0, where E0 , E1
are the energies of the ground state and first excited state, respectively — the
correlation length is finite and given by
1
ξ≡ .
mgap

This is because the dominant contribution comes from the propagation of the
lightest particle/state (apart from the vacuum).
The fixed points however can not have any characteristic scale (such as correlation
length or mass), since they would necessarily change under the renormalization
flow or scaling. Only possibilities are thus:
(1) ξ = 0 ⇐⇒ mgap = ∞ — Correlation functions of local operators vanish, and
we are left with a TQFT. This is what typically happens for generic statistical
and condensed matter systems, since microscopic correlation lengths are
typically ξ ∼ a, which is negligible at large distances. It can be understood
as flowing off to the infinity of the coupling space.
(2) ξ = ∞ ⇐⇒ mgap = 0 — For specially tuned couplings it can happen we flow
into a critical point, where correlation functions do not decay exponentially at
large distances, but rather as a power law. This is the case of the second-order
phase transitions.
⌟
Remark 2.25 (Two-Point Function at Criticality). Since c∗ = const. , also the
anomalous dimension γϕ (c∗ ) ≡ γϕ∗ is constant. The scaling in Remark 2.22 → p.46
can be then easily exponentiated as
[︄ ]︄1/2
ZΛ d−2
s 2 = s∆ϕ .
ZsΛ
Chapter 2. Beyond Basic Perturbation Theory 48

Therefore, for the 2-point function we have scaling symmetry Remark 2.21 → p.46 ,
which together with the translation/rotation invariance gives
x y 1
⟨︃ (︃ )︃ (︃ )︃⟩︃ ⟨︂ ⟩︂ ⟨︂ ⟩︂
ϕ ϕ = s2∆ϕ ϕ(x)ϕ(y) =⇒ ϕ(x)ϕ(y) ∼ .
s s |x − y|2∆ϕ ⌟

Alternatively, we could obtain the same result using the Callan–Symanzik equation
Proposition 2.20 → p.45 .
Example 2.26 (Free Theory, Generalized Free Field, Mean Field Theory). As already
mentioned in Remark 2.19 → p.45 and Remark 2.22 → p.46 , the free massless theory
is a fixed point of the renormalization group flow, and thus should have 2-point
function ⟨︂ ⟩︂ 1 1
ϕ(x)ϕ(y) ∼ free ≡ .
|x − y|2∆ϕ |x − y|d−2
Indeed, this is a solution of the free equation of motion ϕ = 0. The higher-order
correlators are then obtained by Wick’s theorem Example 1.33 → p.9 .
In the context of QFT in AdS to be discussed in Chapter 4 → p.91 , we naturally
encounter Generalized Free Fields (GFFs) — they share the same factorization
property of correlators as the free theory, but have a generic (noninteger) scaling
dimension ∆ϕ .
We will use term Mean Field Theory (MFT) for theories built out of GFFs. Fields
in such theories do not obey simple local equations of motion, so they do not
admit a local Lagrangian description. In particular, MFTs do not have a conserved
energy–momentum tensor. ⌟
Remark 2.27 (Weyl and Conformal Symmetry at Critical Point). The same correla-
tion function was already obtained in Example 1.73 → p.25 using a stronger Weyl
symmetry, which contains arbitrary local scaling transformations. In fact, this is
what often happens at critical points — if a theory is local, and also scale invariant
by virtue of being critical, it generically enjoys further symmetry enhancement —
it becomes conformally invariant. In the next Chapter 3 → p.51 , we will study such
Conformal Field Theories (CFT) in more detail. ⌟

Universality
Now we will look at theories near, but not directly at, a fixed point c∗ of RG flow.
Since by definition β(c∗ ) = 0, we can linearize the flow around the fixed point,
and write
∂c ⃓⃓ ∂βi ⃓⃓
⃓ ⃓
Λ ≈ M δc + O(δc ) ,
•
2
Mij ≡ .
∂Λ ⃓c ≡ c∗ +δc ∂cj ⃓c∗
Assuming the matrix M can be diagonalized with eigenvectors ui and eigenvalues
λi ≡ ∆i − d, we can write
)︄∆i −d
Λ
(︄
∂ui
Λ = λi ui + . . . =⇒ ui (Λ) = ui (Λ0 ) + . . . .
∂Λ Λ0
Chapter 2. Beyond Basic Perturbation Theory 49
?
Recalling Remark 2.19 → p.45 , we would classically expect λi = di − d, but in-
teractions usually induce some anomalous dimension γi ≡ ∆i − di . Similarly
to Remark 2.22 → p.46 , ∆i = di + γi is the scaling dimension of the operator Oi
corresponding to the coupling ui .

Definition 2.28 (Relevant, Irrelevant, Marginal Couplings and Operators).

Depending on the sign of λi ≡ ∆i − d, we distinguish three cases:
• λi > 0 ⇐⇒ ∆i > d — the coupling ui is irrelevant, and decays under the
RG flow (lowering the scale Λ). Any such deformation from the fixed
point is washed away at large distances Λ → 0.
• λi < 0 ⇐⇒ ∆i < d — the coupling ui is relevant, and grows under the
RG flow. Deforming the theory in such direction drives it away from the
fixed point as we go into the IR.
• λi = 0 ⇐⇒ ∆i = d — the coupling ui is marginal, and the fate of such
deformation depends on higher-order terms in the β-functions.
Correspondingly, the operators Oi associated to couplings ui are also called
irrelevant, relevant, or marginal.

Example 2.29 (Mass Term). For a free scalar field ϕ, the mass term corresponds to
the operator ϕ2 with the scaling dimension equal to the classical mass dimension
d −2
∆ϕ2 = dϕ2 = 2 = d − 2.
2
This can be seen from Remark 2.19 → p.45 , where for free theory we have βϕquant
2 = 0,
and only the classical scaling due to explicit power of Λ appears

∂m2
Λ = (dϕ2 − d)m2 = −2m2 .
∂Λ
Clearly, it is a relevant deformation, so as we lower the energy scale Λ, we flow
away from the free Gaussian critical point with m2 = 0. Eventually we reach
trivial theory with mgap = ∞ — see Remark 2.24 → p.47 — since every massive
mode gets integrated out, and we are left with the vacuum. ⌟
Remark 2.30 (Finite Number of Relevant Operators). Since each derivative or
new field adds to the dimension of the operator — anomalous dimensions do
not overwhelm the classical contribution — there can be only finitely many (and
typically just a few) relevant operators.
As a consequence, even though a generic (effective) QFT is parametrized by in-
finitely many couplings, the vast majority is irrelevant and gets quickly suppressed
in the IR. This is the reason why Effective Field Theories are really effective,
as only a handful of couplings govern the low-energy physics. The rest gives
subleading corrections, and if needed, they can be systematically accounted for.⌟
Chapter 2. Beyond Basic Perturbation Theory 50

Remark 2.31 (Universality, Critical Exponents). Another facet of the same idea is
the universality of the critical points. Many systems differing in their microscopic
description — when appropriately tuned — flow to the same critical/fixed point.
This explains why critical behavior at second-order phase transitions (captured by
the critical exponents) is shared by many systems. Depending on the fixed point,
they can be categorized into universality classes. Theories in the same universality
class are described by the same CFT and by the same few relevant operators,
which is also reason for various relationships between the critical exponents. ⌟
Example 2.32 (Ising Model in d = 3, Liquid–Vapor Transition, φ4 Theory). In
d = 3, the Ising model (uniaxial magnet) at critical temperature, critical point of
vapor–liquid transition, and critical ϕ4 theory, all fall into the same universality
class.
While the Z2 symmetry of the Ising model and ϕ4 theory is manifest microscopically,
for the liquid–vapor it is only emergent at the critical point. On the other hand,
water (or ϕ4 theory) has full rotational symmetry, which is not manifest for the
lattice Ising model. ⌟
 51

3
Conformal Field Theory
The importance of Conformal Field Theories (CFTs) in both physics and mathe-
matics can be hardly overstated. Let us briefly mention some of the connections
and motivations to study them (in no particular order):
• At critical points of statistical systems — second order transitions — the
correlation length diverges, and theories become scale-invariant. By this we
mean that fluctuations and their correlations are no longer exponentially
suppressed at long distances, but follow power-laws characteristic to critical
phenomena.
Furthermore, it generically happens that local theories enjoy an additional
symmetry enhancement — they become conformally invariant, and are thus
described by Euclidean CFTs. Note that emergence of such symmetries at
large distances is rather nontrivial for discrete models — imagine the Ising
model on a square lattice — as they initially do not even posses symmetry
under arbitrary rotations.
• Connected to the previous point, the fixed points of the RG flow are usually
described by CFTs. Being such special landmarks in the space of QFTs, they
are important for their study — oftentimes, we can understand a UV-complete
QFT as an RG flow CFTUV CFTIR triggered by deforming the CFTUV .
• Having a larger symmetry group (compared to standard isometries), we obtain
stronger constrains on the theory. This makes non-perturbative “bootstrap”
approaches based mainly on the symmetry and consistency of the underlying
theories much more potent. Such program is in particular successful for d = 2,
where the conformal algebra becomes infinite-dimensional, sometimes allowing
to completely solve particular classes of theories [29].
• String theory falls under the purview of two-dimensional CFTs, since in the
standard formulation it is “just” a specific σ-model on a two-dimensional
CFTworldsheet .
• There are also fruitful relations to TQFT, which can be exploited to study
topological invariants [30].
• Last, but not least, there is the AdS/CFT correspondence, also known as
holography or gauge/gravity duality. The original formulation by Maldacena
[31] related a theory with dynamical gravity in asymptotically AdS spacetime
to a non-gravitational CFT in one less dimension “living” on the boundary.
Since then, it has become much more general.
Even without dynamical gravity, a QFT in fixed AdS background admits
asymptotic observables (boundary correlation functions) satisfying basically
all of the CFT axioms.
 Chapter 3. Conformal Field Theory 52

Remark 3.1 (Gapped QFT vs CFT). The language and methods used to study
CFTs are very different from the ones used for QFTs with a mass gap (or QFTs
with finitely many massless particles in the IR).
In the latter case, due to the mass gap, widely separated particles cease to interact.
We can thus define asymptotic states behaving like free, non-interacting particles,
and study their scattering amplitudes/S-matrix (with appropriate care, it is also
possible to treat massless particles).
On the other hand, in CFTs, it is impossible to disentangle the continuum of
massless excitations — interactions never “turn off” completely at large distances
— so we cannot define asymptotic states and the S-matrix. Instead, we are mainly
interested in the correlation functions of local operators, and utilize the stronger
conformal symmetry to study their structure. ⌟

Remark 3.2 (Coleman–Mandula Theorem). Considering flat space, the spacetime

and internal symmetries of gapped interacting theories (nontrivial S-matrix) can
only combine in a trivial way — the statement of Coleman–Mandula theorem is
that the symmetry group is a direct product G = GPoincaré × Ginternal .
Notable “exceptions” are supersymmetry — includes Grassman (anticommuting)
charges — and conformal symmetry. Thus, in a certain sense, CFTs provide
examples of most symmetric, yet interacting theories. ⌟
To get a good grasp on CFTs, in Section 3.1 → p.52 we will first study the conformal
structure of manifolds in general. Afterward, we focus on flat spaces specifically in
Section 3.2 → p.61 . Initial implications of the conformal symmetry for the correlation
functions of local operators in CFTs will be explored in Section 3.3 → p.69 . In
Section 3.4 → p.78 we will study in more detail the structure of CFTs by choosing
a particular operator formalism — the Radial Quantization. We will see that
CFTs have a special property — quantum states correspond to local operators. Its
particular important consequence, the existence of Operator Product Expansion,
will be discussed in Section 3.5 → p.86 . Finally, in Section 3.6 → p.88 we will briefly
comment on the structure of 4-point correlators.

Exposition of this chapter was strongly influenced by [32] [11] [4], and also
[33] [34] [35] [36].

3.1 Conformal Structure Generalities

We start by studying conformal structure of manifolds in general. Conformal-
structure preserving maps are called conformal transformations. As is usual, the
analysis of such global transformations can be simplified by studying corresponding
infinitesimal actions, which are generated by conformal Killing vectors. Conformal
automorphisms of a manifold form a conformal group, and to not lose out on
transformations which are singular at some points, we will perform the conformal
compactification.
Chapter 3. Conformal Field Theory 53

Conformal Transformations
It is natural that a local field theory invariant under global scaling transformations
will be also invariant under local scaling transformations. By this we mean that
the induced mapping leaves the metric invariant up to a conformal factor, which
can depend on the spacetime point.

Definition 3.3 (Weyl Transformation, Conformal Structure). A Weyl transfor-

mation is an active transformation of the metric on a pseudo-Riemannian
manifold (M, g) ↦→ (M, g̃ ≡ Ω2 g) by scaling it with a (square of) smooth
positive function Ω ∈ C>0
∞
(M), called conformal or Weyl factor.
Two metrics on M related by a Weyl transformation are said to be conformally
equivalent, and the equivalence class c ≡ g/C>0∞ (M) of metrics under Weyl

transformations is called a conformal structure on M.

Remark 3.4. After a Weyl transformation with a conformal factor Ω, the lengths of
vectors at a point x ∈ M are multiplied by Ω(x). Thus, conformal structure enables
us to locally measure angles and ratios of lengths, but not lengths themselves. ⌟
Remark 3.5. Every pseudo-Riemannian manifold has automatically a conformal
structure. However, we can also consider manifolds which intrinsically have just
the conformal structure, with no canonical/preferred metric.
Examples are boundaries of pseudo-Riemannian manifolds after a compactification
à la Penrose [37]. The metric initially does not extend over the boundary where
it diverges, and to make it regular we need to perform a Weyl transformation
which itself vanishes at the boundary. Since there is no natural choice of such
prescription, we are left with the whole equivalence class of metrics, giving us just
the conformal structure. ⌟

Definition 3.6 (Conformal Transformation, Conformal Map). A smooth map

ϕ : (M, c) → (M′ , c′ ) between manifolds equipped with conformal structures
is called a conformal transformation or conformal map if it preserves the
conformal structure — it is a homomorphism of conformal structures.
In the case of pseudo-Riemannian manifolds (M, g) and (M′ , g ′ ), this means
that the metrics (after pullback) are related by a Weyl transformation, that is
!
ϕ∗ g ′ = Ω2 g ≡ e2ω g ,

where ϕ∗ is pullback map corresponding to ϕ, and Ω ≡ eω is the associated

conformal factor. Specially, if Ω = 1 ⇔ ω = 0, the map is called an isometry,
as it preserves the metric itself.

Remark 3.7 (Properties of Conformal Maps, Conformal Equivalence). The identity

map is trivially conformal, and composition of conformal maps yields another
conformal map. If a conformal map is a diffeomorphism, we call it a conformal
Chapter 3. Conformal Field Theory 54

diffeomorphism, and we easily see that its inverse is also conformal. It thus
makes sense to say that two manifolds are conformally equivalent if there exists a
conformal diffeomorphism between them.
In the case of pseudo-Riemannian manifolds, the associated conformal factors
for composition of conformal maps and inverse of conformal diffeomorphism are
straightforward to calculate from definition. Manifolds with conformally equivalent
metrics — see Definition 3.3 → p.53 — are automatically conformally equivalent,
where the conformal diffeomorphism is provided by the identity map. ⌟
Remark 3.8. Since null geodesics are mapped to null geodesics by conformal maps
of a Lorentzian manifold, we can say — ignoring the possible time-reversals —
that conformal transformations preserve the causal structure of the manifold. ⌟
Example 3.9 (Dilatation in Flat Space). Apart from the usual isometries of the
flat space Rp,q (with usual metric) — translations and pseudorotations — other
prominent conformal transformations are the dilatations (scaling transformations).
These are given in the usual Cartesian coordinates by

Dλ : Rp,q −→ Rp,q
x ↦−→ x′ ≡ Dλ (x) ≡ λx ,
′
and can be easily seen to be conformal with Ω = λ since ∂x∂x
= λ1. All of such sim-
ple conformal transformations with constant Ω are illustrated in Figure 3.10 → p.54 .⌟

Figure 3.10 / Simple conformal transformations of a flat

plane R2 — translation, rotation, and scaling/dilatation.

Example 3.11 (Stereographic Projection). Consider a unit n-sphere Sn with metric

induced by the standard embedding in Rn+1 (with usual metric). Denoting the
north pole as N ≡ (1, 0, . . . , 0) — expressed in Cartesian coordinates (x0 , x1 , . . . , xn )
on Rn+1 — we can define the stereographic projection ϕ as the map sending every
point x ∈ Sn \{N} to the intersection of the line through N and x with the
equatorial plane {(0, x1 , . . . , xn )} ≃ Rn . For a visualization, see Figure 3.12 → p.55 .
In detail, using the embedding coordinates it is given by

ϕ: Sn \{N} −→ Rn
(︄ ⃓ n )︄
⃓ ∑︂ (︂ )︂
x≡ x , x , . . . , x ⃓ (xi )2
0 1 n ⃓
= 1 ↦−→ x ≡ ϕ(x) ≡ 1
1−x0
x1 , . . . , xn ,
⃓
i=0
Chapter 3. Conformal Field Theory 55

where we considered affine combination xt ≡ (1 − t)N + tx of N and x in the

embedding space, and found the intersection xt∗ with the equatorial plane at
(1 − t∗ ) + t∗ x0 = 0 ⇔ t∗ = 1−x
1
0 , giving us the desired formula for x as the last n

components of xt∗ . Note that we implicitly understand x0 as a function of xi .

The map ϕ is a diffeomorphism, since x0 can be calculated as a smooth function
of x (outside of N) by

1 n
1 − (x0 )2 1 + x0 x2 − 1
x ≡ |x| =
2 2
(x ) =
i 2
= ⇐⇒ x = 2
0
∑︂
,
(1 − x0 )2 i=1 (1 − x0 )2 1 − x0 x +1
i
giving rest of the components as xi = (1 − x0 )xi = x2x
2 +1 . We can quickly check that

the origin (0, . . . , 0) ∈ R corresponds to the south pole S ≡ (−1, 0, . . . , 0) ∈ Sn ,

n

and the unit circle {x2 = 1} ⊂ Rn is precisely the equator {x0 = 0} ⊂ Sn .

A direct computation verifies that it is conformal, so it maps circles to circles, and
circles through N to straight lines. ⌟

N
Rn+1 ⊃ Sn
x

Rn , {x 0 = 0}
0
x

S ©2025 Jonas Dujava

Figure 3.12 / Stereographic projection Sn \{N} → Rn .

Remark 3.13. Conformal maps in Rd can be viewed through the lens of such
stereographic projections. The idea is to make the inverse stereographic projection
Rd → Sd ↪→ Rd+1 , then some isometry of Rd+1 (rotation + translation), and
finally the stereographic projection back to Rd , for which we need to formulate
stereographic projection for arbitrary positioning of the sphere in Rd+1 . For more
details see Section 1.2 of [38]. ⌟
We will be mainly interested in conformal transformations for d ≥ 3 spacetimes.
Nevertheless, let us briefly mention the specialties of d = 1 and d = 2.
Example 3.14 (Conformal Transformations in d = 1). Since in d = 1 there is no
angle dependence and metrics have only one component, any diffeomorphism is
automatically conformal. For example a diffeomorphism ϕ : R → R — endowing
the real line with the standard metric g = dx2 — is conformal with Ω = dϕ
dx
. ⌟
Chapter 3. Conformal Field Theory 56

Example 3.15 (Conformal Transformations in d = 2 as Holomorphic Maps). In

d = 2 we already obtain nontrivial restrictions on conformal maps. Identifying
R2 with C, it can be quickly verified that any holomorphic or antiholomorphic
map ϕ : D → C locally invertible on open set D ⊂ C is conformal, as illustrated
in Figure 3.16 → p.56 . For more details see for example [4] or [39]. ⌟

∼ z4 ∼ z3

∼ z1 ∼ z2
©2025 Jonas Dujava

Figure 3.16 / Conformal transformations of a square D ≡ [1, 2] × [0, 1] ⊂ R2 ≃ C

induced by holomorphic maps z n : C → C for n = {1, 2, 3, 4}. On the right we display
the shape of z n (D) after some rescaling (such that bottom edge has the same length).

Together with the Riemann mapping theorem, this provides a powerful tool for
studying for example classical electrostatics or ideal hydrodynamics in d = 2, as
is showcased in the following.
Example 3.17 (Parallel Plate Capacitor in d = 2). Consider a (finite) parallel plate
capacitor in R2 . How can we find (at least approximately) the electric potential ϕ
in the region near end of the plates, which satisfies Laplace equation △ϕ = 0 and
boundary conditions ϕ|plate± = ±V on plates?
We can utilize that conformal transformations map solutions of Laplace equation
to solutions of Laplace equation in d = 2, since △ = g− /2 ∂a (g /2 g ab ∂b ) and the
1 1

expression g /2 g ab is invariant with respect to Weyl transformations of the metric

1

in d = 2. This allows us to map an easily found solution with infinite plates to

the one where plates are finite (from one side).
Indeed, the holomorphic map ϕ : C → C given by w ↦→ ϕ(w) ≡ z ≡ 1 + w + exp(w)
transforms the strip S ≡ {|Im(w)| < π} to C\{x ±ı̊π | x ∈ R≤0 }. One can
check that boundaries ∂ ± S = {u ±ı̊π | u ∈ R} of the strip are mapped to the
plate± ≡ {x ±ı̊π | x ∈ R≤0 } by calculating Im(ϕ(u ±ı̊π)) and finding the maximal
value of Re(ϕ(u ±ı̊π)). For example, the region just below ∂ + S and Re(w) < 0
stays below the plate+ , while Re(w) > 0 part wraps back ending just above plate+ .
It is therefore enough to solve Laplace equation on S with boundary conditions
Chapter 3. Conformal Field Theory 57

ϕ |∂ ± S = ±V — just a linear dependence on Im(w) — and then transform the

solution with the conformal map ϕ to obtain the solution of the original problem.
See Figure 3.18 → p.57 for the visualization of resulting field lines and equipotentials.
This works approximately even for d = 3 in the region near edge (not corner)
assuming the distance between plates is much smaller than their extent in the
other two directions. ⌟

Figure 3.18 / Two-dimensional

electric field near one end of a
parallel plate capacitor.
It is obtained by conformal
transformation of field between
two infinite plates.

Legend:
equipotentials
field lines
©2024 Jonas Dujava ©2024 Jonas Dujava
capacitor plate ©2024 Jonas Dujava ©2024 Jonas Dujava

Conformal Group
We usually work with just one underlying manifold as our spacetime, so most
interesting conformal transformations in our case have the same manifold both for
the domain and the target. As already mentioned in Remark 3.7 → p.53 , the identity
©2024 Jonas Dujava ©2024 Jonas Dujava

map is always conformal and composition/inverse of a conformal diffeomorphism ©2024 Jonas Dujava ©2024 Jonas Dujava

stays conformal, so we clearly see that conformal automorphisms form a group.

Definition 3.19 (Conformal Group, Isometry Group). Let M be a manifold with

a conformal structure. The group of all conformal automorphisms Conf(M)
is called the conformal group of M. If M is a pseudo-Riemannian manifold,
all its isometries — conformal automorphisms with Ω = 1 ⇔ ω = 0 — form a
subgroup Iso(M) ⊂ Conf(M) preserving the metric structure.

Remark 3.20. We will usually not be careful to distinguish between Conf(M)

(which includes discrete transformations) and its connected component of the
identity Conf 0 (M). ⌟
Chapter 3. Conformal Field Theory 58

Remark 3.21 (Left Action of Conformal Group). We have a natural left action
▷ : Conf(M) × M → M, (g, x) ↦→ g ▷ x ≡ g(x), where we first think of g as some
abstract group element, and then as a conformal automorphism of M. By action
being left we mean the property g ▷ (h ▷ x) = (gh) ▷ x = g(h(x)) ≡ (g ◦ h)(x). ⌟
Remark 3.22 (Fundamental Vector Field). A Lie algebra element X ∈ Lie(Conf(M))
generates a 1-parametric subgroup of Conf(M) through the exponential map
ϕε ≡ eεX ∈ Conf(M), ε ∈ R. The flow ϕ : R × M → M is in turn generated by
a vector field ξX ≡ dε
D
ϕε |ε=0 ∈ T M, which is called the fundamental vector field
corresponding to X.
Generally, having a group G with a left action on M, it can be shown that the
map Lie(G) ≡ g → T M, X ↦→ ξX assigning the fundamental vector field is a
Lie algebra antihomomorphism, that is ⟦ξX , ξY ⟧ = ξ−[X ,Y ] , where ⟦•, •⟧ is the Lie
bracket of vector fields and [•, •] is the Lie bracket of g. ⌟

Conformal Killing Vector Fields

Since most interesting (for us) are continuous symmetries and corresponding
conservation laws, we will now focus on conformal automorphisms continuously
connected to the identity. As usual, it is convenient to first study the generators
of infinitesimal conformal transformations.

Definition 3.23 (Conformal Killing Vector Field). A vector field ξ ∈ T M is

called a conformal Killing vector field (CKV) if it satisfies (with ϕε being the
infinitesimal flow generated by ξ, that is ξ ≡ dε
D
ϕε |ε=0 )

d
⃓
≡ ϕ∗ε g ⃓⃓ = ∇a ξb + ∇b ξa = 2κ gab ,
⃓
Lξ gab
dε ε=0

for some function κ ∈ C ∞ (M). The second equality is an identity for Lξ g

in terms of the Levi-Civita connection ∇ corresponding to g. In the special
case where κ = 0, we have a Killing vector field corresponding to an isometry.

Remark 3.24. Denoting Ωε ≡ eωε the conformal factor associated to ϕε , we have

κ = dε
d
ωε |ε=0 (since Ω0 = 1 ⇔ ω0 = 0), which need not be positive. Also, by
contracting indices of the CKV equation with the inverse metric g ab , we obtain
1 1
κ = ∇• ξ ≡ ∇n ξ n , (d = Tr(δ) ≡ δkk ≡ gkl g kl ) .
d d ⌟
Remark 3.25 (Lie Algebra of Conformal Killing Vectors). The set of all conformal
Killing vector fields forms a Lie algebra closed under linear combinations and Lie
brackets. This follows from Lξ+ζ = Lξ + Lζ and L⟦ξ,ζ⟧ = [Lξ , Lζ ], where the first
bracket ⟦•, •⟧ is Lie bracket of vector fields, and second one is just a commutator.
We denote set of all CKVs by conf(M), and we have the following sequence

iso(M) ⊂ conf(M) ⊂ TM ,
Chapter 3. Conformal Field Theory 59

where the iso(M) is the Lie subalgebra of Killing vectors (κ = 0), and T M is the
infinite-dimensional Lie algebra of smooth vector fields on M. ⌟

Remark 3.26 (Lie Algebra of Conformal Group versus CKVs). By Remark 3.22 → p.58
we have a natural inclusion Lie(Conf(M)) ↪→ conf(M) of the Lie algebra of the
conformal automorphism group into the Lie algebra of CKVs.
For M compact, it is actually an isomorphism, since the flow of a CKV can be
globally integrated, yielding a 1-parametric subgroup of Conf(M). However, in
general conf(M) can be bigger than CKVs corresponding to global conformal
transformations, since it can happen that the differential equation defining the
flow can blow up for some value of the affine parameter. ⌟

Remark 3.27. Conformally equivalent metrics share their CKVs, since we have
Lξ g ′ ≡ Lξ (e2ω g) = (2ξ(ω) e2ω + 2κ e2ω )g = 2(ξ(ω) + κ)g ′ ≡ 2κ′ g ′ . ⌟

Remark 3.28. Conformal Killing vectors have constant “nature” — they stay
timelike, spacelike, or null — along their integral curves. We can see this by
differentiating f ≡ ξ 2 ≡ g(ξ, ξ) along the associated integral curve xξ (t). We
obtain df dt
= Lξ f = 2κf using Lξ ξ = ⟦ξ, ξ⟧ = 0. By uniqueness of the solution of
smooth first-order differential equations we infer that f has a constant sign, since
if it is zero at any point, it stays zero along the whole integral curve. ⌟

Proposition 3.29 (Further Equations for Conformal Killing Vectors). Let ξ be a

conformal Killing vector field with a conformal factor κ. Then it also satisfies
the following equations

∇a∇b ξc = Rbcna ξn + (∂a κ)gbc + (∂b κ)gca − (∂c κ)gab ,

ξc = −Riccn ξn + (2 − d)∂c κ ,

where ≡ g ab ∇a ∇b and Ricab ≡ Rnanb .

Proof. Writing the covariant derivative of CKV equation in Definition 3.23 → p.58
three times with permuted indices, and summing them in a particular way gives

∇c (∇a ξb + ∇b ξa ) = 2(∂c κ)gab

⎫
(−)⎪
⎪ ⎬ (∇a∇b + ∇b∇a )ξc − (Racnb + Rbcna )ξn
∇a (∇b ξc + ∇c ξb ) = 2(∂a κ)gbc (+) ⇒ [︂ ]︂
⎪ = 2 (∂a κ)gbc + (∂b κ)gca − (∂c κ)gab ,
∇b (∇c ξa + ∇a ξc ) = 2(∂b κ)gca
⎪
⎭
(+)

where we used the Ricci identity (∇a∇b − ∇b∇a )ξc = −Rabnc ξn . Using it once
more to commute indices ba in the term ∇b∇a ξc , and utilizing the Bianchi identity

R[abnc] = 0 ⇐⇒ Rabnc + Rcanb + Rbcna = 0 ⇐⇒ Rabnc − Racnb = −Rbcna

we obtain the desired formula after a simple manipulation.

Chapter 3. Conformal Field Theory 60

Corollary 3.30 (Maximal Number of Conformal Killing Vector Fields). Let M

be a d-dimensional connected pseudo-Riemannian manifold with d ≥ 3. Then
the number of independent conformal Killing vector fields is bounded by

(d + 1)(d + 2)
dim conf(M) ≤ ,
2
and in the case of Killing vectors we have (holds for arbitrary d)

d(d + 1)
dim iso(M) ≤ .
2

Idea of the Proof. We have a system of first-order linear partial differential

equations for CKV data {ξa , ωab ≡ ∇[a ξb] , κ ∼ ∇• ξ, Fa ≡ ∇a κ ∼ ∇a∇• ξ} in
Definition 3.23 → p.58 and Proposition 3.29 → p.59 — for d ≥ 3 it is also possible to
obtain equation for ∇a Fb in terms of other variables — so assuming connectedness
it is enough to specify initial conditions at a single point. This gives us maximally
d + d(d−1)
2
+ 1 + d = (d+1)(d+2)
2
parameters. In the case of ordinary Killing vectors
(isometries) we have κ ≡ 0, so we are left with just d + d(d−1)
2
= d(d+1)
2
parameters.
For further details see [40].
Remark 3.31 (Maximally Conformally-Symmetric Spacetimes). In the following
Section 3.2 → p.61 we will explicitly see that flat spaces Rp,q saturate this bound.
Moreover, by studying the integrability conditions of the equations for CKV data,
it can be shown that any manifold with maximal number of independent CKVs is
automatically conformally flat [40]. ⌟
Remark 3.32. The finite bound for CKVs in Corollary 3.30 → p.60 fails for d ≤ 2.
The technical reason is that it is not possible to constrain ∇a Fb ≡ ∇a ∇b κ, thus
we have more freedom allowing possibly infinite independent CKVs. We will see
this explicitly in context of flat space during the proof of Proposition 3.37 → p.62 . ⌟

Conformal Compactification
As already mentioned in Remark 3.26 → p.59 , not all CKVs can be integrated
to obtain global conformal transformations. Since we want to retain as many
symmetries as possible — and it is anyway mathematically elegant — we will
always try to complete M such that every local conformal transformation can be
extended to a global one.
Such completion would be achieved by embedding M into a compact manifold M′
(by adding measure zero set of points) with the compatible conformal structure,
since then every CKV can be globally integrated [41].

Definition 3.33 (Conformal Compactification). Let (M, c) be a manifold

equipped with a conformal structure. A compact manifold (M′ , c′ ) with
conformal structure is called a conformal compactification of M when:
Chapter 3. Conformal Field Theory 61

(1) There is a conformal embedding ι : (M, c) ↪→ (M′ , c′ ) with dense image

in M′ , that is ι(M) = M′ .
(2) All conformal vector fields on (M, c) extend to ones on (M′ , c′ ), and thus
can be automatically integrated to global conformal automorphisms.

Remark 3.34. Conformal compactifications in this strong sense exist only for
d ≥ 3, since for example in d = 2 it is not possible to continue all local conformal
transformations to global ones.
If the conformal compactification of M exists, it is unique (up to diffeomor-
phism), and we denote it with M. Oftentimes we will omit such distinction
and implicitly mean M, for example when we write Conf(M). Furthermore, by
Remark 3.26 → p.59 we can now identify Lie(Conf(M)) with conf(M), but remember
that corresponding Lie brackets differ by extra minus sign. ⌟

Remark 3.35. Usually we compactify just by adding some points at “infinity”,

which is naturally done for a certain conformally equivalent manifold. It is
then obvious that both conditions in Definition 3.33 → p.60 are satisfied. A prime
example is a sphere Sn as a one-point conformal compactification of a plane Rn
via the inverse stereographic projection, see Example 3.11 → p.54 . There is only one
extra point in Sn , namely the north pole N corresponding to “infinity” of Rn . ⌟

3.2 Flat-Space Conformal Structure

From now on we restrict our attention to flat spaces Rp,q with d ≡ p + q ≥ 3.
For simplicity, we will often consider the Euclidean Rd . As already noted
in Remark 3.31 → p.60 , this is essentially the same as focusing on maximally
conformally-symmetric spacetimes.

Conformal Transformations in Flat Space

Essentially by Definition 3.6 → p.53 , a conformal transformation of a flat space looks
locally like translation + rotation + dilatation.
Remark 3.36 (Jacobian Matrix of Flat-Space Conformal Transformation). For any
conformal transformation g ∈ Conf(Rd ) : x ↦→ x′ ≡ g(x) we can write

∂x′
≡ Ωg (x)Rg (x), Rg (x) ∈ SO(d) or O(d) .
∂x
Alternatively, the scaling factor Ωg is related to the Jacobian by
⃓1
⃓ ∂x′ ⃓ d
⃓ (︄ )︄
1 ∂g(x)
Ωg (x) = ≡ det .
⃓ ⃓ d
⃓ ⃓
⃓ ∂x ⃓ ∂x
⌟
Chapter 3. Conformal Field Theory 62

Due to the simple geometry of Rp,q , we can write down the general form of
conformal Killing vector fields conf(Rp,q ) by explicitly solving the CKV equation.

Proposition 3.37 (General Form of CKV in Flat Space). A general solution of

the conformal Killing equation in flat space Rp,q of dimension d ≡ p + q ≥ 3
can be written as
∂
ξ ≡ ξa ≡ ξ a ∂a = ta pa + 12 ω ab mab + σd + sa ka ,
∂xa
with the generators explicitly given by

pa ≡ ∂a , mab ≡ −xa ∂b + xb ∂a = m[ab] ,

d ≡ x ∂n , ka ≡ 2xa xn ∂n − x2 ∂a ,
n

where t• , ω •• ≡ ω [••] , σ, s• are constants parametrizing the CKVs. In other

words, the Lie algebra of conformal vector fields on Rp,q — on the level of
vector spaces — decomposes into direct sum of subalgebras

conf(Rp,q ) = {translations} ⊕ {pseudorotations} ⊕ {dilatations} ⊕ {SCTs} ,

⏞ ⏟⏟ ⏞ ⏞ ⏟⏟ ⏞ ⏞ ⏟⏟ ⏞ ⏞ ⏟⏟ ⏞
≃ Rp,q ≃ so(p,q) ≃ R ≃ so(1,1) ≃ Rp,q

where SCTs are the so-called Special Conformal Transformations.

Proof. In the flat space we have g ≡ η, ∇ → ∂, and R•••• = 0, which simplifies

the analysis of the conformal Killing equation in Definition 3.23 → p.58 . A chain of
manipulations gives (throughout we will use Cartesian coordinates)

[CKV]ab ≡ ∂a ξb + ∂b ξa = 2κ ηab
∂ b [CKV]ab =⇒ d∂a κ + ξa = 2∂a κ
d̸=1
∂ a ∂ b [CKV]ab =⇒ 2d κ = 2 κ =⇒ κ=0
d̸=2
∂(a ∂ n [CKV]b)n =⇒ d∂a ∂b κ + ∂(a ξb) = 2∂a ∂b κ =⇒ ∂a ∂b κ = 0
⏞ ⏟⏟ ⏞
∝ κ=0

Substituting this into (∂• of) Proposition 3.29 → p.59 , we obtain

∂
∂a ∂b ξc = (∂a κ)ηbc + (∂b κ)ηca − (∂c κ)ηab =⇒
•
∂a ∂b ∂c ξd = 0 ,

therefore ξ • is at most quadratic in x. Expanding it as

ξ a (x) = ta + Aab xb + S abc xb xc , where S abc ≡ S a(bc) ,

substituting it back into [CKV], and comparing order by order we obtain:

(1) No restrictions on t• ∈ Rp,q , so ta ∂a is CKV corresponding to translations.
(2) For linear part we get
2 n
2A(ab) = A n ηab .
d
Chapter 3. Conformal Field Theory 63

Decomposing A•• into traceless symmetric, antisymmetric, and trace parts,

we see that the condition implies traceless symmetric part is zero, so we are
left with
1 n
Aab = A[ab] + A n ηab ≡ ωab + σηab =⇒ Aab xb ∂a = ω ab x[b ∂a] + σxn ∂n .
d
The first part corresponds to pseudorotations, and the second to scaling
transformations also called as dilatations.
(3) For the quadratic part we get
2 n c
4S(ab)c xc = S nc x ηab ≡ 4sc xc ηab ⇐⇒ S(ab)c = ηab sc .
d
We can invert the relation (using Sabc = Sa(bc) ) to get

Sabc = ηab sc + ηac sb − ηbc sa =⇒ S abc xb xc ∂a = sa (2xa xn ∂n − x2 ∂a ) .

This corresponds to Special Conformal Transformations (SCTs).

Remark 3.38. As was mentioned in Remark 3.32 → p.60 , the cases with d ≤ 2
are special. Any vector field in d = 1 is CKV, which is compatible with
Example 3.14 → p.55 . In d = 2 we can have ∂• ∂• κ =
̸ 0, which allows infinite
independent CKVs. For more detail see [4] and [39]. ⌟
Since hopefully everybody knows finite forms of translations and pseudo-rotations,
and we encountered dilatations in Example 3.9 → p.54 , we will look more closely
on SCTs. But first, it is convenient to discuss an important transformation not
continuously connected to the identity component of the conformal group.
Remark 3.39 (Inversion). The (radial) inversion map is given by
I : Rd −→ Rd
x
x ↦−→ I(x) ≡ .
|x|2
It maps 0 to ∞ and vice versa, since I = I −1 . Locally (infinitesimally) we have
∂I a (x) ∂I a ⃓⃓ δba 2xa xb 1 xa xb
⃓ (︄ )︄
a
= = − = δ b − 2 ≡ ΩI (x)Iba ,
∂xb ∂xb ⃓x |x|2 |x|4 |x|2 |x|2
where
⃓ x ⃓2 1
⃓ ⃓
ΩI (x) ≡ = |I(x)|2 = , Iba ≡ (δba − 2x̂a x̂b ) ∈ O(d) ,
⃓ ⃓
⃓ 2⃓
⃓ |x| ⃓ |x|2
so it is rescaling and reflection along the radial direction.
Induced action on a flat metric is
∂I c ∂I d 1
⃓ ⃓ (︂ )︂
c d
(I η)ab ⃓ =
∗ ⃓ = (δ 2x̂c
) 2x̂d
⃓ ⃓
⃓ ηcd a − x̂a δb − x̂b ηcd
∂xa ∂xb |x|4
⃓
x x
1 1
⃓
= 4 (ηab − 4x̂a x̂b + 4x̂a x̂b ) = 4 ηab ≡ ΩI ηab ⃓⃓ ,
2
⃓
|x| |x| x
Chapter 3. Conformal Field Theory 64

confirming that inversion is a conformal transformation.

Alternatively, after the inverse stereographic projection to the sphere Sd ⊂ Rd+1
Example 3.11 → p.54 , it corresponds to the reflection x0 ↦→ −x0 , which is clearly
conformal. ⌟

Remark 3.40 (Finite Special Conformal Transformations). Special Conformal Trans-

formations (SCTs) can be understood as a composition of inversion, translation
by −b ∈ Rd , and inversion again — they are “translations around infinity”. While
SCTs are not defined everywhere on Rp,q , but will be on appropriate conformal
compactification.
Infinitesimally, generators of translations and SCTs are related by

∂I b ⃓⃓
⃓ ⃓ ⃓ (︂ )︂
b
(−Ipa I) ⃓⃓ = −∂a I ⃓⃓ =− = 2
2x̂b
⃓ ⃓
∂ b −|x| δ a − x̂a ∂b
x x
|x|2
∂xa |x|x2
⃓
⃓
= 2xa x ∂ − |x| ∂a ≡ 2 ⃓
• ka ⃓⃓ ,
x

where the generators are understood as differential operators on functions, and I

acts as (If )(x) ≡ f (I(x)). Alternatively, we have

D
⃓ ⃓
(−AdI pa ) ⃓⃓ ≡ I ◦ [ ↦→ − εea ] ◦ I(x) ⃓⃓
⃓ ⃓
x dε ε=0
2
D x − |x| εea
⃓ ⃓
= = −|x| ∂a + 2xa x ∂ ≡ ka ⃓⃓ .
2
⃓ ⃓
⃓ •

dε 1 − 2εxa + |ε| |x| ε=0

2 2 ⃓
x

So, for finite (exponentiated) action of SCTs has the form

eb k = e−Ib pI = I e−b p I = I ◦ [ ↦→ − b] ◦ I ,
• • •

that is (︄ )︄ (︄ )︄
x x
e b •k
(x) = I ◦ [ ↦→ − b] 2 =I −b
|x| |x|2
x
|x|2
−b x − |x|2 b
=⃓ =
1 − 2b x + |b|2 |x|2
⃓2
⃓ x
⃓ |x|2 − b ⃓ •
⃓

Since translations completely preserve the flat metric, using the form of Ω2I we
obtain Ω-factor of a SCT as
1 1 1
(︄ )︄
x
Ωb = ΩI − b ΩI (x) = ⃓ 2 = (︂ )︂2 .
|x|2
⃓2
⃓ x
2 − b⃓
⃓ |x| 1 − 2b x + |b|2 |x|2
•
⃓ |x| ⌟

Conformal Algebra of Flat Space

Having the explicit form of CKVs, we can now compute their Lie brackets.
Chapter 3. Conformal Field Theory 65

Lemma 3.41 (Lie Brackets of CKVs in Flat Space). Lie brackets of CKVs in
flat space Rp,q (d ≡ p + q ≥ 3) are given by
⟦mab , pc ⟧ = −pa ηbc + pb ηac , ⟦mab , kc ⟧ = −ka ηbc + kb ηac ,
⟦mab , mcd ⟧ = mac ηbd − (a ↔ b) − (c ↔ d) + (a ↔ b, c ↔ d) ,
⟦d, d⟧ = 0 , ⟦d, mab ⟧ = 0 , ⟦pa , pb ⟧ = 0 , ⟦ka , kb ⟧ = 0 ,
⟦d, pa ⟧ = −pa , ⟦d, ka ⟧ = ka , ⟦ka , pb ⟧ = −2ηab d + 2mab .

Proof. Can be verified by a straightforward computation. Most of the brackets

with k• can be obtained from the ones with p• by utilizing (at least in Rd )

I ◦ Dλ ◦ I = D1/λ =⇒ AdI d = −d , I ◦ R ◦ I = R =⇒ AdI mab = mab ,

I = 1,
2
AdI pa = −ka , Ad• ⟦ , ⟧ = ⟦Ad• , Ad• ⟧ ,

where R is any rotation. To derive these relations, the relevant places to look are
Example 3.9 → p.54 , Remark 3.39 → p.63 , and Remark 3.40 → p.64 .
It is convenient to introduce alternative basis for the conformal algebra.

Definition 3.42 (Generators of Flat-Space Conformal Algebra). An alternative

basis of conf(Rp,q ) for d ≡ p + q ≥ 3 is provided by the generators lAB ≡ l[AB]
with A, B ∈ {−1, 0, 1, . . . , d} (and a, b ∈ {1, . . . , d}) defined as

l−1,0 = d ,
0 d 1
(pb − kb )
⎛ ⎞
2
l−1,b = 21 (pb − kb ) ,
= ⎜−d 0 (pb + kb ) ⎟
⎜ 1 ⎟
lAB ⎟.
⎜
2
l0,b = 12 (pb + kb ) , ⎝ . ..
.. .
⎠
mab
lab = mab ,
(p+1,q+1)
For capital letters we will use diagonal metric ηAB ≡ ηAB with η−1,−1 ≡ −1,
(p,q)
η0,0 ≡ +1, and ηab ≡ ηab , so we added one timelike and one spacelike
dimension.

Remark 3.43. We usually work in the Lorentzian signature (p, q) = (1, d − 1) or

in the Euclidean signature (p, q) = (0, d) (often obtained by the Wick rotation). It
is then common to use index “0” for the “time” direction, and i ∈ {1, . . . , d − 1}
for “space” directions. Notice that this is not the case in Definition 3.42 → p.65 . ⌟

Theorem 3.44 (Conformal Algebra and Group of Rp,q ). The generators lAB of
conf(Rp,q ) for d ≡ p + q ≥ 3 have the following Lie brackets

⟦lAB , lCD ⟧ = lAC ηBD − (A ↔ B) − (C ↔ D) + (A ↔ B, C ↔ D) ,

so the conformal algebra of Rp,q is

conf(Rp,q ) ∼
= so(p + 1, q + 1) .
Chapter 3. Conformal Field Theory 66

The corresponding conformal group of the conformal compactification Rp,q is

Conf(Rp,q ) ∼
= O(p + 1, q + 1)/Z2 .

Idea of the Proof. First part follows by expanding {lAB } via Definition 3.42 → p.65
in terms of {d, pa , ka , mab }, and then using Lemma 3.41 → p.65 . Second part
can be understood by constructing the conformal compactification using the
embedding/ambient space to be discussed shortly. For a more detailed Theorem
statement and proof, see [39].

Remark 3.45 (Conformal Group of Celestial Sphere). Since sphere Sd−2 is confor-
mally equivalent to flat Euclidean space via stereographic projection, we have
Conf(Sd−2 ) = Conf(Rd−2 ) = SO(1, d − 1), so the conformal group of the celestial
sphere is precisely the Lorentz group. Every light ray is transformed to another
light ray by the Lorentz group, and looking at the corresponding points on the ce-
lestial sphere, they are transformed by the corresponding conformal transformation.
Nice treatment of this can be found in [42]. ⌟

Embedding Space
Both Theorem 3.44 → p.65 and Remark 3.45 → p.66 suggest that viewing given flat
space as embedded in higher-dimensional flat space can lead to a simpler description
of conformal group. This is the idea behind embedding formalism of conformal
field theory. We will just briefly outline the construction, for more details see for
example [11].

On Rp+1,q+1 we have natural linear action of pseudo-orthogonal group SO(p + 1, q +

1) which commutes with dilatations, so it descends also to real projective space...
Real projective spaces are compact, so also are their closed subspaces (example
being the null cone), providing us with natural conformal compactifications of
Rp,q .

For R1,d−1 , by this construction we obtain something like (assume d > 2)

S1 × Sd−1
≃ R1,d−1 ∪ {points at infinity}
Z2

with natural action of SO(2, d), which however contains closed time-like curves.

We can get rid of Z2 (antipodal) identification by projectivizing only with positive

factor λ > 0 (oriented rays) — then it contains two copies of Minkd .

It is then common to unwrap the S1 piece (corresponding to light-like direction)

and consider the universal covering having the form R × Sd−1 , on which acts the
universal (infinitely-sheeted) cover SO(2,
˜︃ d) which unwraps the SO(2) piece of
SO(2, d) ⊃ SO(2) × SO(d).
Chapter 3. Conformal Field Theory 67

Conformal Cross-Ratios
In the following Section 3.3 → p.69 we will look at correlators in flat-space CFTs.
It will be important to understand how the distances between points transform
under conformal transformations, which is content of the following lemma.

Lemma 3.46 (Transformation of Euclidean Distance under Conformal Maps).

For any conformal transformation g ∈ Conf(Rd ) : x ↦→ x′ ≡ g(x) we have
2
|x′ − y ′ | = Ωg (x)Ωg (y)|x − y|2 ,

where Ωg is defined as in Remark 3.36 → p.61 .

Proof. For translations and rotations we have Ω = 1, so it is clearly true. Also

simple are dilatations, where we again have a constant scaling factor Ωλ = λ. It
suffices to show it for the inversion I, and then the relation for SCTs follows from
Remark 3.40 → p.64 and the chain rule for the Jacobian.
Direct computation gives (see Remark 3.39 → p.63 for definitions)
⃓2
1
⃓
⃓ x y ⃓
2
|I(x) − I(y)| ≡ ⃓
⃓ 2 − 2 ⃓⃓ = 2 2 (x|y| − y|x|)2
⃓ |x| |y| ⃓ |x| |y|
(︂ )︂
≡ ΩI (x)ΩI (y) |y|2 − 2x y + |x|2 = ΩI (x)ΩI (y)|x − y|2
•

It is an interesting question how to construct conformal invariants from the

positions of some number of points. The translation and rotation invariance
restricts our attention to the distance between points, and the scaling invariance
(together with special conformal transformations) suggests looking at their ratios.
Starting with 4 points, we indeed find such conformally invariant quantities.

Definition 3.47 (Conformal Cross-Ratios). Given 4 distinct points x1 , x2 , x3 , x4

in Rd , we define their conformal cross-ratios as

x212 x234 x223 x214

u≡ , v≡ = u|1↔3 ,
x213 x224 x213 x224

where xij ≡ xi − xj and xαij ≡ |xij |α for any α ∈ R.

Corollary 3.48 (Conformal Cross-Ratios are Conformal Invariants). The confor-

mal cross-ratios u and v are invariant under conformal transformations.

Proof. Follows directly from Lemma 3.46 → p.67 and the fact that each point xi
appears both in the numerator and denominator of the cross-ratios.
Remark 3.49. We could choose different ways how to arrange the points in the
definition of the conformal cross-ratios, but all of them turn out to be expressible
in terms of u and v. ⌟
Chapter 3. Conformal Field Theory 68

Conformal Frame
We can get better intuition about the cross-ratios by considering them in a specific
conformal frame, where using the conformal transformations we place the points
at special positions.

Lemma 3.50 (3-transitivity of Conformal Group). The action of Conf(Rd ) on

Rd is 3-transitive, that is, for any 3-tuple of distinct points (x1 , x2 , x3 ) ⊂ Rd
there exists a conformal transformation g ∈ Conf(Rd ) such that g(x• ) = y•
for any other 3-tuple of distinct points (y1 , y2 , y3 ) ⊂ Rd .

Proof. We will show that any 3-tuple of distinct points can be mapped to the
standard configuration {0, e, ∞}, where e is a point corresponding to some unit
vector. This is done by composing simple conformal transformations as
(︂ )︂
(x1 , x2 , x3 ) SCT
(x′1 , x′2 , ∞) translation
(0, x′′2 , ∞) ↦−→ xˆ︂′′2 ≡ x′′2 /|x′′2 |
dilatation
(0, xˆ︂′′2 , ∞) rotation
(0, e, ∞) .

In the first two steps we move two points to ∞ and 0 using the fact that SCTs are
translations around ∞, and that ∞ is invariant under usual translations. Since
both {0, ∞} are invariant under dilatations and rotations, in the final steps we
can move the third point to any position on the unit circle.
Remark 3.51. For d = 2 such transformations between 3-tuples of points are
unique, which can be seen from uniqueness of transformation to the specified
standard configuration (0, e, ∞). In higher dimensions, we still have freedom to
make rotations that fix the axis pointing in the direction of e. ⌟
Remark 3.52. Conformal frame point of view clearly shows why we could not
construct nontrivial conformal invariants from 3 or fewer points. ⌟

Proposition 3.53 (Number of Independent Conformal Cross-Ratios). Given a set

of n distinct points in Rd , the number of independent conformal cross-ratios
is given by (with understanding that for n ≤ 3 it is zero)
⎧
⃓{︂
⃓ independent conformal invariants ⃓
}︂⃓ ⎨ n(n−3) for n ≤ d + 2 ,
⃓ build from n distinct points in Rd ⃓ =⎩ 2
nd − (d+2)(d+1)
2
for n ≥ d + 2 .

Proof. For large enough set of points, the whole conformal group Conf(Rd ) acts
effectively on the configuration space Cn (Rd ) of n distinct points in Rd . The
number of independent conformal invariants is then given simply by subtracting
the number of symmetries from the dimension of the configuration space, that is
(d + 2)(d + 1)
dim Cn (Rd ) − dim Conf(Rd ) = nd − .
2
For small enough number of points, not all conformal transformations nontrivially
Chapter 3. Conformal Field Theory 69

act on the n-tuple of points, and thus do not reduce the number of independent
conformal invariants.
We can transform any 3 points to the standard configuration (0, e, ∞) using
Lemma 3.50 → p.68 , leaving us with the subgroup SO(d − 1) ⊂ Conf(Rd ) acting on
the remaining n − 3 points. We find a (possibly nontrivial) stabilizer
[︃ )︂]︃
∼
(︂
Stab SO(d − 1) Rd \{0, e, ∞} = SO(d + 2 − n) for n ≤ d + 2 ,
↶
Cn−3

and adding its dimension to the previously oversubtracted expression we obtain

(d + 2)(d + 1) (d + 2 − n)(d + 1 − n) n(n − 3)

nd − + = .
2 2 2

Remark 3.54. Now we see that conformal cross-ratios u and v in Section 3.2 → p.67
indeed capture all of the conformally-invariant information about the positions
of 4 points in Rd . After performing transformation (x1 , x2 , x3 , x4 ) ↦→ (0, x′2 , e, ∞),
where for simplicity we take e ≡ (1, 0, . . . , 0), we can utilize the remaining rotations
to always bring x′2 to the two-dimensional plane x′2 ≡ (x, y, 0, . . . , 0), so we are
generally left with 2 degrees of freedom — in d = 1 it degenerates just to one.
Using such conformal frame, the cross-ratios are given by the squared distances of
the transformed point x′2 from {0, e}, that is (infinities cancel out)
2 2
u = |x′2 | ≡ zz , v = |x′2 − e| ≡ (1 − z)(1 − z) ,

where x′2 is often parametrized by a complex coordinate z ≡ x +ı̊y ∈ C.

In Euclidean signature the variables z and z are complex conjugates, but when we
analytically continue back to the Lorentzian signature using τ ↦→ ı̊t, they become
two independent real variables. ⌟

3.3 Primary Operators and Descendants

Now that we understand the conformal structure of the flat space, we can start to
study the implications of conformal symmetry on the properties of CFTs. By a
CFT we understand either a Lagrangian QFT with traceless energy–momentum
tensor, or even a “non-local” QFT with conformal symmetry as described below.
In this section we will focus on constraining the form of the correlation functions
by utilizing the methods of Section 1.3 → p.22 .

To each generator of the conformal algebra we can assign a corresponding conserved

charge — a topological surface operator representing the symmetry on the quantum
level. As usual, to simplify the analysis, we will decompose the set of all local
operators into irreducible representations of the conformal group. This will enable
us clearly see the consequences for the CFT correlators.
Chapter 3. Conformal Field Theory 70

Remark 3.55. In the following we will exclusively consider CFTs Wick-rotated to

the Euclidean flat space Rd , unless stated otherwise. ⌟

Quantum Conformal Algebra

When we are in local CFT, we have a local energy–momentum/stress tensor T ••
which is conserved and traceless. As shown in Section 1.3 → p.22 , we can obtain
conserved charges corresponding to the conformal algebra by integrating T ••
against a conformal Killing vector ξ over a (d − 1)–dimensional surface Σ as
∫︂
Qξ ≡ − dSa T ab ξ b .
Σ

While we do not need to commit to any particular operator formalism yet, which
will be the focus of the next Section 3.4 → p.78 , it is natural to consider surfaces
surrounding the origin.

Even if we do not have a local Lagrangian description in our CFT, and thus no
energy–momentum tensor to explicitly write out the charges as surface integrals
of local operators, the conformal symmetry — and in particular the conformal
algebra — is still represented on the quantum level.

Definition 3.56 (Quantum Representation of Conformal Algebra). We denote

the conserved charges corresponding to CKV ξ by Qξ . Moreover, we require
that they form a representation of the conformal algebra conf(Rd ), that is

[Qξ , Qζ ] = Q−⟦ξ,ζ⟧ .

For convenience we introduce the notation

LAB ≡ QlAB , D ≡ Qd , Pa ≡ Qpa , Ka ≡ Qka , Mab ≡ Qmab .

Remark 3.57. We included the minus sign such that it is a proper representation
of Lie(Conf(Rd )), since using Remark 3.22 → p.58 we have [QξX , QξY ] = Q−⟦ξX ,ξY ⟧ =
Qξ[X ,Y ] for any X , Y ∈ Lie(Conf(Rd )). Alternatively, under the correlator we have
[︂ ]︂
Q1 , [Q2 , O] = D2 [Q1 , O] = D2 D1 O ,

where we used that commutator of a charge Qi with the operator O can be

deformed to topological operator on a small sphere around the operator O, which
due to Ward–Takahashi identities Calculation 1.81 → p.27 acts as some differential
operator on the correlator. We first need to shrink the inner commutator, and
only then the outer, so the differential operators act in the reversed order. ⌟

Remark 3.58 (P /K as Raising/Lowering Operators). From Lemma 3.41 → p.65 we

immediately get commutation relations of these charges. For example, we see that
M•• generate the rotations SO(d), under which P• and K• transform as vectors,
 Chapter 3. Conformal Field Theory 71

while D is a scalar/invariant. Of special interest are

[D, Pa ] = Pa , [D, Ka ] = −Ka ,

indicating that P• /K• can be understood as raising/lowering operators for D. ⌟

Remark 3.59. In d = 2 often all local CKVs are included — this in particular
includes CKVs not associated with global conformal transformations, and even
those with singularity at the origin. Corresponding conformal algebra on the
quantum level is generally a central extension of such CKV algebra. ⌟

Definition 3.60 (Quadratic Conformal Casimir). The quadratic Casimir of the

conformal algebra conf(Rd ) ∼
= so(1, d + 1) is given by
1 AB 1 1 1
Cas2 ≡ − L LAB = − M ab Mab − P K − K P + D 2 ,
• •

2 2 2 2
where we write it already represented by quantum charges.

Remark 3.61. Conformal Casimir is a distinguished element of the center of the

universal enveloping algebra of conf(Rd ), that is it commutes with all generators of
the conformal algebra. Thus, by Schur’s lemma, in any irreducible representation
of conf(Rd ) it is represented as a multiple of the identity operator, and the
eigenvalues can be used to classify the representations. ⌟

Conformal Representations of Local Operators

The classification of conformal representations in a certain sense parallels the
classification of one-particle states in Poincaré-invariant QFTs — namely both
utilize the notion of induced representations — so let us first briefly recall the
familiar story from standard QFTs in Mink• .
Remark 3.62 (Classification of One-Particle States in Poincaré Invariant QFTs).
Since the translation subgroup is abelian, we can diagonalize it first, obtaining
the momentum eigenstates. This is naturally connected with the choice of inertial
foliation discussed in Example 1.88 → p.31 . The only thing left it to obtain the
transformation properties under the action of the homogeneous Lorentz subgroup.
After restricting our attention to a fixed value of the translation Casimir P 2 = M 2
— which partly classifies different irreps — we can always use boosts to transform
the state to have some chosen canonical or standard momentum. It is therefore
enough to understand the action of rotations for states with the given canonical
momentum, that is specify a representation of the little group (stabilizer of the
canonical momentum).
Transformation of a general state is then uniquely induced from this little group
action, essentially just by using the group algebra. For more details see [25]. ⌟
Chapter 3. Conformal Field Theory 72

Derivation 3.63 (Conformal Algebra Representations of Local Operators). We want

to perform a similar analysis in the context of local operators in CFTs. Here it is
natural to proceed directly in the position-space [43]:
(1) Since action of translations is transitive, we can fix any point to be the origin.
Then we would like to specify the representation of its stabilizer, which is
generated by {D, M•• , K• }.
(2) We choose to diagonalize the dilatations generated by D, and since the
rotations generated by M•• commute with D, we can specify a representation
of the rotation subgroup SO(d) as well — or its double cover Spin(d) for
spinor representations. So we have

[D, O(0)] = ∆ O(0) ,

[Mab , O(0)] = −Sab O(0) ,

where ∆ is the scaling dimension of the operator, S•• are generators of the
SO(d) representation, and we suppress the corresponding indices carried both
by S and O (and their appropriate contraction). Since we chose S•• to act
from the left, and we want them to have the same commutation relations as
M•• , we have to include a minus sign.
(3) We already know from Remark 3.58 → p.70 that all of K• are lowering operators
for D, which in this case means (using the Jacobi identity)
[︂ ]︂ [︂ ]︂ [︂ ]︂
D, [K• , O(0)] = [D, K• ], O(0) + K• , [D, O(0)] = (∆ − 1)[K• , O(0)] .

As we will see in Remark 3.82 → p.78 , physically-admissible operators must have

∆ ≥ 0, so there must be a lowest-weight state where this process eventually
terminates with
[Ka , O(0)] = 0 ∀a .
Such operators which are annihilated by all K• at the origin are called
primary operators. The rest of operators in a given conformal representation
— also called conformal multiplet — are obtained by acting with momentum
generators Pa (raising operators of D) as
[︃ [︂ ]︂]︃
O(0) ↦−→ PaN , · · · , [Pa1 , O(0)] · · · ≃ ∂a1 · · · ∂aN O(0) ,
∆ ↦−→ ∆ + N ,
which are called descendant operators of level N ≥ 1.
(4) Action at a general point is then obtained via

O(x) ≡ Ux O(0)Ux−1 ≡ ex P O(0) e−x P = e[x P , ] O(0) ,

• • •

and the conformal/operator algebra. In particular, we see that a shifted

primary operator is an infinite linear combination of the primary and its
descendants at the origin. ⌟
Chapter 3. Conformal Field Theory 73

Proposition 3.64 (Action of Conformal Algebra on Primary Operators). Let O

be a primary operator with scaling dimension ∆ and rotation generators S•• .
Then the action of the conformal algebra is then given by

[Pa , O(x)] = pa O(x) ≡ ∂a O(x) ,

[D, O(x)] = (d + ∆)O(x) ,
[Mab , O(x)] = (mab − Sab )O(x) ,
[Ka , O(x)] = (ka + 2∆xa + 2xb Sab )O(x) ,

which can be summarized for any CKV ξ as

(︂ )︂
[Qξ , O(x)] = ξ + ∆ d1 (∂ ξ) + 21 (∂ a ξ b )Sab O(x) .
•

Proof. Using a shorthand notation [Q, O] → QO, we have

(︂ )︂ (︂ )︂
DO(x) = D ex P O(0) = ex P e−x P D ex P O(0) = ex P e[−x P , ] D O(0)
• • • • • •

(︂ )︂
= ex P D + [−x P , D] + ··· O(0) = (x ∂ + ∆) ex P O(0)
• •
• •

⏞ ⏟⏟ ⏞ ⏞ ⏟⏟ ⏞
x•P ∼[P ,P ]=0
≡ (d + ∆)O(x) .

We proceed similarly also for the other generators. The final formula just needs
to be checked individually for each CKV defined in Proposition 3.37 → p.62 .
Alternatively, we could define primary/descendant operators by starting with the
transformation properties under finite conformal maps.

Definition 3.65 (Primary Operator and its Finite Transformation, Descendants).

An operator O is called a primary operator if its transformation under any
conformal map g ∈ Conf(Rd ), x ↦→ x′ ≡ g(x), is given by its scaling dimension
∆ and its representation ρ under rotations SO(d) as
(︂ )︂
Ug O(x)Ug−1 = Ωg (x)∆ ρ Rg (x)−1 • O(x′ ) ,

where we performed the decomposition Remark 3.36 → p.61 into local scaling
and rotation. Often we will keep the g subscript for Ω and R implicit.
The corresponding descendants are operators of the form P• · · · P• O, whose
transformation can be obtained from the transformation of the primary O.

Remark 3.66. It is easy to confirm that transformations of the primary operators

indeed form a representation, that is Ug Uh = Ugh for each g, h ∈ Conf(Rd ).
We can think of the conformal representation corresponding
(︂ to the primary
)︂ O
either as {CO(x) | x ∈ R }, or alternatively as SpanC {P• O(0) | n ∈ N0 } .
d n
⌟

Remark 3.67. The finite transformation Definition 3.65 → p.73 and the infinitesimal
action Proposition 3.64 → p.73 are related by Ugt ≡ etQξ , where the finite conformal
Chapter 3. Conformal Field Theory 74

transformation gt ≡ etξ ∈ Conf(Rd ) is obtained from the flow ϕξ along CKV ξ,

that is x′t ≡ gt (x) = etξ (x) ≡ ϕξt (x).
To give an example, for translations, dilatations, and rotations we have

Tx : ↦→ ex p ( ) = +x Ux ≡ UTx ≡ ex P ,
• •

Dλ : ↦→ eσd ( ) = eσ ≡λ Uλ ≡ UDλ ≡ eσD ≡ λD ,

1 ab 1 ab
Rθ : ↦→ e 2 θ mab
( ) = Rθ ( ) Uθ ≡ URθ ≡ e 2 θ M ab
.
⌟
Remark 3.68 (Scalars, Symmetric Traceless Tensors, Spin). We will primarily con-
sider scalar operators, or more generally the symmetric traceless tensor irreducible
representations of SO(d). A primary operator O∆,J with scaling dimension ∆ and
spin J is understood to be a traceless symmetric tensor of rank J (with J indices).
Scalars have J = 0, vectors have J = 1, and so on. We refer to operators with
nonzero spin J > 0 as “spinning” operators. ⌟
Example 3.69 (Identity Operator). In every CFT there is a primary operator with
∆ = J = 0, so it is invariant under all conformal transformations. This is simply
the identity operator 1, which we already discussed in Example 1.64 → p.20 . As we
will see in Section 3.4 → p.78 , its uniqueness corresponds to the uniqueness of the
vacuum state. ⌟
Example 3.70 (Fundamental Fields). The fundamental fields ϕ over which we
integrate in the path integral are always primary operators, and together with
their descendants/derivatives they form the corresponding conformal multiplet. ⌟
Example 3.71 (Double-Twist Operators). Quite generally, but in particular for
MFTs (see Example 2.26 → p.48 ), given two primaries O1 and O2 — for example
fundamental fields ϕ — there are primaries which are composites of O1 , O2 in the
schematic form
[O1 O2 ]n,J ≃ O1 n ∂ J O2 − (traces) ,
where the derivatives act appropriately on both operators to ensure [O1 O2 ]n,J is a
primary, and the traces are subtracted such that the final operator transforms in
a spin J symmetric traceless representation.
Their appearance from the thermal partition function point of view can be seen
for example in [44]. See Appendix F [45] and Appendix C [46] for derivations
of explicit formulas. ⌟

Anomalous Dimensions
Scaling dimensions of primary operators are fundamental data of the theory which
figure in basically all CFT calculations, in particular the correlation functions.
Also other important quantities such as critical exponents are derived from them.
Therefore, finding their values is one of the major tasks in the study of CFTs.
This is a highly nontrivial task for interacting CFTs. As we already discussed in
Chapter 3. Conformal Field Theory 75

Section 1.1 → p.10 , we can utilize perturbation theory around a certain “free” theory,
where we have a complete knowledge of the scaling dimensions. In such approaches
we calculate quantum corrections due to interactions as a series expansion in some
appropriate parameter.

Definition 3.72 (Anomalous Dimensions). Suppose we are solving the theory

in some kind of perturbative expansion with parameter g. The expansion of
the scaling dimension of some primary operator has the form
∞
∆ = ∆(0) + g k γ (k) ,
∑︂

k=0

where ∆(0) is a known scaling dimension not containing quantum corrections,

and the deviations γ (k) (of k-th order) are the so-called anomalous dimensions.

Remark 3.73. This is a slight generalization of the anomalous dimensions discussed

in Remark 2.22 → p.46 . There we compared ∆ to the classical mass dimensions
corresponding to the free Gaussian fixed point. In this case, we consider ∆(0) to
be the scaling dimension in some CFT from which we are perturbing away. ⌟

Remark 3.74 (Energy–Dimension Correspondence in AdS/CFT). In the context of

QFTs in AdS to be discussed in Chapter 5 → p.100 , the scaling dimensions in the
CFT on the boundary correspond to energies in the bulk AdS.
In particular, the scaling dimensions of boundary operators dual to fundamental
fields in AdS correspond to their bulk masses. In our treatment these will not
obtain any anomalous dimensions, as our choice of renormalization scheme will
fix them to be the same as in the free theory.
However, the double-twist operators discussed in Example 3.71 → p.74 obtain non-
trivial anomalous dimensions due to interactions, which can be interpreted as the
binding energies of the corresponding two-particle states in the bulk. Calculation of
these anomalous dimensions will be the main focus of the final Chapter 5 → p.100 . ⌟

Correlation Functions of Primary Operators

Now that we know transformation properties of local operators in CFTs — in
particular those of the primary operators — we can use them to constrain the
form of their correlation functions. By taking appropriate derivatives we can
obtain also correlators of their descendants.

For simplicity, we will mainly consider only scalar operators.

Lemma 3.75 (Conformal Symmetry of Primary Scalar Correlators). Consider

a set {O1 , . . . , On } of primary scalar operators with scaling dimensions
Chapter 3. Conformal Field Theory 76

{∆1 , . . . , ∆n }. The conformal symmetry of the correlator reads

⟨︂ ⟩︂ ⟨︂ ⟩︂
O1 (x1 ) · · · On (xn ) = (Ug O1 (x1 )Ug−1 ) · · · (Ug On (xn )Ug−1 )
⟨︂ ⟩︂
= Ω(x1 )∆1 · · · Ω(xn )∆n O1 (x′1 ) · · · On (x′n )

for any g ∈ Conf(Rd ), x ↦→ x′ ≡ g(x).

Proof. Follows directly from Definition 3.65 → p.73 and ρ ≡ 1 for scalars.

Proposition 3.76 (Constraints on Correlators of Primary Scalar Operators).

Correlators of scalar primary operators are functions only of the distances
|xij | ≡ |xi − xj | between the insertion points in the general form
⎡ ⎤
⟨︂ ⟩︂ 1 ⎦ (︂ )︂
O1 (x1 ) · · · On (xn ) = ⎣
∏︂
cij f {u, v, . . . } ,
i<j |xij |

where f ( ) is a function of the conformal cross-ratios {u, v, . . . } build from

{x1 , . . . , xn }, and the powers cij ≡ c(ij) must satisfy
1 ∑︂
n
∀i ∈ {1, . . . , n} ∆i = cij .
2 j̸=i

Proof. Translation and rotation invariance implies that scalar correlators can
depend only on the distances between the points — there is no preferred position
or direction. Therefore, they assume a from as stated in the Proposition, or
possibly a sum over such terms with different cij and f .
Since the cross-ratios are conformal invariants, the conformal symmetry of the
correlator Lemma 3.75 → p.75 requires (terms with different cij can not cancel out)

1 ! 1
cij = Ω(x1 ) · · · Ω(xn )∆n
∆1
∏︂ ∏︂
′ cij
i<j |xij | i<j |xij |
1 1
= Ω(x1 )∆1 · · · Ω(xn )∆n
∏︂
,
i<j Ωcij /2 (x i )Ω
c ij /2 (xj ) |xij |cij

where we utilized Lemma 3.46 → p.67 . Since Ω for SCTs depends nontrivially on
the position — see Remark 3.40 → p.64 — the only way to satisfy the conformal
symmetry condition is to have the Ω-factors cancel out for each xi individually.
This is indeed the stated condition for the powers cij .
Picking any allowed {cij } and factoring |xij |−cij out, the rest of the correlator
∏︁

must be a conformal invariant, and thus just a function of the cross-ratios.

Remark 3.77. Here we ignore possible contact terms in the correlators, which
can modify the correlator at coinciding points. ⌟
Conformal symmetry completely fixes (up to overall normalization) the form of
Chapter 3. Conformal Field Theory 77

scalar correlators with n ≤ 3, since there are no conformal cross-ratios build from
3 or fewer points as shown in Proposition 3.53 → p.68 and before. Let us look at
them in more detail.

Corollary 3.78 (One-Point Functions). Let O be any primary operator not

proportional to the identity operator 1. Then its one-point function vanishes
⟨︂ ⟩︂
O(x) = 0 ,

as do one-point functions of all its descendants.

Only operators proportional to the identity operator have a nontrivial one-
point function. Due to the normalization of the correlator we have
⟨︂ ⟩︂
1(x) ≡ 1 .

Proof. With one point insertion there are no distances between the points, and
therefore it is generally not possible to satisfy condition in Proposition 3.76 → p.76
unless the correlator is trivially zero, or ∆ = 0.
The proof is concluded by noticing that the only operator with ∆ = 0 is the
identity 1, and that correlators of descendants are given by derivatives of the
primary one-point functions, which in all cases turn out to be zero.

Remark 3.79 (One-Point Functions in Different Backgrounds). Remember that we

consider CFT in flat space Rd . Things change for example in BCFTs (CFTs with
a boundary) or when we consider the theory on space with compact dimensions
(like torus). Then also other operators can have nontrivial one-point functions,
depending on the distance from the boundary or the sizes of compact dimensions.⌟

Corollary 3.80 (Two-Point Functions of Scalar Primaries). The two-point func-

tion of primary scalar operators is fixed to be of the form
⟨︂ ⟩︂ CO1 O2 δ∆1 ,∆2
O1 (x)O2 (y) = ,
|x − y|2∆1

where CO1 O2 is a constant. In particular, the correlator can be nonzero only

if the scaling dimensions of the two operators are equal, that is ∆1 = ∆2 .

Proof. The condition in Proposition 3.76 → p.76 gives us ∆1 = 12 c12 = ∆2 .

Remark 3.81 (Normalization of Operators). Considering a basis of all (real) scalar

primaries {Oi }, the constants {COi Oj δ∆i ,∆j } form a symmetric matrix. Since it
can be diagonalized, we will usually consider a basis with COi Oj δ∆i ,∆j = δij , where
we appropriately normalized the operators. From now on, any scalar primary
Chapter 3. Conformal Field Theory 78

operator O∆ with scaling dimension ∆ has the two-point function

⟨︂ ⟩︂ 1
O∆ (x)O∆ (y) = .
|x − y|2∆ ⌟
Remark 3.82. Ignoring the identity operator, we must have ∆ > 0, otherwise the
correlator would grow with increasing distance, which would violate the cluster
decomposition principle. In the following Section 3.4 → p.78 we will derive even
stronger bounds on ∆ in unitary theories. ⌟

Corollary 3.83 (Three-Point Functions of Scalar Primaries). The three-point

function of primary scalar operators is fixed to be of the form
⟨︂ ⟩︂ f123
O1 (x1 )O2 (x2 )O3 (x3 ) = ∆1 +∆2 −∆3 ∆1 +∆3 −∆2 ∆2 +∆3 −∆1
.
x12 x13 x23

where f123 is a constant depending on the operators O1 , O2 , O3 .

Proof. The condition in Proposition 3.76 → p.76 gives us 3 equations for 3 unknowns
{c12 , c13 , c23 }, which has a unique solution

∆1 = 12 (c12 + c13 ) c12 = ∆1 + ∆2 − ∆3

∆2 = 1
(c
2 12
+ c23 ) =⇒ c13 = ∆1 + ∆3 − ∆2
∆3 = 1
(c
2 13
+ c23 ) c23 = ∆2 + ∆3 − ∆1 .

Remark 3.84. Since we already fixed the normalization of the operators by

requiring the specific form of the two-point function in Remark 3.81 → p.77 , the
coefficients fijk are important physical data of the theory. We will discuss them
in more detail in Section 3.5 → p.86 . ⌟
Remark 3.85. The proofs in this section utilized the finite form of conformal
symmetry for the correlators, but we could equivalently use Ward–Takahashi
identities for conformal transformations to obtain the same results. ⌟
The higher-point (n ≥ 4) correlators in general have nontrivial dependence on
cross-ratios, and their functional form is therefore not fixed by conformal symmetry.
Still, as we will see in the following sections, they can be expressed in terms of
data present in the two-point and three-point functions.

3.4 Radial Quantization

Up to now we studied the conformal symmetry/algebra and the correlation func-
tions of local operators in CFTs on general grounds without choosing any particular
operator formalism in the sense of Section 1.4 → p.29 . But without choosing some
foliation of the spacetime manifold, and considering the corresponding Hilbert
spaces, it is not clear how to investigate important questions pertaining to the
unitarity/reflection positivity of the theory.
Chapter 3. Conformal Field Theory 79

Radial Foliation and Cylinder Picture

As already alluded to in Section 3.3 → p.69 , in CFTs it is natural to foliate the
spacetime by concentric spheres around a chosen origin. Such radial foliation
is adapted to the action of dilatations and rotations generated by D and M•• ,
which nicely matches with the way we classified the local operator conformal
representations.

As the Hilbert spaces HΣ on all different spheres Σ ≃ Sd−1 around the origin are
actually isomorphic — they are related by the scaling symmetry of the theory
generated by the “radial Hamiltonian” D — we are able to formulate the theory
on a single Hilbert space H. Recall Remark 1.67 → p.21 and Section 1.4 → p.29 for
the general story.

But how do we obtain the underlying Hilbert space, together with the inner
product and corresponding conjugation? The idea is to start with the cylinder.

Idea 3.86 (Cylinder Picture and Radial Quantization). The radial quantization
comes from canonically quantizing the theory on a cylinder R × Sd−1 , and then
mapping the cylinder to the flat space Rd via a Weyl/conformal transformation.
In more detail, by expressing x ∈ Rd ∼ = R≥ × Sd−1 as x ≡ |x|x̂ ≡ rn, and then
rescaling the radial coordinate using ϕ : R → R> , τ ↦→ r = ϕ(τ ) ≡ eτ , we obtain

dr2
(︄ )︄
ϕ (︂ )︂
gRd ≡ dr + r gSd−1 = r
2 2 2
+ gSd−1 ≃ e2τ dτ 2 + gSd−1 ≡ e2τ gR×Sd−1 ,
r2

so after a Weyl transformation we obtain the flat metric on the cylinder R × Sd−1 .
Note that only for CFTs we are able to perform such Weyl rescaling to relate the
theory on flat space to the theory on the cylinder.
In the Figure 3.87 → p.80 we clearly see that spheres of the same radius r ≡ |x|
become constant τ slices, and dilatations x ↦→ λx become shifts of cylinder time
τ ↦→ τ + ln λ. Therefore, D now generates isometries via e−τ D , and can indeed
be interpreted as the Hamiltonian of the theory on the cylinder. Another point
worth stressing is that operator insertions at the origin x = 0 correspond to the
preparation of state at τ = −∞. ⌟

Using the Weyl transformation of a correlator Example 1.73 → p.25 with Ω(x) = eτ ,
we obtain transformation of correlator between cylinder and flat space in the form
(here {yi } are points on the cylinder R × Sd−1 )

(︄ n )︄
⟨︂ ⟩︂ ⟨︂ ⟩︂
O1 (y1 ) · · · On (yn ) = eτi ∆ i
O1 (y1 ) · · · On (yn )
∏︂
gR×Sd−1 e2τ gR×Sd−1
i=1

where the right-side (after change of coordinates with ϕ) is the correlator in flat
space Rd .
Chapter 3. Conformal Field Theory 80

R × Sd−1
Rd
r = eτ

r ϕ

ϕ−1 τ
n τ = ln r
n

r ≡ |x| = 0 τ = −∞
©2025 Jonas Dujava

Figure 3.87 / Conformal map between (flat space Rd \{0} R × Sd−1 cylinder).
By adding two additional points at the infinities of the cylinder we actually obtain a
conformal compactification of the flat space Rd ≃ (R × Sd−1 ) ∪ {±∞}.

Definition 3.88 (Cylinder Operator, Flat Operator). The cylinder operator Ocyl.
inserted at (τ, n) ∈ R × Sd−1 corresponds to the flat operator Oflat inserted
at x ≡ ϕ(τ, n) ≡ eτ n ∈ Rd by

Ocyl. (τ, n) eτ ∆ Oflat (x ≡ eτ n) ,

meaning that correlators of Ocyl. on the cylinder R × Sd−1 can be calculated

in the flat space Rd by considering their flat-space equivalents eτ ∆ Oflat .

Remark 3.89. We often omit subscripts “cyl.”/“flat” and rely on the coordinates
to distinguish which operator or space we have in mind. ⌟

The canonical quantization on the (Lorentzian) cylinder R×Sd−1 gives us a unitary

CFT. In particular, the complex vector space of states H contains a distinguished
vector |vac⟩ ∈ H invariant under all conformal transformations, and H carries a
positive-definite inner product ⟨ , ⟩ (as usual ⟨vac| ≡ ⟨|vac⟩, ⟩ ∈ H∗ ).

After performing the Wick rotation to the Euclidean cylinder we obtain a reflection-
positive Euclidean CFT, where (recall Section 1.4 → p.32 ) the conjugation acts as
[︂ ]︂
b1 ...bl ∗ a1 ...al ∗
a1 ...al
Ocyl. (τ, n)† = Θab11 · · · Θabll Ocyl. (−τ, n) ≡ Θ Ocyl. (τ, n) .

The radial quantization of CFT in Rd borrows all of this structure and thus
becomes a reflection-positive CFT as well (with an appropriate conjugation).
Using Definition 3.88 → p.80 , we see that the cylinder time-reflection (τ ↦→ −τ )
becomes the radial inversion I defined in Remark 3.39 → p.63 .
Chapter 3. Conformal Field Theory 81

Definition 3.90 (Conjugation in Radial Quantization). A local operator with

spin l behaves under radial conjugation as
(︂ )︂ [︂ ]︂
a1 ...al
Oflat (x)†rad = |x|−2∆ Iba11 · · · Iball Oflat
b1 ...bl ∗ x
|x|2
a1 ...al ∗
≡ I Oflat (x) ,

where Iba ≡ δba − 2x̂a x̂b , so the radial inversion operation I in addition to
inverting the position argument also includes the Weyl factor ΩI (x)∆ ≡ |x|−2∆
and a reflection of the indices along the radial direction.

Remark 3.91. We have written above †rad to distinguish the Hermitian conjugation
in the radial quantization (corresponding to D) from the conjugation in the usual
inertial quantization (corresponding to P0 ).
In the following we will often drop the subscript “rad”, as it should be clear when
we are working in the radial quantization. ⌟

Corollary 3.92 (Radial Conjugation of Conformal Generators). The radial

conjugation of a conserved conformal charge Qξ , ξ ∈ conf(Rd ), is given by

Q†ξ = Q−AdI ξ ≡ Q−IξI ,

so in particular

D† = D , †
M•• = −M•• , P•† = K• .

Idea of the Proof. Obtained from the definition of radial conjugation and relations
in the proof of Lemma 3.41 → p.65 , similarly as in Remark 1.100 → p.34 .

Definition 3.93 (Correlators in Radial Quantization Picture, Radial Ordering). In

the radial quantization, a correlation function of local operators is interpreted
as a radially-ordered vacuum expectation value (VEV)
⟨︂ ⟩︂ (︂ )︂
O1 (x1 ) · · · On (xn ) = ⟨vac| R O1 (x1 ) · · · On (xn ) |vac⟩ ,

with the radial ordering of operators inserted at n distinct points defined as

(︂ )︂
R O1 (x1 ) · · · On (xn ) ≡ Oσ(1) (xσ(1) ) · · · Oσ(n) (xσ(n) ) ,

where σ ∈ Sn is a permutation such that |xσ(1) | ≥ · · · ≥ |xσ(n) |.

Remark 3.94. We assume O• (x) and O• (y) commute if x ̸= y and |x| = |y|. ⌟

Remark 3.95. We do not explicitly distinguish between the operator insertion

O(x) and its particular representation by an operator O(x)
ˆ︁ ∈ End(H). It should
be clear from the context which one we mean. ⌟
Chapter 3. Conformal Field Theory 82

Remark 3.96 (Change of Origin). Choosing a different origin would amount to

working with a different Hilbert space and associated operator representation.
Nonetheless, it would be isomorphic to the original one, and calculation of appro-
priately radially-ordered correlators would yield the same results. ⌟

State–Operator Correspondence
Consider the following assignments:

• Operator ↦→ State — An insertion of a local operator O(0) at the origin

x = 0 prepares a state |ψ⟩ ∈ HΣ ≃ H by performing the path integral over
the interior of some sphere Σ centered at the origin (essentially with arbitrary
radius r assuming no other operator insertions).

• State ↦→ Operator — A state |ψ⟩ ∈ HΣ ≃ H on sphere Σ of some radius

can be “back-evolved” by dilatations — which are symmetry transformations
of the theory — to an arbitrarily small sphere Sd−1 ε of radius ε around the
origin, so it defines a local operator O(0) inserted at the origin.

We have thus come to one of the central points of the CFT formalism, which as
we will see has various far-reaching consequences.

Definition 3.97 (State–Operator Correspondence). A Euclidean CFT enjoys

the state–operator correspondence, which provides a linear isomorphism
∼
s : V ≃ O0 Rd → H between local operators at the origin and quantum states
in the radial quantization. This gives us the identification

O(0) |O⟩ ≡ O(0)|vac⟩ ,

where |O⟩ ≡ s(O) is the state obtained by inserting O at the origin.

Remark 3.98 (FQFT Perspective). Point operators in FQFT context were defined
in Definition 1.61 → p.20 and Remark 1.63 → p.20 as a family of states living on
infinitesimally small spheres around a point. Since CFTs have the scaling symmetry,
spaces HSd−1
r
with different radii r are isomorphic, implying a state on any sphere
has a corresponding point operator at the center. ⌟

Remark 3.99 (Hilbert Space Decomposition into Irreducible Representations). Since

vacuum is invariant under conformal transformations, that is Qξ |vac⟩ = 0, the
action of conformal algebra is given by

[Qξ , O(0)] [Qξ , O(0)]|vac⟩ = Qξ O(0)|vac⟩ ≡ Qξ |O⟩ .

Therefore, studying quantum states in CFTs (in radial quantization) is same

as studying the algebra of point operators. The whole analysis performed in
Derivation 3.63 → p.72 automatically carries over also to states in H, so we can
Chapter 3. Conformal Field Theory 83

schematically write the decomposition

[︄ ∞ ]︄
CP•n O(0)
⨁︂ ⨁︂ ⨁︂
H≃ CO(x) ≃ .
all local primary n=0
operators O operators O ⌟

Remark 3.100. It would be more precise to distinguish following spaces

Hsmall ⊂ HHilb ⊂ Hbig ,

by which we mean:
• First, Hsmall is to be identified (by the state–operator correspondence) with
the complex vector space O0 Rd ≃ V of all local observables at the origin. In
particular, we consider here only finite linear combinations of point operators
inserted at the origin. Even though it carries a Hermitian inner product, it is
not a Hilbert space, since it is not complete with respect to it.
• Its completion Hbig on which all operators O(x) act (at arbitrary points), so
it contains all vectors of the form (with {xi } radially-ordered)

O1 (x1 ) · · · On (xn )|vac⟩ ,

which includes certain infinite linear combinations of operators at the origin.

It too does not have a well-defined Hermitian structure, since it contains
vectors with infinite norm — if |x| ≥ 1, the conjugation maps it to |I(x)| ≤ 1,
and the norm would be given by a VEV of non-radially-ordered operators,
which is generally infinite in Euclidean signature.
• Finally, the L2 completion HHilb of Hsmall with respect to its inner product
gives us a Hilbert space. It contains vectors of the form as above, but all
insertion points {xi } must be distinct and have |xi | < 1.
This is similar to the rigged Hilbert space construction in ordinary QM. ⌟

Example 3.101 (Identity Operator Vacuum State). We clearly have

1(0) |1⟩ ≡ 1(0)|vac⟩ ≡ |vac⟩ ,

which we already discussed in Example 1.64 → p.20 . This works even when inserting
1(x) at a general point x, since [P• , 1(0)] ≡ 0 — we can shrink any topological
surface operator to zero, which is equivalent to the invariance of |vac⟩ under
conformal transformations. ⌟

Example 3.102 (Primary State, Descendant States). From Remark 3.99 → p.82 we see
that a primary operator O(0) corresponds to a primary state |O⟩ ≡ O(0)|vac⟩ ∈ H.
The whole conformal multiplet is then given by adding an infinite tower of
corresponding descendant states above the primary
{︂ }︂ {︂ }︂
|O⟩, P• |O⟩, P• P• |O⟩, . . . ≡ |O⟩, |P• O⟩, |P• P• O⟩, . . . .
Chapter 3. Conformal Field Theory 84

Their scaling dimensions (eigenvalues of D) are given as ∆ + N , where N is the

level of the descendant (or N = 0 for the primary). This in particular means that
states at different levels are orthogonal to each other. ⌟

Example 3.103 (Dual State). Considering (for simplicity) a scalar state |O⟩, the
dual state is given by Definition 3.90 → p.81 as
(︂ )︂† (︂ )︂
⟨O| ≡ |O⟩ ≡ ⟨vac|O(0)† = lim |x|−2∆ ⟨vac|O∗ x
|x|2
= y→∞
lim |y|2∆ ⟨vac|O∗ (y) ,
x→0

where in the last step we exchanged x with y ≡ I(x). ⌟

Remark 3.104 (Casimir Eigenvalues for Symmetric Traceless Representations). The

eigenvalue of the Casimir for a symmetric traceless tensor O∆,J is given by

Cas2 |O∆,J ⟩ = λ∆,J |O∆,J ⟩ ,

[︂ ]︂ λ∆,J ≡ ∆(∆ − d) + J(J + d − 2) ,
Cas2 , O∆,J (x) = λ∆,J O∆,J (x) ,

where we could insert operator at arbitrary point since Cas2 commutes with
all generators of the conformal algebra, which also means that whole conformal
multiplet shares the same eigenvalue.
It is thus simplest to calculate the eigenvalue for the primary state. The first term
of Definition 3.60 → p.71 is the Casimir of the SO(d) subalgebra with eigenvalue
J(J + d − 2), and the rest acts on the primary as D(D − d). ⌟

Implications of Unitarity
Having developed a radial quantization formalism, we can now study the implica-
tions of the CFT unitarity — or maybe better “inversion positivity”. We will just
briefly explore some necessary conditions stemming from the non-negativity of
the inner product on H. Following will also illustrate how CFT computations can
be reduced (in radial quantization) just to algebraic manipulations.
Remark 3.105 (Norms of Scalar Primary Operators). In Remark 3.81 → p.77 we
considered a basis of all real scalar primaries {Oi }, and their two-point functions.
Now we know that in radial quantization they correspond to primary states {|Oi ⟩},
and unitarity requires positive-definiteness of the inner product matrix

⟨Oi |Oj ⟩ ≡ lim |y|2∆i ⟨vac|Oi (y)Oj (0)|vac⟩

y→∞
⟨︂ ⟩︂ COi Oj δ∆i ,∆j
= lim |y|2∆i Oi (y)Oj (0) ≡ lim |y|2∆i = COi Oj δ∆i ,∆j .
y→∞ y→∞ |y|2∆i

So COi Oj δ∆i ,∆j is not only symmetric, but also positive-definite, which allows us
to diagonalize it to δij without changing realness of {Oi }. ⌟
What does the positivity imply when we consider the descendant states?
Chapter 3. Conformal Field Theory 85

Calculation 3.106 (Positivity of N = 1 Descendants of Scalar Operators). Consid-

ering a level N = 1 descendant state Pa |O⟩, its squared norm is
⃦ ⃦2
⃦Pa |O⟩⃦ = ⟨O|Pa† Pa |O⟩ = ⟨O|Ka Pa |O⟩
⃦ ⃦
!
= ⟨O|[Ka , Pa ]|O⟩ = ⟨O|2D|O⟩ = 2∆⟨O|O⟩ ≥ 0 ,

where we used Corollary 3.92 → p.81 , Definition 3.56 → p.70 with Lemma 3.41 → p.65 ,
and that primary state satisfies K• |O⟩ ≡ 0 = ⟨O|P• and D|O⟩ ≡ ∆|O⟩.
Assuming nontrivial |O⟩, unitarity implies ∆ ≥ 0, as was already noted in
Remark 3.82 → p.78 . For ∆ = 0, that is identity operator 1 and |1⟩ ≡ |vac⟩, the
descendant becomes null, and unitarity consistently requires P• |1⟩ ≡ 0. ⌟

Similar calculation can be performed for spinning operators/states, see for example
[11] [36] [34], and references therein. In the case of scalars, at level N = 2 one
obtains a stronger condition than ∆ ≥ 0, as mentioned below.

Remark 3.107 (Unitarity Bounds for Traceless Symmetric Tensors). For traceless
symmetric tensors with spin J, the unitarity requires

∆=0 (unit operator), or

⎧
⎨ d−2 J = 0,
∆≥⎩ 2
d −2+J J > 0.

Alternatively, these unitarity bounds — and that they cannot be strengthened —

can be derived directly in the Lorentzian signature [47]. ⌟

Special things happen when the primary saturates the unitarity bound. In such
a case, some descendant state becomes null, and needs to be removed (together
with all of its descendants) from the original conformal multiplet — obtaining
thus a so-called short multiplet.

Example 3.108 (Free Scalar). Example of such short multiplet is a scalar primary
with ∆ = d−2
2
, where the null state is (nice exercise to check explicitly)

P 2 |O⟩ ≡ Pa P a |O⟩ = 0 ⇐⇒ ∂ 2 O(x) = 0 .

This means O satisfies the massless Klein–Gordon equation, and is therefore a

free massless scalar — we indeed recognize the usual value ∆, see for example
Example 2.26 → p.48 . ⌟

Example 3.109 (Conserved Currents). Consider a spin J primary |Oa1 ...aJ ⟩ with a
null state
!
Pn |Ona2 ...aJ ⟩ = 0 ⇐⇒ ∂n Ona2 ...aJ (x) = 0 .
Chapter 3. Conformal Field Theory 86

Then we can calculate

!
0 = Ka Pn |Ona2 ...aJ ⟩ = [Ka , Pn ]|Ona2 ...aJ ⟩ = (2δan D − 2Man )|Ona2 ...aJ ⟩
J
= 2∆|O aa2 ...aJ
2(δab δnn δnb δan )|Oba2 ...aJ ⟩ −2 (δab δnai − δnb δaai )|Ona2 ...b...aJ ⟩
∑︂
⟩− −
i=2
(︂ )︂
= 2 ∆ − (d − 1) − (J − 1) |O aa2 ...aJ
⟩ ⇐⇒ ∆=d −2+J.

In the first line we used that |O•···• ⟩ is a primary and a little bit of conformal
algebra. In the second line we employed the standard form of rotation generators
for each vector index. Finally, we used symmetry and tracelessness of |O•···• ⟩.
We thus verified that conserved currents correspond to a short multiplet saturating
the unitarity bound in Remark 3.107 → p.85 . From the calculation above we clearly
see also the opposite implication — a primary with ∆ = d − 2 + J has necessarily
a null state, making it a conserved current.
Important examples are global symmetry currents with (J = 1, ∆ = d − 1), and
the energy–momentum tensor with (J = 2, ∆ = d). ⌟
In the following section we will mention one additional implication of unitarity —
the reality of the 3-point function coefficients.

3.5 Operator Product Expansion

The state–operator correspondence has another essential consequence.
Idea 3.110 (Operator Product Expansion). Consider an insertion of two local
operators close enough to each other, that we can draw a sphere separating them
from all other operator insertions. They prepare some state on this sphere (in the
corresponding radial quantization), which in turn corresponds to a local operator
at the center — we can express product of two operators as a linear combination
of primaries and descendants at the origin. This is called the Operator Product
Expansion (OPE). ⌟
Remark 3.111. The notion of OPE can be introduced in a more general context
of QFTs, but even when two point operators are brought really close together,
it usually provides just an asymptotic series. Only in CFTs it is elevated to
its powerful status, since it can be shown that conformal symmetry guarantees
nonzero radius of convergence of OPE (for well separated operators) [48] [49].⌟

Operator Algebra
As described in Idea 3.110 → p.86 , we can express the product of two operators as
Oi (x1 )Oj (x2 ) ∼ COi Oj Ok (x1 , x2 , y, ∂y ) Ok (y) ,
∑︂

Ok ∈Oi ×Oj

where Ok runs through all primary operators included in the OPE, which we
denote symbolically by Oi × Oj . The function CO1 O2 O′ also includes contributions
 Chapter 3. Conformal Field Theory 87

of the corresponding descendants through the action of the momentum/derivative

operator Py ≃ ∂y acting on y.
Remark 3.112. We keep all of the indices present for spinning operators implicit.⌟
The OPE is understood as an operator equation, which holds inside the correlation
functions independent of other insertions (assuming they are outside the sphere
surrounding the operators, where the OPE converges). So we have
⟨︂ ⟩︂ ⟨︂ ⟩︂
Oi (x1 )Oj (x2 ) . . . = COi Oj Ok (x1 , x2 , y, ∂y ) Ok (y) . . .
∑︂
,
Ok ∈Oi ×Oj

independent of the precise choice of radial foliation/quantization and the corre-

sponding state–operator correspondence we used to derive/motivate it.
Remark 3.113. When possible, it is convenient for calculations to choose for
example y = x2 , which we assume in the following. ⌟

OPE Coefficients
The conformal symmetry greatly restricts the form of the OPE. For simplicity, we
will take all of the operators Oi , Oj , Ok to be scalars.
The translation invariance gives us that COi Oj Ok (x1 , x2 , ∂) can depend on positions
only through the combination x12 ≡ x1 − x2 . Utilizing also scaling and rotation
invariance — or alternatively consistency of OPE with the action of D and M••
— we obtain expansion for |x12 | ≪ 1 of the form
1 (︂ )︂
COi Oj Ok (x12 , ∂) ∝ ∆ +∆j −∆k
1 + (#)x12 ∂ + . . .
•

x12i
Finally, consistency with the action of K• completely fixes COi Oj Ok up to an
overall constant.
Remark 3.114 (Fixed Form of OPE, OPE Coefficients). More elegant way to see
that OPE is fixed up to a constant is to reduce the computation of 3-point
correlator to 2-point functions as
⟨︂ ⟩︂ ⟨︂ ⟩︂
Oi (x1 )Oj (x2 )Ok (x3 ) = COi Oj Ok′ (x12 , ∂2 ) Ok′ (x2 )Ok (x3 ) .
∑︂

k′

Now we can utilize the fixed forms of 2-point and 3-point functions as shown in
Corollary 3.80 → p.77 (in diagonal form Remark 3.81 → p.77 ) and Corollary 3.83 → p.78 ,
giving us
fijk ! 1
∆i +∆j −∆k ∆i +∆k −∆j ∆j +∆k −∆i
= COi Oj Ok (x12 , ∂2 ) .
x12 x13 x23 x2∆
23
k

We see that COi Oj Ok ≡ fijk Cijk is proportional to the 3-point function coefficient
fijk , which from now on we call OPE coefficient. Matching the small |x12 |/|x23 |
expansion of both sides completely fixes the “reduced” function Cijk in terms of
the operator dimensions ∆i , ∆j , ∆k (if the operators were spinning, then also their
spins), and the spacetime dimension d. ⌟
Chapter 3. Conformal Field Theory 88

Remark 3.115 (Recursive Calculation of Correlators Through OPE). More generally,

using the Operator Algebra/OPE we can recursively reduce calculation of any
correlator eventually down to a sum of two-point functions, or even one-point
functions (then use Corollary 3.78 → p.77 ). ⌟
From previous remarks we see that to calculate any correlator of local operators,
it is enough to know all primary operators and their OPE coefficients. The
contribution of descendants is fixed by the conformal symmetry.

Definition 3.116 (CFT Data). Essential information about a given CFT is

encoded in the corresponding
{︂ }︂
CFT data ≡ (∆i , ρi ), fijk ,

containing the conformal representations of all primary operators {Oi } —

scaling dimensions ∆i and their SO(d) representations ρi — together with all
of the OPE coefficients fijk ≡ ope[Oi Oj Ok ].

Remark 3.117 (Associativity of OPE, Overcompleteness of CFT Data). We can

reduce down the correlator by performing the OPE in different order, but we
should always obtain the same result — the OPE is necessarily “associative”.
This induces relation in the CFT data, which tells us two things — an arbitrary
collection of “CFT data” does not necessarily define a consistent CFT, and when
it does, the CFT data are overcomplete. ⌟

Remark 3.118. The same OPE is even valid on any (locally) conformally flat
background — for example we can consider boundaries or compact dimensions.
Only thing that changes in the calculation of correlators is the form of one-point
functions, as mentioned in Remark 3.79 → p.77 ⌟

Remark 3.119 (Unitarity, Reality of OPE Coefficients). Apart from the unitarity
bounds on the scaling dimensions discussed in Remark 3.107 → p.85 , the unitarity
of the CFT also implies that the OPE coefficients are real [11] [34]. ⌟

3.6 Conformal Blocks

As we already know from Section 3.3 → p.75 , the form of higher-point correlators
is not completely fixed by conformal symmetry. In this section we will focus on
4-point functions of identical real scalars Oϕ with scaling dimension ∆ϕ .

Conformal Block Decomposition

Using Proposition 3.76 → p.76 we can write

⟨︂ ⟩︂ ⟨︂ ⟩︂ g(u, v)
O1 O2 O3 O4 ≡ Oϕ (x1 )Oϕ (x2 )Oϕ (x3 )Oϕ (x4 ) ≡ 2∆ 2∆
,
x12 ϕ x34 ϕ
Chapter 3. Conformal Field Theory 89

where the function g(u, v) of the conformal cross-ratios depends on the dynamics
of the theory. Alternatively, by pairing up the operators and performing the OPE
(assuming points are configured appropriately), we can express it in terms of CFT
data schematically as
⟨︂ ⟩︂ ⟨︂ ⟩︂
(O1 O2 )(O3 O4 ) = fϕϕO fϕϕO′ C• (x12 , ∂2 )C• (x34 , ∂4 ) O• (x2 )O′• (x4 )
∑︂

O,O′
I •• (x24 )
= 2
C• (x12 , ∂2 )C• (x34 , ∂4 )
∑︂
fϕϕO
O x2∆
24
[︂ ]︂ (s) (︂ )︂
ope2 Oϕ Oϕ O G∆,J {xi } .
∑︂
≡
primary O
with ∆,J

(s)
In the last line we defined the conformal block G∆,J , which encodes the contri-
bution of exchanging a single conformal multiplet — a primary O with all of its
descendants — in the s-channel (O1 O2 )(O3 O4 ).
Remark 3.120. Only spin-J symmetric traceless tensors appear in OPE of two
scalar fields. We denoted by I •• the corresponding (unique) 2-point tensor
structure. ⌟
So knowing the CFT data, we can compute 4-point functions. We can also turn this
around — having some alternative way of computing 4-point functions (perhaps
in some perturbative expansion), we can try to extract the CFT data, which is
particularly convenient after performing the conformal block decomposition. This
is what we shall do in the following Chapter 5 → p.100 .

For more details, such as interpretation of the conformal block decomposition

in the radial quantization, and an elegant way of computing conformal blocks
through a differential equation coming from action of the conformal Casimir
Definition 3.60 → p.71 , see [11] [34] and references therein.

Let us make a final remark about yet another 4-point function decomposition,
which we shall discuss more in Section 5.3 → p.113 .

Remark 3.121 (Conformal Block versus Conformal Partial Wave Decomposition).

The decomposition of 4-point function into conformal blocks can be understood as
decomposition into “physical” irreducible representations of the conformal group
Conf(R1,d−1 ) ≃ SO(2,
˜︃ d) unitarily represented on the Hilbert space H, which we
often consider Wick-rotated to the Euclidean signature.
Alternatively, when we have a Euclidean 4-point correlator, it is also natural to
decompose it into irreducible representations of the Euclidean conformal group
SO(1, d + 1). This so-called conformal partial wave decomposition corresponds
to Unitary Irreducible Representations (UIRs) of SO(1, d + 1), but this time
“unitary” corresponds to the Euclidean conformal group, which is different from
the positivity conditions coming from the “physical” Lorentzian signature. ⌟
 Chapter 3. Conformal Field Theory 90

Crossing Symmetry
We can express 4-point function in various channels, so for example we could
perform OPE in the t-channel (O1 O3 )(O2 O4 ) or u-channel (O1 O4 )(O2 O3 ), and
obtain the corresponding conformal block decomposition. All of these expressions
should in the end agree, which leads to non-trivial constraints on the CFT data.
Remark 3.122. It can be shown that crossing symmetry of all 4-point functions
is equivalent to the associativity of the OPE (see Remark 3.117 → p.88 ). ⌟
Remark 3.123. One conclusion of the crossing symmetry is that any CFT must
have an infinite number of primary operators. In the case of MFT, the exchange
of identity operator in the t-channel corresponds to exchanges of the double-twist
operators Example 3.71 → p.74 in the s-channel. ⌟

Conformal Bootstrap
We have various consistency constraints/conditions on a CFT data defining a
consistent (unitary) CFT, which include:
• Crossing Symmetry, which is essentially the OPE associativity,
• Unitarity/Reflection Positivity — implies lower unitarity bounds on scaling
dimensions, and a possibility to choose a real orthonormal operator basis
where all OPE coefficients are real,
• (optionally) Existence of Energy-Momentum Tensor — conserved rank-2
symmetric traceless primary operator (with ∆ = d) with correlators obeying
Ward–Takahashi identities.
One can then try to study these conditions — in particular the crossing equation
— not necessarily by finding exact solutions, but by excluding regions of the CFT
data which are inconsistent. In this way it is possible to obtain non-perturbative
bounds on the scaling dimensions and OPE coefficients, which in some cases can
be surprisingly strong. For a review of the Conformal Bootstrap program, see for
example [34].
Remark 3.124. In some lucky cases, the theory which we are interested in lies at
(or very close to) the boundary of the allowed region of CFT data. Then we are
able to obtain precise data of the given theory — non-trivial example being the
Ising model in d = 3.
If our physical theory of interest is not at the boundary, we can try to input some
model-dependent information to the bootstrap procedure. ⌟
 91

4
QFT in Anti–de Sitter Spacetime
In this chapter we briefly introduce the QFT in AdS spacetime, and mainly focus
on the relation to corresponding CFT on the boundary of AdS.
More details can be found for example in [50] [33] [51] [52].

4.1 Anti–de Sitter Spacetime

We will introduce AdS spacetime in several ways, each offering various insights
into its structure and properties. Often we will work in the Euclidean signature,
that is with EAdS, and sometimes we refer to both as AdS.
Remark 4.1 (Dimensions). We will work with AdSD≡d+1 spacetime, where D is
the number of bulk dimensions, and d is the number of spatial dimensions, or
alternatively the dimension of the asymptotic boundary. ⌟

Maximally Symmetric Space and Solution of Einstein’s Equations

Both Lorentzian and Euclidean AdS are maximally symmetric manifolds of constant
negative curvature — they are homogeneous, isotropic, and admit a maximal
number of Killing vectors. This will be easily seen in the “embedding formalism”.
It is also a vacuum solution (zero energy–momentum tensor Tab = 0) of Einstein
Field Equations Ricab − 12 Rgab + Λgab = 0 with negative cosmological constant
Λ ≡ − ℓ12 d(d−1)
2
< 0, where ℓ is the radius of curvature of AdSd+1 .
Remark 4.2 (Curvature of Maximally Symmetric Spaces). To see this, we can utilize
that a maximally symmetric space(time) M (dim M ≡ D ≡ d + 1) — both
homogeneous and isotropic — has no special point, so the Ricci scalar curvature
R must be constant. Furthermore, we necessarily have [53]
R
Ricab = gab ,
d +1
R 1
Rabcd = ga[c gb]d ≡ − 2 ga[c gb]d ,
d(d + 1) ℓ
otherwise either the homogeneity or isotropy would be violated — there is no
other invariant tensor (with respect to isometries) apart from g which could be
used to construct the curvature.
Einstein’s equations then reduce to relation between R and Λ
d +1 2Λ
R = 2Λ =⇒ Ric = g,
d −1 d −1
so Einstein’s equations are automatically satisfied by choosing the scale ℓ or Λ
appropriately. ⌟
Chapter 4. QFT in Anti–de Sitter Spacetime 92

Remark 4.3 (de Sitter Spacetime). More interesting from the cosmological point
of view is the case of positive cosmological constant Λ > 0 (which seems to be the
case in our Universe), and corresponds to de Sitter spacetime dS. It is used to
model the period of inflation, and also the late-time acceleration of the Universe
(where effect of the matter content is negligible). Lately, there has been progress
in understanding QFT in dS through techniques coming from AdS via analytic
continuation, see for example [54]. ⌟

Hyperboloid/Quadric/Quasi-Sphere
Perhaps the most convenient definition of AdSd+1 is as a (universal cover) of
hyperboloid/quadric/quasi-sphere in the embedding/ambient space R2,d with
metric
η (2,d) ≡ diag(−, −, +, . . . , +) .
⏞ ⏟⏟ ⏞
d

The Anti–de Sitter spacetime (with compactified time coordinate) is defined as

{︂ ⃓ }︂
AdSd+1 /Z ≡ X ∈ R2,d ⃓ X 2 ≡ η (2,d) (X, X) = −ℓ2 ,
⃓

where ℓ is the radius (of curvature) of AdS. We often choose units such that ℓ = 1.
In the Euclidean signature we just work with η (1,d+1) .
Remark 4.4 (Isometries of AdS, AdS as a Homogeneous Space). From such de-
scription we immediately see that the isometry group of AdSd+1 is Iso(AdSd+1 ) =
SO(2, d), since both the metric and the hyperboloid are invariant under its action.
Clearly, action of AdS isometry group is transitive — for example, each point
can be brought to some canonical point, such as X∗ = (−1, 0, 0, . . . , 0). Thus, by
orbit-stabilizer theorem, we have (up to identifications)

SO(2, d) SO(2, d)
AdSd+1 ≡ OrbitX∗ = = .
StabX∗ SO(1, d)

More generally, every so-called homogeneous manifold is a coset space G/H, where
G is the isometry group, and H is isotropy group. ⌟
Remark 4.5 (Closed Timelike Curves). Note that the hyperboloid defined above
has closed non-contractible timelike curves associated with the action of SO(2) ⊂
SO(2, d) in the (−1–0)-plane, for example we have (X −1 )2 + (X 0 )2 = ℓ2 for X i = 0.
Only after decompactifying (unrolling) along these cycles do we obtain the AdS
as we understand it. ⌟
Asymptotically in the limit of large components of X — after dividing the equation
defining the hyperboloid by a large constant factor — the hyperboloid approaches
the light-cone {︂ ⃓ }︂
C = P ∈ R2,d ⃓⃓ P 2 ≡ η (2,d) (P, P ) = 0 .
We can thus define the asymptotic/conformal boundary of AdS as the set of light
(half)rays going through the origin, that is
∂AdS = (C\{0})/∼ ,
Chapter 4. QFT in Anti–de Sitter Spacetime 93

where we identify P ∼ P ′ ⇐⇒ P = λP ′ with λ > 0. This is precisely the

conformal compactification of flat space. For the illustration of this in Euclidean
signature, see Figure 4.6 → p.93 .

EAdSd+1 , X 2 = −1
∂EAdS

P2 = 0

R1,d+1

©2025 Jonas Dujava

Figure 4.6 / EAdSd+1 embedded in R1,d+1 by the equation X 2 = −1.

Its boundary is identified with the null rays P 2 = 0, P ∼ λP , which is
precisely the conformal compactification of Rd .

We usually work with certain representatives of the rays (equivalence classes)

defining ∂AdS by choice of particular section of C. One convenient choice includes
!
the Poincaré section intersecting the cone diagonally, that is fixing P + ≡ P −1 +P 0 =
1, which leads to interpretation of the boundary as a flat Minkowski. A horizontal
section fixed by P −1 = 1 is adapted for global coordinates.
Remark 4.7 (Change of Boundary Defining Section). Changing the section by
choosing different representatives of boundary points induces a conformal transfor-
mation. Indeed, having some parametrization of the boundary point representative
P (lying on a chosen section Σ of the null cone), the induced metric is
⃓ ⃓ ⃓ ⃓
g∂AdS,Σ ≡ g ⃓ ≡ η ⃓ = (dX)2 ⃓ = ηab dX a dX b ⃓ = ηab dP a dP b ≡ (dP )2 .
⃓ ⃓ ⃓ ⃓
AdS Σ Σ Σ

After a change of section Σ ∋ P ↦−→ P ′ ∈ Σ′ with P ∼ P ′ ≡ λ(P )P , using

P 2 ≡ 0 =⇒ d(P 2 ) = 0 =⇒ P dP = 0 •

we obtain
2
g∂AdS,Σ′ ≡ (dP ′ ) = (P dλ + λ dP )2
= P 2 dλ2 + 2P dP λ dλ + λ2 dP 2 = λ2 (dP )2 ≡ λ2 g∂AdS,Σ ,
•

which is precisely a Weyl transformation of the original boundary metric.

Considering that boundary of AdS is in infinity, and we can “really” obtain it only
after conformal compactification of AdS, it is natural that we can determine only
Chapter 4. QFT in Anti–de Sitter Spacetime 94

the conformal class of ∂AdS, which has no canonical choice of metric structure. If
we would like to put a QFT theory on the boundary, it thus seems natural that it
will be a CFT. ⌟
We will further discuss the embedding formalism in Section 4.2 → p.95 .

Coordinate Systems
Now we will briefly discuss some coordinate systems on AdS, even though we will
mainly work directly in the embedding space.

Global coordinates. We can parametrize the whole hyperboloid by (t, r, ω • ) as

(except for the usual polar-type coordinate singularities)
⎧ √
ι : (S1 , R≥0 , Sd−1 ) −→ AdSd+1 /Z ⎪
⎪ X −1 = ℓ2 + r2 cos(t/ℓ),
⎪
⎨ √
( t , r , ω • ) ↦−→ X = ι(t, r, ω • ) ≡ X 0 = ℓ2 + r2 sin(t/ℓ),
⎪
X i = rω i for i ∈ {1, . . . , d} ,
⎪
⎪
⎩

where t ∈ [0, 2π) is angle in (−1–0)-plane, r ∈ R≥0 is radial parameter, and ω is a

unit vector parametrizing Sd−1 , so i (ω i )2 ≡ 1.
∑︁

The induced metric is then

(︄ )︄ (︄ )︄−1
r2 r2
gAdSd+1 = ι (η
∗ (2,d)
) = . . . = − 1 + 2 dt2 + 1 + 2 dr2 + r2 gSd−1 .
ℓ ℓ

Going to the universal cover means we just unwrap the circle S1 ∋ t and take
t ∈ R from the whole real line.
Clearly, AdS is static with ξt ≡ ∂
∂t
being the timelike Killing vector, which
corresponds to J−1,0 = . . . = ℓ ∂t
∂
.

Poincaré/upper-half plane. The idea is to find slicing of AdS which is confor-

mally flat. We define X ± ≡ X d ± X −1 , and write X + ≡ z1 , such that z → 0+ we
go to the asymptotic boundary. For z → ∞ we approach the light-like surface
X + = 0, which is called the Poincaré horizon. If we want to be conformally flat
a 2 2
we choose X a = xz , and finally by X 2 = −1 we must choose X −1 = x +z z
. We can
cover half of AdS/Z by
1 + x2 + z 2
⎧
X −1 =ℓ
⎪
⎪
,
2z
⎪
ι : (R≥0 , R1,d−1 ) −→ AdSd+1 /Z
⎪
⎪
⎪
⎪
xa
⎪
⎨
( z , x ) ↦−→ X = ι(z, x) ≡ ⎪ X a = ℓ for a ∈ {0, . . . , d − 1} ,
⎪ z
1 − x2 − z 2
⎪
⎪
⎪
⎩ Xd = ℓ
⎪
.
⎪
⎪
2z
The metric is then given by
R2 (︂ 2 )︂ R2 (︂ 2 )︂
g= dz + η ≡ dz + η ab dx a
dx b
.
z2 z2
Chapter 4. QFT in Anti–de Sitter Spacetime 95

Similarly, we can cover the whole EAdS by

1 + x2 + z 2
⎧
X −1 =ℓ
⎪
⎪
,
2z
⎪
ι : (R≥0 , Rd ) −→ EAdSd+1
⎪
⎪
⎪
⎪
( z , x ) ↦−→ X = ι(z, x) ≡ X 0 = ℓ 1 − x − z ,
2 2
⎪
⎨
⎪
⎪
⎪ 2z
a
x
⎪
⎪
⎩ Xa = ℓ for a ∈ {1, . . . , d} .
⎪
⎪
⎪
z

The metric has the same form as for Lorentzian Poincaré coordinates, just in the
Euclidean signature, so it can be seen as a warped product of R≥ and Rd . In
particular, this shows EAdSd+1 is conformal to R≥ × Rd , with the boundary at
z = 0 being Rd (if we also consider z = ∞, we have the whole R≥ × Rd ).
These coordinates make explicit the subgroup
SO(1, 1) × ISO(d) ≡ SO(1, 1) × Iso(Rd ) ⊂ Iso(EAdS)
corresponding to the dilatations and Poincaré transformations in Conf(Rd ).

4.2 QFT in AdS

We will not extensively cover the QFT in AdS, so let us quickly mention some
important points:
• Due to the maximal symmetry, similarly as for Minkowski [25], the natural
choice of vacuum is invariant under isometry group Iso(AdS).
• The presence of timelike boundary — the light rays reach it in finite global
coordinate time — means the Cauchy problem is ill-defined, and we need to
additionally fix the boundary conditions [55].
• Particles can be categorized by irreducible representations of Iso(AdSd+1 ), and
since it is exactly the conformal group Conf(Minkd ), it is not surprising that
there is a large overlap between the techniques used in CFT and QFT in AdS.
In particular, a lot can be said just using the group/representation theory.
• The AdS background provides a natural IR regulator — it can be interpreted
as a “gravitational box”, which is however really symmetric. This also means
the states have a discrete spectrum — similarly to how in CFTs the scaling
dimensions of descendants increase discretely — and there is a possibility of
slightly negative mass (BF bound) [56]
• Flat space limit of AdS can be interesting, since it would allow us to port CFT
methods such as conformal bootstrap to study flat-space S-matrix [57].

Embedding Formalism
It is very convenient to work directly in the embedding space, since the action of
isometry group is very simple. Let us mention some important points:
Chapter 4. QFT in Anti–de Sitter Spacetime 96

• It is simple to obtain geodesics in AdS. In particular, the timelike geodesics

turn out to be periodic with a specific period (depends on the chosen length-
scale). This can be understood as a residual effect of obtaining AdS as a
universal cover of the space with closed timelike loops (the geodesics are
essentially the same for both spaces, we just unwrap them).
• It leads to a simple form of bulk-to-boundary propagator, as we will see
in a moment. Moreover, the embedding formalism is also very efficient for
calculating bulk-to-bulk propagators of arbitrary spin.
Remark 4.8 (AdS Isometries, Quadratic Casimir). Generators of AdS isometries
are (their action on AdS functions does not depend on their extension to the
embedding space, since they are “internal”, as they fix the condition X 2 = −R2 )

JAB ≡ XA ∂B − XB ∂A ,

giving us the quadratic Casimir operator

Iso(EAdSd+1 ) 1
Cas2 ≡− JAB J AB = −(X A ∂ B − X B ∂ A )XA ∂B
2
= −X 2 ∂ 2 + (d + 1)X ∂ + X A X B ∂A ∂B
•

= −X 2 ∂ 2 + X ∂ (d + X ∂) ,
• •

where ∂ is the derivative acting in the embedding space R1,d+1 .

Since action of Iso(EAdS) on the boundary is the same as Conf(Rd ), we obtain
Conf(Rd )
⃓
Iso(EAdSd+1 ) ⃓
Cas2 = Cas2 ⃓ ,
∂EAdSd+1 ≃Rd

where the left-hand side is Definition 3.60 → p.71 . ⌟

Calculation 4.9 (d’Alembertian in Embedding Coordinates). By foliating the flat
space R2,d with AdSd+1 hyperboloids of radii R, we have
1/ 1/
gR2,d = − dR2 + gAdSd+1 =⇒ gR22,d = gAdS
2
d+1
∝ Rd+1 .

This gives us the d’Alembertian operator on the flat space as

1 (︂ 1 )︂
R2,d ≡ ∂2 = 1/
∂A g /2 gRAB
2,d ∂B
g2
1 (︂ )︂
= − d+1 ∂R Rd+1 ∂R + AdSd+1
R
1
= − 2 X ∂(d + X ∂) +
• •
AdSd+1 ,
R
where we used X ∂ = R∂R .
•

Comparing with Remark 4.8 → p.96 , we see that (using X 2 = −R2 ≡ −ℓ2 )
Iso(AdS) Conf(∂AdS)
R2 AdS = Cas2 ≃ Cas2 .
Chapter 4. QFT in Anti–de Sitter Spacetime 97

This is a particular case of a more general relation — the d’Alembertian operator

on a homogeneous spacetime M is essentially the Casimir operator of Iso(M). ⌟
Remark 4.10 (Mass–Scaling Dimension Relation). Using the previous calculation
and considering a free Klein–Gordon equation
(︂ )︂
Conf(∂AdS)
AdS + m ϕ(x) = 0 =⇒ m2 R2 = Cas2 = ∆(∆ − d) ,
2
−

we identify the mass-squared m2 (in the units of 1/R ≡ 1/ℓ) with the quadratic
Casimir eigenvalue of the corresponding CFT representation on the boundary.
Here we utilized Remark 3.104 → p.84 for a scalar representation (J = 0). ⌟

Bulk-to-Bulk Two-Point Function/Propagator

Consider a free scalar field on the EAdS background with Euclidean action
∫︂ [︂ ]︂
S[ϕ] = dX 1
2
(∂ϕ)2 + 21 m2 ϕ2
EAdSd+1
∫︂ [︂ ]︂
≡ dd+2 X θ(X −1 )2δ(X 2 + ℓ2 ) 21 (∂ϕ)2 + 12 m2 ϕ2 .
R1,d+1

For homogeneous space we can always absorb the possible coupling to the curvature
into the mass term.
The free bulk-to-bulk two-point function ⟨ϕ(X)ϕ(Y )⟩free ≡ Gbb (X, Y ) is given by
the propagator solving the equation
(︂ )︂
− X + m2 Gbb (X, Y ) = δ(X, Y ) .
Here, X is the EAdS d’Alembertian operator acting on the first argument, where
both X and Y are understood as points of EAdS.
The symmetry of the problem implies that Gbb is a function of only invariant
quantities such as X Y or ζ ≡ (X − Y )2 /R2 , which we call chordal distance.
•

From this point on we set R ≡ ℓ ≡ 1, and all lengths are measured in units of the
AdS radius.
Calculation 4.11 (Free Bulk-to-Bulk Propagator). Using ansatz Gbb (X, Y ) ≡ G(ζ),
expressing d’Alembertian in embedding coordinates using Calculation 4.9 → p.96
and Remark 4.8 → p.96 we obtain (it does not make difference with which form of
ζ ≡ (X − Y )2 = −2(1 + X Y ) we work, we just need to be consistent, and set
•

X 2 = Y 2 ≡ −1 only at the end)

(︂ )︂
X G (X, Y ) = ∂X + X ∂X (d + X ∂X ) G(ζ)
bb 2 • •

[︂ ]︂
= ∂X [(−2Y )G′ (ζ)] + (d + X ∂X ) (−2X Y )G′ (ζ)
• • •

= (−2Y )2 G′′ (ζ) + (d + 1)(−2X Y )G′ (ζ) + (−2X Y )2 G′′ (ζ) • •

(︂ )︂
= −4 + (2 + ζ)2 G′′ (ζ) + (d + 1)(2 + ζ)G′ (ζ)
= z(z − 1)G′′ (z) + (d + 1)(z − 12 )G′ (z) ,
where in the last step we performed substitution z ≡ − ζ4 . Thus, we have obtained
Chapter 4. QFT in Anti–de Sitter Spacetime 98

an equation in the form of the hypergeometric equation (for nonzero z)

[︂ ]︂
− X + ∆(∆ − d) G(z) = 0
(︂ )︂
=⇒ z(1 − z)G′′ (z) + d+1
2
− (d + 1)z G′ (z) − ∆(d − ∆)G(z) = 0
(︂ )︂
z(1 − z)G′′ (z) + c − (a + b + 1)z G′ (z) − abG(z) = 0 ,

where a = ∆, b = d − ∆, c = d+1
2
, or same with ∆ ←→ d − ∆, which leaves the
mass/conformal Casimir eigenvalue unchanged. By requiring the singularity at
X → Y (ζ → 0) to match with the flat space propagator singularity, we obtain
the solution

∆, ∆ − d2 + 12 4
[︄ ⃓ ]︄
C∆ ⃓
bb
G∆ (X, Y ) ≡ G∆ (ζ) ≡ ∆ 2 F1 ⃓−
⃓
ζ 2∆ − d + 1 ⃓ ζ

where the normalization is

Γ(∆)
C∆ = d
(︂ )︂ .
2π Γ ∆ −
2
d
2
+1 ⌟

Remark 4.12 (Free Theory, Perturbation Theory). Calculation of correlators in

perturbation theory follows the general story as discussed in Section 1.1 → p.10 —
we can use same Feynman diagram techniques, just considering EAdS specific
differential operators (and corresponding propagators) and integrals. ⌟

Boundary Limit
Since the correlators are invariant under Iso(EAdSd+1 ) act on the boundary as
Conf(Rd ), by taking their appropriate asymptotic/boundary limit, we obtain
boundary correlators satisfying the CFTd axioms (apart from the presence of the
energy–momentum tensor).

As shown in Remark 4.10 → p.97 , the (physical) mass of the bulk field mϕ is related
to the scaling dimension of the corresponding operator in the boundary theory.
To avoid certain subtleties, in the following we consider the positive “Dirichlet”
branch
√︄
d d2
∆ϕ = ∆+ ≡ + + m2ϕ ,
2 4

which in Poincaré coordinates (z, y) ∈ R≥ × Rd corresponds to the boundary

condition ϕ ∼ z ∆+ as z → 0 (such choice is consistent with the EAdS isometries).

To have well-defined boundary correlators, we define the dual operator by taking

an appropriately scaled boundary limit of the bulk field.
Chapter 4. QFT in Anti–de Sitter Spacetime 99

Definition 4.13 (Dual Boundary Operator). The boundary operator Oϕ dual

to the bulk field ϕ is defined as the limit
1
(︃ )︃
Oϕ (P ) ≡ √ lim s∆ϕ ϕ X ≡ sP + O(1/s) ,
C∆ϕ s→∞

where the bulk point X ∈ EAdS approaches the boundary point P ∈ ∂EAdS.

Remark 4.14 (Bulk-to-Boundary Propagator). Taking the appropriate boundary

limit of one operator in the bulk-to-bulk propagator Calculation 4.11 → p.97 we
obtain the bulk-to-boundary propagator G∆ b∂
ϕ
, which has a much simpler form
√
⟨︂ ⟩︂ C∆ϕ
ϕ(X)Oϕ (P ) ≡ b∂
G∆ (X, P ) = ,
free ϕ
(−2X P )∆ϕ
•

since taking Y ≃ sP to the boundary by s → ∞, we have ζ ≃ X (sP ) → ∞, so •

the hypergeometric function becomes just 1. Note how the boundary limit scaling
in the definition of Oϕ precisely cancels the scaling of ζ −∆ϕ in G∆
bb
ϕ
. ⌟
Remark 4.15 (Boundary-to-Boundary Propagator). Sending also the second operator
to the boundary, one finds the boundary-to-boundary propagator G∆ ∂∂
ϕ
with the
usual form of the 2-point function in flat-space CFT of scalar operator with scaling
dimension ∆ϕ , that is
⟨︂ ⟩︂ 1 1
Oϕ (P1 )Oϕ (P2 ) ∂∂
≡ G∆ (P1 , P2 ) = = ,
free ϕ
(−2P1 P2
• )∆ϕ |y1 − y2 |2∆ϕ

where y1 , y2 ∈ Rd are flat-space coordinates corresponding to the points P1 , P2 ∈

∂EAdSd+1 . We now see that the normalization in the definition of the boundary
operator was chosen such that the boundary 2-point function is conventionally
normalized. ⌟
 100

5
O(N) Model in Anti–de Sitter
This chapter is based on [58]: Jonáš Dujava and Petr Vaško.
Finite-coupling spectrum of O(N) model in AdS. 2025.
arXiv: 2503.16345 [hep-th] GitHub: jdujava/ONinAdS

Abstract. We determine the scaling dimensions in the boundary CFTd corre-

sponding to the O(N ) model in EAdSd+1 . The CFT data accessible to the 4-point
boundary correlator of fundamental fields are extracted in d = 2 and d = 4,
at a finite coupling, and to the leading nontrivial order in the 1/N expansion.
We focus on the non-singlet sectors, namely the anti-symmetric and symmetric
traceless irreducible representations of the O(N ) group, extending the previous
results that considered only the singlet sector. Studying the non-singlet sector
requires an understanding of the crossed-channel diagram contributions to the
s-channel conformal block decomposition. Building upon an existing computation,
we present general formulas in d = 2 and d = 4 for the contribution of a t-channel
conformal block to the anomalous dimensions of s-channel double-twist operators,
derived for external scalar operators with equal scaling dimensions. Up to some
technical details, this eventually leads to the complete picture of 1/N corrections
to the CFT data in the interacting theory.

5.1 Introduction
Recently, various QFT models have been studied in AdS (or dS) spacetime at finite
coupling. Rather than the usual weak coupling expansion, the large N expansion
together with AdS/CFT intuition was employed as an alternative handle to perform
explicit calculations. This movement started with O(N ) and Gross–Neveu Model
[59], and was continued also for scalar QED [60].

While these studies have addressed key questions such as the phase structure,
exact propagators, and the boundary correlators (4-point functions) in the large
N limit, they have primarily focused on the singlet spectrum of the CFT on the
boundary.

Largely unexplored is the non-singlet spectrum corresponding to operators trans-

forming in nontrivial representations of the global internal (large N ) symmetry
group. We will thus primarily focus on extending the analysis of the O(N ) model
in AdS at finite coupling also to the non-singlet sector — namely the rank-2
anti-symmetric and symmetric traceless representations of O(N ) group.

Building on the already established results for the singlet sector, we were able to
extract anomalous dimensions of operators in the non-singlet sector by utilizing
Chapter 5. O(N) Model in Anti–de Sitter 101

the 6j–symbols (also known as the crossing kernel) [61].

More broadly, there are essentially three main pieces of motivation in this line of
research. They were nicely summarized in the Introduction sections of [59] [60],
which the reader is invited to check out for more details. Even after receiving
some attention, numerous issues require deeper investigation:

• Free theory in AdS bulk corresponds to GFF/MFT at the asymptotic boundary,

that is a CFT with all correlation functions simply given just by products of
2-point functions. While a great deal has been said about small deformations
in a bulk coupling, less explored are large deformations of the CFT data
away from MFT. Large N techniques are helpful in this respect, since they
retain more of the nonlinear structure of the exact theory compared to the
ordinary lowest-order perturbation theory. As such, they can shed some light
on interesting phenomena like appearance of new operators in the spectrum
(corresponding to the existence of bound states), AdS analogs of resonances,
or level crossing.

• For a QFT in AdS with a nontrivial phase structure, it is interesting to

determine how CFT data of the dual theory at the boundary change as a bulk
RG-flow connecting the different phases is triggered. The bulk RG-flow is
related to the change in AdS radius ℓ. The framework for studying it, with
a particular emphasis on the flat space limit ℓ → ∞, was laid out in [57].
More detailed studies about constraining the bulk RG-flow by boundary CFT
axioms followed. By now there is already a significant amount of literature,
for an invitation to this topic with further references included, see [62] [63]
[64]. The role of certain 3-point correlation functions (AdS form factors) of
two boundary operators with one bulk field in the change of the AdS length
scale ℓ was investigated in [65]. The important special case of the bulk
operator being the stress tensor was done in [66].

A distinct instance of the above discussion occurs when the flow is initiated
in a free bulk theory corresponding to an MFT at the boundary, which
ubiquitously possesses higher spin currents (all the double-twist families
saturate unitarity bound and form thus short conformal multiplets). Then
one can naturally wonder about their fate once interactions are turned on.
General theorems constraining their existence in arbitrary interacting CFTs
were given in [67] [68]. On the other hand, one can ask a similar question,
just not at the boundary but instead in the bulk. A complete breaking of
bulk higher spin currents by interactions in AdS2 was analyzed in [69].

• In the special case when the bulk theory in AdSd+1 is critical (becoming a
CFTd+1 ), one can perform (in the Euclidean signature) a Weyl symmetry
transformation from EAdS to flat half-space Rd × R≥ . This is easily done
using the Poincaré coordinates, which cover EAdS globally. The CFT with a
Chapter 5. O(N) Model in Anti–de Sitter 102

boundary — BCFTd+1 — on this flat half-space can thus be studied using AdS
methods, whose recent advances have made such approach rather efficient.
Moreover, one can generalize it from boundaries (BCFT) to defects (DCFT).
A Weyl equivalence between an (n − 1)-dimensional defect in Rn+m and
AdSn × Sm allowed [70] to use results of [59] for the singlet spectrum to
extract interesting DCFT data. More concretely, they studied the critical
O(N ) model in the presence of a localized magnetic external field understood
as an 1-dimensional defect, thus AdS2 was relevant. Such data is important
to understand phase transitions of real-world systems.
• Finally, there are attempts to replace the LSZ axioms for the flat-space S-
matrix by the flat-space limit of asymptotic boundary observables in AdS [71]
[72] [73] [74]. Asymptotic boundary correlators for massive QFTs in
AdSd+1 obey CFTd axioms — except for the existence of a stress tensor —
that are mathematically more rigorous than the current ones for the S-matrix
in flat space. These asymptotic correlators are by definition holographic as is
the flat-space S-matrix, and reduce to it in the flat-space limit (in a sense
that still requires a more rigorous definition).
Since the curvature of AdS also acts as an IR regulator, it could potentially cure
the possible IR divergences, which make the flat S-matrix an ill-defined object
[75]. In practice, one often has to recourse to IR safe inclusive observables,
which however require taking the experimental setup into account, and are
thus not clean theoretical observables. More discussion about the IR issues of
the flat S-matrix and another attempt how to solve it can be found in [76].
Our work focuses on a specific topic in the above vast landscape. We examined
large deformations of the free MFT spectrum for the O(N ) model in the unbroken
phase. While primary outcome is the computation and analysis of the non-singlet
spectrum, along the way we also obtained some additional results.
First, we completed the singlet CFT data by extracting the OPE coefficients.
Furthermore, we discussed extension of the spectrum (even the singlet one) to
d = 4, where the necessary regularization complicates the picture. To partly
deal with the associated renormalization scheme ambiguity — finite part of a
counterterm — we resorted to the critical bulk theory describing an ordinary
phase transition.
Formally extending the critical point beyond its upper critical dimension — also
considered in Section 3.1 and Figure 6 [77] — we discovered an intriguing pattern
in the singlet spectrum. It is related with the appearance of emergent operators
at a strong enough coupling, which are not connected with the usual MFT-type
spectrum found at weak coupling. Based on this analysis, a formula for the critical
singlet spectral function in all even (boundary) dimensions reproducing the desired
critical spectrum was proposed.
Chapter 5. O(N) Model in Anti–de Sitter 103

The next natural step would be to deal with the broken phase and especially
with the critical theory in the bulk, whose analysis could result in extraction of
non-singlet DCFT data along the lines of [70]. Hopefully, our results can add a
concrete small piece to the above complex mosaic.

Outline of the Paper. The subsequent sections are organized as follows.

In Section 5.2 → p.103 we define the O(N ) model, mainly focusing on its formula-
tion suitable for a systematic large N expansion. Then we introduce the main
observable — the 4-point boundary correlator of fundamental fields ϕ• in AdS.
The rest of the section is devoted to fixing our conventions, particularly those for
OPE channels, and some comments about renormalization scheme are made as
well.

In Section 5.3 → p.113 we will review relevant generalities of CFTs, focusing on the
extraction of the CFT data from the 4-point correlator. In particular, we will
present general formulas (in d = 2 and d = 4) for the contribution of a single
t-channel conformal block to anomalous dimensions of s-channel double-twist
operators, applicable for arbitrary twists and spins.

These formulas will be utilized in Section 5.4 → p.137 to calculate the leading 1/N
contributions to the non-singlet scaling dimensions of the O(N ) model in AdS at
finite coupling. This computation requires the singlet spectrum as an input, whose
properties are summarized in the previous parts of Section 5.4 → p.123 , together with
the decomposition of the 4-point correlator into O(N ) irreducible representations.
Criticality in the bulk is also briefly discussed in Section 5.4 → p.132 , which suggests
a possible extension beyond its upper critical dimension.

In Section 5.5 → p.138 we present the results for the non-singlet spectrum, mostly
in the form of various plots. The main one is the twist–spin plot showing a
characteristic organization of the spectrum into Regge trajectories. Various limits
are studied, specifically a cross-check regarding the large spin asymptotics.

An executive summary and synthesis of results is given in Section 5.6 → p.148 ,

together with possible future directions.

Various detailed computations and the implementation of main formulas can

be found in the accompanying Notebook — links to a GitHub repository
[jdujava/ONinAdS] .

5.2 Review of O(N) Model in AdS

We start by introducing the O(N ) model in Section 5.2 → p.104 . Some general
features are reviewed, in particular how its large N expansion enables calculations
at finite coupling. In Section 5.2 → p.108 we define our main observable — the
Chapter 5. O(N) Model in Anti–de Sitter 104

boundary 4-point correlator in AdS — which can be viewed as an observable in the

CFT living on the boundary of AdS. In the following Section 5.2 → p.112 we briefly
describe how the spectral representation can be utilized to express the boundary
correlator in a form particularly suitable for the study of the CFT spectrum.
While most of the material in this section is well-known for experts in the field,
with some more specific details thoroughly discussed already in [59], we believe
that an alternative account (stressing some additional points) can be beneficial.
In the meanwhile, we also fix the notation and conventions that will be used
throughout the paper.

Generalities of O(N) Model

The Euclidean action of the O(N ) model on a (d + 1)–dimensional Riemannian
manifold M takes the form (we implicitly contract both spacetime and internal
indices)

1 1
[︄ ]︄
√ λ (︂ • 2 )︂2
∫︂
S[ϕ ] =
•
d
d+1
x g (∂ϕ• )2 + m2 (ϕ• )2 + (ϕ ) , (5.1)
M 2 2 2N

where ϕ• = (ϕ1 , . . . , ϕN ) is an N -tuple of real scalar fields transforming in the

vector representation of the global internal O(N ) symmetry group. In the case of
maximally symmetric spacetimes — in our case EAdS — the possible coupling
to the curvature can be absorbed into the mass term m2 . Results for Lorentzian
AdS can be obtained by analytical continuation, so we will use EAdS/AdS inter-
changeably.
The interaction term was introduced in such a way that the model admits a large
N expansion with the λ coupling fixed and finite. We will review details relevant
to us. It was also outlined in Section 2 [59], and more thorough accounts in flat
space can be found in a specialized review [78] or in the original paper [79].
Feynman rules for (5.1) assign a propagator to each internal line, and the interac-
tion vertex can be diagrammatically represented as (up to numerical factors)

λ
(︃ )︃
∼ + + , (5.2)
N
where the connected lines on the right-hand side indicate Kronecker deltas in the
indices of the corresponding ϕ fields. As usual, all internal vertices throughout
the paper are to be integrated over M with the geometric measure including the
√
g density.
Order of 1/N in a given diagram is not simply given just by the number of
interaction vertices, since after their expansion (5.2) a number of closed index
loops can form. Each such index loop carries an additional factor of N , since all
i=1 circulating in the loop contribute equally.
{ϕi }N
 Chapter 5. O(N) Model in Anti–de Sitter 105

Hubbard–Stratonovich Transformation. For the purposes of large N expansion,

it turns out more efficient to proceed via a reparametrization of the interaction
term under the path integral by introducing an auxiliary Hubbard–Stratonovich
field σ — equivalent on-shell (up to normalization) to the composite operator
(ϕ• )2 — thus obtaining the action

1 1 1 2 1
[︄ ]︄
√
∫︂
SHS [ϕ , σ] =
•
d
d+1
x g (∂ϕ• )2 + m2 (ϕ• )2 − σ + √ σ(ϕ• )2 . (5.3)
M 2 2 2λ N

The interaction vertex (5.2) is now substituted by the following rules

i 2
≡ −λ1 , ≡ √ δ ij , (5.4)
j N

first representing the free σ-propagator, and second the σϕ2 vertex. We suppressed
the position dependence of ≡ ⟨σσ⟩free by writing it — up to factor (−λ)
— as an operator 1 acting on functions through the convolution with the kernel
1(x, y) ≡ δ(x, y). Here we use the δ-distribution normalized with respect to the
geometric measure, so the 1 operator is really acting on functions as an identity.

Effective Action. Now, we want to “integrate out the loops” of the ϕ fields,
which contribute at the leading order in the large N expansion. Since our new
Lagrangian LHS is only quadratic in ϕ, the ϕ-loop subdiagrams have the form of
a loop with arbitrary number of σ lines attached to it, thus inducing non-local
contributions to the 1PI effective interaction vertices between σ fields of the form

σ )︄n ]︄
∞
(−1)n
∞
1 2
[︄(︄
1-loop
[σ] = σ = const. − N Tr
∑︂ ∑︂
Γ n ◦√ σ
n=1 2n (− + m )1
2
n=1 N
σ
2 2
(︄ )︄ (︄ )︄
N N
= Tr ln (− + m )1 + √ σ = − ln Det− 2 (−
2
+ m )1 + √ σ . 2
2 N N
(5.5)
The prefactor of N comes from the already performed trace over the indices in the
closed loop. Also other combinatorial/symmetry factors are properly accounted
for. This can be alternatively seen as a functional determinant coming from
Gaussian integration over ϕ rewritten as contribution to the Euclidean effective
action.

Note, that in (5.5) we understand the argument of Tr ln( ) as an operator acting

on functions, which we can represent as a convolution with an associated kernel.
In particular, the kernel corresponding to (− + m2 )1 is [(− + m2 )1](x, y) ≡
(− x + m2 )δ(x, y), and the action of pointwise multiplication with σ has the
kernel σ(x, y) ≡ σ(x)δ(x, y).
Chapter 5. O(N) Model in Anti–de Sitter 106

We can now introduce a (Euclidean) effective action

Γ [ϕ•, σ] ≡ Γ 0-loop [ϕ•, σ] + Γ 1-loop [σ] ≡ SHS [ϕ•, σ] + (5.5)
1 1 1 2 1
[︄ ]︄
√
∫︂
= d
d+1
x g (∂ϕ• )2 + m2 (ϕ• )2 − σ + √ σ(ϕ• )2
M 2 2 2λ N (5.6)
2
(︄ )︄
N
+ Tr ln (− + m2 )1 + √ σ ,
2 N
which is indeed the effective 1PI action including all of the leading 1/N contri-
butions. Proceeding further, we should first determine the ground state of the
theory by extremizing (preferably minimizing) the effective potential, given by the
effective action Γ evaluated for constant classical fields and divided by the volume
of the spacetime which factorizes. Afterward, one expands the effective action in
terms of the shifted fields δϕ and δσ which vanish at the found extremum.
The coefficients of this expansion are the exact propagators (quadratic terms)
and 1PI vertices (higher order terms) in the leading 1/N approximation. The
linear “tadpole” terms in δϕ and δσ are absent by virtue of expanding around the
extremum of the effective potential. Observables can be then calculated using
these leading forms of exact propagators and 1PI vertices, but all diagrams which
contain ϕ-loops should be omitted, since those are already included in the effective
σ self-interactions.
Alternatively, this can be viewed as a large N saddle point analysis, since all terms
in the effective action are of the same order of magnitude in the neighborhood of
the effective potential extremum — they are all of order O(N ).

Phase Structure. Phases of the O(N ) model have been thoroughly investigated
on flat space of dimension D ≡ d + 1. The dynamics of the theory significantly
differs for the ranges 2 < D < 4 and D ≥ 4. Since we will do computations for
d = 2 ⇔ D = 3 (AdS3 ) and d = 4 ⇔ D = 5 (AdS5 ), we need to briefly summarize
both cases, in order to clearly specify which phase we treat in this work.
In the first case 2 < d + 1 < 4, the flat-space theory is asymptotically free. The
free unstable UV fixed point CFTUV has two relevant operators (the mass term
and the interaction term) that trigger an RG flow. When appropriately tuned, it
ends in a semi-stable interacting (the strength increases as D decreases from 4)
IR fixed point — the Wilson–Fisher CFTIR [80] — with one relevant operator
corresponding to the mass term. The RG flow triggered by it, depending on the
sign of the deformation, leads either to a trivially gapped phase with unbroken
global O(N ) symmetry or to spontaneous symmetry breaking O(N ) → O(N − 1),
whose low energy dynamics is governed by a non-linear sigma model of (N − 1)
Goldstone bosons (with a free fixed point CFTIR in the deep IR).
In the second case d + 1 ≥ 4, the flat-space theory is IR free. For large enough
N , a perturbatively unitary UV fixed point was found in [81]. The authors
Chapter 5. O(N) Model in Anti–de Sitter 107

proposed a UV completion by a cubic theory in D = 6 with an IR fixed point,

whose equivalence to the UV fixed point of the O(N ) model was checked using
ε-expansions. Later, it was shown [82] that this fixed point is non-unitary
beyond perturbation theory. In particular, scaling dimensions receive exponentially
suppressed (in large N ) imaginary parts caused by instantons existing in either
formulation of the fixed point. This phenomenon is known as complex CFT [83].
The gapped phase as well as the spontaneously broken phase exist as in the
previous case.
The phase structure of O(N ) model in AdSd+1 for dimensions ranging in 2 <
d + 1 < 4 was analyzed in Section 3 [59], and differences from flat space were
summarized there. For example, in d = 2 (AdS3 ), there exists a region of the
parameter space where both the unbroken and broken phase seem to coexist. In
this work, we treat almost exclusively the simplest of the phases — the unbroken
phase with the O(N ) symmetry preserved. We leave the broken phase and deeper
analysis of the critical point for future work.
Moreover, the final results will be mainly presented for d = 2 (AdS3 ), where
formulas simplify enough to actually perform the calculations. Nevertheless, at a
certain point a numerical evaluation is necessary. In addition d = 4 (AdS5 ) will
be also included, which however still requires an independent phase structure
analysis.
Choice of the(︂ unbroken O(N )–symmetric
√ )︂ phase specializes to the vicinity of the
saddle point ϕ (x) = 0, σ(x) = N σ⋆ , where σ⋆ is a constant parametrizing the
•
√
vacuum expectation value (VEV) of the σ field as ⟨σ⟩ ≡ N σ⋆ . The expansion of
the effective action Γ (5.6) around this saddle point just gets rid of the σ-tadpole
term, and the ϕ-field mass-squared is shifted to m2ϕ ≡ m2 + 2σ⋆ . Since everything
else stays the same, from now on we take all “free” ϕ-field propagators with
the effective mass-squared m2ϕ , and use σ to refer mostly just to the deviation
δσ ≡ σ − ⟨σ⟩.
Finally, let us remark that the unbroken phase is present when the (effective)
2
mass is above the Breitenlohner–Freedman (BF) bound m2ϕ > − d4 [56] [84], a
point which will be also discussed in Section 5.2 → p.108 .

Exact σ-Propagator.
√ Since the interaction terms in the expansion of Γ are of
the order O(1/ N ) or higher, the leading O(1) form of the exact σ-propagator
can be obtained by inverting the kernel of the quadratic σ-part of Γ expanded
around the saddle point, or equivalently by summing up the geometric series

= + + + ···
∞ ]︃−1
1
[︃
= (−λ1) + (−λ1) ◦ 2B ◦ (−λ1) + · · · = −λ (−2λB)n = − + 2B
∑︂
,
n=0 λ
(5.7)
Chapter 5. O(N) Model in Anti–de Sitter 108

where B is the kernel corresponding to the (half of) “bubble” diagram given by
]︄2
1 1
[︄
B(x, y) ≡ x y ≡ (x, y) . (5.8)
2 (− + m2ϕ )1
√
Note that for each bubble in (5.7) the (1/ N )2 from the vertices canceled out
with the N coming from the index loop. The diagram in (5.8) is really a free
correlator of a scalar composite field ϕ2 ≡ :(ϕi )2 : being the normal-ordered square
of ϕi for any fixed index i, that is ⟨ϕ2 (x)ϕ2 (y)⟩free = 2Gϕ (x, y)2 ≡ 2B(x, y). The
factor of 2 comes from two different ways of contracting legs at both vertices.
From the point of view of the initial action (5.1), the σ-propagator resums an
infinite class of Feynman diagrams contributing in the leading order of 1/N
expansion. For example, the leading s-channel connected contribution to the
4-point function of ϕ fields is given by
i k
∞
1
= (5.9)
∑︂
··· ∼ .
n=0 ⏞ ⏟⏟ ⏞
j l N
n bubbles

To increase clarity, from now on, we will explicitly write out the N -dependence in
front of the diagrams instead of including it in the diagrams themselves.

CFT on Boundary of AdS

Now we will specialize to the case of AdSd+1 spacetime. The isometries of AdS
act on its asymptotic/conformal boundary as conformal transformations, and by
performing an appropriate boundary limit of the correlators one obtains boundary
correlators satisfying the CFTd axioms (apart from the presence of the stress
tensor operator) [85].
The boundary operator Oϕi ≡ Oϕi corresponding (or dual) to the scalar field
ϕi has a scaling dimension ∆ϕ satisfying the equation m2ϕ = ∆ϕ (∆ϕ − d) [86].
We measure all dimensionful quantities in units of the AdS radius ℓ, which we
throughout set to ℓ ≡ 1.
To avoid certain subtleties, in the following we consider the positive “Dirichlet”
branch √︄
d d2
∆ϕ = ∆+ ≡ + + m2ϕ , (5.10)
2 4
which in Poincaré coordinates (z, y) ∈ R≥ × Rd corresponds to the boundary
condition ϕ ∼ z ∆+ as z → 0. The dual operator is then defined by the boundary
limit
1
(︃ )︃
Oϕi (P ) ≡ √ lim s∆ϕ ϕi X ≡ sP + O(1/s) , (5.11)
C∆ϕ s→∞
where P is a point on the boundary of EAdSd+1 (a future directed null vector in
R1,d+1 ) and X is a point of EAdS in the embedding formalism [51] approaching
 Chapter 5. O(N) Model in Anti–de Sitter 109

P in the limit s → ∞, where the O(1/s) term enables the EAdS condition
X 2 = −ℓ2 ≡ −1 to be satisfied while P 2 = 0. The normalization constant C∆ϕ is
given by
Γ(∆ϕ )
C∆ϕ = d
(︂ )︂ . (5.12)
2π 2 Γ ∆ϕ − d2 + 1

Free propagator of ϕ• fields in EAdS is given by the bulk-to-bulk propagator G∆

bb
ϕ
expressible in terms of the chordal distance ζ(X, Y ) ≡ (X − Y )2 = −2 − 2X Y •

between two points in AdS

⟨︂ ⟩︂
ϕi (X)ϕj (Y ) ≡ δ ij G∆
bb
ϕ
(X, Y )
free
(5.13)
⎡ ⃓ ⎤
C∆ϕ ∆ , ∆ϕ − d2 + 12 4
⃓
= ⎣ ϕ
⃓
2 F1
⃓− ⎦.
ζ(X, Y )∆ϕ 2∆ϕ − d + 1 ⃓
⃓ ζ(X, Y )

Taking the appropriate boundary limit of one operator — in the sense of (5.11)
— one obtains the bulk-to-boundary propagator G∆ b∂
ϕ
, which has a much simpler
form √
⟨︂ ⟩︂
ij b∂ C∆ϕ δ ij
ϕ (X)Oϕ (P )
i j
≡ δ G∆ϕ (X, P ) = . (5.14)
free (−2X P )∆ϕ •

Sending also the second operator to the boundary, one finds the boundary-to-
boundary propagator G∆ ∂∂
ϕ
with the usual form of the 2-point function in flat-space
CFT of scalar operator with scaling dimension ∆ϕ , that is

⟨︂ ⟩︂ δ ij δ ij
Oϕi (P1 )Oϕj (P2 ) ≡ δ ij G∆
∂∂
(P1 , P2 ) = = , (5.15)
free ϕ
(−2P1 P2 )∆ϕ
•
|y1 − y2 |2∆ϕ

where y1 , y2 ∈ Rd are flat-space coordinates corresponding to the points P1 , P2 ∈

∂EAdSd+1 . We now see that the normalization in the definition of the boundary op-
erator (5.11) was chosen such that the boundary 2-point function is conventionally
normalized.

Main Observable — Boundary 4-Point Correlator. Let us define the main

observable that we will use to probe the spectrum of the CFTd emerging at the
asymptotic/conformal boundary of AdSd+1 . It is the boundary limit of the 4-point
correlator ⟨ϕϕϕϕ⟩ of fundamental fields ϕ• in the vector representation of the
global O(N ) symmetry, that is

⟨︂ ⟩︂ ⟨︂ ⟩︂ i k
O1i O2j O3k O4l ≡ Oϕi (P1 ) Oϕj (P2 ) Oϕk (P3 ) Oϕl (P4 ) ≡ , (5.16)
j l

where P• are points (suppressed in diagrams) lying on the boundary of AdSd+1 ,

which is represented diagrammatically by a circle. Considering it up to the order
Chapter 5. O(N) Model in Anti–de Sitter 110

1/N in the large N expansion, but to all orders in the coupling λ, its Witten
diagram representation is given by
⎛ ⎞
i k i k i k i k
=⎝ + +
⎜ ⎟
⎠
j l j l j l j l
⎛ ⎞ (5.17)
i k i k i k
1 ⎜ 1
(︃ )︃
+ + + ⎠+O
⎟
,
N j N2
⎝
l j l j l

where we used the leading forms of the 1PI σϕ2 vertex (5.4) and exact σ-propagator
(5.7). The N -dependence is now explicitly written in front of the diagrams. Lines
with one or both ends on the boundary are bulk-to-boundary (5.14) or boundary-
to-boundary (5.15) propagators, respectively. Using these ingredients, one can
easily compose explicit expressions for diagrams figuring in (5.17), where as usual
we integrate over the bulk points.

Convention for Channels. Our naming convention for both the OPE and also
Witten diagrams of the 4-point correlator is

s-channel t-channel u-channel

(︂ )︂(︂ )︂ (︂ )︂(︂ )︂ (︂ )︂(︂ )︂ (5.18)
O1i O2j O3k O4l , O1i O3k O2j O4l , O1i O4l O2j O3k .

They correspond to the diagrams in (5.17), in respective order at each order of

1/N . This is in agreement with the conventions used in [59] and (2.53) [87], but
t-channel and u-channel are swapped in (3.5)/(3.8) [61]. As we will comment at
appropriate places in Section 5.3 → p.113 , one just needs to be careful with including
an additional factor of (−1)J for 6j–symbol compared to their final formulas.

Renormalization Scheme. In (5.17) we drew only nontrivial Witten diagrams

contributing to the 4-point correlator. Those just modifying the normalization
of the ϕ-field and shifting its mass — connected with the normalization of the
dual boundary operator Oϕ and shifting the scaling dimension via the relation
m2ϕ = ∆ϕ (∆ϕ − d) — were omitted, so let us comment on them here. Intuition
about UV renormalization can be borrowed from flat space, since when distances
among points are infinitesimally small, any spacetime looks approximately flat.
Explicit renormalization in AdS in weak coupling perturbation theory, especially
for a ϕ4 -theory relevant for us, was done in [88] [89] [90].
We work with the 4-point correlator computed up to (and including) the order 1/N .
We will describe the renormalization of the classical action (5.3), but sometimes it
is useful to reiterate what various steps mean from the perspective of the original
action (5.1).
First, let us once again qualify the term “free” in connection with the ⟨ϕϕ⟩
propagator. The discussion is intimately related with quantization around the
Chapter 5. O(N) Model in Anti–de Sitter 111
√
nontrivial saddle point σ(x) = N σ⋆ of the effective action (5.6). This step
produces an O(1) exact ϕ-propagator (in the large N expansion), which is from
the point of view of (5.1) given by a resummation of all “cactus” diagrams.
Those renormalize just the mass m of the free ϕ-field and shift it to the value
m2ϕ = m2 + 2σ⋆ . Thus, the “free” ϕ-propagator is associated with a free equation
of motion, but with a shifted mass.
Such ϕ-propagator is exact at O(1) as we stated, therefore it is sufficient for all
diagrams that are already of order O(1/N ), in particular for the exchanges of the
σ-field in the second line of (5.17).
However, it is not sufficient for the disconnected diagrams of order O(1) in the
first line of (5.17). In those diagrams, the renormalization of the ϕ-propagator
needs to be carried out to order O(1/N ). Such a renormalization can affect both
the normalization of the ϕ-field and a shift in its mass. Luckily, the type of the
propagator subjected to it is the boundary-boundary one, which is completely
fixed by conformal symmetry on the boundary. We choose the two counterterms in
such a way that it remains in the “free” form (5.15). This finishes renormalization
of the ϕ-propagator.
Now we start discussing renormalization of the ⟨σσ⟩ bulk-bulk propagator. Its
form (5.7) is O(1) exact and from the perspective of (5.3) renormalizes the “mass-
squared” λ−1 of the free σ-field. In particular,(︂the spectral
)︂−1
representation — see
next Section 5.2 → p.112 — of the σ-propagator λ−1 + 2B˜︁ requires a nontrivial
renormalization in d ≥ 3, where the bubble diagram acquires a UV divergence. The
bubble function must be then somehow regularized, and with the UV divergence
absorbed by a counterterm we are left with the finite combination λ−1 + 2B. ˜︁
The poles of such renormalized spectral function determine the physical scaling
dimensions of boundary operators associated with the bulk field σ. We will
comment on fixing the subtraction scheme employed in defining the regularized
bubble function B˜︁ in Section 5.4 → p.132 .
Let us also note that the renormalization of the σ-propagator described above,
when reinterpreted from the point of view of (5.1), would correspond to renormal-
ization of the 1PI vertex ϕ4 associated with the coupling λ. Classically it starts
at order O(1/N ), and since all “bubble-chain” diagrams contribute at the same
order, they need to be resummed. This is done in (5.9), which precisely leads to
the O(1) exact form (5.7) of the σ-propagator.
√
Finally, let us comment about the 1PI vertex σϕ2 , which is of order O(1/ N )
classically. It appears twice in the σ-exchange diagrams, thus making them of
order O(1/N ). Consequently, to the maximal order we consider, it does not require
any renormalization, as all corrections to its classical part would be subleading in
the large N expansion.
This finishes the description of the renormalization. The renormalization scheme
 Chapter 5. O(N) Model in Anti–de Sitter 112

defined above might be called on-shell, since it is fully determined in terms of

physical scaling dimensions ∆ϕ , ∆ˆ︁σ0 — σˆ︁0 being the lowest-dimensional boundary
operator associated to the σ-field — and the standard normalization of the bound-
ary operator Oϕ , leading to the canonical form of the CFT 2-point function (5.15).

Utilizing Spectral Representation

To proceed further with the calculation of the 4-point correlator (5.16), mainly
the evaluation of bulk integrals in the σ-exchange diagrams, it is convenient
to employ the so-called spectral representation of the σ-propagator. It is an
integral expansion into harmonic functions — eigenfunctions of the AdS Laplacian
— which form a continuous basis for integrable functions depending only on the
geodesic distance between two points. This is in precise analogy to the (radial)
Fourier representation of the propagator in flat space, which is also adapted to
the isometries (translations and rotations) of the theory. In the limit of large AdS
radius, it actually reduces to the flat-space (radial) Fourier transform.
While we will not explicitly use this technology — the relevant results from [59]
will be cited in Section 5.4 → p.125 — it is rather crucial in calculation and deserves a
brief mention. More details can be found in Appendix 4.C [50] Appendix B [45]
and Appendix B [59].
Suppose we know the spectral representation B(∆)
˜︁ of the bubble function B(x, y)
defined in (5.8), that is we have a decomposition into harmonic functions Ω∆
∫︂ ∫︂
d∆ ˜︁
B(x, y) = dν B(∆)Ω
˜︁
∆ (x, y) ≡ B(∆)Ω∆ (x, y) , (5.19)
R d
2
+ı̊R ı̊
where we use ∆ ≡ d2 +ı̊ν ⇔ ν ≡ −ı̊(∆ − d2 ) interchangeably. While it is common
to use the spectral parameter ν to index the harmonic functions and also to denote
the functional dependence of spectral representations, we will mostly prefer using
the dimension ∆ directly for later notational convenience.
Ignoring the index structure, the calculation can be schematically represented as
i k )︄ i
(︄ k
−1
∫︂
∆
= 4 dν
j l R λ−1 + 2B(∆)
˜︁
j l
i k
1
(︄ )︄√︂
∫︂
ν 2 ∆ ∆˜︁
= 4 dν − C∆ C∆
λ−1 + 2B(∆) π j
˜︁
R
(5.20)
˜︁
l
d∆ 1 Γ··· · · · Γ···
2
∫︂ (︄ )︄(︄ )︄
= − ∆ ∆
˜︁
d
2
+ı̊R≥0 2πı̊ λ−1 + 2B(∆)
˜︁ · · · Γ···2 · · · Γ···
d∆ 1 Γ···2 · · · Γ···
∫︂ (︄ )︄(︄ )︄ ⃓ ⟩︃
⃓ ∆,0
≡ − ⃓ .
d
2
+ı̊R≥0 2πı̊ λ + 2B(∆)
−1 ˜︁ · · · Γ··· · · · Γ··· ⃓
2

Line by line, we utilized the following properties and relations — for more details
see Appendix C [59]:
Chapter 5. O(N) Model in Anti–de Sitter 113

(1) Similarly to the Fourier transform, the spectral representation transforms a

convolution of functions into a product of their spectral functions. Thus, the
“operator” inversion figuring in the σ-propagator (5.7) is represented just as
a numeric inversion of the corresponding spectral function. The overall factor
of 4 comes from the two σϕ2 vertices (5.4), we implicitly integrate over the
black bulk points, and the wavy line represents the harmonic function.

(2) The harmonic function can be further rewritten by means of the split rep-
resentation [51] [91], which up to a certain prefactor is given by two
bulk-to-boundary propagators — one with dimension ∆ and the other with
the shadow dimension ∆ ˜︁ ≡ d − ∆ — with the black boundary point im-
plicitly integrated over ∂AdS. Remember that the rest of bulk-to-boundary
propagators have the dimension ∆ϕ .

(3) Now we can perform the bulk integrations. For fixed boundary points, both
of the bulk integrals individually must result in a multiple of the unique
CFT 3-point structure of the corresponding scalar operators — these are
diagrammatically represented by the blue blobs. We are thus left with a
convolution of two such 3-point structures, since we have yet to integrate over
the boundary point.

(4) Such a convolution is actually the shadow representation of the Conformal

Partial Wave (CPW) [92] [93] [94] [95] [96] [61], which we denote by
| ⟩.

In the end, the computation results in the Conformal Partial Wave decomposition
of the s-channel σ-exchange diagram. The t-channel/u-channel diagrams are given
by the same expression, just with the CPWs in the corresponding channels.

We will discuss such decompositions in more detail in the next section. The
associated coefficient/weight function (also called the spectral function) is the
main
⟨︂ object of
⟩︂ interest. It encodes the CFT data accessible to the 4-point correlator
Oϕ Oϕ Oϕ Oϕ .

5.3 CFT Generalities — 4-Point Correlators

As the boundary limits of correlators in AdS enjoy a conformal symmetry, it is
only natural to study them using the language and methods of CFTs. Calculation
of the exchange diagram contribution to the 4-point correlator — directly relevant
for the O(N ) singlet CFT spectrum — naturally resulted in its conformal partial
wave decomposition, which we will review in Section 5.3 → p.115 .

Things get more involved when one wants to extract the non-singlet spectrum. We
will see in Section 5.4 → p.123 that crossed-channel diagrams — take for example
the t-channel — turn out to be essential for this task. Even though it is not viable
Chapter 5. O(N) Model in Anti–de Sitter 114

to calculate directly the s-channel decomposition of the t-channel interaction

diagram, its t-channel decomposition is basically known after resolving the singlet
spectrum. In Section 5.3 → p.116 we will discuss how to translate the t-channel
decomposition into the s-channel one via the 6j–symbols.

To fully prepare for the extraction of the leading 1/N corrections to the non-singlet
spectrum, in Section 5.3 → p.118 we will present the general formulas (in d = 2 and
d = 4) for the contribution of a t-channel conformal block to the anomalous
dimensions of s-channel double-twist operators, applicable for external scalar
operators with equal scaling dimensions.

For simplicity — and because it is the relevant case for us — in the following
we assume all external operators to be scalars Oϕ with equal scaling dimensions
∆ϕ . To avoid clutter, we will suppress throughout the possible global symmetry
index structure of the operators, and the dependence of correlators/conformal
blocks/conformal partial waves on the scaling dimensions and positions of external
operators.

It is enough to concentrate only on the nontrivial crossing between s-channel and

t-channel, since the u-channel is then given by including an additional factor of
(−1)J . Here J is the spin of the exchanged operator belonging to the symmetric
traceless representation of the rotation group SO(d). This follows from a simple
relation between t-channel and u-channel conformal blocks for equal external
scaling dimensions (59) [34], or can be deduced directly from the property of
the OPE coefficients for two scalar operators O1 , O2 and a spin J operator OJ —
ope[OJ O1 O2 ] = (−1)J ope[OJ O2 O1 ] — see (25) [34].

To summarize, the starting point is MFT, where it is known that the identity
operator in the t-channel induces a double-twist family of operators On,J in
(MFT)
the s-channel decomposition, with scaling dimensions ∆n,J ≡ 2∆ϕ + 2n + J.
The question that we want to answer in the following is: Given a t-channel
contribution of an operator O′ with scaling dimension ∆′ and spin J ′ , what
anomalous dimensions (and corrections to OPE coefficients) does it induce for
the double-twist family in the s-channel decomposition? Supposing that the
exchange contribution is of order O(1/N ), we will compute the leading form of the
(MFT)
corrected scaling dimensions ∆n,J ≡ ∆n,J + γn,J ≡ 2∆ϕ + 2n + J + γn,J , where
the anomalous dimensions γn,J are also of order O(1/N ). Thus, as expected, this
family of operators reduces to the MFT ones in the limit N → ∞. The dimension
∆′ of the t-channel exchanged primary does not need to be parametrically close
to an MFT value in the large N expansion. In fact, it actually turns out that to
contribute nontrivially at leading order, it should not be.
Chapter 5. O(N) Model in Anti–de Sitter 115

Conformal Block and Conformal Partial Wave Decompositions

By grouping the operators in pairs and using the Operator Product Expansion
(OPE) twice, we can write the 4-point correlator as a discrete sum over con-
tributions of exchanged conformal families, weighted by the (squares of) OPE
coefficients. Such contributions are called Conformal Blocks (CBs), and they
correspond to an exchange of a physical primary operator together with all of its
descendants. Since only symmetric traceless tensors appear in the OPE of two
scalars, the exchanged families are labeled by their scaling dimension ∆⋆ and spin
J⋆ , giving us the Conformal Block Decomposition

i k [︂ ]︂⃓
⃓ (s)
⟩︃ (︂ )︂
= ope2 Oϕ Oϕ O⋆ ⃓⃓G∆ ,J = (5.21)
∑︂
analogously in
t-channel ,
⋆ ⋆
j l primary O⋆
with ∆⋆ ,J⋆

[︂ ]︂
where ope Oϕ Oϕ O⋆ is the OPE coefficient for operator O⋆ appearing in the
⃓ ϕ × O⟩︂ϕ OPE, and the corresponding s-channel conformal block is denoted by
O
⃓ (s)
⃓G∆⋆ ,J⋆ . We suppress the dependence of CBs on conformal representations and
positions of the external operators.

Alternatively, there is an expansion into Unitary Irreducible Representations

(UIRs) of the (Euclidean) conformal group SO(d + 1, 1), including mainly the
Principal Series (traceless symmetric) representations with integer spin J ∈ N0 but
“unphysical” complex dimensions ∆ ∈ d2 +ı̊R≥ . The associated eigenfunctions of
conformal Casimir are called Conformal Partial Waves (CPWs), which furthermore
form a complete basis of “normalizable” functions [93]. They come from the
harmonic analysis of the Euclidean conformal group SO(1, d+1), and are analogous
to plane waves, which are eigenfunctions of translations in the flat space.

Ignoring non-normalizable contributions such as the exchange of the identity

operator (see [96] for some discussion), we have the Conformal Partial Wave
Decomposition

i k ⃓i k ⃓⃓
⎡ ⃓ ⎤
∞ ∫︂
d∆ ∆ ⃓⃓
⟩︃ (︂ )︂
= ⎦⃓ ∆,J = , (5.22)
∑︂
analogously in
Specs
2πı̊
⎣
t-channel
j l J=0
d
2
+ı̊R≥ J ⃓⃓ j l
⃓

where the s-channel conformal partial wave is denoted by | ⟩, and the spectral
function Specs [ · · · ] represents the coefficients of the corresponding CPWs. Again,
we suppress the dependence of CPWs on conformal representations and positions
of the external operators.

If we were to perform decomposition in the t-channel, we would use t-channel

spectral function Spect [ · · · ] and t-channel CPWs | ⟩. In [61] [96] they use
either ρ or I/n for our Specs , where the normalization n is defined in (5.28).
Notice that the integration in (5.22) is over the half-line ∆ ∈ d2 +ı̊R≥ , since
Chapter 5. O(N) Model in Anti–de Sitter 116

the shadow scaling dimensions ∆ ˜︁ ≡ d − ∆ ∈ d

2
+ı̊R≤ correspond to equivalent
representations of the conformal group.
These two decompositions are closely related, as CPWs are (up to normalization)
shadow-symmetric combinations of CBs
⃓ ⟩︃ ⃓ ⟩︃ ⃓ ⟩︃
⃓ ∆,J ⃓ (s) ⃓ (s)
⃓
⃓ = ˜︁ ⃓G∆,J
K∆,J ⃓ + K∆,J ⃓G∆,J
⃓ ,
(5.23)
˜︁
⃓ ⟩︃ ⃓ ⟩︃ ⃓ ⟩︃
⃓ (t) ⃓ (t)
= ˜︁′ ,J ′ ⃓G∆′ ,J ′ + K∆′ ,J ′ ⃓G∆
⃓
⃓
⃓ ∆ ′,J ′
K∆ ⃓ ⃓
˜︁′ ,J ′ ,

where the normalization coefficients are given by (2.16) [61] (A.5), (A.6) [96]
⎞2
π 2 Γ∆− d2 Γ∆+J−1 ⎝ Γ∆+J
d
⎛

= (5.24)
˜︁
2
K∆,J ⎠ .
(−2)J Γ∆−1 Γd−∆+J Γ∆+J
2

In the t-channel we will usually use primed quantities for improved distinction.
Also, to make the expressions compact, we will often utilize the abbreviation
Γ ≡ Γ( ).
Since CPWs are shadow-symmetric, it is natural to choose the spectral function to
be shadow-symmetric as well. Substituting (5.23) into (5.22), using the aforemen-
tioned shadow-symmetry of Spec to extend the integration from half-line d2 +ı̊R≥
to d2 +ı̊R together with taking only the first conformal block term in CPW, we
obtain an integral decomposition in terms of conformal blocks as
i k ⃓i k
⎡ ⃓ ⎤
d∆ ∆ ⃓⃓
⃓ ⟩︃
∑︂ ∫︂ ⃓ (s)
= Specs ⎦ K ⃓G
˜︁ ⃓ ∆,J . (5.25)
2πı̊
⎣
d J ⃓⃓ j ∆,J
+ı̊R
j l J 2 l
After enclosing the integration in the right-half plane, the residue theorem enables
us to write it as a discrete sum over physical poles of Specs , thus obtaining the usual
conformal block decomposition (5.21). Certain subtleties of this procedure, in
particular appearance and cancellation of additional spurious poles, are discussed
in Appendix B [96].
We see that (at least the accessible part of) the CFT data are encoded in the
spectral function Specs , namely the scaling dimensions of primary operators
are given by the positions of Specs poles, and the corresponding squared OPE
coefficients in the s-channel are given by the (minus, since the contour is clockwise)
residues as
⃓i k
⎛ ⎡ ⃓ ⎤⎞
⎣∆⃓
[︂ ]︂
ope2 Oϕ Oϕ O⋆ = − Res∆=∆⋆ ⎝K∆,J (5.26)
⃓
Spec s
⎦⎠ .
˜︁ ⋆ J⋆ ⃓⃓ j l

CPW Orthogonality and Completeness, 6j–Symbol

As already mentioned, CPWs (along the principal series) form a complete basis
of functions, which furthermore are orthogonal with respect to an appropriate
Chapter 5. O(N) Model in Anti–de Sitter 117

conformally-invariant pairing (1.3) [96], or alternatively a closely related inner

product (A.27) [96]. We will use the bra-ket/inner-product notation in the
following, that is
⟨︃ ⃓ ⟩︃
∆,J ¯ J¯
∆,
= n∆,J 2πδ(ν − ν̄) δJ J¯ , (5.27)
⃓
⃓
⃓

where ∆ ≡ d2 +ı̊ν and ∆ ¯ ≡ d +ı̊ν̄ with ν, ν̄ ≥ 0 are scaling dimensions in the

2
principal series, and n∆,J is the normalization (2.35) [61]

˜︁ K∆,J vol(S ) (2J + d − 2)π ΓJ+1 ΓJ+d−2

d−2
K∆,J
n∆,J ≡ d . (5.28)
2 vol(SO(d − 1)) 2d−2 ΓJ+
2
d
2

Note that it includes an extra 2−d compared to (A.14), (A.15) [96]. We can thus
express the completeness relation (for normalizable functions) in terms of CPWs
as
d∆ ⃓⃓ 1
∑︂ ∫︂ ⃓ ⟩︃ ⟨︃ ⃓
∆,J ∆,J
1= (5.29)
⃓
⃓.
J
d
2
+ı̊R≥ 2πı̊ ⃓ n∆,J ⃓

Consider now a certain contribution to 4-point correlator, for which we are able to
calculate its t-channel spectral function Spect , that is we know the decomposition
i k ⃓i k ⃓⃓
⎡ ⃓ ⎤
d∆′ ∆′
∑︂ ∫︂ ⟩︃
= (5.30)
⃓
Spect⎣ ′ ⃓ ⎦⃓ ∆′,J ′
j l J′
d
2
+ı̊R≥ 2πı̊ J ⃓
⃓j l
⃓

into t-channel CPWs. To extract the associated contribution to the CFT data,
we first need to translate this decomposition into the s-channel. By inserting
the completeness relation (5.29) into the t-channel decomposition, or by directly
utilizing the orthogonality (5.27), we obtain the s-channel spectral function as
⟨︃ ⃓ ⟩︃
∆,J ⃓
⃓i k ⃓i k
⃓ ⃓
∆′,J ′
⎡ ⎤ ⎡ ⎤
∆ d∆′ ∆′
⃓
∑︂ ∫︂
=
⃓ ⃓ ⃓
Specs ⃓ Spect⎣ ′ ⃓ ⎦,
2πı̊
⎣ ⎦
J ⃓⃓ j l J′
d
2
+ı̊R≥
l n∆,J J ⃓
⃓j
(5.31)
where the (s ← t)–channel translation is being performed (up to the normalization
n∆,J ) by the “Clebsch-Gordan coefficient” for the conformal group called 6j–symbol
(3.5) [61]
⟨︃ ⃓ ⟩︃
∆,J ⃓
6j–symbol ≡ ⃓
⃓ ∆′,J ′
⟨︃ ⃓ ⟩︃ ⟨︃ ⃓ ⟩︃
(t) (t)
= K∆
∆,J
G∆′ ,J ′ + K∆′ ,J ′
∆,J
(5.32)
⃓ ⃓
˜︁′ ,J ′
⃓
⃓
⃓
⃓ G∆
˜︁′ ,J ′ ,
⏞ ⏟⏟ ⏞ ⏞ ⏟⏟ ⏞
∆ϕ ∆ϕ
B[∆,J],[∆′ ,J ′ ] B[∆,J],[∆′ ,J ′ ]
˜︁

where B is the notation used in (3.35), (3.41) [61], and corresponds to the s-
channel spectral function of a single t-channel conformal block.
Chapter 5. O(N) Model in Anti–de Sitter 118

Alternatively, we could start with the t-channel conformal block decomposition of

the contribution (5.30), given by sum over t-channel conformal blocks with some
coefficients
i k ⃓
⃓ (t)
⟩︃
= (5.33)
∑︂
CO⋆′ ⃓⃓G∆′ ,J ′ .
⋆ ⋆
j l primary′ O⋆′
′

with ∆⋆ ,J⋆

Again, utilizing the orthogonality (5.27), the corresponding s-channel spectral

function is given by
⃓i k
⎡ ⃓ ⎤
∆ ⃓⃓ 1
⟨︄ ⃓
⃓ i k⟩︄
∆,J
Specs =
⎣ ⎦ ⃓
⃓
J ⃓⃓ j l n∆,J ⃓
j l
1 (5.34)
⟨︃ ⃓ ⟩︃
∆,J (t)
=
∑︂ ⃓
CO⋆′ ⃓ G∆′ ,J ′ ,
primary O⋆′
n∆,J ⃓ ⋆ ⋆

with ∆′⋆ ,J⋆′

⏞ ⏟⏟ ⏞
s ←− t
≡ CrK⟨∆,J|∆′ ,J ′ ⟩
⋆ ⋆

where we introduced the “crossing kernel” CrKs←t notation for the s-channel
spectral function of a single t-channel conformal block (including the proper
normalization).
Strictly speaking, due to the limited validity range of the Lorentzian inversion
formula utilized to calculate CrKs←t in [96], the t-channel conformal block inver-
sion in (5.34) seems to work only for J > J ′ . We will touch on this issue a little
bit more in Section 5.5 → p.141 , also see remarks in Section 4 [61].
From extensive studies of the Lorentzian inversion formula [97] [96] [98],
it is well known that CrKs←t has double zeros in ∆′ at the locations of MFT
double-twist dimensions, therefore only non-MFT operators contribute to the
s-channel spectral function Specs . Direct consequence of this fact can be explicitly
seen in the formulas for the anomalous dimensions presented in the following
Section 5.3 → p.118 .

Contribution of t-Channel CBs to Anomalous Dimensions

Suppose we solve the theory in some kind of perturbative expansion — we will be
formulating everything in the context of large N expansion, but it is applicable
more generally. We therefore organize correlators and CFT data into expansion
in powers of 1/N , for example consider that certain part of 4-point correlator is
given by
i k i k i k
1
= + + ... , (5.35)
j l j l N j l

where the first disconnected term is the t-channel identity contribution, and the
second term is some interaction contributing at O(1/N ) order, for which we are
Chapter 5. O(N) Model in Anti–de Sitter 119

able to calculate its t-channel conformal block decomposition. Our goal is to

extract the respective correction to the CFT spectrum.

Disconnected t-Channel Spectral Function. Spectral function of the t-channel

identity (t-channel disconnected 4-point MFT correlator) in the s-channel was
computed in its general form in [46] and later reproduced within an elegant
harmonic analysis formalism [87]

⃓i k J−1 Γ∆−1 Γd
2
Γd +J Γ∆+J ΓJ+∆
2
ΓJ−∆ +∆ϕ Γ∆+J−d +∆ϕ
⎡ ⃓ ⎤
∆ ⃓⃓ ⎦=
2 2
−∆ϕ 2
˜︁ 2 2 2
Specs⎣ ,
˜︁ Γ∆ϕ ΓJ+1 Γ∆− d2 Γ∆+J−1 ΓJ+2 ∆
˜︁ Γ2d+J−∆ −∆ϕ Γ∆+J+d −∆ϕ
2 2
J ⃓⃓ j l S∆,J
2 2
(5.36)
where S∆,J
˜︁ ≡ (−2) K∆,J
J
˜︁ .
(︂ )︂
As expected, it contains poles — which come from Γ J−∆
2
+ ∆ϕ — located at
(MFT)
the dimensions ∆n,J = 2∆ϕ + 2n + J corresponding to the family of MFT
(MFT)
double-twist operators schematically given by On,J = [Oϕ n ∂ J Oϕ − traces].
Corresponding squared OPE coefficients can be easily calculated using (5.26).
There are also some spurious poles coming from Γ(∆ ˜︁ + J) ≡ Γ(d + J − ∆), but
these are resolved as discussed in Appendix B [96].

Thus, from the s-channel conformal block decomposition point of view, at the
leading O(1) order the t-channel identity gives rise to the aforementioned double-
twist operators. At the following O(1/N ) order, the connected interaction term
modifies the CFT data, in particular it induces anomalous dimensions γn,J of these
operators. This is what we will focus on in the following.

Contribution of t-Channel Exchange. Recalling (5.26), the appearance of

an operator O⋆ in the s-channel conformal block decomposition of the 4-point
correlator is reflected in the spectral function as a simple pole of the form (now
given as series in 1/N )
(︂ )︂
⃓i k
⎡ ⃓ ⎤
1
−C⋆ N ∆⃓
(︂ )︂ ∈ K˜︁ Specs⎣ ⃓⃓ ⎦. (5.37)
∆ − ∆⋆ 1 N
∆,J J ⃓j l

Defining the leading corrections to the squared OPE coefficients and scaling
dimensions as

1 1 (1) 1
[︂ ]︂ (︃ )︃ (︃ )︃
2
ope Oϕ Oϕ O⋆ ≡ C⋆ = C⋆(MFT) + C⋆ + O , (5.38)
N N N 2 )︃
1 1 (1) 1
(︃ )︃ (︃
∆⋆ = ∆(MFT)
⋆ + γ⋆ + O , (5.39)
N N N2
Chapter 5. O(N) Model in Anti–de Sitter 120

the expansion of (5.37) to the first order in 1/N reads

(︂ )︂ ⎡ ⎤
−C⋆ N1 −C⋆
(MFT)
1 −C⋆
(1) (MFT) (1)
−C⋆ γ⋆
(︃
1
)︃
(︂ )︂ = (MFT)
+ ⎢ (MFT)
+ (︂ )︂2 ⎦ + O
⎥
.
∆ − ∆⋆ ∆ − ∆⋆ N ∆ − ∆⋆ N2
⎣
1 (MFT)
N ∆ − ∆⋆
(5.40)

Anomalous dimensions are thus encoded in the coefficients of double poles in the
spectral function, divided by the corresponding MFT squared OPE coefficients.
Since the MFT squared OPE coefficients are given by residues of the t-channel
identity, we can express the (O(1/N ) part of) anomalous dimensions as

∆ ⃓⃓ i k]︄
⎛ [︄ ⃓ ⎞
⎜ Specs ⎟
J ⃓j
⃓
(1)
⎜
l ⎟
γn,J = Res∆=2∆ϕ +2n+J ⎜ . (5.41)
⎜ ⎟
∆ ⃓⃓ i k]︄ ⎟
[︄ ⃓
⎜ ⎟
⎝
Specs ⎠
J ⃓j
⃓
l

If we wanted to extract the O(1/N ) contributions to the OPE coefficients, we

would just calculate the (minus) residue without dividing by the spectral function
of t-identity.
For simplicity, consider that the interaction term in (5.41) is composed of a single
t-channel conformal block with scaling dimension ∆′ and spin J ′ . Utilizing (5.34),
the corresponding contribution to the anomalous dimensions is given by
⎛ ⎞
s ←− t
⃓ ⎜ CrK⟨∆,J|∆′ ,J ′ ⟩ ⎟
(1) ⃓
= Res∆=2∆ϕ +2n+J ⎜ (5.42)
⎜ ⎟
γn,J ⃓⃓t-channel ⃓i
[︄ ⃓ ]︄ ⎟ .
k ⎟
Specs ∆ ⃓⃓
⎜
exchange ⎝ ⎠
∆′ ,J ′
J ⃓j l

Simple poles of the expression inside residue (5.42) — or equivalently, double poles
of 6j–symbol or CrKs←t — are not present for generic non-equal external scaling
dimensions. Nonetheless, for pairwise-equal external dimensions the crossing
kernel develops double poles (3.48) and below [61], and hence the anomalous
dimensions obtain nontrivial contributions.

General Formulas for d = 2 and d = 4. The 6j–symbol and thus CrKs←t was
explicitly computed in (3.36), (3.42) [61] for d = 2 and d = 4 by methods relying
on the Lorentzian inversion formula [97]. Compared to our conventions specified
in (5.18), their t-channel and u-channel are swapped, so we include a factor of
(−1)J when taking over their results for 6j–symbol. Moreover, they calculated
also the t-channel conformal block contributions to the leading-twist (n = 0)
anomalous dimensions (3.55), (3.56) [61].
Here we present the general formulas applicable for arbitrary (assuming J > J ′ )
double-twist operators On,J in d = 2 and d = 4. For the detailed calculation see
Chapter 5. O(N) Model in Anti–de Sitter 121

the accompanying Notebook . The final formulas read

∆′ −J ′
(︂ [︂ ]︂)︂
(1) ⃓
⃓
d=2 2 Γ∆2 ϕ Γ∆2 ϕ +J+n sin2 π ∆ϕ − 2 ΓJ+n+1
γn,J ⃓⃓t-channel = − ×
exchange Γ2(∆ϕ +J+n) π2 Γ2∆ϕ +J+n−1
∆′ ,J ′ ⎡ ⎤
1 Γ ′ ′
⎣ ∆ +J˜︁ F(d=4)
(︂ )︂
′ Ω + ′ ˜︁′ ≡ −J ′ ⎦ ,
× ′ +J ∆′ +J ′ J ←→ J
1 + δ0,J ′ Γ∆2 ′ +J˜︁′ n, 2
∆ ˜︁ ∆ϕ +J+n, 2 ,∆ϕ
2
(5.43)
∆′ −J ′
(︂ [︂ ]︂)︂
⃓
(1) ⃓⃓ d=4 2 Γ∆2 ϕ Γ∆2 ϕ +J+n sin π ∆ϕ −
2
2 ΓJ+n+1
γn,J ⃓t-channel = ×
exchange Γ2(∆ϕ +J+n) π2 Γ2∆ϕ +J+n−1
∆′ ,J ′
J + n + 1 2∆ϕ + J + n − 2
× ×
⎡
J + 1 2∆ϕ + J + 2n − 2 ⎤
Γ ′ ˜︁′ (d=4) (︂ )︂
× ⎣ ∆2 +J Fn, ∆′ +J˜︁′ Ω∆ϕ +J+n, ∆′ +J ′ ,∆ϕ −1 − J ′ ←→ J˜︁′ ≡ −2 − J ′ ⎦ .
Γ∆′ +J˜︁′ 2 2
2
(5.44)
The function Ω appearing in the above formulas is given in (3.38) [61], which
after simplification for equal external dimensions takes the form
⎡ ⃓ ⎤
Γ2h′ Γh+p−1
2
Γh′ −h−p+1 h, h, h + p − 1, h+p−1
⃓
Ωh,h′ ,p = ⃓ 1⎦
⃓
4 F3
⎣
Γh′ Γh′ +h+p−1
2
2h, h + h + p − 1, h − h′ + p
′ ⃓
(5.45)
⃓
(︄ )︄
+ h ←→ h′ , p ←→ 2 − p .

To further simplify the expressions, we defined (for d = 2 and d = 4)

(−1)n Γd−2∆ϕ −n Γd2 −∆ϕ −n Γα+n

2
(d) d∈{2,4}
Fn,α = ×
n! Γd−2∆ϕ −2n Γd2 −∆ϕ Γα−n
⎡ 2
⃓ ⎤ (5.46)
−n, −n, d2 − ∆ϕ − n, d2 − ∆ϕ − n
⃓
⃓ 1⎦ .
⃓
× 4 F3 ⎣
d − 2∆ϕ − 2n, 1 − α − n, α − n ⃓
⃓
(d)
Using F0,α = 1, we can quickly check that the original leading-twist (n = 0) results
are indeed reproduced.
Note that d = 2 formula (3.55) [61] considers exchange of a t-channel operator
′
with general quantum numbers (h′ , h ), while we directly wrote the exchange of
′
symmetric traceless tensor STTJ ′ with spin J ′ = |h′ − h |, which for nonzero J ′
′ ′
reduces as STTJ ′ = (h′ , h ) ⊕ (h , h′ ) Section 2 [96].
Interestingly, both formulas (5.43)/(5.44) resemble each other quite well with
minor modifications. Note the appearance of terms related by the “spin shadow”
affine Weyl reflection given by J˜︁′ ≡ 2 − d − J ′ [98], which is already present at the
level of conformal blocks. It would be interesting to see if such similar structure
persists for higher dimensions as well, or perhaps whether a simple master formula
for general (even) dimensions can be derived.
Chapter 5. O(N) Model in Anti–de Sitter 122

Summary. Now we come back to the case of the interaction term having a
nontrivial t-channel conformal block decomposition. The overall contribution to
the anomalous dimensions of double-twist operators is then simply given by a sum
of contributions for each appearing conformal block — in d = 2/d = 4 given by
(5.43)/(5.44) — weighted by the corresponding (squared OPE) coefficients, that is
⃓
(1) ⃓
γn,J = COint⋆′ (5.47)
∑︂
γn,J ⃓⃓t-channel .
primary O⋆′ exchange of O⋆′

The 1/N factor associated with interaction/exchange diagram — see (5.35) — is

included in the weighting coefficients C int , so the anomalous dimensions γn,J are
of order O(1/N ).

Note the appearance of the sin(π[ · · · ]) factor in both of the formulas (5.43)/(5.44).
They have therefore zeros at the MFT dimensions ∆′ = 2∆ϕ + 2n′ + J ′ , n′ ∈ N0 ,
so only non-MFT operators in the crossed channels contribute to the anomalous
dimensions. This directly reflects the property of CrKs←t mentioned at the end of
Section 5.3 → p.116 .

The only theory-specific information is thus dimensions and spins of non-MFT

operators together with corresponding squared OPE coefficients, or equivalently the
poles and residues of t-channel spectral function of t-channel exchange. Conformal
symmetry then dictates the form of the anomalous dimensions through the
structure of 6j–symbol/CrKs←t .

Large Spin Asymptotics. Finally, we will briefly discuss the large spin asymp-
totics of the anomalous dimensions, and in particular their n-dependence. More
details (with references) can be found in Section 5.5 → p.144 .

One can check that large J asymptotics of the contributions (5.43)/(5.44) to the
anomalous dimensions are given by
cn
⃓
(1) ⃓ ′ ′ ′
γn,J ⃓⃓t-channel ∼ J −τ ≡ J −(∆ −J ) =⇒ γn,J ≃ − + ... , (5.48)
exchange J τmin
∆′ ,J ′
so the leading asymptotics of anomalous dimensions (5.47) are governed by the
exchanged operator with the lowest twist τmin = min τ ′ ≡ min(∆′ − J ′ ).

The point we want to make here is that whatever the coefficient c0 for the leading-
twist operators is — for explicit formula see (5.87), which is actually independent
of the dimension — inspection of (5.43)/(5.44) leads to

d∈{2,4} (d)
cn = Fn, ∆′ +J ′ c0 , (5.49)
2

where we take the operator with (∆′ , J ′ ) to be the one with the lowest twist, and
(d)
Fn,α was defined in (5.46).
Chapter 5. O(N) Model in Anti–de Sitter 123

This follows from the fact that the “spin-shadow” terms — with Ω∆ϕ +J+n, ∆′ −J ′ ,∆ϕ
2
in d = 2 and Ω∆ϕ +J+n, ∆′ −J ′ −1,∆ϕ −1 in d = 4 — are dominant for large J.
2

Since rest of their prefactors are the same for n = 0, the dimension-independence
of c0 results from their identical asymptotics. Going now to nonzero n, the
modifications are:
(1) in d = 4 we have extra rational factors which go to 1 for J → ∞,
(d)
(2) there is extra factor of Fn, ∆′ +J ′ before the “spin-shadow” dominant term,
2

(3) in the rest of formula we replace J ↦→ J + n, which does not change the
′
asymptotics of the form J −τ .
The conclusion is that only (2) changes the asymptotics, leading to (5.49).
(d)
Just to be clear, we have defined Fn,α in (5.46) only for d ∈ {2, 4} by bundling
similar factors appearing in both dimensions. Whether such elegant structure for
general-twist anomalous dimensions persists also for higher dimensions is an open
question.

5.4 Spectrum of O(N) Model in AdS

Naturally, as the first step, in Section 5.4 → p.123 we decompose the 4-point boundary
correlator into irreducible representations of the global symmetry group O(N ).
This effectively solves the O(N ) index structure, and enables us to focus on each
irrep separately.
The singlet spectrum was investigated in [59]. In Section 5.4 → p.125 we summarize
their main results and also extend them in some ways — in addition to d = 2 we
consider also d = 4, where a new operator can appear at strong enough coupling.
We showcase the coupling dependence of both the anomalous dimensions and
OPE coefficients, and also discuss the striking pattern of the spectrum when the
bulk is tuned to the criticality. The role of singlet spectrum does not end here,
since it provides crucial input for the non-singlet spectrum via methods presented
in Section 5.3 → p.116 and Section 5.3 → p.118 . We treat their specific implementation
for non-singlet spectrum of O(N ) model in Section 5.4 → p.137 .

Decomposition into O(N) Irreducible Representations

We want to study various operators appearing in the Oϕ• × Oϕ• OPE, where Oϕ• is
the CFTd operator dual to the elementary field ϕ• in the EAdSd+1 bulk. It is well
known that primaries appearing in the MFT limit (in our case N → ∞) — apart
from the identity operator — are the double-twist operators schematically given by
••
On,J = [Oϕ• Oϕ• ]n,J ≡ [Oϕ• n ∂ J Oϕ• − traces]. They come from the crossed-channel
identities, as was already mentioned in Section 5.3 → p.118 .
Chapter 5. O(N) Model in Anti–de Sitter 124

In order to alleviate the struggle of carrying the O(N ) indices around, we will
organize the CFT operators by their O(N ) irreps, and correspondingly decompose
the 4-point correlator. Without loss of generality, we choose the s-channel as the
one in which we perform both the OPE and the O(N )-irrep decomposition.

Double-Twist Operators. Since each Oϕ• transforms in the vector representation

V of O(N ), the Oϕ• ×Oϕ• OPE decomposition under O(N ) follows from the standard

V ⊗ V = ⏞⏟⏟⏞
1 ⊕ 2V ⊕ 2V , (5.50)
⋀︂ ⨀︂

⏞ ⏟⏟ ⏞ ⏞ ⏟⏟ ⏞
(S) (AS) (ST)

where the O(N ) irreps appearing are the singlet (S), the anti-symmetric (AS), and
the symmetric traceless (ST) representation. The corresponding projectors are
given by
(︂ )︂ij 1 ij (︂ )︂ij
[i j]
(︂ )︂ij
(i j) 1 ij
P(S) = δ δkl , P(AS) = δ[k δl] , P(ST) = δ(k δl) − δ δkl . (5.51)
kl N kl kl N
Thus, the double-twist operators can be organized into these three O(N ) irreps,
and are schematically given as
1 ∑︂ i i
⎧ (︂ )︂
(S)
⎪ (S) On,J ≡ P(S) [Oϕ• Oϕ• ]n,J = [O O ]n,J ,
⎪
⎪
⎪
⎪
⎪
⎪ N i ϕ ϕ
⎪
⎨ (︂ )︂
[ij] [i j]
••
On,J · · ·⎪(AS) On,J ≡ P(AS) [Oϕi Oϕj ]n,J = [Oϕ Oϕ ]n,J ,
⎪
1 ij ∑︂ k k
⎪
⎪
{ij}
(︂ )︂
(i j)
⎩(ST) On,J ≡ P(ST) [Oϕi Oϕj ]n,J = [Oϕ Oϕ ]n,J − [Oϕ Oϕ ]n,J .
⎪
δ
⎪
⎪
⎪
N k
(5.52)

Decomposition of the 4-Point Correlator. From the form of the σϕ2 interaction
vertex (5.4) that couples only two identical ϕ• fields, the correlator (5.17) clearly
takes the form
⎛ ⎞ ⎛ ⎞
i k i k i k i k i k
1 1
= δ ij δ kl ⎝ + ik jl ⎜
⎠+δ δ ⎝ +
⎜ ⎟ ⎟
l N j l N j
⎠
j l j l j l
⎛ ⎞
i k i k
1 1
(︃ )︃
+ δ il δ jk ⎜ + ⎠+O
⎟
.
N j N2
⎝
j l l
(5.53)
Diagrams on the right-hand side without indices are meant to be evaluated by
taking any (but fixed) field ϕi on all external legs. We have thus decoupled the
global group-theoretic index structure of the correlator from the dynamics carried
by exchange of the σ-field.
By irreducibility (Schur’s lemma or equivalently the fact that a product of non-
equal projectors vanishes), only matching O(N ) irreps on both sides of the s-
channel decomposition can combine to contribute nontrivially to the 4-point
Chapter 5. O(N) Model in Anti–de Sitter 125

correlator. Therefore, its O(N )-irrep decomposition reads

δ ij δ kl
i k ⏟ ⏞⏞ ⏟ i k i k i k
= N P(S)
ijkl
S + P(AS)
ijkl
AS + P(ST)
ijkl
ST , (5.54)
j l j l j l j l
⏞ ⏟⏟ ⏞ ⏞ ⏟⏟ ⏞ ⏞ ⏟⏟ ⏞
A(S) A(AS) A(ST)

where indices of projectors (5.51) were raised using the Kronecker delta δ •• .
Together with the singlet projector P(S) we introduced an explicit factor of N ,
such that together they are of order O(1), same as non-singlet projectors.
Solving (5.53) and (5.54) for A(S) , A(AS) and A(ST) yields
⎛ ⎞ ⎛ ⎞
i k i k i k i k i k
1 ⎜
S =⎝ ⎠+ + + + ... , (5.55)
⎜ ⎟ ⎟
N j
⎝ ⎠
j l j l l j l j l
⎛ ⎞ ⎛ ⎞
i k i k i k i k i k
1 ⎜
AS =⎝ − ⎠+ − ⎠ + ... , (5.56)
⎜ ⎟ ⎟
N
⎝
j l j l j l j l j l
⎛ ⎞ ⎛ ⎞
i k i k i k i k i k
1 ⎜
ST =⎝ + ⎠+ + + ... , (5.57)
⎜ ⎟ ⎟
N j
⎝ ⎠
j l j l j l l j l

which represent the three projections of the correlator onto O(N ) irreps. Since the
whole correlator is itself of order O(1), their leading order is also O(1). Next are
then O(1/N ) corrections — mainly coming from the interactions — which we are
about to study. As we know, the boundary CFT spectrum decomposes into the
same O(N ) irreps, and each sector can be analyzed by its associated projection of
the correlator.
Authors of [59] paid thorough attention to the singlet spectrum encoded in the
singlet projection A(S) = (5.55). The goal of this paper is to supplement their
efforts by the analysis of the remaining rank-2 non-singlet sectors governed by
A(AS) = (5.56) and A(ST) = (5.57).
However, before embarking on this journey, we need to recall the results of singlet
sector, as it will serve as an input for computing the anomalous dimensions for the
remaining two O(N ) irreps via crossing relations discussed in Section 5.3 → p.118 .

Singlet Spectrum
The leading O(1) term in singlet sector (5.55) is just the s-channel identity
contribution — two identical scalar operators Oϕi Oϕi always contain the identity
operator 1 in their OPE.
Substantially more interesting is the subleading O(1/N ) term, where the t-channel
and u-channel disconnected diagrams meet together with the s-channel exchange
Chapter 5. O(N) Model in Anti–de Sitter 126

diagram. To understand this correction to the MFT picture, we need to analyze

the spectral function
i k i k i k
⎡ ⎤

Specs⎣ + + ⎦. (5.58)
j l j l j l

Disconnected Contributions. The first two disconnected Witten diagrams

correspond in the boundary CFT to GFF/MFT contributions, which were already
showcased in Section 5.3 → p.119 . Slightly more generally (for even/odd combination
of t-channel and u-channel) we have
⃓i k i k
⎡ ⃓ ⎤
∆ ⃓⃓
(︄ )︄
(︂ )︂
Specs⎣ ± ⎦ = 1 ± (−1) J
(5.36) , (5.59)
J ⃓⃓ j l j l

with the two terms differing just by the overall sign (−1)J , see the beginning of
Section 5.3 → p.113 .
In the O(1/N ) singlet sector spectral function (5.58) we encounter the even
combination of (5.59), so only even J survive. Thus, the O(1/N ) disconnected part
(MFT)
of the singlet sector generates poles at the MFT dimensions ∆n,J = 2∆ϕ +2n+J
for even spins J, and the associated squared OPE coefficients are 2/N times the
expression one would get just from the t-channel identity (5.36). Since the squared
OPE coefficients
√ have a factor of 1/N , the OPE coefficients themselves are of
order O(1/ N ).
The conformal block decomposition of the disconnected contributions to (5.55) is
therefore
⎛ ⎞
i k i k
1 ⎜ ∞
∞ ∑︂
2 (MFT) ⃓⃓ (s)
⃓ ⟩︃
+ = (5.60)
∑︂
C G ,
⎟
N j N n,J ⃓ 2∆ϕ +2n+J,J
⎝ ⎠
l j l n=0 J=0
J even

(MFT)
where Cn,J are MFT squared OPE coefficients coming from the t-channel
identity.

Exchange of σ-Field. The next contribution in (5.58) comes from the s-channel
exchange diagram, whose calculation was schematically outlined in (5.20). The
corresponding spectral function — including all numeric and Γ-factors — is
obtained by comparing the result of (4.18) [59] with our convention for the
integral CB decomposition (5.25), yielding
2
⃓i k Γ∆2 ϕ − ∆ Γ∆2 ϕ − ∆˜︁ Γ∆2 Γ∆˜︁
⎡ ⃓ ⎤
∆ ⃓⃓ 1
Specs⎣ ⎦ = −δJ,0 2 2 2 2
. (5.61)
J ⃓⃓ j l λ−1 + 2B(∆)
˜︁ 4π d Γ∆2 ϕ Γ1−
2
d Γ
+∆ϕ ∆− 2 ∆− 2d Γ˜︁ d
2

Compared to [59], we use ∆ϕ for external scaling dimension instead of theirs ∆,

and we usually prefer writing formulas directly in terms of ∆ ≡ d2 +ı̊ν instead of
Chapter 5. O(N) Model in Anti–de Sitter 127

ν. We easily see that (5.61) is shadow-symmetric — ∆ ←→ ∆ ˜︁ — provided that

the bubble function B˜︁ is shadow-symmetric as well, which holds generally for
spectral representations of functions.

The first thing to notice is that (5.61) has support only on spin J = 0 operators.
Already at this point we can see that J > 0 operators in the singlet sector have zero
anomalous dimensions — a statement valid to the O(1) order actually considered,
since the squared OPE coefficients already have a factor of 1/N , and we do not
consider O(1/N 2 ) corrections. Concluding this observation, to the order O(1/N ),
(S)
the singlet J > 0 operators appearing in the Oϕ• × Oϕ• OPE are On,J with even J,
(S) (MFT)
MFT dimensions ∆n,J>0 = ∆n,J ≡ 2∆ϕ + 2n + J, and squared OPE coefficients
as in (5.60).

Bootstrap Idea — Consistency )︂ Spectrum. The case of J = 0 is

(︂ of the
more intricate. The factor of Γ ∆ϕ − 2 generates a set of double-poles at the
2 ∆

MFT dimensions ∆ = 2∆ϕ + 2n, which is something one would expect from
(S)
perturbative corrections to the scaling dimensions of On,0 from the interaction.
Indeed, expanding (5.61) in λ, and comparing it to the perturbative expansion of
poles in Specs — as in (5.40), where instead of 1/N we now consider series in λ —
one sees that O(λ) term coming from single insertion of λϕ4 vertex has double
(S)
poles and generates O(λ) anomalous dimensions for On,0 .

Furthermore, at O(λk ) we would expect appearance of (k + 1)–degree poles at

MFT dimensions. Such a behavior can come only from higher and higher powers
of bubble function B˜︁ in the λ-expansion of (5.61). It is thus plausible to anticipate
that B˜︁ has poles at MFT dimensions, such that diagrams (5.9) with more bubbles
contribute to anomalous dimensions at corresponding orders of λ. At the same
time, if B˜︁ happened to have additional poles at non-MFT dimensions, it would
indicate appearance of new operators already at the O(λ2 ) order. Such operator
spectrum changes are generally not expected to happen perturbatively for weak
coupling, so we can conclude that the B˜︁ function should have poles only at MFT
dimensions.

However, in the resummed situation(︂ — that

)︂ is finite λ — in addition to MFT
double-poles still coming from Γ ∆ϕ − 2 there appear poles coming from zeros
2 ∆

of the λ−1 + 2B˜︁ denominator. These poles depend continuously on the coupling
λ, so in order for this to be consistent — without sudden appearance of a whole
bunch of new operators once we turn on the coupling — the remnants of the MFT
poles must be canceled between the disconnected and exchange diagrams. For
such cancellation, the Γ2 double-poles must be accompanied by simple zeros of
the (λ−1 + 2B)˜︁ −1 , that is simple poles of B
˜︁ at MFT dimensions, as was already
anticipated. Such arguments allowed the authors of [59] to “bootstrap” the
bubble function B˜︁ (or more precisely its spectral representation).
Chapter 5. O(N) Model in Anti–de Sitter 128

Bubble Function. This requirement of precise cancellation of the MFT poles —

together with the nice behavior at infinity for low enough dimension d — leads to
the following sum over the ∆ poles (4.26), (4.27) [59]

1 ∞
1 Γd +n Γ∆ϕ +n Γ∆ϕ − d + 1 +n Γ2∆ϕ − d +n 1
(︄ )︄
= + ∆←
→∆
∑︂
2 2 2 2
B(∆)
˜︁ ˜︁
∆ϕ − + n Γd2 Γ∆ϕ + 12 +n Γ∆ϕ − d2 +1+n Γ2∆ϕ −d+1+n n!
d ∆
4(4π) 2 n=0 2

Γ∆ϕ Γ∆ϕ − d + 1 Γ2∆ϕ − d

= 2 2
d
2
×
⎛ 4(4π) ⎡ 2
⃓ ⎤⎞
˜︁ ⎣ ∆ϕ − 2 , 2 , ∆ϕ , ∆ϕ − 2 + 2 , 2∆ϕ − 2
∆ d d 1 d ⃓
⎜Γ ⃓ 1 ⎦⎟
⃓
⎜ ∆ϕ − ∆2 5 F4
∆ϕ − 2 + 1, ∆ϕ + 2 , ∆ϕ − 2 + 1, 2∆ϕ − d + 1 ⃓⃓ ⎟
∆ 1 d
×⎜
⎜
⎟
⎟,
⎜ (︃ )︃ ⎟
exchange
+ ∆ ←−−−→ ∆
⎝ ˜︁ ⎠

(5.62)
where after extracting some Γ-factors in front we recognized the regularized
generalized hypergeometric function 5 F
˜︁ . The bubble function B(∆)
4
˜︁ — even
for unequal masses — has been previously computed also by different methods
in [99] (see also [100]).

The summands in (5.62) behave asymptotically as nd−4 for n → ∞, so the sum

becomes divergent for d + 1 ≥ 4. Same as in flat space, this corresponds to the
UV divergence of the bubble diagram, and one must regularize it somehow. By
subtracting sufficient number of terms (up to and including degree d − 3) in the
Taylor expansion of the summands around ν = 0 ⇔ ∆ = d2 , the series becomes
convergent and can be summed up in principle. Note that only even powers of
ν ≡ −ı̊(∆ − d2 ) are present, since (5.62) is shadow-symmetric. At the end, to
account for the subtraction, an even polynomial in ν of the corresponding degree
with arbitrary coefficients (in principle depending on ∆ϕ ) must be added back to
the regularized sum. This was also discussed for the spin 1 bubble function in
Section 3.3 [60].

The sum simplifies in even dimensions (see Notebook for details). For example,
taking d = 2, the bubble function does not need any regularization, and it evaluates
to
(︂ )︂ (︂ )︂
ı̊ ψ ∆ϕ − − ψ ∆ϕ −
1+ı̊ν 1−ı̊ν
d=2 2 2 1 ψ∆ϕ − ∆2 − ψ∆ϕ − ∆˜︁2
B(∆)
˜︁ = ≡ , (5.63)
4π 2ν 4π ∆˜︁ − ∆

where ψz ≡ ψ(z) ≡ d
dz
ln Γ(z) is the digamma function.

In the case of d = 4, the bubble function is UV divergent and requires a reg-

ularization as we described above. Due to the shadow-symmetry, it enough to
subtract the constant term in the expansion of the summands around ∆ = d2 , and
Chapter 5. O(N) Model in Anti–de Sitter 129

the result is
[︃ (︃ )︃(︃ )︃]︃
ν 2(2∆ϕ − 5)ν −ı̊ 4(∆ϕ − 2) + ν 2 2
ψ∆ϕ − ∆ − ψ∆ϕ − ∆˜︁
d=4
= + a0 (∆ϕ ) .
2 2
B(∆)
˜︁
128π 2 (1 + ν 2)
(5.64)
There are no clear physical constraints which would enable us to fix the subtraction
ambiguity in the calculation of the regularized bubble function, in the case of
d = 4 being just the undetermined constant a0 . This can be seen from the fact
that λ itself is not renormalization scheme independent, and only the combination
λ−1 + 2B˜︁ in (5.61) has invariant meaning directly connected with the shape of
the physical spectrum.
In the following we will set a0 ≡ − 16π
1
2 . In Section 5.4
→ p.132
we will discuss the
rationale behind this choice. Choosing different a0 can always be reabsorbed in
the coupling λ.
One question still remains — what is the range of physically admissible values of
λ? Without a more detailed analysis of the phase structure in d = 4 we are unable
to provide a definitive answer, and will assume that any non-negative value of λ
is allowed.

Analysis of the Singlet Sector

Now that we presented all relevant formulas for the singlet sector, we can analyze
it in more detail. Since the J > 0 operators are unaffected by the s-channel
exchange diagram, in the following we will focus solely on the J = 0 operators.
In particular, we can extract the scaling dimensions of the singlet operators
contributing in the subleading O(1/N ) order from the poles of the spectral
function (5.58), and also the corresponding squared OPE coefficients from the
residues of these poles.

Scaling Dimensions in the Singlet Sector. As we already discussed, due to

the interplay of the disconnected diagrams and the s-channel exchange, the J = 0
(MFT)
MFT poles ∆n,0 ≡ 2∆ϕ + 2n are canceled in the complete spectral function
(S)
(5.58). Instead, they are replaced by their finite-shifted counterparts ƥ,0 , which
are roots of
(︂ )︂
(S)
λ−1 + 2B˜︁ ∆•,0 = 0 , (5.65)

such that (5.61) has a pole at the corresponding location. We use • to indicate
the place for indexing this new (infinite) family of singlet J = 0 non-MFT primary
operators denoted suggestively as σˆ︁• , since they are induced by the exchange of
the σ-field. As we will see soon, not all of them need to be continuously connected
to the MFT operators.
Chapter 5. O(N) Model in Anti–de Sitter 130

For generic λ and ∆ϕ , the key equation (5.65) is transcendental, and its roots
have to be found numerically. Nevertheless, this can be easily done to a very
high precision using standard numerical methods, for example with the help of
Wolfram Mathematica.
Examples of the bubble functions B˜︁ in d = 2 and d = 4, together with some
particular choices of λ, are displayed in Figure 5.66 → p.131 . Intersection points in
the plots correspond to singlet scalar operators σˆ︁• . At small λ, we can clearly
(S)
identify them as finite deformations of the MFT operators On,0 ≃ [Oϕ• n Oϕ• ](S) .
With increasing λ, the anomalous dimensions grow, and eventually become of
order O(1) in both λ and 1/N .
There is a subtlety in d = 4, where for a strong enough coupling λ a new operator
possibly appears that is not continuously connected with the MFT spectrum.
Such an operator would be then associated to a bound state in AdS. As we will see,
its emergence is crucial for the bulk theory to be critical, which will be discussed
in Section 5.4 → p.132 .
(S)
At large conformal dimensions ∆ ≫ 1, which concerns the operators On,0 with
n ≫ 1, the bubble functions have asymptotics (dots include subleading 1/∆ terms)
(︂ )︂ (︂ )︂
d=2 cot π(∆ϕ − ∆
2
) d=4 ∆ cot π(∆ϕ − ∆
2
)
B(∆)
˜︁ = + ... , B(∆)
˜︁ = + ... .
8∆ 128π
(5.67)
For finite λ, the inspection of (5.65) gives us following asymptotic values of the
anomalous dimensions (governed by the infinities or zeros of cot(· · ·) in d = 2 or
d = 4, respectively)
⎧
⎨0 for d = 2 ,
(S)
(0 < λ < ∞) lim γn,0 = (5.68)
n→∞ ⎩1 for d = 4 .

Of course, for λ = 0 all anomalous dimensions vanish, and for λ = ∞ the different
(S)
∆ scaling in (5.67) does not play a role, and we have limn→∞ γn,0 = 1 for both
d = 2 and d = 4.
In summary, the complete singlet spectrum given by the poles of the spectral
function (5.58) consists of non-MFT scalar (J = 0) operators supplemented
by MFT operators supported at even spins J ≥ 2. The corresponding singlet
twist–spin plot is displayed in Figure 5.69 → p.132 , with the twist being defined as
(S) (S) (S)
τn,J ≡ ∆n,J − J ≡ 2∆ϕ + 2n + γn,J .
The precise dependence of the singlet scalar anomalous dimensions on the coupling
is plotted in Figure 5.70 → p.133 . The plot mainly focuses on the strong coupling
region, and shows in what fashion the finite constant values of the anomalous
dimensions are approached at infinite coupling. The weak coupling regime is not
entirely captured in these plots, but the asymptotics in that region are simple.
Chapter 5. O(N) Model in Anti–de Sitter 131

Figure 5.66 / Graphical representation of scalar singlet spectrum equation (5.65) in

d = 2 and d = 4. The graph of (twice) the bubble function B ˜︁ (5.63)/(5.64) is drawn
by solid black lines. The term −λ−1 for λ = 4 and λ = 20 is represented by the orange
and red lines, respectively. Their intersection points with black lines correspond to
the J = 0 operators σ ˆ︁• in the singlet sector. Most of these operators (all in d = 2) are
associated with the MFT spectrum as can be seen by their asymptotic convergence to
MFT values (gray dashed lines) for λ → 0. In the case of d = 4, there is however a
potential emergent operator at strong enough coupling in the region ∆ ≲ 2∆ϕ , where
the graph of B˜︁ is reaching slightly below the horizontal axis. The plot was evaluated at
indicated external scaling dimensions, corresponding to a massive ϕ-field with Dirichlet
boundary conditions.

Anomalous dimensions are approximately linear in the coupling there, which

follows from standard perturbation theory, namely the ϕ4 contact Witten diagram
contributes to the J = 0 anomalous dimensions at order O(λ).

Squared OPE Coefficients in the Singlet Sector. Having identified the locations
(S)
of physical poles ƥ,0 in the spectral function (5.61), we can now compute the
(S)
squared OPE coefficients of the corresponding singlet operators O•,0 .

Excluding the isolated case where λ is tuned precisely to the critical value when
a new operator emerges at ∆ = d2 , all physical poles at solutions of (5.65) are
Chapter 5. O(N) Model in Anti–de Sitter 132

Figure 5.69 / Twist–spin plots of the singlet spectrum for d = 2 and d = 4 given by the
poles of the complete spectral function (5.58). Only spin J = 0 operators get O(1)
anomalous dimensions in large N expansion. The plots correspond to Figure 5.66 → p.131 ,
so the non-MFT scaling dimensions of scalar operators are precisely given by the
intersection points of black and red lines in that figure.

simple. Since the rest of the remaining factors in (5.61) together with K∆,0
˜︁ are
holomorphic at these poles, the corresponding squared OPE coefficients are given
by (5.26) as
⃓i k
⎛ ⎡ ⃓ ⎤⎞
1 ∆ ⃓⃓ 1
[︂ (︂ )︂ ]︂ (︃ )︃
(S)
ope2 P(S) Oϕ• Oϕ• O•,0 = − Res∆=∆(S) K∆,0
⎝
˜︁ N Specs 0 ⃓
⎣ ⎦⎠ + O
•,0 ⃓j l N2
⃓
1 1 Γ∆2 ϕ − ∆ Γ∆2 ϕ − ∆˜︁ Γ∆4 1
⃓ (︃ )︃
⃓
= 2 2 2
+O ,
⃓
N 2B ′(∆) 4π d2 Γ∆2 ϕ Γ1−
2
Γ d Γ∆ ⃓ N2
⃓
˜︁ ⃓
d ∆−
+∆ ϕ 2 (S)
2 ∆=∆•,0
(5.71)
′
where ≡
B˜︁ denotes the derivative of the bubble function. Note that we
dB
˜︁
d∆
included the factor of 1/N with which the exchange diagram enters into the
correlator.
Plots for the coupling dependence of squared OPE coefficients (5.71) for some of
(S)
the J = 0 singlet operators On,0 associated to MFT ones in the λ → 0 limit — we
do not show possible new operator in d = 4 — is displayed in Figure 5.72 → p.134 .

Criticality in the Bulk

By appropriately tuning the couplings of the theory, a critical point in the bulk
of AdS can be reached. Its existence and the evidence for the bulk conformal
symmetry was first discussed in Section 5 [59]. Since the critical theory in
EAdSd+1 describes an interacting BCFTd+1 by performing a Weyl transformation
Chapter 5. O(N) Model in Anti–de Sitter 133

Figure 5.70 / Coupling dependence of the singlet scalar anomalous dimensions at

indicated fixed external scaling dimensions coinciding with previous plots. In the
free limit λ = 0, they correspond to primary MFT operators of the schematic form
[Oϕ• n Oϕ• ](S) (we do not include here the possible new operator in d = 4). For any
fixed n, the anomalous dimensions in the singlet sector are all positive and approach
a constant value (bounded above by 1) at sufficiently strong coupling. Positiveness
of anomalous dimensions indicates that the interaction in bulk AdS has repulsive
character in the singlet sector. Different behavior when increasing n for constant λ in
d = 2 and d = 4 follows from the large conformal dimension asymptotics discussed
in (5.68).

to a flat half-space Rd × R≥ , we start by recalling the results for the large N

critical O(N ) model obtained there [101].
Chapter 5. O(N) Model in Anti–de Sitter 134

Figure 5.72 / Coupling dependence of squared OPE coefficients of singlet scalar opera-
tors appearing in the Oϕ• × Oϕ• OPE. They are of order 1/N as is clear from (5.71), so
we plot the expression multiplied by this factor. In the free limit λ → 0, they reduce to
MFT values (dashed lines) given by minus residues of (5.36), as they should. All OPE
coefficients approach constant values at strong coupling, with O(1) corrections to the
corresponding MFT values. No regularities can be inferred from these plots, neither
in d = 2 nor in d = 4. Corrections of both signs occur, their strength is not ordered
by scaling dimensions of the singlet scalar operators, moreover their magnitudes can
cross.

As already mentioned in Section 5.2 → p.104 , an appropriately tuned O(N ) model

in Rd+1 with dimension ranging in 2 < d + 1 < 4 flows in the IR to an interacting
CFTIR describing the second-order phase transition separating the broken and the
unbroken phase. Considering now the theory on a flat half-space, we can obtain
 Chapter 5. O(N) Model in Anti–de Sitter 135

different BCFTs in the IR by imposing different conformal boundary conditions

— for the definition of the ordinary, special, and extraordinary transitions of
the O(N ) model on half-space see [102] [103] [77]. Our choice of boundary
conditions in AdS corresponds to the Dirichlet boundary conditions leading to the
ordinary transition.
As explained for example in [77], the ordinary transition can be reached by
setting ∆ϕ = d − 1 and sending the coupling λ → ∞. In the critical point, the
scaling dimensions of boundary operators σˆ︁n induced by the bulk σ-field are given
by ∆ˆ︁σn = d + 1 + 2n. The leading boundary operator σˆ︁0 corresponds to the
displacement operator generally present in any BCFT with a protected scaling
dimension ∆ˆ︁σ0 = d + 1 — see Section 3.2 [104].
As was shown already in [59] for d = 2 (AdS3 ), the singlet spectrum — or
equivalently the spectral representation of the σ-propagator (5.7) or the bubble
function (5.63) — simplifies greatly in the critical point (∆ϕ = 1, λ → ∞).
Recalling the equation governing the singlet spectrum (5.65), we just need to find
the zeros of the simplified bubble function
(︂ )︂
⃓
d=2 cot π
2
∆
=− (5.73)
⃓
B(∆) ⃓⃓
˜︁ .
∆ϕ =1 8(∆ − 1)
Thus, the induced boundary singlet scalar operators σˆ︁n indeed turn out to have
the expected scaling dimensions ∆ˆ︁σn = 3 + 2n.
Although the preceding discussion assumed 2 < d + 1 < 4, and in particular d = 2,
we found a similar striking pattern exhibited also for other even d, suggesting
the continuation of the ordinary transition above its upper critical dimension
d + 1 = 4. Here we however need to discuss the subtraction ambiguity in the
definition of bubble function (5.62), which for example in d = 4 is just the constant
a0 in (5.64).
This turns out to be tied with the appearance of a new operator in d = 4. Taking
∆ϕ = d − 1 = 3, all singlet scalar operators continuously connected with the MFT
spectrum have scaling dimensions ∆ > 2∆ϕ = 6, so the only candidate for the
displacement operator with ∆ˆ︁σ0 = 5 is the one coming from the branch reaching
slightly below the horizontal axis for small ∆, see Figure 5.66 → p.131 .
If we want to obtain the critical behavior at λ → ∞, the requirement of ∆ˆ︁σ0 = 5
fixes the subtraction constant as a0 = − 16π
1
2 — we choose it independent of ∆ϕ ,

such that the critical value of the coupling λ⋆ ≡ 16π 2 when the new operator
emerges is fixed. It just so happens that the rest of operators σˆ︁n with n ≥ 1
have scaling dimensions ∆ˆ︁σn = 5 + 2n, so they complete the family of boundary
operators induced by the bulk σ-field.
A similar story is true also for higher dimensions. Let us first present the observed
formula for simplified bubble functions at critical point in even dimensions, and
Chapter 5. O(N) Model in Anti–de Sitter 136

comment on the fixing of the subtraction ambiguity after. We found

(︂ )︂ ⎛ ⎞
cot π
∆ 2(d−2)
⃓ /︄ d−1
d even 2
=− (∆ − a) (∆ − b)⎠ , (5.74)
⃓ ∏︂ ∏︂
B(∆) ⃓⃓ (︂ )︂ ⎜
˜︁ ⎟
d
22d−1 π −1
Γ
⎝
∆ϕ =d−1 2
d
2 a=4−d b=1
a even b odd

so for example we have explicitly

⎧ (︂ )︂
⎪
d=2 cot π
2
∆ 1
=
⎪
− ,
⎪
⎪
8 (∆ − 1)
⎪
⎪
⎪
⎪
⎪
⎪ (︂ )︂
cot ∆
⎪
π
(∆ − 0)(∆ − 2)(∆ − 4)
⃓ ⎪
⎨
= =d=4 (5.75)
⃓
B(∆) ⃓⃓ 2
− ,
˜︁
∆ϕ =d−1 ⎪
⎪
⎪⎪ 128π (∆ − 1)(∆ − 3)
⎪ (︂ )︂
cotπ
∆
⎪
(∆ + 2)(∆ − 0)(∆ − 2)(∆ − 4)(∆ − 6)(∆ − 8)
⎪
⎪
d=6 2
⎩=
⎪
− .
⎪
⎪
4096π 2 (∆ − 1)(∆ − 3)(∆ − 5)
⎪

Plots of the bubble functions at the critical point can be seen in Figure 5.76 → p.136 .

Figure 5.76 / Bubble functions for ∆ϕ = d − 1 in even dimensions up to d = 8. The

critical point is reached by sending the coupling λ → ∞, that is by looking at the zeros
of the bubble function. The scale of y-axis is adapted in each dimension to make the
nuanced appearance of d2 − 1 new operators more visible. All zeros are at dimensions
∆ = d + 1 + 2n, the first corresponding to the displacement operator, and possibly
some other emergent operators below the MFT dimension 2∆ϕ = 2(d − 1). The rest
are finite deformations of the MFT operators.

Expressions for d = 2 and d = 4 in (5.75) can be directly obtained from general

formulas (5.63) and (5.64) (with the appropriate choice of a0 already discussed)
by substituting ∆ϕ = d − 1. In the case of d = 6, the subtraction ambiguity
is a polynomial of the form a0 + a1 ν 2 . There are now two emergent operators,
and requiring their dimensions to be ∆ˆ︁σ0 = 7 and ∆ˆ︁σ1 = 9 fixes the subtraction
constants a0 and a1 . As before, continuous deformations of MFT operators then
complete the family σˆ︁n with scaling dimensions ∆ˆ︁σn = 7 + 2n.
Chapter 5. O(N) Model in Anti–de Sitter 137

Similar procedure was applied to all even dimensions up to d = 12. Although we

were not able to perform an explicit resummation of the critical bubble function for
dimensions d ≥ 8 at general ∆ to explicitly compare with (5.74), direct evaluation
at any chosen ∆ confirms the expected structure.
The product in the numerator of (5.74) generates shadow-symmetric zeros at even
integers below
(︂ the
)︂ MFT dimension ∆ = 2∆ϕ = 2(d − 1), canceling the “spurious”
poles of cot 2 ∆ . The product in the denominator generates shadow-symmetric
π

poles at odd integers below the displacement

(︂ )︂ operator dimension ∆ = d + 1,
canceling the “spurious” zeros of cot 2 ∆ . We are thus left with poles at MFT
π

dimensions and zeros at ∆ˆ︁σn = d + 1 + 2n.

Non-Singlet Spectrum
The non-singlet part of the spectrum is described by the spectral decomposition
of (5.56) and (5.57). Both (ST) and (AS) cases can be treated simultaneously, as
we have
i k i k i k
⎡ ⎤ ⎛ ⎡ ⎤ ⎡ ⎤ ⎞
(︂ )︂ 1
Specs⎣ ST
AS
⎦ = 1 ± (−1) ⎝Specs⎣ J ⎦+ Specs⎣ ⎦ + ...⎠,
j l j l N j l
(5.77)
where similarly to (5.59) we used that u-channel diagrams are related to the t-
channel ones by a factor of (−1)J . Now we will apply methods of Section 5.3 → p.116
and Section 5.3 → p.118 just focusing on t-channel diagrams of (5.77), and the
subsequent addition of (︂ u-channel)︂ diagrams merely multiplies all of the squared
OPE coefficients by 1 ± (−1)J . We instantly see, that non-singlet sectors
(ST)/(AS) contain only even/odd spins, respectively.
Decomposition of the t-channel exchange diagram into t-channel conformal blocks
is clearly the same as the s-channel conformal block decomposition of the s-channel
exchange diagram. Alternatively, the t-channel spectral function Spect of t-channel
exchange diagram is given by the right-hand side of (5.61).
Even though (5.61) has two types of poles/operators — MFT ones originating
from the Γ2 -factor and the non-MFT ones located at solutions of (5.65) — only
the non-MFT poles/operators contribute to the s-channel spectral function Specs
of t-channel exchange diagram. This was discussed after equation (5.34), which
we now use to write the s-channel spectral function of the t-channel exchange
diagram (including the 1/N prefactor) as
⃓i k ⃓i k
⎡ ⃓ ⎤ ⎡ ⃓ ⎤
1 ∆ ⃓⃓
∫︂
d∆′ 1 ⎣∆
′
= s ←− t ⃓
Specs CrK⟨∆,J|∆′ ,0⟩ K˜︁′ Spec ⃓
2πı̊ 0
⎣ ⎦ ⎦
∆ ,0 N t
N J ⃓⃓ j l
d
2
+ı̊R
⃓
⃓j l
[︂ ]︂
(S)
= ope2 Oϕ Oϕ O•,0 s ←− t
(5.78)
∑︂
CrK⟨∆,J |∆(S) ,
•,0 ,0⟩
(S)
O•,0
Chapter 5. O(N) Model in Anti–de Sitter 138

where the sum runs only over the non-MFT operators exchanged in the t-channel,
(S)
all of which are scalar singlets O•,0 with dimensions given by roots of (5.65) and
corresponding squared OPE coefficients given by (5.71).

Anomalous Dimensions in the Non-Singlet Sector. We are precisely in the

setting of Section 5.3 → p.118 , and the formula for the anomalous dimensions of
double-twist operators from the t-channel exchange (5.47) applied to the non-
singlet sector of O(N ) model reads
[︂ ]︂ ⃓
(ST)/(AS) (S) (1) ⃓
= 2
(5.79)
∑︂
γn,J ope Oϕ Oϕ O•,0 γn,J ⃓⃓t-channel ,
(S) (S)
O•,0 exchange of O•,0

(1)
where again only the non-MFT operators contribute. Definition of γn,J was
given in (5.42) and explicit expressions in d = 2 and d = 4 were presented
in (5.43) and (5.44). Remember that the squared OPE coefficients (5.71) of
the exchanged non-MFT operators are of order O(1/N ), and thus are also the
(ST)/(AS)
anomalous dimensions γn,J . Even though the same formula (5.79) is applicable
for both (ST)/(AS) sectors, only operators with even/odd spins J actually appear
in the (ST)/(AS) sector, respectively.

The whole next section is dedicated to the analysis of the O(1/N ) non-singlet
spectrum corrections governed by (5.79).

5.5 Analysis of the Non-Singlet Sector

Methods established in previous sections allow us to compute the complete non-
singlet spectrum occurring in the Oϕ• × Oϕ• OPE to the order 1/N considered.
Almost all ingredients in the master formula (5.79) determining the non-singlet
spectrum contributions are known analytically, exception being the precise values
(S)
of singlet scalar scaling dimensions ƥ,0 , which need to be found numerically.

Therefore, in Section 5.5 → p.139 we first describe how we chose to truncate the
sum initially going over infinitely many singlet operators. See the accompanying
Notebook for the code implementing all of the calculations.

We then analyze the obtained non-singlet spectrum in a series of plots. The

principal one — a twist–spin plot — is discussed in Section 5.5 → p.140 and verifies
a known Regge trajectory structure of the spectrum. Asymptotic behaviors of two
important sections of this plot are investigated further. First, Section 5.5 → p.143
studies the asymptotics at large twist for a fixed spin. Second, the large spin
asymptotics for a fixed twist family is investigated in Section 5.5 → p.144 , where a
theorem determining the asymptotic behavior is verified as a consistency cross-
check of the obtained spectral data.
Chapter 5. O(N) Model in Anti–de Sitter 139

Finally, dependence on the external scaling dimension and the coupling is examined
in Section 5.5 → p.141 . In particular, strong and weak coupling asymptotics are
explored, from which it is recognized that scaling dimensions of (ST) scalar
operators were not computed correctly due to a known limitation of the Lorentzian
inversion formula entering the derivation of (5.79).

Numerical Calculation of Anomalous Dimensions

As anticipated, the practical problem with the sum (5.79) is its infinite support
on numerical solutions to (5.65) labeled by non-negative integers. The exchanged
(S)
t-channel singlet J ′ = 0 operators will be labeled by k as Ok,0 , and we reserve n to
label the non-singlet twist family whose anomalous dimensions we are computing.
For simplicity, we exclude the possible contribution of the emergent operator in
d = 4.
For numerical evaluation, we chose to simply truncate the sum over exchanged
singlet operators at some maximal value kmax . This corresponds to a pole of
(S)
the t-channel spectral function of some maximal scaling dimension ∆kmax . To be
completely explicit, let us reproduce here the practical truncated formula for the
non-singlet anomalous dimensions
k∑︂
max [︂ ]︂ ⃓
(ST)/(AS) (S) (1) ⃓
γn,J = ope 2
Oϕ Oϕ Ok,0 γn,J ⃓⃓t-channel . (5.80)
(S)
k=0 exchange of Ok,0

Before discussing its behavior on the choice of kmax , or equivalently the convergence
(1)
of original sum (5.79), let us make a remark about γn,J provided in (5.43)/(5.44).
Its evaluation sometimes involves performing a difference of two terms that are nu-
merically huge but almost equal. This forced us to keep a high WorkingPrecision
in the Notebook , namely we set it to a value of 100, sufficient for our range
of computations.
We are still left with a legitimate question — how large should kmax be, such
that the anomalous dimensions (5.80) of non-singlet operators are calculated with
sufficient precision?

Convergence. Obtaining the asymptotic behavior of the summands in (5.79) is

not straightforward — in particular for general n — so we chose a more pedestrian
approach. To assess the convergence, we plotted individual terms in (5.80) and
analyzed their behavior for large k. See for example the n = 0 case shown
in Figure 5.81 → p.140 .
By creating more examples of such log–log scale plots, where a power law falloff
can be fitted conveniently, we observed an asymptotic behavior ∼ k −2(1+J) for
large k. Hence, the convergence improves drastically with increasing spin, and not
many t-channel terms are required to obtain sufficiently precise results. Based
Chapter 5. O(N) Model in Anti–de Sitter 140

on this analysis, we chose the value kmax = 50 for spinning (J ≥ 1) non-singlet

operators and kmax = 200 for scalar (J = 0) operators, where the convergence is
much slower.

As will be explained later, (ST) operators with J = 0 are problematic for yet
another reason — a limitation present already in the derivation of (5.79) — so we
did not attempt to employ resummation techniques that would take care of the
neglected tail of the sum.

Figure 5.81 / Log–log plot of magnitudes of individual terms in (5.80) helping to assess
the convergence of the sum calculating the non-singlet anomalous dimensions. Fixed
external parameters are summarized in the boxes. For large k, the summands falloff
as ∼ k −2(1+J) .

Twists in the Non-Singlet Sector

Roughly since [105] [106] [107], it was known that the proper organizing
principle for a CFT spectrum is in terms of twist/Regge trajectories labeled by
n ∈ N0 that are analytic in spin. This idea was definitely confirmed by [97].
Therefore, we present the main plots of the non-singlet spectrum in the twist–spin
plane, for a fixed value of the coupling and external scaling dimension. The plot
for d = 2 (AdS3 ) is shown in Figure 5.83 → p.142 , while the one for d = 4 (AdS5 ) is
in Figure 5.84 → p.143 .

To prepare the plots we had to choose also a particular value for N . It was fixed to a
rather small unphysical value N = 20 1
, otherwise corrections of twists by anomalous
dimensions (multiplied by a factor 1/N ) from their MFT values would be barely
visible. So the functional dependence of the anomalous dimensions on (n, J) is
faithfully displayed, but their absolute value is exaggerated by extrapolating them
far beyond the validity of the large N expansion.
Chapter 5. O(N) Model in Anti–de Sitter 141

Both plots clearly exhibit the discussed organization of the spectrum into twist
trajectories labeled by n ∈ N0 that are concave as a function of the spin J and
asymptote to the MFT spectrum for J → ∞. This asymptotic shape is valid
above some minimal spin Jcrit that depends on a concrete theory (in our case
Jcrit = 1).
These claims are by now a theorem valid in any CFT, which was established in
a series of papers [106] [107] [108] [109] [110]. Applied to our setting, it
states that if Oϕi is in the spectrum, then so must be (at least for J ≥ 1) the
double-twist families of schematic form [Oϕi n ∂ J Oϕj ].
For the MFT theory at N → ∞, they all have twists τn,J ≡ ∆n,J − J = 2∆ϕ + 2n.
Deforming from the strict large N limit, they receive anomalous dimensions and
the large spin asymptotics of their twist trajectories was shown in (1.7) [106]
and (1) [107] to be of the form
cn
τn,J ≃ 2∆ϕ + 2n − + ··· . (5.82)
J τmin
This theorem serves us as a check of consistency done in Section 5.5 → p.144 , where
we also discuss which “minimal twist” operator governs the asymptotics in our
case.

Dependence of Anomalous Dimensions on Coupling

The main advantage of the large N approach is that it outputs observables as exact
functions of the coupling λ. We will exploit the known functional dependence
in this section. However, we should emphasize that the coupling dependence is
complicated — entering implicitly via numeric solutions to (5.65) — thus the best
we can do is sample over a finite set of coupling values.
Of particular interest are the limiting cases. At which rate do the anomalous
dimensions approach weak/strong coupling? Such question is best answered by a
log–log plot, where a power law approach is captured by the slope of the graph.
We explore this question for d = 2 in Figure 5.85 → p.144 . It shows the dependence
of anomalous dimensions on the coupling for (ST)/(AS) operators with the lowest
spins J = 0/J = 1 belonging to the first two leading Regge trajectories. As is
clear from the plot, non-singlet anomalous dimensions approach a constant value
at very strong coupling. On the other hand, the slope at very weak coupling is
two, thus non-singlet anomalous dimensions start to decrease at a quadratic rate
from their vanishing values in the free theory. This is expected for the J = 1 (AS)
operators, as the O(λ) contact Witten diagram does not contribute to anomalous
dimensions of J ≥ 1 operators.
However, it is worrisome for the (ST) J = 0 operators, since the contact diagram
affects their anomalous dimensions, whose weak coupling asymptotics should be
Chapter 5. O(N) Model in Anti–de Sitter 142

Figure 5.83 / The twist–spin plot of non-singlet spectrum in d = 2 at finite coupling

λ ≈ 90. The chosen dimension ∆ϕ corresponds to a massive ϕ-field in AdS3 with
Dirichlet boundary conditions. The plot shows operators in both non-singlet irreps
of O(N ) — (ST) operators are supported on even spins, while (AS) operators on odd
spins. Anomalous dimensions for J = 0 are comparatively huge compared to those
for J ≥ 1, so we excluded the scalar operators from this plot (furthermore, they are
not reliably computed by (5.79) due to a limitation in its derivation). The essential
feature of the spectrum — its organization into Regge trajectories concave in spin —
is discussed in the main text.

thus linear. This might be connected with the fact that the Lorentzian inversion
formula, that was used in [61] to derive the key formulas on which we build
in Section 5.3 → p.113 , is guaranteed to apply only for spin J ≥ 1 operators (since
one inverts spin J ′ = 0 operators in the t-channel). Based on this weak coupling
discrepancy, we must honestly admit that anomalous dimensions of (ST) J = 0
operators might get additional corrections that are not accounted for by the
Lorentzian inversion formula.
Chapter 5. O(N) Model in Anti–de Sitter 143

Figure 5.84 / The twist–spin plot of non-singlet spectrum in d = 4 at finite coupling

λ ≈ 59. The chosen dimension ∆ϕ corresponds to a massive ϕ-field in AdS5 with
Dirichlet boundary conditions. The plot shows operators in both non-singlet irreps
of O(N ) — (ST) operators are supported on even spins, while (AS) operators on odd
spins.

Dependence of Anomalous Dimensions on Trajectory Label n

Inspecting a cut through Figure 5.83 → p.142 at fixed spin J ≥ 1, it appears that
the anomalous dimensions are growing (in absolute value) as a function of the
twist trajectory label n. If the growth continued even beyond a couple of leading
Regge trajectories shown in that figure, that would be an unexpected behavior.
Therefore, it is interesting to study the anomalous dimensions as a function of the
twist trajectory label at fixed spin.

A representative plot of this dependence is shown in Figure 5.86 → p.145 for the case
d = 2 and spin J = 2 corresponding to the Figure 5.83 → p.142 , but conclusions are
generic. The anomalous dimensions increase at first until reaching a maximum for
a critical twist family n∗ — slightly higher than shown in Figure 5.83 → p.142 — and
then start monotonically decreasing. This asymptotic fall off for very subleading
Chapter 5. O(N) Model in Anti–de Sitter 144

(1)
Figure 5.85 / Dependence of anomalous dimensions γn,J on the coupling in d = 2. The
top row shows the J = 0 operators in the (ST) irrep for the first two Regge trajectories
— n = 0 and n = 1. The bottom row displays the same for the J = 1 operators in the
(AS) irrep. Note, that the plots are in a log–log scale and cover a range of indicated
external scaling dimensions starting at the Breitenlohner–Freedman bound, above
which Dirichlet boundary conditions are applicable.

Regge trajectories (n → ∞) is in fact what one would expect. Recalling (5.79), it

is clear that dependence on n resides only in the second piece inside the sum, given
by (5.43) or (5.44). Those are known analytic functions whose large n asymptotics
can in principle be determined exactly.

Large Spin Asymptotics

Similarly to the previous subsection, also the large spin asymptotics — taking n
fixed — is fully specified by the second piece in (5.79), given by analytic formulas
(5.43) and (5.44) for d = 2 and d = 4, respectively.

The large spin asymptotics were already derived in [106] [107], as presented
in (5.82). The authors even computed the coefficient c0 determining the leading-
twist deviation from MFT at large spin (1.8) [106]. It is given in terms of OPE
data associated with the operator of minimal twist τmin ≡ ∆min − Jmin and spin
Jmin . Since it turns out to be important for determining the correct “minimal
(R)
twist” operator O∆min ,Jmin governing the large spin asymptotics, we reproduce it
Chapter 5. O(N) Model in Anti–de Sitter 145

Figure 5.86 / Dependence of anomalous dimensions on the twist trajectory label n in

d = 2. A fixed spin J = 2 was chosen — operators are transforming in the (ST) irrep
of O(N ) — and λ and ∆ϕ correspond to values in Figure 5.83 → p.142 . The absolute
value of the anomalous dimension increases with n at first, until it reaches a maximum
at around n∗ ≈ 8 and then starts decreasing. This is in fact a general behavior for any
J ≥ 1.

here in the form

2 Γτmin +2Jmin Γ∆2 ϕ [︂ (︂ )︂
(R)
]︂
c0 = ope2 P(R) Oϕ• Oϕ• O∆min ,Jmin . (5.87)
Γτ2min +Jmin Γ∆2 ϕ − τmin
2 2

We take all operators involved to have 2-point functions normalized to unity,

and compared to the cited formula we left out a factor of 2−Jmin due to different
normalization of conformal blocks.
Using (5.43)/(5.44) we were able to derive cn even for non-leading (n > 0) twist
families, see (5.49) for the expression in terms of c0 .
Prefactors in (5.87) can be traced back to (︂(5.43)/(5.44), in particular Γ∆−2ϕ − τmin
[︂ ′ ′
]︂)︂ 2
is related via Γ-function identities to sin2 π ∆ϕ − ∆ −J 2
after identification
τmin ↔ ∆′ − J ′ . These factors are responsible for producing double zeros at values
of MFT operators, which will be important in the following.
Let us now recall possible candidates for the minimal twist operators, and discuss
their respective orders at which they contribute in (5.87). The candidates are
clearly the three minimal twist families [Oϕi ∂ J Oϕj ](R) , (R) ∈ {(S), (AS), (ST)} (up to
a possible emergent operator in d = 4 that would take over the role of a minimal
twist operator at sufficiently strong coupling):
(S) The discussion in singlet sector splits into scalar and spinning operators
(see Figure 5.69 → p.132 ).
• Scalar operators get O(1) deformations from MFT scaling dimensions,
thus their coefficients cn are O(1/N ) due to the squared OPE coefficients
in the singlet sector (5.71).
• Spinning operators have MFT scaling dimensions (at O(1) order), there-
fore their cn coefficients are strongly suppressed by the double zeros
 Chapter 5. O(N) Model in Anti–de Sitter 146

emphasized above. Given they receive some O(1/N ) anomalous dimen-

sions by extending the computation of the correlator to order O(1/N 2 ),
their contributions to coefficient cn turn out to be O(1/N 3 ) — a O(1/N 2 )
suppression coming from the double-pole factor, and an additional O(1/N )
from their squared OPE coefficients.
(ST)
Both types of non-singlet operators have O(1/N ) anomalous dimensions,
(AS)
therefore their contributions to cn coefficients are of order O(1/N 2 ), since
their leading OPE coefficients are O(1).
The important corollary of this analysis is that although there are the (ST)/(AS)
leading-twist families with τ (ST)/(AS) < 2∆ϕ needed for unitarity of the boundary
CFT, their contribution to large spin asymptotics (5.82) is strongly suppressed as
O(1/N 2 ). Hence, the large spin asymptotics are governed by the singlet scalar
(S)
operator σˆ︁0 ∼ [Oϕ• Oϕ• ](S) of “minimal” non-MFT twist τmin = 2∆ϕ + γ0,0 .
(︂ )︂
(S)
By evaluating (5.87, 5.49) for τmin = 2∆ϕ + γ0,0 , Jmin = 0 , the formula (5.82)
provides a complete theoretical prediction for the large spin asymptotics of the
(ST)/(AS) spectrum, that can be checked against data generated by our code.
We turn to describing the setup of these tests and presenting their results in the
rest of this section. Starting from (5.82), taking its logarithm and plugging in the
relation τn,J ≃ 2∆ϕ + 2n + γn,J , we considered the following functional dependence
for the asymptotics

log10 |γn,J | ∼ log10 cn − τmin log10 J . (5.88)

The left-hand side was evaluated for log10 J ∈ [0, 4] with a step 15 , and last few
data points were then fitted. The results for first two twist families are shown
in Figure 5.89 → p.147 , but we performed the same analysis also for higher twist
families with similar results.
The leading asymptotics J −τmin given by the slope of the linear fit in these log–log
(fit)
plots is the same for all Regge trajectories. As explained, the value τmin determined
(S)
by the fits should be compared against the scaling dimension ∆0,0 of the non-MFT
“minimal” twist operator σˆ︁0 . We confirmed almost a perfect match, which only
improves with higher J. Similarly, the coefficients cn show a good agreement with
theoretical values (5.87) (and (5.49)).
In spirit, this verifies that general formulas (5.43)/(5.44) have an appropriate form
reproducing the large spin asymptotics theorem. The specific results for the O(N )
model then necessarily follow.
It is worthwhile to highlight another feature implemented in Figure 5.89 → p.147 .
The black points represent “complete” anomalous dimensions (5.80) summed up
to kmax = 20, while the colored (blue, yellow, green) points correspond in order to
contributions of the first few members {σˆ︁0 , σˆ︁1 , σˆ︁2 } of the scalar singlet family. It is
Chapter 5. O(N) Model in Anti–de Sitter 147

Figure 5.89 / Log–log plots displaying the large spin asymptotics of non-singlet anoma-
lous dimensions for first two twist trajectories (n = 0 and n = 1) in d = 2 and d = 4.
Choices of the parameters used in each plot are shown in the upper right boxes. The
last few data (black) points were fitted by a linear function (5.88) (dashed line) with
parameters given in the bottom left boxes. The slopes of dashed lines represent the
calculated minimal twists, and are in good agreement with the theory value, thus
verifying the large spin asymptotics theorem.

satisfying to visually see how the “minimal” twist operator σˆ︁0 completely saturates
the asymptotics, and how the subdominant contributions are suppressed at large
J. Even if one resummed the whole infinite tower of remaining operators σˆ︁n≥1 ,
their asymptotics would still remain subleading compared to the one associated
with σˆ︁0 .

Finally, we should comment that the theorem is applicable for a generic CFTd
only in d > 2. The problem with d = 2 is that the stress tensor is in the same
conformal Virasoro multiplet as the identity operator (it is its first nontrivial
descendant at level two) and thus has the same twist τ = 0. Without a twist
gap between the identity and an operator of minimal twist, the theorem does not
apply. Yet, we are dealing with a holographic theory without dynamical gravity
in the bulk, and therefore with a non-local dual CFTd on the boundary without
a stress tensor. This implies that the whole Virasoro multiplet of the identity
operator reduces just to the identity operator, and explains why we are able to use
the asymptotic behavior (5.82) also in d = 2, which Figure 5.89 → p.147 confirms.
Chapter 5. O(N) Model in Anti–de Sitter 148

5.6 Summary and Outlook

This project arose from the desire to examine associativity of the OPE (crossing)
in a CFT that is at finite distance (in the sense of its spectral and OPE data) from
an MFT. Holographic theories are a fruitful playground, since especially in d > 2
there are not many other interacting examples that can be handled analytically,
without resorting to the numerical conformal bootstrap.

We chose the O(N ) model at finite coupling in the bulk, as it is a reasonably simple
theory and its input data for crossing, in the form of the singlet spectrum, were
already available thanks to [59]. The set of techniques employed is applicable to
any other theory as long as one has partial knowledge of its CFT data.

In the course of this research we obtained two classes of results which we sum-
marize in the following. We remind the interested reader that various detailed
computations and the implementation of main formulas can be found in the
accompanying Notebook .

Model-Independent — t-Channel CB Contribution to Anomalous Dimensions.

Imagine a situation when one has certain control of the CFT data in one channel,
say the t-channel. To complete the analysis, one should try to deduce from
it CFT data in the nontrivial crossed s-channel (as u-channel is related in a
simple way). In particular, for analyzing the CFT spectrum it is important to
know how a single t-channel conformal block contributes to s-channel anomalous
dimensions. This contribution needs to be further weighted by t-channel (squared)
OPE coefficients to obtain the final result for s-channel anomalous dimensions,
hence both scaling dimensions and OPE coefficients are needed as input in one of
the crossing channels.

This computation requires the knowledge of conformal blocks, whose analytic

expressions are known in d = 2 and d = 4 dimensions. For these spacetime
dimensions, [61] provided a closed-form formula for anomalous dimensions of
leading-twist (n = 0) double-twist operators [Oϕ• Oϕ• ]n,J , once CFT data associated
with the exchange of an operator in the crossed channel are specified. In this
paper we present the generalization of these formulas to arbitrary twists n ∈ N0 ,
namely see (5.43) and (5.44). Moreover, we found an elegant formula (5.49) for
the coefficients of the corresponding large spin asymptotics.

Quite generally, once the direct channel spectrum is resolved, contributions

from the crossed channel can be incorporated by the techniques presented in
Section 5.3 → p.113 . We have demonstrated them in the context of the O(N ) model,
but they are applicable also for other CFTs, perhaps after some generalization.
The closest candidates are CFTs on the boundary of AdS in the large N expansions
of the Gross–Neveu model or the scalar QED.
Chapter 5. O(N) Model in Anti–de Sitter 149

Model-Specific — Non-Singlet Spectrum of O(N) Model at Finite Coupling.

The main output of this paper is the structure of the OPE of two boundary
operators Oϕ• associated with the fundamental fields ϕ• in the bulk, considering
the phase with unbroken O(N ) global symmetry. It was obtained at a finite
coupling λ and up to the first nontrivial order of 1/N in the large N expansion.

It can be viewed as a deformation of the two limiting cases where the quartic
interaction in the action (5.1) is turned off — either by setting N = ∞ or λ = 0.
In both cases, all correlators are just products of two-point functions, and the
corresponding MFT OPE can be schematically decomposed into O(N ) irreps
(denoted by the superscripts) as

j] (AS) {i j} (ST)
[︂ ]︂ [︂ ]︂
N =∞ [i
Oϕi × Oϕj ∼ 1(S) ⊕ Oϕ n J
∂odd Oϕ ⊕ Oϕ n J
∂even Oϕ ,
MFT MFT
1 [︂ • ]︂(S)
Oϕi × Oϕj
λ=0
∼ 1(S) ⊕ √ Oϕ ∂even Oϕ•
n J
⊕ (5.90)
N MFT

j] (AS) {i j} (ST)
[︂ ]︂ [︂ ]︂
[i n J n J
⊕ Oϕ ∂odd Oϕ ⊕ Oϕ ∂even Oϕ .
MFT MFT

Both of them are obtained from the disconnected parts of the 4-point function
decomposition in (5.55, 5.56, 5.57). As the singlet sector double-twist operators
have a factor
√ of 1/N in their squared OPE coefficients, OPE coefficients are of
order O(1/ N ). For N = ∞ they vanish, and the singlet double-twist operators
are transferred fully to the (ST) part, see also (5.52).

At finite λ and large (but finite) N , this picture is modified — only leading
corrections in large N are shown — to the following schematic form

1 [︂ ]︂(S) 1 [︂ • ]︂(S)
Oϕi × Oϕj ∼ 1(S) ⊕ √ σˆ︁• ∼ “Oϕ• n
Oϕ• ” ⊕√ Oϕ ∂even Oϕ•
n J>0
⊕
N fin N MFT

j] (AS) {i j} (ST)
[︂ ]︂ [︂ ]︂
[i
⊕ Oϕ n J
∂odd Oϕ ⊕ Oϕ n J
∂even Oϕ . (5.91)
1/N 1/N

Discussion of singlet operators in the first line of (5.91) splits according to the spin.
Scalar boundary operators σˆ︁• are induced by the composite interacting σ-field in
the bulk — they are given by the poles of the spectral function associated with
the exact σ-propagator. As indicated, their scaling √ dimensions get finite shifts
from MFT, and their OPE coefficients are O(1/ N ) to the order considered.
Spinning singlet operators do not receive such O(1) corrections to their MFT
scaling dimensions. The formula for scaling dimensions of the O(N ) singlet sector
was already given in [59].

The non-singlet operators in the second line of (5.91) get 1/N shifts to their
scaling dimensions.
√ Leading OPE coefficients in this sector are of order O(1), and
acquire O(1/ N ) deformations. Both of these corrections correspond to a 1/N
modification of the correlator. In this work we did not compute the corrections to
the OPE coefficients, even though the setup is ready for it. It is just a matter of
Chapter 5. O(N) Model in Anti–de Sitter 150

computing residues of fairly complicated functions at known poles, and we might

come back to it in the future.
The main new contribution of this work are the scaling dimensions for the non-
singlet double-twist families transforming in the symmetric traceless (ST) and
anti-symmetric (AS) irreps of the O(N ) group. They were obtained for d = 2
(AdS3 ) and d = 4 (AdS5 ), as a function of the twist/Regge trajectory label n, the
spin J, the external scaling dimension ∆ϕ , and the coupling λ.
It is important to realize that they are given as a sum of partly factorized
expressions — the model-dependent squared OPE coefficients, times a model-
independent contribution of a crossed channel block, which however still needs to
be evaluated at model-dependent values. The relevant formula is (5.79), which
we reproduce here with a particular emphasis on the functional dependence
[︂ ]︂ ⃓
(ST)/(AS) (S) (1)
(∆ϕ , λ) = 2
(∆ϕ , λ) γn,J (∆ϕ ) ⃓⃓ t-channel
∑︂ ⃓
γn,J ope Oϕ Oϕ O•,0 .
(S) (S)
O•,0 (∆ϕ ,λ) exchange of O•,0 (∆ϕ ,λ)

(5.92)

The dependence on the external scaling dimension ∆ϕ and the coupling λ is

complicated, in particular due to an implicit dependence in γ (1) via solutions
of a transcendental equation (5.65). However, the dependence on the Regge
trajectory label n and the spin J is completely isolated in the model-independent
piece. Especially, asymptotic behavior for large J or n can be in principle
analytically determined from (5.43) or (5.44), and is displayed in Figure 5.89 → p.147
and Figure 5.86 → p.145 . The large spin asymptotics are verified by a theorem
independently predicting it.
The most significant projection of these data — after choosing some fixed coupling
and desired external scaling dimension — is a plot in the twist–spin plane. A
combined plot of the anti-symmetric and symmetric traceless double-twist families
is shown in Figure 5.83 → p.142 for d = 2, and in Figure 5.84 → p.143 for d = 4. Both
plots confirm organization of the non-singlet double-twist spectrum into Regge
trajectories concave in spin.
Unlike in the singlet sector where anomalous dimensions are strictly positive for
J = 0 operators — indicating a repulsive interaction in the bulk — and zero
otherwise, anti-symmetric and symmetric traceless double-twist families have
negative anomalous dimensions — indicating an attractive interaction in the bulk.
In the process of computing non-singlet anomalous dimensions, the (squared)
OPE coefficients (5.71) in the scalar singlet sector were needed. We plotted
them in Figure 5.72 → p.134 as a function of the coupling. In the limit λ → 0
they approach MFT values as they must, which provides another successful test
of implementation. Compared to anomalous dimensions displaying fixed sign
and convexity properties, OPE coefficients show a rather unorganized behavior.
Chapter 5. O(N) Model in Anti–de Sitter 151

Different σˆ︁• operators get corrected by different signs and their relative strength
changes with the coupling.
Based on the self-consistency scrutiny of the spectrum summarized in the previous
paragraphs and also additional checks done in [59] we believe that the results
are reliable and ready to be used in applications (at least in d = 2 which has a
well understood phase structure and no UV divergence).

Future Directions. There are still various unresolved questions and possible
future directions. One concerns the anomalous dimensions of J = 0 (ST) operators,
for which we have indications that they were not properly captured by the
6j–symbol. Another is the computation of the OPE coefficients for the non-singlet
operators, which is primarily just a technical challenge. Since in principle we can
calculate anomalous dimensions in the non-singlet sector for any n and J, this
opens a possibility to perform a more detailed analysis of the spectrum and its
properties.
Following [111], it would be interesting to calculate non-singlet spectrum of the
O(N ) model in de Sitter (dS) background.
A further valuable extension of our work concerns the critical theory in the bulk.
Its careful investigation particularly attracts our attention for a future project —
either in the BCFTd+1 setting [77] linked directly to AdSd+1 or in the DCFTn+m
setting [70] corresponding to a theory on AdSn × Sm (and requiring thus the
Kaluza–Klein reduction on the sphere). We hope to report on some additional
observables in these theories that could be inferred from the results obtained in
this work.
 152

C
Conclusion
Since we have covered a broad range of topics in this thesis, let us take a moment
to summarize the main narrative.
In Chapter 1 → p.1 , we began by reviewing the QFT formalism, taking the path
integral as a powerful tool for both constructing and analyzing QFTs. We then
formalized its key features within the FQFT framework, which gave us a firmer
grip on the formulation of QFTs in general backgrounds. This allowed us to better
appreciate the role of symmetries, and clarified the conceptual meaning of various
operator formalisms associated with different spacetime foliations.
In Chapter 2 → p.37 , we discussed various types of effective actions. In particular,
we introduced the notion of the RG flow, which has profound implications for
the structure of QFTs and physics in general. It provides a framework for un-
derstanding physical behavior at different energy scales, and explains the origin
of universality observed at second-order phase transitions of statistical systems,
which correspond to the fixed points of the RG flow described by CFTs.
In Chapter 3 → p.51 , we built the foundations of (d ≥ 3) CFTs from the ground up.
We first looked at the conformal structure of manifolds in general, and later we
focused on flat space. Identifying the conformal group and algebra, we studied
their representations and the initial implications of symmetry for correlation
functions. Finally, a particularly important feature of CFT was discussed — the
existence of state–operator correspondence, and the Operator Product Expansion.
In Chapter 4 → p.91 , we turned our attention to the AdS space. Since the isometry
group of AdS acts on its asymptotic boundary as a flat-space conformal group,
QFTs in AdS admit a set of asymptotic observables that are conformally invari-
ant, so much of the CFT methodology can be used to study them. This falls
under the umbrella of the AdS/CFT correspondence, which more generally (after
taking the bulk metric to be dynamic) attempts to define a Quantum Gravity
in asymptotically AdS spacetime through a dual (holographic) CFT living on its
boundary.
In Chapter 5 → p.100 , we studied the O(N ) model in AdS at finite coupling in the
large N expansion. Much of the formalism developed in the previous chapters
was used, either directly or indirectly. The main focus was to extend the known
results from the analysis of the singlet sector also to the non-singlet sector, thereby
obtaining a more complete picture of the spectrum at order 1/N . For a thorough
summary of the results, we refer to Section 5.6 → p.148 .
The ideas presented here reflect a deep interplay between various fields, such as
Mathematics, Quantum Field Theory, Statistical Physics, and Gravity. I hope
you found it as fascinating as I did.
 153

N
Notation & Conventions
Miscellaneous
√
Imaginary unit is denoted as ı̊ ≡ −1 (taking principal square root), and the
Euler’s number or natural logarithm base as “e”.
Throughout we use units where the speed of light is unity. Sometimes, we also set
reduced Planck’s constant to be ℏ ≡ 1.
We will work in the “mostly plus” signature, that is in R1,d−1 we have
d−1
⏟ ⏞⏞ ⏟
(1,d−1)
η•• = diag(−1, +1, +1, . . . , +1) ,

and in Rd ≡ R0,d we have

d
⏟ ⏞⏞ ⏟
(0,d)
η•• = diag(+1, +1, . . . , +1) .

Corresponding Lorentz groups are SO(1, d − 1) and SO(d) ≡ SO(0, d), respectively.

Differential Geometry
While we assume familiarity with various concepts from differential geometry,
including Lie groups, we briefly summarize the notation we will use in this thesis.
Tangent bundle is denoted as T M, the cotangent bundle as T ∗M, and the bundle
of (p, q)-tensors is denoted as Tqp M. Corresponding vector, covector, and tensor
fields are denoted as T M ≡ Sect T M, T ∗M ≡ Sect T ∗M, and T qpM ≡ Sect Tqp M,
respectively.
Bold indices denote abstract indices [42] [112], for example ξ a ∈ T M, gab ∈
T2 M. Repeated indices denote contraction. Indices are raised or lowered automat-
ically using the metric tensor g•• or its inverse g •• , for example A • B ≡ Aa B a ≡
gab Aa B b = g ab Aa Bb .
Indices which are not bold denote components of a tensor in a given coordinate
system (choice of which is usually obvious from context), for example gab ≡ gab .
Repeated indices denote summation.
As already shown, we sometimes omit abstract/coordinate indices, and indicate
the contraction/summation by a dot •/ . For example, we
•
(︂ have x
)︂ ≡ x x ≡ xa xa ,
2 •
a

gab ≡ gkl da x db x ⇔ g ≡ gkl dx dx where gkl ≡ g ∂xk , ∂xl , or ξ ≡ ξ ∂xk

k l k l ∂ ∂ a k ∂

with ξ k ≡ ξk ≡ ξ • dxk ≡ dxk (ξ).

When we want to indicate the type of the object, or the place where some
argument/object should be inserted, we use for example the following placeholders
whose meaning should be clear from the context — a• , b• , A , ⟦•, •⟧.
Notation & Conventions 154

The generator of the flow ϕ ∗ (tangent vector field to the curves generated by the
flow) is defined as ξ ≡ dε
D
ϕε |ε=0 ≡ Dϕ ε
|
dε ε=0
∈ T M, where we distinguished D to
be capital and bold — compared to the ordinary derivative — to indicate that
the resulting object is a vector field.
1
Metric density (of weight 1) is denoted as g /2 ≡ |Det g| √︂
/2
∈ Dens1M, which
1

in given coordinate system can be expressed as g /2 = |det g•• | dd x, where

1

⃓ ⃓
dd x ≡ ⃓dx0 ∧ dx1 ∧ . . . ∧ dxd−1 ⃓ is the coordinate density.
⃓ ⃓

In the “tetrad” formalism, we will denote the vector field frame as {eak } ⊂ T M,
where a is the abstract index and k is the “enumerating” index. The dual (covector)
tetrad frame will be denoted as {eka } ⊂ T ∗M. It is chosen such that eak ela = δkl
and eak ekb = δba .
Frame is always taken to be orthonormal, that is gab = ηij eia ejb , where ηij are
coordinates of the canonical metric in Rp,q (with corresponding signature). We
can easily see (using |det η•• | = 1) that metric density is then e = g /2 = |Det e•• | ∈
1

Dens1M.
We will always use Levi-Civita connection ∇, which is torsion-free (Tor[∇] = 0)
and metric-compatible (∇g = 0). Alternatively, we use ∂ when acting on scalars,
since their covariant derivative does not depend on the connection.
 155

R
References
Back-references to the pages where the publication was cited are given by • .

[1] Jonáš Dujava. TEXtured — LATEX Template. 2024.

GitHub: jdujava/TeXtured
url: https://overleaf.com/latex/templates/textured/zwtzzwgddbsh iii

[2] Arthur Jaffe and Frank Quinn. “Theoretical Mathematics”: Toward a Cultural
Synthesis of Mathematics and Theoretical Physics. Bulletin of the American
Mathematical Society, 1993.
arXiv: math/9307227 vii

[3] Richard P. Feynman. Space-Time Approach to Non-Relativistic Quantum

Mechanics. Reviews of Modern Physics, 1948.
doi: 10.1103/RevModPhys.20.367 1

[4] Pavel Mnev. Lecture Notes on Conformal Field Theory. 2022.

url: https://nd.edu/~pmnev/f22/CFT_2022_notes.pdf 2 , 13 , 52 , 56 , 63

[5] David Skinner. Quantum Field Theory II. 2018.

url: https://damtp.cam.ac.uk/user/dbs26/AQFT.html 2 , 37 , 39

[6] Pierre Deligne, Pavel Etingof, Daniel S. Freed, Lisa C. Jeffrey, David
Kazhdan, John W. Morgan, David R. Morrison, and Edward Witten
(editors). Quantum Fields and Strings: A Course for Mathematicians •
Volume 1. American Mathematical Society, 1999.
url: https://math.ias.edu/QFT/fall/
url: https://bookstore.ams.org/qft-1-2-s 2, 5

[7] Pierre Deligne, Pavel Etingof, Daniel S. Freed, Lisa C. Jeffrey, David
Kazhdan, John W. Morgan, David R. Morrison, and Edward Witten
(editors). Quantum Fields and Strings: A Course for Mathematicians •
Volume 2. American Mathematical Society, 1999.
url: https://math.ias.edu/QFT/spring/
url: https://bookstore.ams.org/qft-1-2-s 2

[8] Andrew Neitzke. Applications of QFT to Geometry. 2021.

GitHub: neitzke/qft-geometry
url: https://gauss.math.yale.edu/~an592/teaching/392C-qft-geometry 2

[9] Maxim Kontsevich and Graeme Segal. Wick Rotation and the Positivity of
Energy in Quantum Field Theory. The Quarterly Journal of Mathematics,
2021.
doi: 10.1093/qmath/haab027 arXiv: 2105.10161 [hep-th] 13 , 18
References 156

[10] Pavel Mnev. Quantum Field Theory: Batalin–Vilkovisky Formalism and Its
Applications. American Mathematical Society, 2019.
url: https://bookstore.ams.org/ulect-72 13 , 39

[11] David Simmons-Duffin. Conformal Field Theory. 2018.

GitHub: davidsd/ph229
url: http://theory.caltech.edu/~dsd/ph229ab/2017-2018/
13 , 52 , 66 , 85 , 88 , 89

[12] Nils Carqueville and Ingo Runkel. Introductory Lectures on Topological

Quantum Field Theory. Banach Center Publications, 2018.
doi: 10.4064/bc114-1 arXiv: 1705.05734 [math] 13

[13] Robbert H. Dijkgraaf. A Geometrical Approach to Two-Dimensional Confor-

mal Field Theory. PhD thesis. Utrecht University, 1989.
url: https://dspace.library.uu.nl/handle/1874/210872 13

[14] Graeme Segal. The Definition of Conformal Field Theory. In: Differential
Geometrical Methods in Theoretical Physics. Springer, 1988.
doi: 10.1007/978-94-015-7809-7_9 13

[15] Michael Atiyah. Topological Quantum Field theories. Publications Mathé-

matiques de l’Institut des Hautes Études Scientifiques, 1988.
doi: 10.1007/BF02698547 13

[16] Mykola Dedushenko. Snowmass White Paper: The Quest to Define QFT.
International Journal of Modern Physics A, 2023.
doi: 10.1142/S0217751X23300028 arXiv: 2203.08053 [hep-th] 14

[17] David Poland and Leonardo Rastelli. Snowmass Topical Summary: Formal
QFT. In: Snowmass 2021. 2022.
arXiv: 2210.03128 [hep-th] 14

[18] John C. Baez. Quantum quandaries: A Category theoretic perspective. 2004.

arXiv: quant-ph/0404040 17

[19] John C. Baez and Mike Stay. Physics, Topology, Logic and Computation: A
Rosetta Stone. In: New Structures for Physics. ed. by Bob Coecke. Springer,
2010.
doi: 10.1007/978-3-642-12821-9_2 17

[20] Raymond F. Streater and Arthur S. Wightman. PCT, Spin and Statistics,
and All That. Princeton University Press, 2000.
url: https://press.princeton.edu/isbn/9780691070629 36

[21] Konrad Osterwalder and Robert Schrader. Axioms for Euclidean Green’s
functions. Communications in Mathematical Physics, 1973.
doi: 10.1007/BF01645738 36
References 157

[22] James Glimm and Arthur Jaffe. Quantum Physics • A Functional Integral
Point of View. Springer, 1987.
doi: 10.1007/978-1-4612-4728-9 36

[23] Sidney Coleman. Aspects of Symmetry. Cambridge University Press, 1985.

doi: 10.1017/CBO9780511565045 37

[24] John Cardy. Scaling and Renormalization in Statistical Physics. Cambridge

University Press, 1996.
doi: 10.1017/CBO9781316036440 37

[25] Steven Weinberg. The Quantum Theory of Fields • Volume I: Foundations.

Cambridge University Press, 1995.
doi: 10.1017/CBO9781139644167 37 , 71 , 95

[26] Steven Weinberg. The Quantum Theory of Fields • Volume II: Modern
Applications. Cambridge University Press, 1996.
doi: 10.1017/CBO9781139644174 37 , 39

[27] Jean Zinn-Justin. Quantum Field Theory and Critical Phenomena. Oxford
University Press, 2021.
doi: 10.1093/oso/9780198834625.001.0001 37 , 39

[28] Michael E. Peskin and Dan V. Schroeder. An Introduction to Quantum

Field Theory. CRC Press, 2018.
doi: 10.1201/9780429503559 37

[29] Alexander A. Belavin, Alexander M. Polyakov, and Alexander B. Zamolod-

chikov. Infinite Conformal Symmetry in Two-Dimensional Quantum Field
Theory. Nuclear Physics B, 1984.
doi: 10.1016/0550-3213(84)90052-X 51

[30] Edward Witten. Quantum Field Theory and the Jones Polynomial. Commu-
nications in Mathematical Physics, 1989.
doi: 10.1007/BF01217730 51

[31] Juan Maldacena. The Large N Limit of Superconformal Field Theories and
Supergravity. Advances in Theoretical and Mathematical Physics, 1998.
doi: 10.4310/ATMP.1998.v2.n2.a1 arXiv: hep-th/9711200 [hep-th] 51

[32] David Simmons-Duffin. The Conformal Bootstrap. In: TASI 2015, New
Frontiers in Fields and Strings. World Scientific, 2017.
doi: 10.1142/9789813149441_0001 arXiv: 1602.07982 [hep-th] 52

[33] Marco Meineri and João Penedones. Conformal Field Theory and Gravity.
2019.
url: https://epfl.ch/labs/fsl/wp-content/uploads/2022/05/CFTandGravity.pdf
52 , 91

[34] David Poland, Slava Rychkov, and Alessandro Vichi. The Conformal Boot-
References 158

strap: Theory, Numerical Techniques, and Applications. Reviews of Modern

Physics, 2019.
doi: 10.1103/RevModPhys.91.015002 arXiv: 1805.04405 [hep-th]
52 , 85 , 88 , 89 , 90 , 114

[35] Philippe Francesco, Pierre Mathieu, and David Sénéchal. Conformal Field
Theory. Springer, 1996.
doi: 10.1007/978-1-4612-2256-9 52

[36] Slava Rychkov. EPFL Lectures on Conformal Field Theory in D ⩾ 3 Dimen-

sions. Springer, 2016.
doi: 10.1007/978-3-319-43626-5 arXiv: 1601.05000 [hep-th] 52 , 85

[37] Roger Penrose and Wolfgang Rindler. Spinors and Space-Time • Volume 2:
Spinor and Twistor Methods in Space-Time Geometry. Cambridge University
Press, 1986.
doi: 10.1017/CBO9780511524486 53

[38] Bert Schellekens. Conformal Field Theory. 2017.

url: https://www.nikhef.nl/~t58/CFT.pdf 55

[39] Martin Schottenloher. A Mathematical Introduction to Conformal Field

Theory. Springer, 2008.
doi: 10.1007/978-3-540-68628-6 56 , 63 , 66

[40] Carlos Batista. On Spaces with the Maximal Number of Conformal Killing
Vectors. 2017.
arXiv: 1711.01337 [gr-qc] 60

[41] T. Bröcker and K. Jänich. Introduction to Differential Topology. Cambridge

University Press, 1982.
url: https://cambridge.org/9780521284707 60

[42] Roger Penrose and Wolfgang Rindler. Spinors and Space-Time • Volume
1: Two-Spinor Calculus and Relativistic Fields. Cambridge University Press,
1984.
doi: 10.1017/CBO9780511564048 66 , 153

[43] Gerhard Mack and Abdus Salam. Finite-component field representations of

the conformal group. Annals of Physics, 1969.
doi: 10.1016/0003-4916(69)90278-4 72

[44] Per Kraus, Stathis Megas, and Allic Sivaramakrishnan. Anomalous dimen-
sions from thermal AdS partition functions. Journal of High Energy Physics,
2020.
doi: 10.1007/JHEP10(2020)149 arXiv: 2004.08635 [hep-th] 74

[45] João Penedones. Writing CFT correlation functions as AdS scattering ampli-
tudes. Journal of High Energy Physics, 2011.
doi: 10.1007/JHEP03(2011)025 arXiv: 1011.1485 [hep-th] 74 , 112
References 159

[46] A. Liam Fitzpatrick and Jared Kaplan. Unitarity and the Holographic
S-Matrix. Journal of High Energy Physics, 2012.
doi: 10.1007/JHEP10(2012)032 arXiv: 1112.4845 [hep-th] 74 , 119

[47] David Simmons-Duffin. TASI Lectures on Conformal Field Theory in Lorentz-

ian Signature. 2020.
GitLab: davidsd/lorentzian-cft-notes 85

[48] Duccio Pappadopulo, Slava Rychkov, Johnny Espin, and Riccardo Rattazzi.
OPE Convergence in Conformal Field Theory. Physical Review D, 2012.
doi: 10.1103/PhysRevD.86.105043 arXiv: 1208.6449 [hep-th] 86

[49] Matthijs Hogervorst and Slava Rychkov. Radial Coordinates for Conformal
Blocks. Physical Review D, 2013.
doi: 10.1103/PhysRevD.87.106004 arXiv: 1303.1111 [hep-th] 86

[50] João Penedones. High Energy Scattering in the AdS/CFT Correspondence.

PhD thesis. University of Porto, 2007.
arXiv: 0712.0802 [hep-th] 91 , 112

[51] Miguel S. Costa, Vasco Gonçalves, and João Penedones. Spinning AdS
propagators. Journal of High Energy Physics, 2014.
doi: 10.1007/JHEP09(2014)064 arXiv: 1404.5625 [hep-th] 91 , 108 , 113

[52] Martin Ammon and Johanna Erdmenger. Gauge/Gravity Duality: Founda-

tions and Applications. Cambridge University Press, 2015.
doi: 10.1017/CBO9780511846373 91

[53] Steven Weinberg. Gravitation and Cosmology • Principles and Applications

of the General Theory of Relativity. John Wiley and Sons, 1972.
url: https://wiley.com/Gravitation+and+Cosmology-p-9780471925675 91

[54] Lorenzo Di Pietro, Victor Gorbenko, and Shota Komatsu. Analyticity and
Unitarity for Cosmological Correlators. Journal of High Energy Physics,
2022.
doi: 10.1007/JHEP03(2022)023 arXiv: 2108.01695 [hep-th] 92

[55] S. J. Avis, Christopher J. Isham, and D. Storey. Quantum Field Theory in

Anti–de Sitter Space-Time. Physical Review D, 1978.
doi: 10.1103/PhysRevD.18.3565 95

[56] Peter Breitenlohner and Daniel Z. Freedman. Stability in Gauged Extended

Supergravity. Annals of Physics, 1982.
doi: 10.1016/0003-4916(82)90116-6 95 , 107

[57] Miguel F. Paulos, João Penedones, Jonathan Toledo, Balt C. van Rees,
and Pedro Vieira. The S-matrix bootstrap. Part I: QFT in AdS. Journal of
High Energy Physics, 2017.
doi: 10.1007/JHEP11(2017)133 arXiv: 1607.06109 [hep-th] 95 , 101
References 160

[58] Jonáš Dujava and Petr Vaško. Finite-coupling spectrum of O(N) model in
AdS. 2025.
arXiv: 2503.16345 [hep-th] GitHub: jdujava/ONinAdS 100

[59] Dean Carmi, Lorenzo Di Pietro, and Shota Komatsu. A Study of Quantum
Field Theories in AdS at Finite Coupling. Journal of High Energy Physics,
2019.
doi: 10.1007/JHEP01(2019)200 arXiv: 1810.04185 [hep-th]
100 , 101 , 102 , 104 , 107 , 110 , 112 , 123 , 125 , 126 , 127 , 128 , 132 , 135 , 148 , 149 , 151

[60] Ankur, Dean Carmi, and Lorenzo Di Pietro. Scalar QED in AdS. Journal
of High Energy Physics, 2023.
doi: 10.1007/JHEP10(2023)089 arXiv: 2306.05551 [hep-th] 100 , 101 , 128

[61] Junyu Liu, Eric Perlmutter, Vladimir Rosenhaus, and David Simmons-
Duffin. d-dimensional SYK, AdS Loops, and 6j Symbols. Journal of High
Energy Physics, 2019.
doi: 10.1007/JHEP03(2019)052 arXiv: 1808.00612 [hep-th]
101 , 110 , 113 , 115 , 116 , 117 , 118 , 120 , 121 , 142 , 148

[62] Matthijs Hogervorst, Marco Meineri, João Penedones, and Kamran Salehi
Vaziri. Hamiltonian truncation in Anti-de Sitter spacetime. Journal of High
Energy Physics, 2021.
doi: 10.1007/JHEP08(2021)063 arXiv: 2104.10689 [hep-th] 101

[63] António Antunes, Miguel S. Costa, João Penedones, Aaditya Salgarkar,

and Balt C. van Rees. Towards bootstrapping RG flows: sine-Gordon in AdS.
Journal of High Energy Physics, 2021.
doi: 10.1007/JHEP12(2021)094 arXiv: 2109.13261 [hep-th] 101

[64] António Antunes, Edoardo Lauria, and Balt C. van Rees. A bootstrap study
of minimal model deformations. Journal of High Energy Physics, 2024.
doi: 10.1007/JHEP05(2024)027 arXiv: 2401.06818 [hep-th] 101

[65] Nat Levine and Miguel F. Paulos. Bootstrapping bulk locality. Part I: Sum
rules for AdS form factors. Journal of High Energy Physics, 2024.
doi: 10.1007/JHEP01(2024)049 arXiv: 2305.07078 [hep-th] 101

[66] Marco Meineri, João Penedones, and Taro Spirig. Renormalization group
flows in AdS and the bootstrap program. Journal of High Energy Physics,
2024.
doi: 10.1007/JHEP07(2024)229 arXiv: 2305.11209 [hep-th] 101

[67] Juan Maldacena and Alexander Zhiboedov. Constraining Conformal Field

Theories with A Higher Spin Symmetry. Journal of Physics A, 2013.
doi: 10.1088/1751-8113/46/21/214011 arXiv: 1112.1016 [hep-th] 101

[68] Juan Maldacena and Alexander Zhiboedov. Constraining conformal field

References 161

theories with a slightly broken higher spin symmetry. Classical and Quantum
Gravity, 2013.
doi: 10.1088/0264-9381/30/10/104003 arXiv: 1204.3882 [hep-th] 101

[69] António Antunes, Nat Levine, and Marco Meineri. Demystifying integrable
QFTs in AdS: No-go theorems for higher-spin charges. 2025.
arXiv: 2502.06937 [hep-th] 101

[70] Gabriel Cuomo, Zohar Komargodski, and Márk Mezei. Localized magnetic
field in the O(N) model. Journal of High Energy Physics, 2022.
doi: 10.1007/JHEP02(2022)134 arXiv: 2112.10634 [hep-th] 102 , 103 , 151

[71] Balt C. van Rees and Xiang Zhao. Quantum Field Theory in AdS Space
instead of Lehmann-Symanzik-Zimmerman Axioms. Physical Review Letters,
2023.
doi: 10.1103/PhysRevLett.130.191601 arXiv: 2210.15683 [hep-th] 102

[72] Joseph Polchinski. S-Matrices from AdS Spacetime. 1999.

arXiv: hep-th/9901076 102

[73] Mirah Gary and Steven B. Giddings. The Flat space S-matrix from the
AdS/CFT correspondence? Physical Review D, 2009.
doi: 10.1103/PhysRevD.80.046008 arXiv: 0904.3544 [hep-th] 102

[74] Eliot Hijano. Flat space physics from AdS/CFT. Journal of High Energy
Physics, 2019.
doi: 10.1007/JHEP07(2019)132 arXiv: 1905.02729 [hep-th] 102

[75] Sarthak Duary, Eliot Hijano, and Milan Patra. Towards an IR finite S-matrix
in the flat limit of AdS/CFT. 2022.
arXiv: 2211.13711 [hep-th] 102

[76] Andrew Strominger. Lectures on the Infrared Structure of Gravity and Gauge
Theory. Princeton University Press, 2017.
arXiv: 1703.05448 [hep-th] 102

[77] Simone Giombi and Himanshu Khanchandani. CFT in AdS and boundary
RG flows. Journal of High Energy Physics, 2020.
doi: 10.1007/JHEP11(2020)118 arXiv: 2007.04955 [hep-th] 102 , 135 , 151

[78] Moshe Moshe and Jean Zinn-Justin. Quantum field theory in the large N
limit: A Review. Physics Reports, 2003.
doi: 10.1016/S0370-1573(03)00263-1 arXiv: hep-th/0306133 104

[79] Sidney R. Coleman, R. Jackiw, and H. David Politzer. Spontaneous Sym-

metry Breaking in the O(N) Model for Large N. Physical Review D, 1974.
doi: 10.1103/PhysRevD.10.2491 104

[80] Kenneth G. Wilson and Michael E. Fisher. Critical exponents in 3.99

dimensions. Physical Review Letters, 1972.
References 162

doi: 10.1103/PhysRevLett.28.240 106

[81] Lin Fei, Simone Giombi, and Igor R. Klebanov. Critical O(N) models in
6 − ε dimensions. Physical Review D, 2014.
doi: 10.1103/PhysRevD.90.025018 arXiv: 1404.1094 [hep-th] 106

[82] Simone Giombi, Richard Huang, Igor R. Klebanov, Silviu S. Pufu, and
Grigory Tarnopolsky. The O(N) Model in 4 < d < 6 : Instantons and
complex CFTs. Physical Review D, 2020.
doi: 10.1103/PhysRevD.101.045013 arXiv: 1910.02462 [hep-th] 107

[83] Victor Gorbenko, Slava Rychkov, and Bernardo Zan. Walking, Weak first-
order transitions, and Complex CFTs. Journal of High Energy Physics,
2018.
doi: 10.1007/JHEP10(2018)108 arXiv: 1807.11512 [hep-th] 107

[84] Igor R. Klebanov and Edward Witten. AdS / CFT correspondence and
symmetry breaking. Nuclear Physics B, 1999.
doi: 10.1016/S0550-3213(99)00387-9 arXiv: hep-th/9905104 107

[85] Michael Dütsch and Karl-Henning Rehren. A Comment on the dual field
in the scalar AdS / CFT correspondence. Letters in Mathematical Physics,
2002.
doi: 10.1023/A:1021601215141 arXiv: hep-th/0204123 108

[86] Edward Witten. Anti–de Sitter Space and Holography. Advances in Theo-
retical and Mathematical Physics, 1998.
doi: 10.4310/ATMP.1998.v2.n2.a2 arXiv: hep-th/9802150 [hep-th] 108

[87] Denis Karateev, Petr Kravchuk, and David Simmons-Duffin. Harmonic

Analysis and Mean Field Theory. Journal of High Energy Physics, 2019.
doi: 10.1007/JHEP10(2019)217 arXiv: 1809.05111 [hep-th] 110 , 119

[88] Igor Bertan and Ivo Sachs. Loops in Anti–de Sitter Space. Physical Review
Letters, 2018.
doi: 10.1103/PhysRevLett.121.101601 arXiv: 1804.01880 [hep-th] 110

[89] Igor Bertan, Ivo Sachs, and Evgeny D. Skvortsov. Quantum φ4 Theory in
AdS4 and its CFT Dual. Journal of High Energy Physics, 2019.
doi: 10.1007/JHEP02(2019)099 arXiv: 1810.00907 [hep-th] 110

[90] Máximo Bañados, Ernesto Bianchi, Iván Muñoz, and Kostas Skenderis.
Bulk renormalization and the AdS/CFT correspondence. Physical Review D,
2023.
doi: 10.1103/PhysRevD.107.L021901 arXiv: 2208.11539 [hep-th] 110

[91] A. Liam Fitzpatrick, Jared Kaplan, João Penedones, Suvrat Raju, and
Balt C. van Rees. A Natural Language for AdS/CFT Correlators. Journal of
High Energy Physics, 2011.
doi: 10.1007/JHEP11(2011)095 arXiv: 1107.1499 [hep-th] 113
References 163

[92] Gerhard Mack. Group Theoretical Approach to Conformal Invariant Quantum

Field Theory. In: Renormalization and Invariance in Quantum Field Theory.
ed. by Eduardo R. Caianiello. Springer, 1974.
doi: 10.1007/978-1-4615-8909-9_7 113

[93] Vladimir K. Dobrev, Gerhard Mack, Valentina B. Petkova, Svetlana G.

Petrova, and Ivan T. Todorov. Harmonic Analysis • On the n-Dimensional
Lorentz Group and Its Application to Conformal Quantum Field Theory.
Springer, 1977.
doi: 10.1007/BFb0009678 113 , 115

[94] Francis A. Dolan and Hugh Osborn. Conformal Partial Waves: Further
Mathematical Results. 2011.
arXiv: 1108.6194 [hep-th] 113

[95] David Simmons-Duffin. Projectors, Shadows, and Conformal Blocks. Journal

of High Energy Physics, 2014.
doi: 10.1007/JHEP04(2014)146 arXiv: 1204.3894 [hep-th] 113

[96] David Simmons-Duffin, Douglas Stanford, and Edward Witten. A spacetime

derivation of the Lorentzian OPE inversion formula. Journal of High Energy
Physics, 2018.
doi: 10.1007/JHEP07(2018)085 arXiv: 1711.03816 [hep-th]
113 , 115 , 116 , 117 , 118 , 119 , 121

[97] Simon Caron-Huot. Analyticity in Spin in Conformal Theories. Journal of

High Energy Physics, 2017.
doi: 10.1007/JHEP09(2017)078 arXiv: 1703.00278 [hep-th] 118 , 120 , 140

[98] Petr Kravchuk and David Simmons-Duffin. Light-Ray Operators in Confor-

mal Field Theory. Journal of High Energy Physics, 2018.
doi: 10.1007/JHEP11(2018)102 arXiv: 1805.00098 [hep-th] 118 , 121

[99] Jacques Bros, Henri Epstein, Michel Gaudin, Ugo Moschella, and Vin-
cent Pasquier. Anti de Sitter Quantum Field Theory and a New Class of
Hypergeometric Identities. Communications in Mathematical Physics, 2012.
doi: 10.1007/s00220-011-1372-0 arXiv: 1107.5161 [hep-th] 128

[100] Sergio L. Cacciatori, Henri Epstein, and Ugo Moschella. Loops in anti de
Sitter space. Journal of High Energy Physics, 2024.
doi: 10.1007/JHEP08(2024)109 arXiv: 2403.13142 [hep-th] 128

[101] David M. McAvity and Hugh Osborn. Conformal field theories near a
boundary in general dimensions. Nuclear Physics B, 1995.
doi: 10.1016/0550-3213(95)00476-9 arXiv: cond-mat/9505127 133

[102] Max A. Metlitski. Boundary criticality of the O(N) model in d = 3 critically

revisited. SciPost Physics, 2022.
References 164

doi: 10.21468/SciPostPhys.12.4.131 arXiv: 2009.05119 [cond-mat.str-el]

135

[103] Jaychandran Padayasi, Abijith Krishnan, Max A. Metlitski, Ilya A. Gruz-

berg, and Marco Meineri. The extraordinary boundary transition in the 3d
O(N) model via conformal bootstrap. SciPost Physics, 2022.
doi: 10.21468/SciPostPhys.12.6.190
arXiv: 2111.03071 [cond-mat.stat-mech] 135

[104] Pedro Liendo, Leonardo Rastelli, and Balt C. van Rees. The Bootstrap
Program for Boundary CFTd . Journal of High Energy Physics, 2013.
doi: 10.1007/JHEP07(2013)113 arXiv: 1210.4258 [hep-th] 135

[105] Miguel S. Costa, Vasco Gonçalves, and João Penedones. Conformal Regge
theory. Journal of High Energy Physics, 2012.
doi: 10.1007/JHEP12(2012)091 arXiv: 1209.4355 [hep-th] 140

[106] Zohar Komargodski and Alexander Zhiboedov. Convexity and Liberation at

Large Spin. Journal of High Energy Physics, 2013.
doi: 10.1007/JHEP11(2013)140 arXiv: 1212.4103 [hep-th] 140 , 141 , 144

[107] A. Liam Fitzpatrick, Jared Kaplan, David Poland, and David Simmons-
Duffin. The Analytic Bootstrap and AdS Superhorizon Locality. Journal of
High Energy Physics, 2013.
doi: 10.1007/JHEP12(2013)004 arXiv: 1212.3616 [hep-th] 140 , 141 , 144

[108] Luis F. Alday. Large Spin Perturbation Theory for Conformal Field Theories.
Physical Review Letters, 2017.
doi: 10.1103/PhysRevLett.119.111601 arXiv: 1611.01500 [hep-th] 141

[109] Sridip Pal, Jiaxin Qiao, and Slava Rychkov. Twist Accumulation in Confor-
mal Field Theory: A Rigorous Approach to the Lightcone Bootstrap. Commu-
nications in Mathematical Physics, 2023.
doi: 10.1007/s00220-023-04767-w arXiv: 2212.04893 [hep-th] 141

[110] Balt C. van Rees. Theorems for the Lightcone Bootstrap. 2024.
arXiv: 2412.06907 [hep-th] 141

[111] Lorenzo Di Pietro, Victor Gorbenko, and Shota Komatsu. Cosmological

Correlators at Finite Coupling. 2023.
arXiv: 2312.17195 [hep-th] 151

[112] Pavel Krtouš. Relationships between Scalar Field and Relativistic Particle
Quantizations. PhD thesis. University of Alberta, 1997.
url: https://utf.mff.cuni.cz/~krtous/papers/THESIS.pdf 153