Henriette Elvang, Yu-Tin Huang - Scattering Amplitudes in Gauge Theory and Gravity-Cambridge University Press (2015) - 1

Scattering Amplitudes in Gauge Theory and Gravity
Providing a comprehensive, pedagogical introduction to scattering amplitudes in gauge

theory and gravity, this book is ideal for graduate students and researchers. It offers a
smooth transition from basic knowledge of quantum field theory to the frontier of modern
research.
Building on basic quantum field theory, the book starts with an introduction to the spinor
helicity formalism in the context of Feynman rules for tree-level amplitudes. The material
covered includes on-shell recursion relations, superamplitudes, symmetries of N = 4 super
Yang–Mills theory, twistors and momentum twistors, Grassmannians, and polytopes. The
presentation also covers amplitudes in perturbative supergravity, 3d Chern–Simons matter
theories, and color-kinematics duality and its connection to “gravity = (gauge theory)2 .”
Basic knowledge of Feynman rules in scalar field theory and quantum electrodynamics
is assumed, but all other tools are introduced as needed. Worked examples demonstrate the
techniques discussed, and over 150 exercises help readers absorb and master the material.
Henriette Elvang is Associate Professor in the Department of Physics, University of Michigan.

She has worked on various aspects of high energy theoretical physics, including black holes
in string theory, scattering amplitudes, and the structure of gauge theories.
Yu-tin Huang is Assistant Professor at National Taiwan University. He is known for his work in
the study of scattering amplitudes beyond four dimensions, most notably in 3-dimensional
Chern–Simons matter theory.
“In recent years, a series of surprising insights and new methods have transformed the
understanding of gauge and gravitational scattering amplitudes. These advances are im-
portant both for practical calculations in particle physics, and for the fundamental structure
of relativistic quantum theory. Elvang and Huang have written the first comprehensive
text on this subject, and their clear and pedagogical approach will make these new ideas
accessible to a wide range of students.”
Joseph Polchinski, University of California, Santa Barbara
“This book provides a much needed text covering modern techniques that have given
radical new insights into the structure of quantum field theory. It gathers together a very
large body of recent literature and presents it in a coherent style. The book should appeal to
the wide body of researchers who wish to use quantum field theory as a tool for describing
physical phenomena or who are intending to gain insight by studying its mathematical
structure.”
Michael B. Green, University of Cambridge
Scattering Amplitudes in
Gauge Theory and Gravity
HENRIETTE ELVANG
University of Michigan, Ann Arbor
YU-TIN HUANG
National Taiwan University
University Printing House, Cambridge CB2 8BS, United Kingdom
Cambridge University Press is a part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of
education, learning, and research at the highest international levels of excellence.
www.cambridge.org
Information on this title: www.cambridge.org/9781107069251

C H. Elvang and Y. Huang 2015
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2015
Printed in the United States of America by Edwards Brothers Malloy
A catalog record for this publication is available from the British Library
Library of Congress Cataloging in Publication data
Elvang, Henriette, 1976– author.
Scattering amplitudes in gauge theory and gravity / Henriette Elvang, University of Michigan,
Ann Arbor, Yu-tin Huang, Institute for Advanced Study and National Taiwan University.
pages cm
Includes bibliographical references and index.
ISBN 978-1-107-06925-1 (hardback)
1. Scattering amplitude (Nuclear physics) 2. Gauge fields (Physics) – Mathematics.
I. Huang, Yu-tin, author. II. Title.
QC20.7.S3E48 2015
539.7 58 – dc23 2014032234
ISBN 978-1-107-06925-1 Hardback
Cambridge University Press has no responsibility for the persistence or accuracy of
URLs for external or third-party internet websites referred to in this publication,
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
For mom, dad, and Coco.
Yu-tin
Thank you, Mormor.

Henriette
Contents
Preface page xi
1 Introduction 1
Part I Trees
2 Spinor helicity formalism 15

2.1 Dirac spinors 15
2.2 Spinor helicity notation 16
2.3 Examples from Yukawa theory 19
2.4 Massless vectors and examples from QED 25
2.5 Yang–Mills theory, QCD, and color-ordering 30
2.6 Little group scaling 37
2.7 MHV classification 41
2.8 Analytic properties of tree amplitudes 44
3 On-shell recursion relations at tree-level 50

3.1 Complex shifts and Cauchy’s theorem 50
3.2 BCFW recursion relations 53
3.3 When does it work? 62
3.4 MHV vertex expansion (CSW) 65
4 Supersymmetry 69
4.1 N = 1 supersymmetry: chiral model 69
4.2 Amplitudes and supersymmetry Ward identities 73
4.3 N = 1 supersymmetry: gauge theory 75
4.4 N = 4 SYM: on-shell superspace and superamplitudes 77
4.5 Super-BCFW and all tree-level amplitudes in N = 4 SYM 86
4.5.1 MHV superamplitude from super-BCFW 87
4.5.2 NMHV superamplitude and beyond 89
5 Symmetries of N = 4 SYM 95
5.1 Superconformal symmetry of N = 4 SYM 95
5.2 Twistors 99
5.3 Emergence of dual conformal symmetry 102
5.4 Momentum twistors 107
viii Contents
Part II Loops
6 Loop amplitudes and generalized unitarity 117

6.1 UV and IR divergences 119
6.2 Unitarity method 124
6.3 1-loop amplitudes from unitarity 126
6.4 1-loop amplitudes in planar N = 4 SYM 132
6.5 Higher-loop amplitudes in planar N = 4 SYM 135
7 BCFW recursion for loops 138

7.1 Loop-integrands 138
7.2 BCFW shift in momentum twistor space 141
7.3 Momentum twistor BCFW at tree-level 145
7.4 Momentum twistor BCFW for planar loop-integrands 149
7.5 Example: 4-point 1-loop amplitude from recursion 156
7.6 Higher loops 159
8 Leading Singularities and on-shell diagrams 161

8.1 1-loop Leading Singularities 162
8.3 On-shell diagrams 173
Part III Topics
9 Grassmannia 181
9.1 Yangian invariance and cyclic symmetry 182
9.2 The Grassmannian 185
9.3 Yangian invariants as residues in the Grassmannian 188
9.3.1 MHV amplitudes 188
9.3.2 6-point NMHV amplitudes 190
9.4 From on-shell diagrams to the Grassmannian 194
10 Polytopes 200
10.1 Volume of an n-simplex in CPn 202
10.2 NMHV tree superamplitude as the volume of a polytope 206
10.3 The boundary of simplices and polytopes 209
10.4 Geometric aftermath 213
11 Amplitudes in dimensions D = 4 216

11.1 Helicity formalism in D = 4 216
11.2 Scattering amplitudes in D = 6 219
11.3.1 D = 3 kinematics 222
Contents ix
11.3.2 3d SYM and Chern–Simons theory 224

11.3.3 Special kinematics and poles in amplitudes 225
11.3.4 D = 3 superconformal algebra 226
11.3.5 N = 8 superconformal theory: BLG 228
11.3.6 N = 6 superconformal theory: ABJM 230
11.3.7 BCFW recursion in 3d 233
11.3.8 ABJM and dual conformal symmetry 238
11.3.9 Loops and on-shell diagrams in ABJM 239
11.3.10 The orthogonal Grassmannian 245
12 Supergravity amplitudes 249

12.1 Perturbative gravity 249
12.2 Supergravity 254
12.3 Superamplitudes in N = 8 supergravity 256
12.4 Loop amplitudes in supergravity 259
12.5 N = 8 supergravity: loops and counterterms 261
12.6 Supergravity divergences for various N , L, and D 269
13 A colorful duality 272

13.1 The color-structure of Yang–Mills theory 273
13.2 Color-kinematics duality: BCJ, the tree-level story 276
13.3 Color-kinematics duality: BCJ, the loop-level story 283
13.3.1 1-loop 4-point N = 4 SYM 284
13.3.2 2-loop 4-point N = 4 SYM 285
13.3.3 3-loop 4-point N = 4 SYM 286
13.3.4 Summary 288
13.4 Implications for UV behavior of supergravity 288
13.5 Extensions 293
14 Further reading 294
Appendix Conventions for 4d spinor helicity formalism 297
References 299
Index 319
Preface
This book grew out of a need to have a set of easily accessible notes that introduced the
basic techniques used in modern research on scattering amplitudes. In addition to the key
tools, such a review should collect some of the small results and intuitions the authors had
acquired from their work in the field and which had not previously been exposed in the
literature. As the authors quickly realized, such an introduction would bring the reader only
part of the way towards some of the most exciting topics in the field, so they decided to
add “a little extra” material. While doing so – and this took quite a while – the authors
remained in full and complete denial about writing a book. It was only at the end of
the process that they faced their worst fears: the review was becoming a book. You now
hold the result in your hands. Because the authors were not writing a book, they actually
thoroughly enjoyed the work. Their hope is that you will enjoy it too and that you will find it
useful.
It is a pleasure to thank our friends and collaborators who have worked with us and helped
us learn the subject of scattering amplitudes: Ratin Akhoury, Nima Arkani-Hamed, Zvi
Bern, Freddy Cachazo, John Joseph Carrasco, Simon Caron-Huot, Tim Cohen, Scott Davies,
Tristan Dennen, Lance Dixon, Dan Freedman, David Kosower, Johannes Henn, Harald Ita,
Henrik Johansson, Michael Kiermaier, Sangmin Lee, Arthur Lipstein, Thomas Lam, David
McGady, Timothy Olson, Cheng Peng, Jan Plefka, Radu Roiban, Mark Srednicki, Warren
Siegel, David Speyer, and Jaroslav Trnka. H.E. is grateful for the hospitality offered by
Stanford/SLAC during her visit in February/March 2013 and KITP/UCSB during January–
March 2014.
A few people have suffered early drafts of this book and we are indebted to them for their
helpful comments/suggestions/corrections: Cindy Keeler, Timothy Olson, Sam Roland,
David Speyer, Sri Suresh, Jonathan Walsh, and John Ware. Their careful readings caught
multiple typos and helped us improve the presentation. We would also like to thank Michael
Enciso, Karol Kampf, Joe Minahan, Stephen Naculich, and especially Stefan Theisen as
well as the Harvard amplitude reading group (communications from Marat Freytsis) for
detailed feedback on the manuscript.
Y.-t.H. would especially like to thank Warren Siegel, Peter van Nieuwenhuizen, and Zvi
Bern for their continuing support and guidance without which his career would have taken
a much different path. H.E. thanks Gary Horowitz, Dan Freedman, and Joe Polchinski for
the education, support, and friendship for which she is deeply grateful. She would also like
to thank her colleagues at the University of Michigan for their support for this project. Both
xii Preface
authors have benefited tremendously from amplitude discussions with Zvi Bern, Lance
Dixon, and Nima Arkani-Hamed.
The authors would also like to express their thanks to Cambridge University Press editor
Simon Capelin for his support, enthusiasm, and patience throughout the project.
Yu-tin Huang and Henriette Elvang

Introduction 1
Scattering experiments are crucial for our understanding of the building blocks of nature.
The standard model of particle physics was developed from scattering experiments, in-
cluding the discovery of the weak force bosons W ± and Z 0 , quarks and gluons, and most
recently the Higgs boson.
The key observable measured in particle scattering experiments is the scattering cross-
section σ . It encodes the likelihood of a given process to take place as a function of the
energy and momentum of the particles involved. A more refined version of this quantity is
the differential cross-section dσ/d: it describes the dependence of the cross-section on
the angles of the scattered particles.
Interpretation of data from scattering experiments relies heavily on theoretical predic-
tions of scattering cross-sections. These are calculated in relativistic quantum field theory
(QFT), which is the mathematical language for describing elementary particles and their
interactions. Relativistic QFT combines special relativity with quantum physics and is
a hugely successful and experimentally well-tested framework for describing elementary
particles and the fundamental forces of nature. In quantum mechanics, the probability
distribution |ψ|2 = ψ ∗ ψ for a particle is given by the norm-squared of its complex-
valued wavefunction ψ. Analogously, in quantum field theory, the differential cross-
section is proportional to the norm-squared of the scattering amplitude A, dσ/d ∝ |A|2 .
The amplitudes A are well-defined physical observables: they are the subject of this
book.
Scattering amplitudes have physical relevance through their role in the scattering cross-
section. Moreover, it has been realized in recent years that amplitudes themselves have
a very interesting mathematical structure. Understanding this structure guides us towards
more efficient methods to calculate amplitudes. It also makes it exciting to study scattering
amplitudes in their own right and explore (and exploit) their connections to interesting
branches of mathematics, including combinatorics and geometry.
The purpose of this book is two-fold. First, we wish to provide a pedagogical introduction
to the efficient modern methods for calculating scattering amplitudes. Second, we survey
several interesting mathematical properties of amplitudes and recent research advances.
The hope is to bridge the gap between a standard graduate course in quantum field theory
and the modern approach to scattering amplitudes as well as current research. Thus we
assume readers to be familiar with the basics of quantum field theory and particle physics;
however, here in the first chapter we offer an introduction that may also be helpful for a
broader audience with an interest in the subject.
2 Introduction
What is a scattering amplitude?

Let us consider a few examples of scattering processes involving electrons e− , positrons
e+ , and photons γ :
Compton scattering e− + γ → e− + γ ,
Møller scattering e− + e− → e− + e− ,
(1.1)
Bhabha scattering e− + e+ → e− + e+ ,
e− e+ annihilation e− + e+ → γ + γ .
These processes are described in the quantum field theory that couples Maxwell’s elec-
tromagnetism to electrons and positrons, namely quantum electrodynamics (QED). Each
process is characterized by the types of particles involved in the initial and final states
as well as the relativistic momentum p and energy E of each particle. This input is the
external data for an amplitude: a scattering amplitude An involving a total of n initial
and final state particles takes the list {E i , pi ; typei }, i = 1, 2, . . . , n, of external data and
returns a complex number:

n-particle amplitude An : E i , pi ; typei → An E i , pi ; typei ∈ C. (1.2)
“Particle type” involves more than saying which particles scatter: it also includes a speci-
fication of the appropriate quantum numbers of the initial and final states, for example the
polarization of a photon or the spin-state of a fermion.
In special relativity, the energy E and momentum p of a physical particle must satisfy
the relationship E 2 = | p|2 c2 + m 2 c4 with m the rest mass of the particle. This is simply
the statement that E 2 − | p|2 c2 is a Lorentz-invariant quantity in Minkowski space. Thus,
one of the constraints on the external data for an n-particle amplitude is that the relativistic
on-shell condition,
E i2 − | pi |2 c2 = m i2 c4 , (1.3)
holds for all particles i = 1, . . . , n. The initial and final state electrons and positrons in the
QED processes (1.1) must satisfy the on-shell condition (1.3) with m i equal to the electron
mass, m e = 511 keV/c2 , while m i is zero for the photon since it is massless.
It is convenient to work in “natural units” where the speed of light c and Planck’s con-
stant h̄ = h/2π are set to unity: c = 1 = h̄. Proper units can always be restored by dimen-
sional analysis. We combine the energy and momentum into a 4-momentum piμ = (E i , pi ),
with μ = 0, 1, 2, 3, and write pi2 = −E i2 + | pi |2 such that the on-shell condition (1.3)
becomes
pi2 = −m i2 . (1.4)
1 Introduction 3
μ
Conservation of relativistic energy and momentum requires the sum of initial momenta pin
μ
to equal the sum of final state momenta pout . It is convenient to flip the signs of all incoming
momenta so that the conservation of 4-momentum simply reads

n
piμ = 0 . (1.5)
i=1
Thus, to summarize, the external data for the amplitude involve a specification of 4-
μ
momentum and particle type for each external particle, { pi , typei }, subject to the on-shell
constraints pi2 = −m i2 and momentum conservation (1.5). We often use the phrase on-shell
amplitude to emphasize that the external data satisfy the kinematic constraints (1.4) and
(1.5) and include the appropriate polarization vectors or fermion spin wavefunctions.
Feynman diagrams
Scattering amplitudes are typically calculated as a perturbation series in the expansion of
a small dimensionless parameter that encodes the coupling of interactions between the
particles. For QED processes, this dimensionless coupling is the fine structure constant1
α ≈ 1/137. In the late 1940s, Feynman introduced a diagrammatic way to organize the
perturbative calculation of the scattering amplitudes. It expresses the n-particle scattering
amplitude An as a sum of all possible Feynman diagrams with n external legs:

An = (Feynman diagrams) = + + + · · · (1.6)
tree-level 1-loop 2-loop
The sum of diagrams is organized by the number of closed loops. For a given number
of particles n, the loop-diagrams are suppressed by powers of the coupling, e.g. in QED
an L-loop diagram is order α 2L compared with the tree-level. Hence the loop-expansion
is a diagrammatic representation of the perturbation expansion. The leading contribution
comes from the tree-diagrams; their sum is called the tree-level amplitude, Atree
n . The next
1-loop
order is 1-loop; the sum of the 1-loop diagrams is the 1-loop amplitude An , etc. The
full amplitude is then
An = Atree
n + An
1-loop
+ A2-loop
n + ··· (1.7)
It is rare that the loop-expansion is convergent; typically, the number of diagrams grows so
quickly at higher-loop order that it overcomes the suppression from the higher powers in the
small coupling. Amplitudes can be Borel summable, but this is not a subject we treat here.
Instead, we pursue an understanding of the contributions to the amplitude order-by-order
in perturbation theory.
1 In SI units, the fine structure constant is α = e2 /(4π 0h̄c), where 0 is the vacuum permittivity.
4 Introduction
The quantum field theories we are concerned with in this book are defined by Lagrangians
that encode the particles and how they interact. The Lagrangian for QED determines a
simple basic 3-particle interaction between electrons/positrons and the photon:
. (1.8)
From this, one can build diagrams such as
, , and . (1.9)
Reading the diagrams left to right, the first diagram describes the absorption and subsequent
emission of a photon (wavy line) by an electron (solid line), and as such it contributes at the
leading order in perturbation theory to the Compton scattering process e− + γ → e− + γ .
The second and third diagrams in (1.9) are the two tree-level diagrams that give the leading-
order contribution to the Bhabha scattering process e− + e+ → e− + e+ . The first of those
two diagrams encodes the annihilation of an electron and a positron to a photon and the
subsequent e+ e− pair-creation. The second diagram is the exchange of a photon in the
scattering of an electron and a positron.
A Feynman diagram is translated to a mathematical expression via Feynman rules. These
rules are specific to the particle types and the theory that describes their interactions. A
Feynman graph has the following essential parts:
• External lines: the Feynman diagram has an external line for each initial and final state
particle. The rule for the external line depends on the particle type. For spin-1 particles,
such as photons and gluons, the external line rule encodes the polarizations. For fermions,
it contains information about the spin state.
• Momentum labels: every line in the Feynman diagram is associated with a momentum. For
the external lines, the momentum is fixed by the external data. Momentum conservation
is enforced at any vertex. For a tree-graph, this fully determines the momenta on all
internal lines. For an L-loop graph, L momenta are undetermined and the Feynman rules
state that one must integrate over all possible values of these L momenta.
• Vertices: the vertices describe the interactions among the particles in the theory. Vertices
can in principle have any number of lines going in or out, but in many theories there
are just cubic and/or quartic vertices. The Feynman rules translate each vertex into a
mathematical rule; in the simplest case, this is just multiplication by a constant, but more
generally the vertex rule can involve the momenta associated with the lines going in or
out of the vertex.
• Internal lines: the Feynman rules translate every internal line into a “propagator” that
depends on the momentum associated with the internal line. (A propagator is the Fourier
transform of a Green’s function.) The mathematical expression depends on the type of
internal line, i.e. what kind of virtual particle is being exchanged.
1 Introduction 5
One converts a Feynman diagram to a mathematical expression by tracing through the

diagram and picking up factors, in successive order of appearance, from the rules for
L-loop
external and internal lines and vertices. The Feynman rules then state that i An is
the sum of all possible L-loop diagrams with n external lines specified by the external
data.
Example. Consider the simplest case of a massless spin-0 scalar particle. The Feynman
rules for external scalar lines are simply a factor of 1 while an internal line with
momentum label P contributes −i/P 2 ; this is the massless scalar propagator. Let us
assume that our scalars only interact via cubic vertices for which the Feynman rule is
simply to multiply a factor of ig, where g is the coupling. A process involving four
scalars then has the tree-level contribution
p p
p2 p3 p p 2 3
2 3
4 (1, 2, 3, 4) =
Atree + +
p1 p p p
4 1 4 p p
1 4

1 1 1
= g2 + + . (1.10)
( p 1 + p 2 )2 ( p 1 + p 3 )2 ( p1 + p4 ) 2
Here ( p1 + p2 )2 = −( p10 + p20 )2 + | p1 + p2 |2 etc.

Note that the amplitude (1.10) is symmetric under exchange of identical bosonic external
states. This is a manifestation of Bose symmetry. Similarly, Fermi statistics requires that an
amplitude must be antisymmetric under exchange of any two identical external fermions.
Example. A 1-loop diagram in our simple scalar model takes the form
p2 p3

d 4 1
= g4

. (1.11)
(2π )4 2 − p1 2 − p1 − p2 2 + p4 2
p1 p4
We have used momentum conservation p1 + p2 + p3 + p4 = 0 to simplify the fourth

propagator. The 1-loop 4-scalar amplitude in this model is obtained from the sum over
inequivalent box diagrams (1.11).
Loop-integrals can be divergent both in the large- and small-momentum regimes. Such
“ultraviolet” and “infrared divergences” are well-understood and typically treated using a
scheme such as dimensional regularization in which one replaces the measure d 4 in the
loop-integral by d D with D = 4 − 2. The divergences can then be cleanly extracted in the
expansion of small . The treatment of ultraviolet divergences is called “renormalization”;
it is a well-established method and the resulting predictions for scattering processes have
been tested to high precision in particle physics experiments.
6 Introduction
Example in QED
As an example of a tree-level process in QED, consider the annihilation of an electron–
positron pair to two photons:
γ
θ
e− + e+ → γ + γ e– e +. (1.12)
γ
The leading-order contribution to the amplitude is a sum of two tree-level Feynman diagrams
constructed from the basic QED vertex (1.8). It is

−
i Atree
4 e + e+ → γ + γ = + . (1.13)
The two diagrams (1.13) are translated via the QED Feynman rules to Lorentz-invariant
contractions of 4-momentum vectors, photon polarizations, and fermion wavefunctions. In
the standard formulation, the expression for the amplitude is not particularly illuminating.
However, squaring it and subjecting it to a significant dose of index massage therapy, one
finds that the scattering cross-section takes a rather simple form. In the high-energy limit,
where the center of mass energy dominates the electron/positron rest mass, E CM m e c2 ,
the result for the differential cross-section is
2 ECM m e c2
A4 −−−−−−→ α 2 1 1 + cos θ + O(α 4 ) .
2
dσ 1
= 2 2 1 − cos2 θ
(1.14)
d 64π 2 E CM E CM
The sum indicates a spin-sum average. The cross-section (1.14) depends on the center of
mass energy E CM , the scattering angle θ indicated in (1.12), and the fine structure constant
α. (In natural units, α = e2 /4π .) The result (1.14) serves to illustrate the salient properties
of the differential cross-section: the dependence on particle energies and scattering angles
as well as powers of a small fundamental dimensionless coupling constant α. In particular,
the expression in (1.14) is the leading-order contribution to the scattering process, and
higher orders, starting at 1-loop level, are indicated by O(α 4 ).
Since the focus in this book is on scattering amplitudes, and not the cross-sections,
it is worthwhile to present an expression for the amplitude of the annihilation process
e− + e+ → γ + γ . It turns out to be particularly simple in the high-energy regime E CM
m e c2 , where the masses can be neglected and the momentum vectors of the incoming
electron and positron can be chosen to be light-like (i.e. null, pi2 = 0). In that case, the sum
of the two diagrams (1.13) reduces to a single term that can be written compactly as

− +
2 24
2
Atree e + e → γ + γ = 2e . (1.15)
4
1323
Here, the angle brackets i j are closely related to the particle 4-momenta via |i j|2 ∼
μ
2 pi . p j , where the Lorentz-invariant dot-product is pi . p j ≡ pi p jμ = − pi0 p 0j + pi . p j . The
1 Introduction 7
4-momenta p1 and p2 are associated with the incoming electron and positron while p3 and
p4 are the 4-momenta of the photons. The angle bracket notation i j is part of the spinor
helicity formalism which is a powerful technical tool for describing scattering of massless
particles in four spacetime dimensions. We will introduce the spinor helicity formalism in
Chapter 2 and derive the expression (1.15) in full detail from the Feynman rules of QED.
You may be surprised to note that the expression (1.15) is not symmetric in the momenta
of the two photons. This is because the representation (1.15) selects distinct polarizations
for the photons.
Using momentum conservation and the on-shell condition pi2 = 0, one can show that
the norm-squared of the expression (1.15) is proportional to ( p1 . p3 )/( p1 . p4 ). The sum in
the cross-section (1.14) indicates a sum of the polarizations of the final state photons and
an average over the spins of the initial state electron and positron. This procedure yields a
second term with p3 ↔ p4 . Hence, what goes into the formula (1.14) is

Atree e− + e+ → γ + γ 2 = 2e4 p1 . p3 + p1 . p4 . (1.16)
4
p1 . p4 p1 . p3
μ μ
An explicit representation of the 4-momenta is p1 = (E, 0, 0, E) and p2 = (E, 0, 0, −E)
μ μ
for the initial states and p3 = (E, 0, E sin θ, E cos θ ) and p4 = (E, 0, −E sin θ, −E cos θ )
for the final state photons. Momentum conservation is simply p1 + p2 = p3 + p4 . Using
2
that the center of mass energy is E CM = −( p1 + p2 )2 = 4E 2 , one finds that the high-energy
result for the differential cross-section in (1.14) follows from (1.16).
It is typical that amplitudes involving only massless particles are remarkably simpler
than those with massive particles. This can be regarded as a high-energy regime, as in our
example above. Because of the simplicity, yet remarkably rich mathematical structure, the
focus of many developments in the field has been on scattering amplitudes of massless
particles.
The S-matrix
We have introduced the scattering amplitude as a function that takes as its input the
μ
constrained external data { pi , typei } and produces a complex number that is traditionally
calculated in terms of Feynman diagrams. The scattering process can also be considered
as an operation that maps an initial state |i to a final state | f , each being a collection of
single-particle states characterized by momenta and particle types. The scattering matrix S,
also called the S-matrix, is the unitary operator that “maps” the initial states to final states.
In other words, the probability of an initial state |i changing into a final state | f is given
by | f |S|i|2 . Separating out the trivial part of the scattering process where no scattering
occurs, we write
S = 1 + iT . (1.17)
Then the amplitude is simply A = f |T |i and this is the quantity that is calculated by
Feynman diagrams. Solving the S-matrix for a given theory means having a way to generate
all scattering amplitudes at any order in perturbation theory.
8 Introduction
Beyond Feynman diagrams

Armed with Feynman rules, we can, in principle, compute scattering amplitudes to our
hearts’ content. For instance, starting from the QED Lagrangian one can calculate the tree-
level differential cross-section for the processes (1.1). It is typical for such a calculation
that the starting point – the Lagrangian in its most compact form – is not too terribly
complicated. And the final result for dσ/d can be rather compact and simple too, as
we have seen in (1.14). But the intermediate stages of the calculation often explode in an
inferno of indices, contracted up-and-down and in all directions – providing little insight
into the physics and hardly any hint of simplicity.
In a QFT course, students are exposed (hopefully!) to a lot of long character-building
calculations, including the 4-particle QED processes in (1.1). But students are rarely asked
to use standard Feynman rules to calculate processes that involve more than four or five
particles, even at tree-level: for example, e− + e+ → e− + e+ + γ or e− + e+ → e− +
e+ + γ + γ . Why not? Well, one reason is that the number of Feynman diagrams tends to
grow very quickly with the number of particles involved: for gluon scattering at tree-level
in quantum chromodynamics (QCD) we have
g+g→g+g 4 diagrams
g+g→g+g+g 25 diagrams (1.18)
g + g → g + g + g + g 220 diagrams
and for g + g → 8g one needs more than one million diagrams [1]. Another very important
point is that the mathematical expression for each diagram becomes significantly more
complicated as the number of external particles grows. So the reason students are not asked
to calculate multi-gluon processes from Feynman diagrams is that it would be awful, un-
insightful, and in many cases impossible.2 There are tricks for simplifying the calculations;
one is to use the spinor helicity formalism as indicated in our example for the high-energy
limit of the tree-level process e− + e+ → γ + γ . However, for a multi-gluon process
even this does not directly provide a way to handle the growing number of increasingly
complicated Feynman diagrams. Other methods are needed and you will learn much more
about them in this book.
Thus, although visually appealing and seemingly intuitive, Feynman diagrams are not
a particularly physical approach to studying amplitudes in gauge theories. The underlying
reason is that the Feynman rules are non-unique: generally, individual Feynman diagrams
are not physical observables. In theories, such as QED or QCD, that have gauge redundancy
in their Lagrangian descriptions, we are required to fix the gauge in order to extract the
Feynman rules from the Lagrangian. So the Feynman rules depend on the choice of gauge
and therefore only gauge-invariant sums of diagrams are physically sensible. The on-
shell amplitude is at each loop-order such a gauge-invariant sum of Feynman diagrams.
2 Using computers to do the calculation can of course be very helpful, but not in all cases. Sometimes numerical
evaluation of Feynman diagrams is simply so slow that it is not realistic to do. Moreover, given that there are
poles that can cancel between diagrams, big numerical errors can arise in such evaluations. Therefore compact
analytic expressions for the amplitudes are very useful in practical applications.
1 Introduction 9
Furthermore, field redefinitions in the Lagrangian change the Feynman rules, but not the
physics, hence the amplitudes are invariant. So field redefinitions and gauge choices can
“move” the physical information between the diagrams in the Feynman expression for the
amplitude and this can hugely obscure the appearance of its physical properties.
It turns out that despite the complications of the Feynman diagrams, the on-shell ampli-
tudes for processes such as multi-gluon scattering g + g → g + g + · · · + g can actually
be written as remarkably simple expressions. This raises the questions: “why are the on-
shell amplitudes so simple?” and “isn’t there a better way to calculate amplitudes?”. These
are questions that have been explored in recent years and a lot of progress has been made
towards improving calculational techniques and gaining insight into the underlying mathe-
matical structure.
What do we mean by “mathematical structure” of amplitudes? At the simplest level, this
means the analytic structure. For example, the amplitude (1.10) has a pole at ( p1 + p2 )2 =
0 – this pole shows that a physical massless scalar particle is being exchanged in the
process. Tree amplitudes are rational functions of the kinematic invariants, and understand-
ing their pole structure is key to the derivation of on-shell recursive methods that provide
a very efficient alternative to Feynman diagrams. These recursive methods allow one to
construct on-shell n-particle tree amplitudes from input of on-shell amplitudes with fewer
particles. The power of this approach is that gauge redundancy is eliminated since all input
is manifestly on-shell and gauge invariant.
Loop amplitudes are integrals of rational functions and therefore have more complicated
analytic structure, involving for example logarithmic branch cuts. The branch cuts can
be exploited to reconstruct the loop amplitude from lower-loop and tree input, providing
another class of efficient calculational techniques. On-shell approaches that exploit the
analytic structure of amplitudes constitute a major theme in this book.
Thus, it is very useful to understand the analytic structure of amplitudes, but it is not
the only mathematical structure of interest. We are also interested in the symmetries that
leave the amplitudes invariant and how they constrain the result for the scattering process.
Some amplitudes are invariant under symmetries that are not apparent in the Lagrangian
and hence are not visible in the Feynman rules. Pursuing these symmetries and seeking
to write the amplitudes in terms of variables that trivialize the constraints on the external
data and simplify the action of the symmetries will often lead to new insights about the
amplitudes; we will see this repeatedly as we develop the subject.
Finally, let us mention one of the highlights of what we mean by mathematical structure.
It turns out that certain amplitudes naturally “live” in more abstract spaces in which they
can be interpreted as volumes of geometric objects known as “polytopes”; these are higher-
dimensional generalizations of polygons. The different compact representations found for
a given amplitude turn out to correspond to different triangulations of the corresponding
polytope. It is remarkable that a physical observable that encodes the probability for a
scattering process as a function of energies and momenta has an alternative mathematical
interpretation as the volume of a geometric object in a higher-dimensional abstract space!
It is such surprising mathematical structures that are pursued in the third part of this
book.
10 Introduction
Some of the keywords for the topics we explore are:
1. spinor helicity formalism;

2. on-shell recursion relations (BCFW, CSW, all-line shifts, . . . );
3. on-shell superspace, superamplitudes, Ward identities;
4. twistors, zone-variables, momentum twistors;
5. dual superconformal symmetry and the Yangian;
6. generalized unitarity, maximal cuts;
7. Leading Singularities and on-shell blob-diagrams ;
8. the Grassmannian, polytopes, amplituhedrons, and mathematicians;

9. gravity = (gauge theory)2 , KLT and BCJ relations;
and much more.

The study of on-shell amplitudes may suggest a paradigm that can be phrased loosely
as “avoiding the (full) Lagrangian” with all its ambiguities of field redefinitions and gauge
choices, and instead focusing on how kinematics, symmetries, unitarity, and locality impact
the physical observables. Or, more strongly, we may ask if the hints from the simplicity
of amplitudes allow us to find another approach to perturbative quantum field theory: one
might hope for a novel formulation that captures the physics of the full perturbative S-
matrix. Such a new formulation could make amplitude calculations much more efficient
and one might hope that it would lead to new insights even beyond amplitudes, for example
for correlation functions of gauge-invariant operators and perhaps even for non-perturbative
physics.
But now we are getting ahead of ourselves. The purpose of this book is to provide a
practical introduction to on-shell methods for scattering amplitudes of massless particles.
The choice of massless particles is because their processes are the simplest and cur-
rently the best understood. The above list of keywords will be explained and discussed in
detail.
This book
Our presentation assumes a basic pre-knowledge of quantum field theory, in particular
of the Dirac equation, simple scalar and QED Feynman rules and tree-level processes.
Chapter 2 introduces the spinor helicity formalism in the context of tree-level Feynman
rules and builds up the notation and conventions via explicit examples. Yang–Mills theory
is introduced as are the needed tools for studying gluon amplitudes. Various fundamental
properties of scattering amplitudes are discussed, including little group scaling and the
analytic structure of tree amplitudes.
In Chapter 3 we take the first step towards modern on-shell methods for calculating
scattering amplitudes, namely recursion relations for tree amplitudes. The material in
Chapters 2 and 3 could be part of any modern course on quantum field theory, but tradi-
tionally it is not. So we hope that our presentation will be useful, either as part of a course
or for self-study.
1 Introduction 11
Chapter 4 introduces supersymmetry and superamplitudes. In Chapter 5 we explore

the symmetries of superamplitudes in the most intensely studied theory in the field of
amplitude-explorations, namely N = 4 super Yang–Mills theory. We do not assume prior
knowledge of supersymmetry or N = 4 super Yang–Mills theory, but will introduce the
concepts and tools as needed in Chapters 4 and 5.
Part I of the book (Chapters 2–5) focuses exclusively on tree-level processes.
Part II (Chapters 6–8) offers an upgrade to loops. Chapter 6 is an introduction to
loop amplitudes and to the unitarity method, Chapter 7 discusses recursion relations for
loop-integrands, and on-shell diagrams are introduced in Chapter 8.
Part III (Chapters 9–14) is devoted to a selection of topics. You may view Part III as
an introduction to current research-topics on amplitudes. The subjects covered include the
interesting connections between amplitudes and the Grassmannian (Chapter 9), which is
the space of k-dimensional planes in n-dimensional space, and the enticing interpretation
of amplitudes as volumes of polytopes (Chapter 10). While most of the book is concerned
with scattering processes in four spacetime dimensions, Chapter 11 takes us on a tour of
amplitudes of higher- and lower-dimensional theories. Chapter 12 explores amplitudes in
perturbative gravity and supergravity while we end in Chapter 13 with the remarkable con-
nection between scattering amplitudes of gravitons and those of gluons. Chapter 14 contains
a brief outline of subjects not otherwise covered in this book and a list of suggestions for
further reading.
One should keep in mind that the subject of scattering amplitudes has two main moti-
vations. One is its practical application in particle physics: some of the on-shell methods
we cover here are indeed already implemented in numerical codes for processes relevant
in particle physics experiments. The other motivation is the fascinating internal mathe-
matical beauty of the subject. The physical relevance and mathematical structure are both
important, neither should be underestimated. They complement and benefit each other.
The style of presentation in this book is detailed and concrete, so that you can learn
the tools. The purpose is to be pedagogical – but in this as well as other matters, there
is no substitute for getting your own hands dirty. Therefore you will find many exercises
scattered throughout the text. We hope you will enjoy them.
Conventions
The subject of amplitudes is often viewed as quite technical and notationally intense. We
will try to avoid a long deadly-boring introduction about γ -matrix conventions and about
which indices go up and down and who is dotted and who is not. Suffice it here to say that we
work in four dimensions (except in Chapters 11–13), our metric convention is mostly-plus
ημν = diag(−1, +1, +1, +1), and we follow the spinor and Clifford algebra conventions in
Srednicki’s quantum field theory textbook [2]. For easy access, and to make our presentation
reasonably self-contained, some conventions are collected in the short Appendix.
PART I
TREES
Spinor helicity formalism 2
The spinor helicity formalism is a highly convenient and powerful notational tool for
amplitudes of massless particles in 4-dimensional theories. We are going to introduce the
spinor helicity formalism in the context of the basic Feynman rules that you are familiar with
from Yukawa theory and QED. So we start with Dirac spinors and build up the formalism
based on simple scattering problems.
2.1 Dirac spinors
The Lagrangian for a free massive 4-component Dirac field is
L = i γ μ ∂μ − m . (2.1)
Our conventions for the Dirac conjugate and the γ μ s are given in the Appendix. The
equation of motion for gives the Dirac equation
(−i ∂/ + m) = 0 , (2.2)
with ∂/ = ∂μ γ μ . Multiplying the Dirac equation by (i ∂/ + m) gives the Klein-Gordon equa-

tion, (−∂ 2 + m 2 ) = 0. It is solved by a plane-wave expansion
(x) ∼ u( p) ei p.x + v( p) e−i p.x , (2.3)
provided the momentum is on-shell, p 2 ≡ p μ pμ = −m 2 . This (x) will also solve the
Dirac equation (2.2) if
( p/ + m)u( p) = 0 and p + m)v( p) = 0 ,

(−/ (2.4)
where p/ = pμ γ μ . The equations (2.4) are the momentum space form of the Dirac equation.
Each of the equations in (2.4) has two independent solutions which we will label by a
subscript s = ±. We can now write the general free field expansion of as

(x) = bs ( p) u s ( p) ei p.x + ds† ( p) vs ( p) e−i p.x ,
dp (2.5)
s=±
= d 3 p is the 3d Lorentz-invariant momentum measure. For one finds a

3
where dp (2π) 2E p
†
similar result involving d± ( p) and b± ( p).
16 Spinor helicity formalism
† †
In canonical quantization, (b± ( p), d± ( p)) and (b± ( p), d± ( p)) become fermionic creation
and annihilation operators. They take care of the Grassmann nature of (x), so that u ± ( p)
and v± ( p) are commuting 4-component spinors that solve (2.4).
The next step in canonical quantization is to define the vacuum |0 such that b± ( p)|0 =
†
d± ( p)|0 = 0. One-particle states are then defined as | p; ± ≡ d± ( p)|0, etc. As described
in QFT textbooks, this leads to the Feynman rules for external fermions, namely that
they come equipped with wavefunctions v± ( p) for an outgoing anti-fermion (e.g. e+ )
and u ± ( p) for an outgoing fermion (e.g. e− ). The latter comes from the expansion of .
We can choose a basis such that in the rest-frame u ± and v± are eigenstates of the z-
component of the spin-matrix; then ± denotes spin up/down along the z-axis. For massless
fermions, ± denotes the helicity, which is the projection of the spin along the momentum
of the particle. It will be our interest here to study the wavefunctions u ± ( p) and v± ( p)
further.
The wavefunction v± ( p) solves the Dirac equation (2.4) and u ± ( p) satisfies
u ± ( p)( p/ + m) = 0. Starting with a momentum 4-vector p μ = ( p 0 , pi ) = (E, pi ) with
p μ pμ = −m 2 , let us use the gamma-matrix conventions (A.8) in the Appendix to write

0 pa ḃ
p/ = , (2.6)
pȧb 0
with
0
μ − p + p3 p1 − i p2
pa ḃ ≡ pμ (σ )a ḃ = , (2.7)
p1 + i p2 − p0 − p3
and similarly pȧb ≡ pμ (σ̄ μ )ȧb . We have σ μ = (1, σ i ) and σ̄ μ = (1, −σ i ) with σ 1,2,3 the
Pauli-matrices (A.2). The momentum bi-spinors pa ḃ and pȧb can be thought of as 2×2
matrices. The determinant is Lorentz-invariant,
det p = − p μ pμ = m 2 . (2.8)
In most of this book, we study scattering processes for massless particles. You can think of
this as the high-energy scattering limit in which the fermion mass is neglected. So let us
now specialize to the case of massless spinors.
2.2 Spinor helicity notation
When m = 0, the Dirac equation for the wavefunction 4-component spinors reads
/p v± ( p) = 0 , ū ± ( p) /p = 0 . (2.9)
We focus on v± ( p) and u ± ( p) as the wavefunctions associated with outgoing anti-fermions

and fermions. As mentioned above, in the massless case, we can choose a basis such that the
subscript ± indicates the helicity h = ±1/2. Crossing symmetry exchanges (incoming ↔
outgoing), (fermions ↔ anti-fermions) and flips the sign of the helicity, so in the massless
case the wavefunctions are related as u ± = v∓ and v ± = u ∓ .
2.2 Spinor helicity notation 17
We write the two independent solutions to the Dirac equation (2.9) as

| p]a 0
v+ ( p) = , v− ( p) = , (2.10)
0 | pȧ
and

u − ( p) = 0 , p|ȧ , u + ( p) = [ p|a , 0 . (2.11)
The angle and square spinors are 2-component commuting spinors (think 2-component
vectors) written in a very convenient Dirac bra-ket notation. By virtue of (2.6) and (2.9),
they satisfy the massless Weyl equation,
pȧb | p]b = 0 , pa ḃ | pḃ = 0 , [ p|b pbȧ = 0 , p|ḃ p ḃa = 0 . (2.12)
Raising and lowering their indices is done with the 2-index Levi-Civita symbols defined
in (A.3):
[ p|a = ab | p]b , | pȧ = ȧ ḃ p|ḃ . (2.13)
Now,
the angle and square spinors are the core of what we call the
“spinor helicity formalism.”
As you see, these bra-kets are nothing to be scared of: they are simply 2-component
commuting spinors that solve the massless Weyl equation.
It is one of the powers of the spinor helicity formalism that we do not need to find explicit
representations for the angle and square-spinors; we can simply work abstractly with | p and
| p] and later relate the results to the momentum vectors. We will see examples of how this
works in this chapter. Let us now note two important properties of the spinor bras and kets.
Angle vs. square spinors: reality conditions. The spinor field is the Dirac conjugate of
. Applying Dirac conjugation to the momentum space Dirac equations (2.9), we find that
ū ∓ = v̄± is related to v± via this conjugation provided the momentum p μ is real-valued,
i.e. the components of p μ are real numbers. Thus for real momenta
p μ real : [ p|a = (| pȧ )∗ and p|ȧ = (| p]a )∗ . (2.14)
On the contrary, for complex-valued momenta p μ , the angle and square spinors are inde-
pendent.1 It may not seem physical to take p μ complex, but it is a very very very useful
strategy. We will see this repeatedly.
1 One can keep p μ real and change the spacetime signature to (−, +, −, +); in that case, the angle and square
spinors are real and independent.
Spinor completeness relation. The spin-sum completeness relation with m = 0 reads

u − u − + u + u + = − /p . (See for example Equation (38.23) of the QFT textbook [2].) With
the help of crossing symmetry u ∓ = v ± , this can be written in spinor helicity notation as
− /p = | p[ p| + | p] p| . (2.15)
There is a small abuse of notation in writing (2.15): the LHS is a 4 × 4 matrix and the
RHS involves products of 2-component spinors. The relation should be read in terms of
matching the appropriate L- and R-spinor indices via (2.6), namely:
pa ḃ = −| p]a p|ḃ , pȧb = −| pȧ [ p|b . (2.16)
The relations (2.16) may look new but they should not shock you. After all, it is taught in
some algebra classes that if a 2 × 2 matrix has vanishing determinant, it can be written as a
product of two 2-component vectors, say λa and λ̃ḃ : i.e. det p = 0 ⇔ pa ḃ = − λa λ̃ḃ . In
fact, this is often the starting point of introductions to the spinor helicity formalism. In this
presentation, we will suppress the λa and λ̃ḃ notation in favor of the more intuitive Dirac
bra-kets, λa → | p]a and λ̃ȧ → p|ȧ .
It is useful for keeping your feet on the ground to work out an explicit solution for
| p and | p] for a given 4-momentum p μ . The following exercise guides you to do just
that.
Exercise 2.1
Consider the momentum vector
p μ = (E, E sin θ cos φ, E sin θ sin φ, E cos θ ) . (2.17)
Express pa ḃ and pȧb in terms of E, sin θ2 , cos θ2 and e±iφ .

√ cos θ2
Show that the helicity spinor | p = 2E
ȧ
solves the massless Weyl
sin θ2 eiφ
equation. Find expressions for the spinors p|ȧ , | p]a , and [ p|a and check that they
satisfy pa ḃ = −| p]a p|ḃ and pȧb = −| pȧ [ p|b .
You have probably noted that the angle and square spinors are only defined up to an overall
scaling that leaves p μ invariant. This is called the little group scaling and it plays a central
role which we explore much more in Section 2.6.
We are now in dire need of some examples! Before we move ahead, it is convenient to
summarize the external line Feynman rules for outgoing massless (anti)fermions:

• Outgoing fermion with h = +1/2: u + ( p) ←→ [ p|a , 0 ;

• Outgoing fermion with h = −1/2: u − ( p) ←→ 0 , p|ȧ ;

| p]a
• Outgoing anti-fermion with h = +1/2: v+ ( p) ←→ ;
0

0
• Outgoing anti-fermion with h = −1/2: v− ( p) ←→ .
| pȧ
Note the useful mnemonic that positive helicity of an outgoing particle is associated with
square spinors while negative helicity comes with angle-spinors. Finally, let us note that
for massless fermions we usually do not bother much to distinguish fermion-anti-fermion
due to the simple crossing rules. In the amplitudes, we consider all the external particles
to be outgoing, so think of the rules here as the difference between the arrow on a fermion
line pointing into the diagram (anti-fermion) or out of the diagram (fermion).
Exercise 2.2
The helicity of a massless particle is the projection of the spin along the momen-
tum 3-vector p, so the helicity operator can be written = S · p/| p|, where the
spin Si = 12 i jk S jk (i, j, k = 1, 2, 3) is defined by the spin matrix S μν = 4i [γ μ , γ ν ].
For simplicity, you can pick a frame where p μ is along the z-axis. Use the results
of Exercise 2.1 to show that the chiral basis (2.10), (2.11) is also a helicity basis,
i.e. show that v± = −h ± v± for h ± = ± 12 .
2.3 Examples from Yukawa theory
Consider a Dirac fermion interacting with a real scalar φ via a Yukawa coupling:
L = i γ μ ∂μ − 12 (∂φ)2 + gφ . (2.18)
The interaction term gives the simple 3-vertex Feynman rule ig. For a diagram with two
outgoing Dirac fermions connecting to the rest of the particles in the process via an internal
scalar line, the usual Feynman rules give
−i
= ig u h 1 ( p1 )vh 2 ( p2 ) × × (rest), (2.19)
( p 1 + p 2 )2
with the spinor indices contracted and the gray blob representing the rest of the diagram.
We focus on the spinor product: choosing specific examples for the helicities we find

0
u + ( p1 ) v− ( p2 ) = [ p1 |a , 0 = 0, (2.20)
| p2 ȧ

0
u − ( p1 ) v− ( p2 ) = 0 , p1 |ȧ = p1 |ȧ | p2 ȧ ≡ p1 p2 . (2.21)
| p2 ȧ
In the first case, the diagram vanishes. In the second case, we introduce the angle spinor
bracket p1 p2 . Together with its best friend, the square spinor bracket [ p1 p2 ], it is a
key ingredient for writing amplitudes in spinor helicity formalism. So let us introduce the
spinor brackets properly: for two light-like vectors p μ and q μ , we define spinor brackets
p q = p|ȧ |qȧ , [ p q] = [ p|a |q]a . (2.22)
Since indices are raised/lowered with the antisymmetric Levi-Civita tensors (A.3),
cf. (2.13), these brackets are antisymmetric:
p q = −q p , [ p q] = −[q p] . (2.23)
There are no p|q]-brackets, because the indices cannot be contracted directly to form a
Lorentz-scalar.
For real momenta, the spinor products satisfy [ p q]∗ = q p.
It is a good exercise (hint: use (2.16) and (A.7)) to derive the following important
relation:
p q [ p q] = 2 p · q = ( p + q)2 . (2.24)
In amplitudes with light-like momenta p1 , p2 , . . . we use the short-hand notation |1 =

| p1 , |2 = | p2 , etc. Applying (2.24) to our Yukawa example (2.19), we find
1 1 1
g12 × × (rest) = g12 × × (rest) = g × (rest) . (2.25)
2 p1 . p2 12[12] [12]
The cancellation of the 12-factors is the first tiny indication of simplifications that await
us in the following.
Example. Let us calculate the 4-fermion tree amplitude A4( f¯h f h 2 f¯h 3 f h 4 ) in Yukawa
1
theory. In our notation, f denotes an outgoing fermion and f¯ an outgoing anti-fermion.

The superscripts indicate the helicity. When specifying the helicity of each particle,
we call the amplitude a helicity amplitude.
The s-channel diagram for the 4-fermion process is
1 4
−i
= ig u 4 v3 × × ig u 2 v1 . (2.26)
( p 1 + p 2 )2
2 3
Here we are using a shorthand notation vi = vh i ( pi ) and u i = u h i ( pi ). Our observations
in the previous example show that the diagram (2.26) vanishes unless particles 1 and 2
have the same helicity, and 3 and 4 have the same helicity. So suppose we take particles
1 and 2 to have negative helicity and 3 and 4 positive. Then the u-channel diagram
vanishes and the diagram (2.26) is the only contribution to the 4-fermion amplitude.
Translating the uv-products to spinor brackets we find

1 1 [34]
i A4 f¯− f − f¯+ f + = ig 2 [43] 21 = ig 2 [34] 12 = ig 2 .
2 p1 . p2 12[12] [12]
(2.27)
The result is a nice simple ratio of two spinor brackets. Now, it is both fun and useful
to note that by momentum conservation, we have (using (2.24))
12[12] = 2 p1 . p2 = ( p1 + p2 )2 = ( p3 + p4 )2 = 2 p3 . p4 = 34[34] . (2.28)
Using this in the second equality of (2.27) we get another expression for the same
amplitude:

12
A4 f¯− f − f¯+ f + = g 2 . (2.29)
34
This hints at another important lesson: there are various relationships among spinor
brackets, implied for example by momentum conservation as in (2.28), and they allow
for multiple equivalent forms of the same physical amplitude.
Example. Next, let us see what new features appear when we calculate the 4-point
tree amplitude with two scalars and two fermions in Yukawa theory. Two diagrams
contribute to the process:
1 4 4 1

i A4 φ f¯h 2 f h 3 φ = +
2 3 2 3
−i(/ p1 + p/2 )
= (ig)2 u 3 v2 + (1 ↔ 4) . (2.30)
( p 1 + p 2 )2
If the fermions have the same helicity, each diagram has a numerator factor
u ± ( p3 )γ μ v± ( p2 ) = 0. Hence

A4 φ f¯− f − φ = A4 φ f¯+ f + φ = 0. (2.31)
Thus, the fermions need to have opposite helicity to give a non-vanishing result: for
example

μ

0 (σ μ )a ḃ |2]b
u − ( p3 )γ v+ ( p2 ) = 0 , 3|ȧ ≡ 3|γ μ |2] . (2.32)
(σ̄ μ )ȧb 0 0
Note the abuse of notation in the definition above of the angle-square bracket p|γ μ |k]:
it combines the 2-component spinors with the 4×4 gamma-matrix. The meaning should
be clear, though, in that the 2-component spinors project out the matching sigma-matrix
from γ μ . The spinor bracket [ p|γ μ |k is defined similarly. For same-helicity fermions
we have p|γ μ |k = 0 = [ p|γ μ |k].
Angle-square brackets appear often, so it is useful to record the following properties:
[k|γ μ | p = p|γ μ |k] , (2.33)
[k|γ μ | p∗ = [ p|γ μ |k (for real momenta). (2.34)
We often use p|P|k] ≡ Pμ p|γ μ |k]. The notation implies that p μ and k μ are light-like,
but no assumptions are made about P μ . However, if P μ is also light-like, then
p|P|k] = p|ȧ P ȧb |k]b = p|ȧ (−|Pȧ [P|b ) |k]b = − p P[Pk] , (P 2 = 0) .

(2.35)
Finally, the Fierz identity
1|γ μ |2]3|γμ |4] = 213[24] (2.36)
will come in handy in several applications.
Exercise 2.3
Prove the Fierz identity (2.36).
Exercise 2.4
Show that k|γ μ |k] = 2k μ and k|P|k] = 2 P · k.
It is often useful to use a shorthand for the momenta “sandwiched” in the angle-square
brackets, for example 1|3|4] = 1| p3 |4] or 2|1 + 4|5] = 2|( p1 + p4 )|5].
With our new tools, we return now to the tree amplitude with two scalars and two
fermions.
Example. Choosing opposite helicities for the fermions in (2.30), we have

3|1 + 2|2]
A4 φ f¯+ f − φ = −g 2 + (1 ↔ 4)
( p 1 + p 2 )2
3|1|2]
= −g 2 + (1 ↔ 4) (using the Weyl Eq. p2 |2] = 0)
( p 1 + p 2 )2
−31[12]
= −g 2 + (1 ↔ 4) (using (2.35))
12[12]
13
= −g 2 + (1 ↔ 4) , (2.37)
12
so that the result is

13 34
A4 φ f¯+ f − φ = −g 2 + . (2.38)
12 24
Note Bose symmetry under exchange of the scalar particle momenta.
In amplitude calculations, momentum conservation is imposed on n particles as

n μ
i=1 pi = 0 (consider all particles outgoing). This is encoded in the spinor helicity for-
malism as

n
n
|i[i| = 0, i.e. qi[ik] = 0, (2.39)
i=1 i=1
for any light-like vectors q and k. For example, you can (and should) show that for n = 4
momentum conservation implies 12[23] = −14[43]. In (2.28), we already found the
identity 12[12] = 34[34] valid when p1 + p2 + p3 + p4 = 0.
With all momenta outgoing, the Mandelstam variables are defined as
si j = −( pi + p j )2 , si jk = −( pi + p j + pk )2 , etc. (2.40)
In particular, we have s = s12 , t = s13 , and u = s14 for 4-particle processes.

To see some of the power of the spinor helicity formalism, let us now calculate the spin
sum

A4 φ f¯ f φ 2 = A4 φ f¯h 2 f h 3 φ 2 (2.41)
h 2 ,h 3 =±
for the 2-scalar 2-fermion process in the previous example. To really appreciate the differ-
ence in formalism, it is educational to first do the calculation the standard way, using the
spinor-completeness relations and evaluating the gamma-matrix traces:
Exercise 2.5
Use standard techniques to show that |A4 (φ f¯ f φ)|2 = 2g 4 (s − t)2 /(st).
[Hint: This is very similar to the massless limit of the example e− ϕ → e− ϕ in Chapter
48 of Srednicki [2], but we include no 12 -factors from averages here.]
Having resharpened your pencils after doing this exercise, let us now do the spin sum in
the
spinor helicity
formalism. We already know from (2.31) that the helicity amplitudes
A4 φ f f φ vanish unless the spinors have opposite helicity, hence
¯h2 h3

A4 φ f¯ f φ 2 = A4 φ f¯− f + φ 2 + A4 φ f¯+ f − φ 2 . (2.42)
The first term is calculated easily using the result (2.38) for the helicity amplitude and the
reality condition (2.14):

− + 2
A4 φ f¯ f φ = g 4 13 + 34 [13] [34]
+
12 24 [12] [24]

13[13] 34[34] 13[34] 34[13]
= g4 + + + . (2.43)
12[12] 24[24] 12[24] 24[12]
In the first two terms, we can directly translate the spinor products to Mandelstam variables
using (2.24). For the last two terms, the momentum conservation identity (2.39) comes in
handy, giving 12[24] = −13[34] and 24[12] = −34[13]. Thus (2.43) gives

− + 2
A4 φ f¯ f φ = g 4 t + s − 2 = g 4 (s − t) .
2
(2.44)
s t st
The second term in (2.42) gives exactly the same, so |A4 (φ f¯ f φ)|2 = 2g 4 (s − t)2 /(st),
in agreement with the result of the standard calculation – but with use of much less
pencil-power!
Exercise 2.6

Calculate the 4-fermion “all-minus” amplitude A4 f¯− f − f¯− f − in Yukawa theory.
Exercise 2.7
Calculate the spin sum |A4 ( f¯ f f¯ f )|2 for the 4-fermion process in Yukawa theory.
Exercise 2.8
Consider a model with a Weyl-fermion ψ and a complex scalar φ:
L = iψ † σ̄ μ ∂μ ψ − ∂μ φ̄ ∂ μ φ + 12 g φ ψψ + 12 g ∗ φ̄ ψ † ψ † − 14 λ |φ|4 . (2.45)
Show that

24
12
A4 φφ φ̄ φ̄ = −λ , A4 φ f − f + φ̄ = −|g|2 , A4 f − f − f + f + = |g|2 .
34 34
(2.46)
We do not put a bar on the f s here because in this model the 4-component fermion
field is a Majorana fermion so there is no distinction between f and f¯. The amplitudes
(2.46) serve as useful examples when we discuss supersymmetry in Chapter 4.
We end this section by discussing one more identity from the amplitudes tool-box: the
Schouten identity is a fancy name for a rather trivial fact: three vectors in a plane cannot be
linearly independent. So if we have three 2-component kets |i, | j, and |k, we can write
one of them as a linear combination of the two others:
|k = a|i + b| j for some a and b. (2.47)
One can “dot” in spinors ·| and form antisymmetric angle brackets to solve for the
coefficients a and b. Then (2.47) can be cast in the form
|i jk + | jki + |ki j = 0 . (2.48)
This is the Schouten identity. It is often written with a fourth spinor r | “dotted-in”:
ri jk + r jki + r ki j = 0 . (2.49)

A similar Schouten identity holds for the square spinors: [ri][ jk] + [r j][ki] +
[r k][i j] = 0.
Exercise 2.9
[12][34]2
Show that A5 ( f − f¯− φφφ) = g 3 + (3 ↔ 5) + (4 ↔ 5) in Yukawa
[13][14][23][24]
theory (2.18).
2.4 Massless vectors and examples from QED
The external line rule for outgoing spin-1 massless vectors is simply to “dot-in” their
polarization vectors. They can be written in spinor helicity notation as follows:
μ p|γ μ |q] μ q|γ μ | p]

− ( p; q) = − √ , + ( p; q) = − √ , (2.50)
2 [q p] 2 q p
where q = p denotes an arbitrary reference spinor. Note that the massless Weyl equation
μ
ensures that pμ ± ( p) = 0. It can be useful to write the polarizations as
√ √
2 2
/− ( p; q) = | p[q| + |q] p| , /+ ( p; q) = | p]q| + |q[ p| . (2.51)
[qp] qp
The arbitrariness in the choice of reference spinor reflects gauge invariance, namely that
one is free to shift the polarization vector with any constant times the momentum vector:
μ μ
± ( p) → ± ( p) + C p μ . This does not change the on-shell amplitude An , as encoded in the
μ
familiar Ward identity pμ An = 0. For each external vector boson, one has a free choice of
the corresponding reference spinor qi = pi ; however, one must stick with the same choice
in each diagram of a given process. When summing over all diagrams, the final answer for
the amplitude is independent of the choices of qi .2
Exercise 2.10
Consider the momentum p μ = (E, E sin θ cos φ, E sin θ sin φ, E cos θ ). In Exercise
2.1, you found the corresponding angle and square spinors | p and | p]. In this exercise,
we establish the connection between the polarization vectors (2.50) and the more
familiar polarization vectors
μ e∓iφ
˜± ( p) = ± √ 0, cos θ cos φ ± i sin φ, cos θ sin φ ∓ i cos φ, −sin θ . (2.52)
2
μ
Note that for θ = φ = 0, we have ˜± ( p) = ± √12 (0, 1, ∓i, 0).
2 In practical calculations, independence of the reference spinor provides a useful check on the result for an
amplitude. For example, in a numerical evaluation one has to get the same result for different choices of the
polarization reference spinors.
(a) Show that ˜± ( p)2 = 0 and ˜± ( p) · p = 0.

μ μ μ
(b) Since ˜± ( p) is null, (˜± ( p))a ḃ = (σμ )a ḃ ˜± ( p) can be written in as a product of a
μ
square and an angle spinor. To see this specifically, first calculate (˜± ( p))a ḃ and
μ
then find an angle spinor r | such that (˜+ ( p))a ḃ = −| p]a r |ḃ .
√
[Hint: you should find that r p = − 2.]
√
(c) Next, show that it follows from (2.50) that (+ ( p; q))a ḃ = q p
2
| p]q|.
μ μ
(d) Now suppose there is a constant√c+ such that + ( p; q) = ˜+ ( p) + c+ p μ . Show that
this relation requires r p = − 2 (as is consistent with the solution you found for
r | in part (b)) and then show that c+ = −rq/ pq.
μ μ μ μ
Since + ( p; q) = ˜+ ( p) + c+ p μ , the polarization vectors + ( p; q) and ˜+ ( p) are
equivalent. You can show the same for the negative helicity polarization. It should
be clear from this exercise that the arbitrariness in the reference spinors q in the
polarizations (2.50) is directly related to the gauge invariance reflected in the possibility
of adding any number times p μ to the polarization vectors.
We now calculate some amplitudes in QED to illustrate the use of the spinor helicity
formalism. The QED Lagrangian
1
L = − Fμν F μν + i γ μ (∂μ − ie Aμ ) (2.53)
4
describes the interaction of a massless3 fermion with a photon via the interaction Aμ γ μ .
The vertex rule is ieγ μ .
Example. Consider the 3-particle QED amplitude A3( f h f γ h 3 ); here f = e− and

1 ¯h 2
+
f = e denote electron and positron outgoing states, respectively, in the massless
¯
limit. Choose, as an example, helicities h 1 = −1/2, h 2 = +1/2 and h 3 = −1. We
then have

μ
i A3 f − f¯+ γ − = u − ( p1 )ieγμ v+ ( p2 ) − ( p3 ; q)
3|γ μ |q] √ 13[2q]
= ie1|γμ |2] √ = 2ie ,
2 [3 q] [3 q]
using in the last step the Fierz identity (2.36). Thus

13[2q]
A3 f − f¯+ γ − = ẽ . (2.54)
[3 q]
√ √
We have absorbed the 2 into the definition of the coupling e as ẽ ≡ 2e.
When we introduced the polarization vectors (2.50) in spinor helicity formalism, we men-
tioned that the on-shell amplitude should be independent of the reference spinor q. Here,
there are no other diagrams and naively it appears that (2.54) depends on |q]. However, it
is actually independent of |q] – and this brings us to discuss several important aspects:
3 Think of this as the high-energy scattering limit in which we consider electrons/positrons massless.
• First, let us see how to eliminate |q] from (2.54). Multiply (2.54) by 1 = 12/12. In
the numerator, we then have 1312[2q]. Now use (2.35), momentum conservation
p2 = − p1 − p3 , and the massless Weyl equation to get
12[2q] = −1|2|q] = 1|1 + 3|q] = 1|3|q] = 13[3q] . (2.55)
The square bracket [3q] cancels against the same factor in the denominator of (2.54),
and we are left with

132
A3 f − f¯+ γ − = ẽ . (2.56)
12
This is clearly independent of |q].
• Second, note that the result (2.56) depends only on angle brackets, not square brackets.
This is no coincidence, but a consequence of 3-particle special kinematics. Note that if
μ μ μ
three light-like vectors satisfy p1 + p2 + p3 = 0, then
12[12] = 2 p1 . p2 = ( p1 + p2 )2 = p32 = 0, (2.57)
so either 12 or [12] must vanish. Suppose 12 is non-vanishing; then by (2.39) and
the massless Weyl equation we have 12[23] = −1| p2 |3] = 1|( p1 + p3 )|3] = 0. So
[23] = 0. Similarly, [13] = 0. Thus [12] = [23] = [31] = 0 which means that these three
square spinors are proportional:
|1] ∝ |2] ∝ |3] . (2.58)
Alternatively, 3-particle kinematics could hold with square brackets non-vanishing and
|1 ∝ |2 ∝ |3 . (2.59)
To summarize, special 3-particle kinematics is the statement that for three on-shell
massless momenta satisfying momentum conservation, the associated angle and square
spinors must satisfy either (2.58) or (2.59). As a consequence:
1. A non-vanishing on-shell 3-particle amplitude with only massless particles can only
depend on either angle brackets or square brackets of the external momenta, never
both.
2. Since for real momenta, angle and square spinors are each other’s complex conjugates,
on-shell 3-particle amplitude of only massless particles can only be non-vanishing
in complex momenta (unless it is a constant, as in φ 3 -theory).4 Although they do not
occur in nature, the massless complex momentum 3-point amplitudes are extremely
useful for building up higher-point amplitudes recursively. In many cases, the on-shell
3-point amplitudes are the key building blocks. More about this in Chapter 3.
• Finally, let us comment on choices of q in (2.54). Naively, it might seem that choosing
|q] ∝ |2] gives zero for the amplitude; this would be inconsistent with our q-independent
4 Or using a (−, +, −, +) spacetime signature, cf. footnote 1.

non-vanishing result (2.56). However, this choice gives [3q] ∝ [23], so the denominator
therefore vanishes by special kinematics. One could say that the zero [22] in the numerator
is canceled by the zero [23] in the denominator, or simply that |q] ∝ |2] is not a legal
μ
choice since it makes the polarization vector − ( p3 ; q) divergent.
At this stage it is natural to ask how, then, we know if a given 3-point amplitude of
massless particles should depend on angle brackets or square brackets. This has a good
answer, which we reveal in Section 2.6. For now, let us carry on exploring QED amplitudes
in the spinor helicity formalism.
Example. Consider the QED Compton scattering process: e−γ → e−γ . By crossing
symmetry, we can view this as the amplitude A4 ( f¯ f γ γ ) with all particles outgoing
and labeled by momenta 1,2,3,4:
3 4 4 3
i A4 ( f¯ f γ γ ) = +
1 2 1 2
−i(/ p1 + p/3 )
= (ie)2 u 2 /4 /3 v1 + (3 ↔ 4) . (2.60)
( p 1 + p 3 )2
Note that we have an odd number of gamma-matrices sandwiched between two

spinors. If f¯ and f have the same helicity, then such spinor products vanish,
e.g. 2|γ μ γ ν γ ρ |1 = 0. So we need the fermions to have opposite helicity for the
process to be non-vanishing.
Suppose the photons both have negative helicity. Then the first diagram in (2.60)
involves (/3− v1+ ) ∝ |3[q3 1] using (2.51). By picking |q3 ] ∝ |1], this diagram van-
ishes. Similarly, we can choose |q4 ] ∝ |1] to make the second diagram vanish. So
A4 ( f¯+ f − γ − γ − ) = 0.
Exercise 2.11
As a spinor helicity gymnastics exercise, show that A4 ( f¯+ f − γ − γ − ) = 0 without
making any special choices of the reference spinors q3 and q4 .
Now consider A4 ( f¯+ f − γ + γ − ). We have

+ − + −
2e2 24[q4 | − |1]1| − |3]3| |q3 [31]
A4 f f γ γ =
¯
13[13]q3 3[q4 4]

2e 2q3 [3| − |1]1| − |4]4| |4[q4 1]
2
+ . (2.61)
14[14]q3 3[q4 4]
Let us choose q3 = q4 = p1 . Then the second diagram in (2.61) vanishes and we get

24[13]31[31] 24[13]
A4 f¯+ f − γ + γ − = −ẽ2 = −ẽ2 , (2.62)
13[13]13[14] 13[14]
√
where ẽ = 2e. Momentum conservation lets us rewrite this using 23[13] =
−24[14], giving

242
A4 f¯+ f − γ + γ − = ẽ2 . (2.63)
1323
The amplitude A4 ( f¯+ f − γ − γ + ) is obtained by interchanging the momentum labels 3

and 4 in (2.63).
Exercise 2.12
Show that the amplitude A4 ( f¯+ f − γ + γ − ) is independent of q3 and q4 by deriving
(2.63) without making a special choice for the reference spinors q3 and q4 .
Exercise 2.13
Via crossing symmetry, the Compton scattering process can be regarded as electron-
positron annihilation e− + e+ → γ + γ . Compare (2.63) with (1.15) and fill out the
missing steps in that introductory example to obtain the differential scattering cross-
section (1.14).
Exercise 2.14
Calculate the tree-level Bhabha scattering process e− e+ → e− e+ using spinor helicity
formalism.
For further experience with spinor helicity formalism, consider massless scalar-QED: the
interaction between the complex scalar ϕ and the gauge field Aμ is captured via the covariant
derivative Dμ = ∂μ − ie Aμ as
1 1
L = − Fμν F μν − |Dϕ|2 − λ|ϕ|4
4 4
1 1
= − Fμν F − |∂ϕ| + ie Aμ (∂μ ϕ ∗ )ϕ − ϕ ∗ ∂μ ϕ − e2 Aμ Aμ ϕ ∗ ϕ − λ|ϕ|4 . (2.64)
μν 2
4 4
The Feynman rules give a scalar–scalar–photon 3-vertex ie( p2 − p1 )μ (both momenta
outgoing), a scalar–scalar–photon–photon 4-vertex −2ie2 ημν , and a 4-scalar vertex −iλ.
We can think of ϕ and ϕ ∗ as the spin-0 supersymmetric partners of the electron/positron
and we loosely call them selectrons/spositrons, though we are not assuming that our model
is part of a supersymmetric theory. A process like ϕ + γ → ϕ + γ is then the spin-0
analogue of Compton scattering. Here, we consider the extreme high-energy regime in
which the mass of the selectron/spositron is taken to be zero.
Exercise 2.15
Calculate the 3-particle amplitude A3 (ϕ ϕ ∗ γ − ). Show that it is independent of the
reference spinor of the photon polarization vector and write the result in a form that
only involves angle brackets.
Use complex conjugation to write down the amplitude A3 (ϕ ϕ ∗ γ + ).
Exercise 2.16
Consider the amplitude A4 (ϕ ϕ ∗ γ γ ). Show that for any choice of photon helicities, one
can always pick the reference spinors in the polarizations such that the scalar–scalar–
photon–photon contact term gives a vanishing contribution to the on-shell 4-point
amplitude.
Exercise 2.17
Calculate A4 (ϕ ϕ ∗ γ γ ) and massage the answer into a form that depends only on either
angle or square brackets and is manifestly independent of the reference spinors.
Exercise 2.18
Calculate the spin sum |A4 (ϕ ϕ ∗ γ γ )|2 .
Exercise 2.19
Calculate A4 (ϕ ϕ ∗ ϕ ϕ ∗ ). The answer can be expressed in terms of the Mandelstam
variables, but show that you can bring it to the following form:

∗ ∗ 132 242
A4 (ϕ ϕ ϕ ϕ ) = −λ + ẽ 1 +
2
. (2.65)
12233441
This result will resurface in Section 3.3.
2.5 Yang–Mills theory, QCD, and color-ordering
Gluons are described by the Yang–Mills Lagrangian

1
L = − Tr Fμν F μν , (2.66)
4
with Fμν = ∂μ Aν − ∂ν Aμ − √ig2 [Aμ , Aν ] and Aμ = Aaμ T a . The gauge group is G = SU (3)
for QCD, but we will keep the number of colors N general and take G = SU (N ). The
gluon fields are in the adjoint representation, so the color-indices run over a, b, . . . =
1, 2, . . . , N 2 − 1. The generators T a are normalized5 such that Tr T a T b = δ ab and
[T a , T b ] = i fãbc T c .
To extract Feynman rules from (2.66), one needs to fix the gauge redundancy. An
amplitude-friendly choice is Gervais–Neveu gauge for which the gauge-fixing term is
Lgf = − 12 Tr(Hμ μ )2 with Hμν = ∂μ Aν − √ig2 Aμ Aν [2]. In this gauge, the Lagrangian
√
5 A more common normalization
√ is Tr T a T b = 12 δ ab and [T a , T b ] = i f abc T c . So we have fãbc = 2 f abc , in
analogue
√ with ẽ = 2e in QED in Section 2.4. It serves the same purpose here, namely as compensation for
the 2 in the polarization vectors (2.50), so the on-shell amplitudes can be written without such factors.
takes the form6

1 √ g2 μ ν
L = Tr − ∂μ Aν ∂ μ Aν − i 2g ∂ μ Aν Aν Aμ + A A Aν Aμ . (2.67)
2 4
η
The Feynman rules then give a gluon propagator δ ab pμν2 . The 3- and 4-gluon vertices involve
fãbc and fãbx f˜xcd +perms, respectively, each dressed up with kinematic factors that we
will get back to later. The amplitudes constructed from these rules can be organized into
different group theory structures each dressed with a kinematic factor. For example, the
color factors of the s-, t-, and u-channel diagram of the 4-gluon tree amplitude are
cs ≡ fã1 a2 b f˜b a3 a4 , ct ≡ fã1 a3 b f˜b a4 a2 , cu ≡ fã1 a4 b f˜b a2 a3 , (2.68)
and the 4-point contact term generically gives a sum of contributions with cs , ct , and cu
color factors. The Jacobi identity relates the three color factors:
cs + ct + cu = 0 . (2.69)
So there are only two independent color-structures for the tree-level 4-gluon amplitude. Let
us now see this in terms of traces of the generators T a . Note that
i fãbc = Tr(T a T b T c ) − Tr(T b T a T c ) . (2.70)
The products of generator-traces in the amplitudes can be Fierz’ed using the completeness
relation
1 j l
(T a )i j (T a )k l = δi l δk j − δi δk . (2.71)
N
For example, for the 4-gluon s-channel diagram we have

f˜ a1 a2 b f˜ b a3 a4 − Tr T a1 T a2 T a3 T a4 + Tr T a1 T a2 T a4 T a3

= Tr T a1 T a3 T a4 T a2 − Tr T a1 T a4 T a3 T a2 . (2.72)
Here we have used the cyclic property of the traces to deduce the four color-structures.
Similarly, the three other diagrams contributing to the 4-gluon amplitude can also be written
in terms of single-trace group theory factors. So that means that we can write the 4-gluon
tree amplitude as

Afull,tree
4 = g 2 A4 [1234] Tr T a1 T a2 T a3 T a4 + perms of (234) , (2.73)
where the partial amplitudes A4 [1234], A4 [1243] etc. are called color-ordered amplitudes.
We use the squared parenthesis in A4 [1234] to distinguish the color-ordered amplitude
from the non-color-ordered amplitude A4 (1234). Each partial amplitude is gauge invariant.7
6 We ignore ghosts, since our focus here is on tree-level amplitudes.

7 This follows from a partial orthogonality property of the single-traces [1].
The color-structure generalizes to any n-point tree-level amplitude involving any particles
that transform in the adjoint of the gauge group: we write

Afull,tree
n = g n−2 An 1 σ (2 . . . n) Tr T a1 T σ (a2 · · · T an ) , (2.74)
perms σ
where the sum is over the (overcomplete) trace-basis of (n − 1)! elements that takes into
account the cyclic nature of the traces. For loop-amplitudes, one also needs to consider
multi-trace structures in addition to the simple single-trace – for more about this, see [3, 4].
We have factored out the coupling constant g to avoid carrying it along explicitly in all the
color-ordered amplitudes.
In the Gervais–Neveu gauge, the Feynman vertex rules for calculating the color-ordered
amplitudes are:
√
μ μ μ
• 3-gluon vertex V μ1 μ2 μ3 ( p1 , p2 , p3 ) = − 2 ημ1 μ2 p1 3 + ημ2 μ3 p2 1 + ημ3 μ1 p3 2 ,
• 4-gluon vertex V μ1 μ2 μ3 μ4 ( p1 , p2 , p3 , p4 ) = ημ1 μ3 ημ2 μ4 .
The color-ordered amplitude An [12 . . . n] is calculated in terms of diagrams with no lines

crossing and the ordering of the external lines fixed as given 1, 2, 3, . . . , n. The gluon
polarization vectors are given in (2.50), (2.51).
Let us consider the simplest case, namely the 3-gluon amplitude. From the 3-vertex rule,
we get
√
A3 [1 2 3] = − 2 (1 2 )(3 p1 ) + (2 3 )(1 p2 ) + (3 1 )(2 p3 ) . (2.75)
Let us now pick gluons 1 and 2 to have negative helicity while gluon 3 gets to have positive
helicity. Translating to spinor helicity formalism (using the Fierz identity (2.36)) we have
12[q1 q2 ]q3 1[13] + 2q3 [q2 3]12[2q1 ] + q3 1[3q1 ]23[3q2 ]
A3 1− 2− 3+ = − .
[q1 1][q2 2]q3 3
(2.76)
We must now consider 3-particle special kinematics, (2.58) or (2.59). If |1 ∝ |2 ∝ |3, all
three terms vanish in the numerator of (2.76). So select 3-particle kinematics |1] ∝ |2] ∝ |3].
Then the first term vanishes and we are left with
2q3 [q2 3]12[2q1 ] + 1q3 [q1 3]23[3q2 ]
A3 1− 2− 3+ = − . (2.77)
[q1 1][q2 2]q3 3
To simplify this, first use momentum conservation to write 12[2q1 ] = −13[3q1 ]. Then
[q1 3][q2 3] factors out and we get

− − + [q1 3][q2 3] 13q3 2 + 1q3 23
A3 1 2 3 = . (2.78)
[q1 1][q2 2]q3 3
After a quick round of Schouten’ing, this simplifies to

− − + [q1 3][q2 3] − 123q3 12[q1 3][q2 3]
A3 1 2 3 = = . (2.79)
[q1 1][q2 2]q3 3 [q1 1][q2 2]
So we have got rid of q3 . To eliminate q1 and q2 , use momentum conservation [q1 3]23 =
−[q1 1]21 and [q2 3]13 = −[q1 1]12. The result for the 3-gluon amplitude is a remark-
ably simple ratio of angle brackets:
123
A3 1− 2− 3+ = . (2.80)
2331

The result for the “googly” gluon amplitude A3 1+ 2+ 3− is
[12]3
A3 1+ 2+ 3− = . (2.81)
[23][31]
Exercise 2.20
Fill in the details to derive the amplitude (2.81).
Exercise 2.21
Calculate − ( p, q) · − (k, q ), + ( p, q) · + (k, q ), and − ( p, q) · + (k, q ).
Show that ± ( p, q) · ± (k, q ) vanishes if q = q .
How can you make − ( p, q) · + (k, q ) vanish?
Exercise 2.22
Use the previous exercise to show that for any choice of gluon helicities, it is always
possible to choose the polarization vectors such that the contribution from the 4-gluon
contact term to the 4-gluon amplitude vanishes.
Exercise 2.23
Use a well-chosen set of reference spinors to show that the entire 4-gluon amplitude
vanishes if all four gluons have the same helicity.
Exercise 2.24

Calculate the color-ordered 4-gluon tree amplitude A4 1− 2− 3+ 4+ using the color-
ordered Feynman rules and a smart choice of reference spinors. Show that the answer
can be brought to the form
124
A4 1− 2− 3+ 4+ = . (2.82)
12233441
Note the cyclic structure of the denominator.
The result (2.82) for the 4-gluon amplitude is an example of the famous Parke–Taylor
n-gluon tree amplitude [5]: for the case where gluons i and j have helicity −1 and all
the n − 2 other gluons have helicity +1, the Parke–Taylor formula for the gluon tree
amplitude is
i j4
An 1+ . . . i − . . . j − . . . n + = . (2.83)
1223 · · · n1
We prove this formula in Chapter 3.

It is relevant to point out that the number of Feynman diagrams that generically contribute
to an n-gluon tree amplitude is8
n=3 4 5 6 7 ...
#diagrams = 1 3 10 38 154 ...
A fun little trivia point you can impress your friends with in a bar (oh, we mean at the
library), is that the number of trivalent graphs that contribute to the n-gluon tree process is
counted by the Catalan numbers.
It should be clear that even though you have now learned some handy tricks for how to
choose the polarization vectors to reduce the difficulty of the calculation, it would be no
fun to try to compute these higher-point gluon amplitudes by brute force. But despite the
complications of the many diagrams and their increased complexity, the answer is just
the simple Parke–Taylor expression (2.83) for the − − + + · · · + helicity case. And that
is the answer no matter which fancy field redefinitions we might subject the Lagrangian to
and no matter which ugly gauge we could imagine choosing. It is precisely the point of the
modern approach to amplitudes to avoid such complications and get to an answer such as
(2.83) in a simple way.
Exercise 2.25
Rewrite the expression (2.82) to show that the 4-gluon amplitude can also be written
[34]4
A4 1− 2− 3+ 4+ = . (2.84)
[12][23][34][41]
Exercise 2.26
Convince yourself that in general if all helicities are flipped h i → −h i , then the result-
ing amplitude An [1h 1 2h 2 . . . n h n ] is obtained from An [1−h 1 2−h 2 . . . n −h n ] by exchanging
all angle and square brackets.
The color-ordered amplitudes have a number of properties worth noting:
1. Cyclic: It follows from the trace-structure that An [12 . . . n] = An [2 . . . n 1], etc.

2. Reflection: An [12 . . . n] = (−1)n An [n . . . 2 1]. Convince yourself that this is true.
3. The U (1) decoupling identity:
An [123 . . . n] + An [213 . . . n] + An [231 . . . n] + · · · + An [23 . . . 1 n] = 0 .
(2.85)
8 This can be seen by direct counting, but see also the analysis in [6].
The vanishing of this sum of n − 1 color-ordered amplitudes is also called the photon
decoupling identity; it follows from taking one of the generators T a proportional to
the identity matrix.
Exercise 2.27
Use (2.83) to show explicitly that (2.85) holds for n = 4 for the case where gluons 1
and 2 have negative helicity and 3 and 4 have positive helicity.
The trace-basis (2.74) is overcomplete and that implies that there are further linear relations
among the partial tree-level amplitudes: these are called the Kleiss–Kuijf relations [7, 8]
and they can be written [9]

An [1, {α}, n, {β}] = (−1)|β| An [1, σ, n] , (2.86)
σ ∈OP({α},{β T })
where {β T } denotes the reverse ordering of the labels {β} and the sum is over ordered
permutations “OP,” namely permutations of the labels in the joined set {α} ∪ {β T } such that
the ordering within {α} and {β T } is preserved. The sign on the RHS is determined by the
number of labels |β| in the set {β}.
To make (2.86) a little less intimidating, consider the 5-point case as an example. Taking
the LHS of (2.86) to be A5 [1, {2}, 5, {3, 4}], we have {α} ∪ {β T } = {2} ∪ {4, 3}, so the sum
over ordered permutations is over σ = {243}, {423}, {432}. Thus the Kleiss–Kuijf relation
reads
A5 [12534] = A5 [12435] + A5 [14235] + A5 [14325] . (2.87)
Exercise 2.28
Show that for n = 4, the Kleiss–Kuijf relation (2.86) is equivalent to the U (1) decou-
pling relation.
Exercise 2.29
Start with A5 [1, {2, 3}, 5, {4}] to show that the Kleiss–Kuijf relation gives
A5 [12345] + A5 [12354] + A5 [12435] + A5 [14235] = 0 . (2.88)
Show then that (2.88) together with the U (1) decoupling relation implies that (2.87).
The Kleiss–Kuijf relations combine with the cyclic, reflection, and U (1) decoupling identi-
ties to reduce the number independent n-gluon tree amplitudes to (n − 2)!. However, there
are further linear relationships, called the (fundamental) BCJ relations – named after Bern,
Carrasco, and Johansson [9] – that reduce the number of independent n-gluon color-ordered
tree amplitudes to (n − 3)!. Examples of 4-point and 5-point BCJ amplitude relations are
s14 A4 [1234] − s13 A4 [1243] = 0 , (2.89)
s12 A[21345] − s23 A[13245] − (s23 + s24 )A[13425] = 0 . (2.90)
In Chapter 13, we show that the number of independent color-ordered tree amplitudes
under Kleiss–Kuijf relations is (n − 2)! and we also discuss the origin of BCJ amplitude
relations.
Exercise 2.30
Use the Parke–Taylor formula (2.83) to verify (2.88), (2.89), and (2.90).
Exercise 2.31
Let us get a little preview of the BCJ relations. Suppose we use the color-basis (2.68)
to write the full 4-point gluon amplitude as
n s cs n t ct n u cu
Afull,tree
4 = + + (2.91)
s t u
for some numerator factors n i that in general depend on the kinematic variables and the
polarizations. Write each ci in terms of the three traces Tr(T a1 T a2 T a3 T a4 ) and those
with orderings 1243 and 1324. (Make sure to check that the Jacobi identity (2.69)
holds.) Then use your expressions to convert (2.91) to a basis with those three traces.
Now use the cyclic and reflection properties of the trace and the color-ordered
amplitudes to write the full amplitude Afull,tree
4 in (2.73) in terms of the traces with the
same three orderings 1234, 1243, and 1324.
Comparing the resulting expressions for Afull,tree
4 , read off the relationship between
the numerator factors n i and the color-ordered amplitudes. You should find
ns nu
A4 (1234) = − + (2.92)
s u
and two similar expressions for A4 (1243) and A4 (1324). Show that it follows directly
from these expressions that the color-ordered amplitudes satisfy the n = 4 photon
decoupling relation (2.85).
Note that the numerator factors n i are not unique. Suppose that there is a choice of
numerator factors n i that satisfy the same relation as the color factors ci ,
ns + nt + nu = 0 . (2.93)
Show that (2.93) implies that the color-ordered amplitudes satisfy the BCJ relation
(2.89).
The existence of numerator factors n i that satisfy the same identity (2.93) as the
corresponding color factors is called color-kinematics duality. It has been of great
interest and applicability in studies of amplitudes in both gauge theory and gravity, and
we will discuss it further in Chapter 13.
We end this section with a quick look at interactions between gluons and fermions.
Adding
g
L = i γ μ Dμ = i γ μ ∂μ + √ Aμ γ μ (2.94)
2
to the Yang–Mills Lagrangian (2.66), we now acquire a fermion–fermion–gluon 3-vertex

i √g2 γ μ . If the fermion represents a quark, transforms in the fundamental of the gauge
group SU (N ). In that case the trace-structure of the amplitudes is a little different, for
example for the case of scattering 2 quarks with n gluons, we get (T a1 T a1 . . . T an )i j .
If we want to study the interactions of gluons with their supersymmetric partners, the
gluinos, then the fermion field must transform in the adjoint so we replace with λ = λa T a
and include a trace in the Lagrangian. The trace-structure for gluon–gluino scattering is
exactly the same as for gluon scattering.
We have by now seen enough examples of how to use spinor helicity formalism in the
context of standard Feynman rules. It is about time that we get a little fancier. Therefore we
postpone further discussion of Yang–Mills and super Yang–Mills amplitudes until we have
developed a few more tools.
2.6 Little group scaling
We have introduced | p and | p] as solutions to the massless Weyl equation, p| p = 0 and

p| p] = 0 for p 2 = 0. Their relation to p μ was given in (2.16) as pa ḃ = −| p]a p|ḃ . It is
very useful to observe that these relations are invariant under the scaling
| p → t| p , | p] → t −1 | p] . (2.95)
This is called little group scaling. Recall that the little group is the group of transformations
that leave the momentum of an on-shell particle invariant. For a massless particle, we can
go to a frame where p μ = (E, 0, 0, E). Rotations in the x y-plane leave the vector invariant,
so the little group representations9 are characterized by S O(2) = U (1). In the angle and
square spinor representation of the momentum, the little group transformation is realized
as the scaling (2.95): for real momenta, t has to be a complex phase such that | p]∗ = p| is
preserved. For complex momenta, the angle and square spinors are independent so we can
be more generous and let t be any non-zero complex number.
Now let us consider what an amplitude is made of: each Feynman diagram consists of
propagators, vertices, and external line rules. When only massless particles are involved,
the amplitude can always be rewritten in terms of angle and square brackets. But note that
neither propagators nor vertices can possibly scale under little group transformations. Only
the external line rules scale under (2.95):
• The scalar rule is a constant factor 1: it does not scale under (2.95).
• The wavefunctions for (Weyl) fermions are angle (h = −1/2) and square spinors
(h = +1/2): they scale as t −2h .
9 More precisely, the little group is E(2), the group of transformations that map a 2d plane into itself. This is
similar to the more familiar SU (2) group, whose generators J+ and J− can be identified as the two translation
generators of the little group and Jz can be identified with the rotation generator. Thus, just as in SU (2), where
representations are characterized by their Jz eigenvalue, representations of the little group E(2) are characterized
by their spin under the 2-dimensional rotation group S O(2) = U (1).
• Polarization vectors for spin-1 bosons: you can directly check (2.50) to see that under
μ
little group scaling of | p and | p], the polarization vectors ± ( p; q) scale as t −2h for
h = ±1. They are invariant under scaling of the reference spinor.
Thus, for an amplitude of massless particles10 only, we have the following powerful result.
Under little group scaling of each particle i = 1, 2, . . . , n, the on-shell amplitude transforms
homogeneously with weight −2h i , where h i is the helicity of particle i:

An {|1, |1], h 1 }, . . . , {ti |i, ti−1 |i], h i }, . . . = ti−2h i An . . . {|i, |i], h i } . . . . (2.96)
132
As an example, consider the QED amplitude (2.56), A3 ( f − f¯+ γ − ) = ẽ 12
. For the neg-
−2(−1)
ative helicity photon (particle 3) we get = t32 t3 .
Likewise, one confirms the scaling
(2.96) for the two fermions. In fact, all massless 3-particle amplitudes are completely fixed
by little group scaling! Let us now see how.
3-particle amplitudes
Recall that by 3-particle special kinematics, (2.58) and (2.59), an on-shell 3-point amplitude
with massless particles depends only on either angle or square brackets of the external
momenta. Let us suppose that it depends on angle brackets only. We can then write a
general Ansatz

A3 1h 1 2h 2 3h 3 = c 12x12 13x13 23x23 , (2.97)
where c is some constant independent of the kinematics. Little group scaling (2.96) fixes
−2h 1 = x12 + x13 , − 2h 2 = x12 + x23 , − 2h 3 = x13 + x23 . (2.98)
This system is readily solved to find x12 = h 3 − h 1 − h 2 etc. so that

A3 1h 1 2h 2 3h 3 = c 12h 3 −h 1 −h 2 13h 2 −h 1 −h 3 23h 1 −h 2 −h 3 . (2.99)
This means that the helicity structure uniquely fixes the 3-particle amplitude up to an overall
constant! This may remind you of a closely related fact, namely that in a conformal field
theory, the 3-point correlation functions are determined uniquely (up to a multiplicative
constant) by the scaling dimensions of the operators.
We have already confirmed (2.99) for A3 ( f − f¯+ γ − ). So let us do something different.
Consider a 3-gluon amplitude with two negative and one positive helicity gluons. By (2.99),
the kinematic structure is uniquely determined:

123
A3 g1− g2− g3+ = g . (2.100)
1323
μν ν;μ
10 For a spin-2 graviton, the helicity ±2 states are encoded in polarizations that can be chosen as e± = ± ±
this choice automatically makes the polarizations symmetric and traceless. For a spin-3/2 gravitino, the wave-
μ
functions can be taken to be v± ± . It follows from these external line rules that gravitons and gravitinos also
obey the scaling t −2h under little group scaling.
This matches our calculation (2.80). We have included here the Yang–Mills coupling g
for the purpose of the discussion to follow. Note how the little group argument fixes the
kinematic form of 3-gluon amplitude; this is much simpler than the direct Feynman diagram
calculation we did in Section 2.5.
But – there is perhaps a small glitch in our little group scaling argument. We assumed
that the amplitude depended only on angle brackets. What if it only depended on square
brackets? Then the scaling would have been the opposite, so we would have found

[13][23]
A3 g1− g2− g3+ = g . (2.101)
[12]3
To distinguish between (2.100) and (2.101), we use dimensional analysis. From (2.24)
we note that both angle and square brackets have mass-dimension 1. Thus the momentum
dependence in (2.100) is (mass)1 ; this is compatible with the fact that it comes from the
A A∂ A-interaction in Tr Fμν F μν . However, in (2.101), the momentum dependence has
mass-dimension (mass)−1 , so it would somehow have to come from an interaction of the
form g A A ∂ A. Of course, we have no such interaction term in a local Lagrangian; hence
we discard the expression (2.101) as unphysical.
The combination of little group scaling and locality uniquely fixes the massless 3-particle
amplitudes. As we will see in Chapter 3, in some theories the 3-particle amplitudes can
actually determine all other tree-level amplitudes!
While we are considering dimensional analysis, it is worth making a couple of other
observations. First, note that while the Yang–Mills coupling g is dimensionless, the coupling
g in the g A A ∂ A has dimension (mass)2 . This means that the RHS of (2.101) has mass-
dimension 1, just as the correct expression (2.100). This is sensible since the two amplitude-
expressions had better have the same mass-dimension. In general,
an n-particle amplitude in D = 4 must have mass-dimension 4 − n. (2.102)
This follows from dimensional analysis since the cross-section must have dimensions of
area. You can also check it by direct inspection of the Feynman diagrams.
One more comment about (2.100): you may be worried that the expression on the RHS
is not Bose-symmetric in the exchange of the identical gluons 1 and 2. Fear not. The full
3-point amplitude of course comes dressed with a fully-antisymmetric group theory factor
f a1 a2 a3 : this restores Bose symmetry. As discussed in Section 2.5, the kinematic structure
in (2.100) is exactly that of the color-ordered 3-point amplitude A3 [1− 2− 3+ ].
Exercise 2.32
Write down spinor helicity formulas for the possible color-ordered 3-point amplitudes
with two gluinos (massless spin-1/2 in the adjoint representation of the gauge group)
and one gluon. The result will encode the gluino–gluino–gluon interactions in super
Yang–Mills theory: work “backwards” to find out what that interaction term looks like
in the Lagrangian.
Exercise 2.33
Let us play a little game. Suppose someone gives you the following amplitudes for
scattering processes involving massless particles:
[13]4
(a) A5 = ga , (2.103)
[12][23][34][45][51]
14242
(b) A4 = gb , (2.104)
122334
127 [12]
(c) A4 = gc . (2.105)
13142324342
With all particles outgoing, what are the helicities of the particles?
What is the dimension of the couplings gi relevant to the interactions?
In each case, try to figure out which theory could produce such an amplitude.
Example. What about a gluon amplitude with all-negative helicities? Well, let us do it.
The formula (2.99) immediately tells us that

A3 g1− g2− g3− = a 121323 . (2.106)
The kinematic part has mass-dimension 3 and this reveals that: (i) the coupling a must
have mass-dimension −2 for the whole amplitude to have mass-dimension 4 − 3 = 1;
and (ii) this must come from a Lagrangian interaction term with three derivatives,
i.e. (∂ A)3 . Furthermore, the kinematic terms are antisymmetric under exchanges of
gluon-momenta, so Bose symmetry tells us that the couplings must be associated
with antisymmetric structure constants – as is of course the case for a non-abelian
gauge field. Thus, there is a natural candidate, namely the dimension-6 operator
Tr F μ ν F ν λ F λ μ . Indeed this operator produces the amplitude (2.106). We can also
conclude that in pure Yang–Mills theory or in QED, A3 (g1− g2− g3− ) = 0.
Exercise 2.34
Let us look at gravity scattering amplitudes. If we expand the Einstein–Hilbert action
√
1
2κ 2
d 4 x −g R around flat space gμν = ημν + κ h μν , we obtain an infinite series of
2-derivative interactions involving n fields h μν for any n. This makes it very compli-
cated to calculate graviton scattering amplitudes using Feynman rules. (Gravitons are
massless spin-2 particles; in 4d they have two helicity states, h = ±2.) For now just
focus on the 3-point amplitude: use little group scaling to write down the result for
the on-shell 3-graviton amplitudes. Check the mass-dimensions. Compare your answer
with the 3-gluon amplitudes. Graviton scattering is discussed in much further detail
in Chapter 12.
Exercise 2.35
Consider in gravity an operator constructed from some contraction of the indices of
three Riemann-tensors; we denote it schematically as R 3 . If we linearize the met-
ric around flat space, gμν = ημν + κ h μν , then we can calculate graviton scattering
associated with R 3 . What is the mass-dimension of the coupling associated with R 3 ?

Use little group scaling to determine A3 (h − − − − − +
1 h 2 h 3 ) and A3 (h 1 h 2 h 3 ). Compare the
expressions with 3-gluon scattering processes.
Exercise 2.36
Consider a dimension-5 Higgs-gluon fusion operator H Tr Fμν F μν , where H is a
Higgs scalar field. Use little group scaling to determine the 3-particle amplitudes of
this operator in the limit of a massless Higgs, m H = 0. (For more about on-shell
methods and Higgs-gluon fusion, see [10].)
Example. Consider a 3-point amplitude with three scalars. We learn from (2.99) that
there can be no momentum dependence in the amplitude, A3 (φφφ) = constant. This
is of course compatible with a φ 3 -interaction, but what about throwing in some deriva-
tives, as in a non-linear sigma model? – something like φ ∂μ φ ∂ μ φ. The Feynman rules
for this 3-scalar vertex give a vertex ∼ ( p1 · p2 + p2 · p3 + p3 · p1 ) which on-shell is
proportional to (s + t + u) = 0. So the on-shell 3-point amplitude corresponding to
φ ∂μ φ ∂ μ φ vanishes. We can see this directly from the Lagrangian since the interaction
term is readily rewritten as 12 ∂μ (φ 2 ) ∂ μ φ, which by partial integration gives − 12 φ 2 φ.
This clearly vanishes on-shell for a massless scalar that obeys the Klein–Gordon
equation φ = 0.
Exercise 2.37
Now your turn: why does the 3-particle on-shell amplitude for three distinct massless
scalars, e.g. φ1 ∂μ φ2 ∂ μ φ3 , vanish? Try to see how this works both from the amplitude
perspective and from the Lagrangian.
There is a lesson to learn from the above example. The expression for an amplitude can
be manipulated using momentum conservation and on-shell conditions pi2 = 0. If the
amplitude is the on-shell matrix element of a (local) operator in the Lagrangian, then these
manipulations correspond to a rewriting of the operator by, respectively, dropping terms
that are total derivatives or vanish on-shell (i.e. on the equations of motion).
2.7 MHV classification
In this section we return to the study of gluon scattering amplitudes. The Yang–Mills
Lagrangian contains two types of interaction terms: schematically
Tr Fμν F μν −→ A A∂ A + A4 . (2.107)
In a typical gauge, such as a Feynman gauge or Neveu–Gervais gauge (2.67), this gives rise
to Feynman rules with two types of interaction vertices: the cubic vertex which depends
linearly on momentum and the quartic vertex which is momentum-independent. Since the
Yang–Mills coupling is dimensionless, the cubic vertex is O(mass1 ) and the quartic is
O(mass0 ).
Consider tree diagrams with only cubic vertices, i.e. trivalent tree-graphs, with n external
legs. If you start with a 3-point vertex (n = 3) you can easily convince yourself that every
time you add an extra external line, you have to add both a new vertex and a new propagator
to keep the graph trivalent. Hence the numbers of vertices and propagators both grow
linearly with n, and it takes just a few examples to see that the number of vertices is n − 2
and the number of propagators is n − 3. Since the cubic vertices are O(mass1 ) and the
propagators are O(mass−2 ), we find that the mass-dimension of the diagrams, and hence
of the tree amplitude, is
(mass)n−2
An ∼ ∼ (mass)4−n . (2.108)
(mass2 )n−3
This confirms the statement (2.102) that n-point amplitudes in D = 4 dimensions have
mass-dimension 4 − n. Any diagram with a mix of cubic and quartic vertices has the same
mass-dimension of (mass)4−n . But note that the number of powers of momenta in the
numerator cannot exceed n − 2; this point will be useful shortly.
Consider now the schematic form of a gluon tree amplitude:

(i . j ) (i .k j ) (ki .k j )
An ∼ 2 . (2.109)
diagrams
PI
The ki s stand for various sums of external momenta p j . Each diagram has a numerator that
is a polynomial in Lorentz scalar products of polarizations and momentum vectors. The
denominators are products of momentum invariants from the propagators.
Perhaps you know the statement that in pure Yang–Mills theory all-plus tree gluon
amplitudes vanish, An (1+ 2+ . . . n + ) = 0? We have already observed this for n = 3, 4 in
Exercise 2.23 and in the example (2.106). Let us show it for all n. First recall from Exercise
2.21 that the polarization vector dot-products are
i+ · j+ ∝ qi q j , i− · j− ∝ [qi q j ] , i− · j+ ∝ iq j [ jqi ] . (2.110)
Suppose we choose all qi to be the same q. Then i+ · j+ = 0. This means that the only
non-vanishing way for the positive helicity gluon polarizations to enter in the numerator of
(2.109) is as i+ · k j . For an all-plus gluon amplitude, we need to absorb the Lorentz indices
of all n polarization vectors this way, so that requires n powers of momenta in the numerator.
But as we argued below (2.108), no more than n − 2 powers of momenta are possible in
the numerator of any gluon tree diagram. Hence we conclude that An (1+ 2+ . . . n + ) = 0.
If we had not known that we should write down a smart choice of the polarization vectors,
but had worked with general expressions, we would have had to work very hard to prove
that the sum of combinatorially many n-point tree diagrams in the all-plus amplitude adds
up to zero.
Next, let us flip one of the helicities and consider an amplitude An (1− 2+ . . . n + ). This
time, choose q2 = q3 = · · · = qn = p1 . This achieves i+ · j+ = 0 and 1− · j+ = 0. So
again we would need n factors of i+ · k j in the numerators of (2.109) and as before we
conclude that the tree-level amplitude vanishes: An (1− 2+ . . . n + ) = 0.
We have shown that

tree-level gluon amplitudes: An 1+ 2+ . . . n + = 0 and An 1− 2+ . . . n + = 0 .
(2.111)
The same holds with all helicities flipped:

tree-level gluon amplitudes: An 1− 2− . . . n − = 0 and An 1+ 2− . . . n − = 0 .
(2.112)
In supersymmetric Yang–Mills theories, the results (2.111) and (2.112) hold true at any
loop-order. In fact, supersymmetry provides an elegant argument for the vanishing of these
so-called “helicity violating” amplitudes. We present this derivation in Chapter 4.
In pure Yang–Mills theory, the amplitudes (2.111) and (2.112) are actually non-vanishing
at loop-level (and have a quite interesting structure which is briefly discussed in Section
6.1). Can you see how the argument based on (2.109) and (2.110) is changed at 1-loop
level?
Let us move on and flip one more helicity: An (1− 2− 3+ . . . n + ). We choose the reference
qi s to maximize the number of vanishing dot-products of polarization vectors. The choice
q1 = q2 = pn and q3 = q4 = · · · = qn = p1 implies that all i · j = 0 vanish, except 2− ·
i+ for i = 3, . . . , n − 1. The polarization vector of gluon 2 can only appear once, so the
terms in (2.109) can take the schematic form

− − + (2− · i+ )( j · kl )n−2
+
An 1 2 3 . . . n ∼ 2 . (2.113)
diagrams
PI
Since only one product of iμ s can be non-vanishing, n − 2 factors of ( j · kl ) were needed,
and this exactly saturates the number of momentum vectors possible by dimensional analysis
(2.108). Note also that with our choice of polarization vectors, any diagram that contributes
to An (1− 2− 3+ . . . n + ) is trivalent.
Thus we conclude – based on dimensional analysis and thoughtful choices of the po-
larization vectors – that An (1− 2− 3+ . . . n + ) is the “first” gluon tree amplitude that can be
non-vanishing, in the sense that having fewer negative helicity gluons gives a vanishing
amplitude. More negative helicity states are also allowed, but one needs at least two positive
helicity states to get a non-vanishing result. The only exception is for n = 3, as we have
seen in (2.81).
The gluon amplitudes with two negative helicities and n − 2 positive helicities are called
maximally helicity violating – or simply MHV for short. The name “maximally helicity
violating” comes from thinking of 2 → (n − 2) scattering. By crossing symmetry,

an outgoing gluon with negative helicity is an incoming gluon with positive helicity.
positive negative
So with all outgoing particles, the process An [1+ 2+ 3+ . . . n + ] crosses over to 1− 2− →

3+ . . . n + in which the outgoing states all have the opposite helicity of the incoming
states; it is “helicity violating.” The process 1− 2− → 3− 4+ . . . n + is a little less helicity

violating and it crosses to An [1+ 2+ 3− 4+ . . . n + ]. We know from the above analysis that
both these helicity violating processes vanish at tree-level in pure Yang–Mills theory. The
process 1− 2− → 3− 4− 5+ . . . n + – equivalent to An [1+ 2+ 3− 4− 5+ . . . n + ] – is the most
we can “violate” helicity and still get a non-vanishing answer at tree-level: therefore it is
maximally helicity violating. As we have learned, the MHV gluon tree amplitudes are given
by the Parke–Taylor formula,
i j4
An 1+ . . . i − . . . j − . . . n + = . (2.114)
1223 · · · n1
The MHV amplitudes play a key role in many studies of scattering amplitudes in (super)
Yang–Mills theory. They are the simplest amplitudes in Yang–Mills theory. The next-to-
simplest amplitudes are called Next-to-MHV, or NMHV, and this refers to the class of
amplitudes with three negative helicity gluons and n − 3 positive helicity gluons. This
generalizes to the notation N K MHV amplitudes with K +2 negative helicity gluons and
n− K −2 positive helicity gluons. When an amplitude has (n−2) gluons of negative helicity
and two of positive helicity, it is called anti-MHV. Anti-MHV amplitudes are obtained from
the MHV amplitudes by flipping all helicities and exchanging angle and square brackets:
[i j]4
An 1− . . . i + . . . j + . . . n − = . (2.115)
[12][23] · · · [n1]
This is the anti-MHV Parke–Taylor formula.

Finally, let us mention that the MHV classification also applies to gravity amplitudes.
We encountered the 4-graviton amplitude in Exercise 2.33 and saw how little group scaling
and dimensional analysis fixed the 3-graviton amplitude in Exercise 2.34. In gravity, an
N K MHV graviton amplitude has K +2 negative helicity gravitons and N − K −2 positive
helicity gravitons. Graviton amplitudes will show up occasionally throughout the book, but
are treated more systematically in Chapter 12.
2.8 Analytic properties of tree amplitudes
A tree-level amplitude is a rational function of Lorentz invariants, such as pi μ p μj and pi μ μj ,

and it can have additional structure from color factors. In the spinor helicity formalism, a
tree amplitude of massless particles is a rational function of the (complex-valued) kinematic
invariants i j and [i j], it scales homogeneously under little group scaling (2.96), and it
has definite mass-dimension (2.102). But these are not the only properties of an amplitude:
in addition, the physics places constraints on the analytic structure of the rational function.
Tree amplitudes must have a certain pole structure in order to legitimately describe a
physical scattering process. Locality tells us that any pole of a tree-level amplitude must
correspond to a propagating particle going on-shell, in other words all interactions are
either local or mediated by a physical particle. This can be traced back to the locality of the
Lagrangian. Its locality implies that there exists a choice of fields and a gauge such that the
Feynman rules give interaction vertices that are constants or polynomial in the momenta,
but which do not have any poles. Thus, in such a gauge, the only poles in the tree amplitude
come from the propagators; examples are the Feynman gauge or Gervais–Neveu gauge in
the standard formulation of Yang–Mills theory. Since the amplitude is gauge-independent,
it must be true in any gauge that the only poles are those that arise from propagators going
on-shell, i.e. from the exchanges of physical particles.11 Thus the only poles that can appear
in a tree amplitude of massless particles are of the form
j 1
−→ . (2.116)
( pi + p j + · · · + pk )2
k
For color-ordered tree amplitudes, the propagators contain only momenta of consecutive
external legs, i.e.
1
. (2.117)
( pi + pi+1 + · · · + pi+m )2
μ
It may appear that the above statements hold only if we use formal polarization vectors i ;
for if we express them in terms of spinor variables as in (2.50), then one would invariably
introduce factors such as pq or [ pq] in the denominators of the Feynman diagrams. While
this is indeed true, poles from such factors depend on the choice of reference spinors q,
so this means that the location of such a pole is gauge dependent. Since the amplitude is a
gauge-invariant quantity, this dependence must drop out in the sum over all diagrams that
contribute to the tree amplitude. In other words, the residue of such a pole must vanish in
the full tree amplitude. It is an example of a spurious pole, i.e. an unphysical pole that does
not correspond to an exchange of a physical particle.
Field redefinitions and (obscure) gauge choices can give rise to expressions for the
amplitudes in which the individual Feynman diagrams have unphysical poles. Of course,
the amplitude is independent of field redefinitions and gauge choices, so the residues of
such spurious poles must cancel among the diagrams to give zero. In the spinor helicity
formalism, spurious poles arise from denominator factors of the form
1
. (2.118)
[i| pi + · · · + p j |k
In the amplitude such spurious poles must have vanishing residue.
11 We encountered locality in Section 2.6 when we discussed the constraints on 3-point amplitudes from little
group scaling, dimensional analysis, and locality of the Lagrangian. Also, when we discussed the helicity
structure of gluon amplitudes in Section 2.7, we assumed that we were using a standard gauge, such as the
Feynman gauge or Gervais–Neveu gauge, in which the only poles come from the propagators and the Feynman
diagrams are of the form (2.109).
Note that the locality requirement does not exclude poles of the form 1/i j or
1/[i j], where i and j are external legs. This is because we can always write 1/i j =
[i j]/( pi + p j )2 and 1/[i j] = i j/( pi + p j )2 . The exception is the 3-particle amplitudes;
but special 3-particle kinematics implies that the apparent poles in, for example, the Parke–
123
Taylor expression 2331 have vanishing residues. Thus, the locality requirement for a tree
amplitude is that any pole with non-vanishing residue must correspond to the propagator
of a physical particle going on-shell.
There are special limits of the external momenta in which tree amplitudes behave in
a well-defined way. These are soft and collinear singularities. Let us consider them in
turn.
Soft singularities
Soft singularities occur when the momentum of a massless particle goes to zero. Feynman
diagrams of the form
(2.119)
Soft
develop a singularity in this limit because the propagator, shown in gray, goes on-shell. As
all other diagrams are finite, the amplitude becomes dominated by this subset of diagrams.
With the propagator going on-shell, the leading part of the amplitude factorizes into a
singular factor times an (n−1)-point amplitude.
+ −
As an example, consider the color-ordered amplitude An [. . . , f i−1 , i − , f i+1 , . . .] in
Yang–Mills theory coupled to adjoint fermions f . Particle i is a negative-helicity gluon
and i −1 and i +1 are fermions of helicity +1/2 and −1/2, respectively. In the soft-gluon
limit, pi → 0, the diagrams relevant for the soft-gluon singularity are12
i+1
i+1 i
+
i
i−1 i−1
i + 1, i[q, i + 1] [i − 1, q]i, i − 1
= − An−1 + An−1
[qi] si,i+1 [qi] si,i−1
[i + 1, i − 1]
= − × An−1 . (2.120)
[i + 1, i][i, i − 1]
+ −
The fermion external wavefunctions for legs i −1 and i +1 in An−1 [. . . , f i−1 , f i+1 , . . .]
come from the numerator of the fermion propagator. Note how the reference spinor |q]
of the gluon polarization vector drops out in the sum to leave the overall singular factor
[i+1,i−1]
[i+1,i][i,i−1]
gauge-independent. This factor is called the soft function (or soft factor), and
12 Dropping coupling constants and color factors.

it is singular since the spinor |i] is near-zero. In general, for soft gluons in Yang–Mills
theory, the amplitude behaves as
pi →0 ±

An . . . , i − 1, i ± , i + 1, . . . −−−−→ Si−1,i,i+1 × An−1 . . . , i − 1, i + 1, . . . ,
(2.121)
±
where the soft functions Si−1,i,i+1 depend on the helicity of the soft gluon and are
− [i + 1, i − 1] + i + 1, i − 1
Si−1,i,i+1 =− , Si−1,i,i+1 = . (2.122)
[i + 1, i][i, i − 1] i + 1, ii, i − 1
The soft functions are independent of the helicities of the adjacent “hard” legs; the helicities
of the hard particles are inherited by An−1 .
Exercise 2.38
As a test to see that the soft function is indeed independent of the helicities of the hard
legs, replace the adjoint fermions with two adjoint scalars and show that the same soft
function (2.122) is produced in the soft-gluon limit.
Exercise 2.39
Show that the MHV amplitude (2.114) has the correct soft behavior (2.121).
Note that (2.121) and (2.122) were derived from Feynman rules. If you suspect that the
behavior of the soft limit is not universal but depends on the theory at hand, you are certainly
correct. Consider the 6-point amplitude in λφ 4 theory. The single soft-scalar limit does not
lead to a singularity, because no internal propagator goes on-shell, as is clear from diagram
(a):
(2.123)
(a) (b)
However, if we take a double-soft limit in which two scalars simultaneously go soft, as in

diagram (b) above, one indeed finds a propagator singularity and the result is a singular
factor times a residual 4-scalar amplitude.
To summarize, the tree-level amplitude must be a rational function of kinematic invariants
such that only propagator-like singularities are present. For Yang–Mills theory coupled to
matter fields, the amplitude must factorize as (2.121) when a gluon goes soft. Thus soft
limits are useful for testing that the result of an amplitude calculation is indeed correct.
Collinear singularities
Another type of amplitude singularity arises when two external momenta become collinear.
Diagrams in which the two external legs attach to a common 3-vertex are then divergent
because the internal propagator goes on-shell:
i+1 P
(2.124)
i
In the collinear limit, the momentum configuration can be parameterized as

pi = (1 − z) p P , pi+1 = z p P , (2.125)
where p P is on-shell, p 2P = 0. The parameter z cancels in the sum pi + pi+1 , so momentum
conservation for the original n on-shell momenta, p1 + p2 + · · · pi + pi+1 + · · · + pn =
0, becomes momentum conservation for n − 1 on-shell momenta p1 + p2 + · · · + p P +
· · · + pn = 0, as appropriate for an (n−1)-point amplitude.
Let
us examine the collinear limit of the MHV amplitude (2.114). Consider
An 1− 2− 3+ . . . n + and take legs 1 and n to be collinear:
123 1||n z2 P23
An 1− 2− 3+ . . . n + = −−→ ,
23 · · · n1 z(z − 1)n1 23 · · · n − 1, P
(2.126)
√ √
where |1 = z|P, |n = 1 − z|P (and similarly for the square spinors) using (2.125).
Thus, in the collinear limit, the n-point MHV amplitude factorizes into the (n−1)-point
amplitude An−1 [P − 2− 3+ . . . (n−1)+ ] multiplied by a singular splitting function

z2
Split+ n + , 1− = . (2.127)
z(1 − z) n1
We write this as
1||n

An 1− 2− 3+ . . . n + −−→ Split+ n + , 1− An−1 P − 2− 3+ . . . (n−1)+ . (2.128)
This splitting function is singular because n1 is zero in the collinear limit. Note that –
unlike the soft function – the splitting function depends on the helicities of the collinear
particles as well as the helicity of the particle with momentum P.
Exercise 2.40
For the MHV amplitude, show that the splitting function for the collinear limit with
two adjacent positive helicity gluons i and i + 1 is

1
Split− i + , (i +1)+ = . (2.129)
z(1 − z) i, i + 1
Show that the results (2.128) and (2.129) are consistent with little group scaling.
For more general splitting functions, see the review article [11] and references therein.
At this stage, you might have noticed that all singularities of the MHV gluon amplitude
precisely correspond to 2-particle collinear singularities. Why are there no multi-particle
singularities of the form (2.117)? The reason is simply that the “residue” of such a pole
should factorize into two amplitudes with n ≥ 4-point, since the intermediate momentum
becomes on-shell:
. (2.130)
However, for the two gluon amplitudes in the residue to be non-vanishing, each must have
two negative helicity gluons among its external states. The original n lines have two gluons
with negative helicity and we can only get one from the on-shell internal line in (2.130).13
So this means that there are not enough negative helicity gluons for the factorization
channel (2.130) to have non-vanishing residue for an MHV amplitude. Hence the MHV
amplitude only has the 2-particle factorization channels for the collinear limits. For Nk MHV
amplitudes, multi-particle poles do appear.
The factorization property of tree-level scattering amplitudes has important and inter-
esting consequences. Given that the residues of multi-particle poles are products of lower-
point on-shell amplitudes, can one use this to set up a recursive method for calculating
higher-point amplitudes? Indeed, this idea is central to the on-shell approach to calculate
amplitudes. This is the next subject: go ahead to Chapter 3.
13 A negative helicity gluon going out from the left subamplitude of the diagram (2.130) is, via crossing symmetry,
a positive helicity gluon when viewed as an outgoing state from the right subamplitude.
3 On-shell recursion relations at tree-level
Recursion relations provide a method for building higher-point amplitudes from lower-point
information. The 1988 off-shell recursion relations by Berends and Giele [12] construct
n-point parton amplitudes from building blocks that are lower-point amplitudes with one
leg off-shell (see the review articles [1, 3]). This off-shell method continues to be useful as
an algorithm for efficient numerical evaluation of scattering amplitudes. In this review, we
focus on the “modern” recursive methods whose building blocks are on-shell amplitudes.
These on-shell recursion relations are elegant in that they use input only from gauge-
invariant objects and they have proven very powerful for elucidating the mathematical
structure of scattering amplitudes.
In the on-shell approaches, a key idea is to use the power of complex analysis to exploit the
analytic properties of on-shell scattering amplitudes. The derivation of on-shell recursion
relations is a great example of this, as we shall see in Section 3.1. The most famous on-shell
recursion relations are the BCFW recursion relations of Britto, Cachazo, Feng, and Witten
[13, 14], but there are other versions based on the same idea as BCFW, namely the use
of complex deformations of the external momenta. We describe the idea here, first in a
very general formulation in Section 3.1, then specialize the results to derive the BCFW
recursion relations in Section 3.2. We illustrate the BCFW method with a selection of
examples, including an inductive proof of the Parke–Taylor formula (2.83). Section 3.3
contains a discussion of when to expect existence of recursion relations in general local
QFTs. Finally, in Section 3.4, we present the CSW construction (Cachazo-Svrcek-Witten
[15]), also called the MHV vertex expansion.
3.1 Complex shifts and Cauchy’s theorem
An on-shell amplitude An is characterized by the momenta of the external particles and their
type (for example a helicity label h i for massless particles). We focus here on massless par-
n
ticles so pi2 = 0 for all i = 1, 2, . . . , n. Of course, momentum conservation i=1 piμ = 0
is also imposed.
Let us now introduce n complex-valued vectors riμ (some of which may be zero) such
that

n
μ
(i) ri = 0 ,
i=1
(ii) ri · r j = 0 for all i, j = 1, 2, . . . , n. In particular ri2 = 0 , and
(iii) pi · ri = 0 for each i (no sum).
3.1 Complex shifts and Cauchy’s theorem 51
These vectors ri are used to define n shifted momenta
μ μ μ
p̂i ≡ pi + z ri with z ∈ C . (3.1)
Note that
n μ
(A) By property (i), momentum conservation holds for the shifted momenta: i=1 p̂i = 0.
(B) By (ii) and (iii), we have p̂i = 0, so each shifted momentum is on-shell.
2
μ μ
(C) For a non-trivial1 subset of generic momenta { pi }i∈I , define PI = i∈I pi . Then P̂I2
is linear in z:
2
P̂I2 = p̂i = PI2 + z 2 PI · R I with RI = ri , (3.2)
i∈I i∈I
because the z 2 term vanishes by property (ii). We can write
PI2 PI2
P̂I2 = − (z − z I ) with zI = − . (3.3)
zI 2PI · R I
As a result of (A) and (B), we can consider our amplitude An in terms of the shifted
μ μ
momenta p̂i instead of the original momenta pi . In particular, it is useful to study the
shifted amplitude as a function of z; by construction, it is a holomorphic function Ân (z).
μ
The amplitude with unshifted momenta pi is obtained by setting z = 0, An = Ân (z = 0).
We specialize to the case where An is a tree-level amplitude. In that case, the analytic
structure of Ân (z) is very simple. As discussed in Section 2.8, the tree amplitude does not
have any branch cuts – there are no logs, square-roots, etc. at tree-level. The tree amplitude
is a rational function of the kinematic variables. As such, its analytic structure is captured
by its poles and these are determined by the exchanges of physical particles. Thus, the
shifted amplitude Ân (z) is a rational function of z and, for generic external momenta, its
poles are all simple poles in the z-plane. To see that the poles are simple, consider the
Feynman diagrams: the only way we can get poles is from the shifted propagators 1/ P̂I2 ,
where P̂I is a sum of a non-trivial subset of the shifted momenta. By (C) above, 1/ P̂I2 gives
a simple pole at z I , and for generic momenta, PI2 = 0 so z I = 0. For generic momenta,
no Feynman tree diagram can have more than one power of a given propagator 1/ P̂I2 ; and
poles of different propagators are at different locations in the z-plane. Hence, for generic
momenta, Ân (z) only has simple poles and they are all located away from the origin z = 0.
Note the underlying assumption of locality, i.e. that the amplitudes can be derived from
some local Lagrangian so that the propagators determine the poles.
Let us now study the holomorphic function Ânz(z) in the complex z-plane. Pick a contour
that surrounds the simple pole at the origin. The residue at this pole is nothing but the
unshifted amplitude, An = Ân (z = 0). Deforming the contour to surround all the other
1 Non-trivial means at least two and no more than n −2 momenta such that PI2 = 0.
poles, Cauchy’s theorem tells us that

Ân (z)
An = − Resz=z I + Bn , (3.4)
zI
z
where Bn is the residue of the pole at z = ∞. By taking z → 1/w it is easily seen that Bn
is the O(z 0 ) term in the z → ∞ expansion of An .
Now, so what? Well, at a z I -pole the propagator 1/ P̂I2 goes on-shell. In that limit, the
shifted amplitude factorizes into two on-shell parts,
z near z I 1 zI 1
Ân (z) −−−−→ ÂL (z I ) ÂR (z I ) = − ÂL (z I ) 2 ÂR (z I ) . (3.5)
P̂I2 z − zI PI
In the second step we used (3.3). This makes it easy to evaluate the residue at z = z I :
^ PÎ ^
Ân (z) 1 ^ ^.
−Resz=z I = ÂL (z I ) 2 ÂR (z I ) = L R (3.6)
z PI
^ ^
Note that – as opposed to Feynman diagrams – the momentum of the internal line in (3.6)
is on-shell, P̂I2 = 0, and the vertex-blobs represent shifted on-shell amplitudes evaluated
at z = z I ; we call them subamplitudes. The rule for the internal line in the diagrammatic
representation (3.6) is to write the scalar propagator 1/PI2 of the unshifted momenta. Each
subamplitude necessarily involves fewer than n external particles, hence all the residues at
finite z can be determined in terms of on-shell amplitudes with less than n particles. This
is the basis of the recursion relations.
The contribution Bn from the pole at infinity has no similar general expression in terms
of lower-point amplitudes; there has been some work on how to compute Bn systematically
(see for example [16, 17]), but currently there is not a general constructive method. Thus,
in most applications, one assumes – or, much preferably, proves – that Bn = 0. This is most
often justified by demonstrating the stronger statement that
Ân (z) → 0 for z → ∞. (3.7)
If (3.7) holds, we say that the shift (3.1) is valid (or good).
For a valid shift, the n-point on-shell amplitude is completely determined in terms of
lower-point on-shell amplitudes as
1 ^^ PÎ ^
An = ÂL (z I ) 2 ÂR (z I ) = ^. (3.8)
L R
diagrams I
PI diagrams I ^
^
The sum is over all possible factorization channels I . There is also implicitly a sum over
all possible on-shell particle states that can be exchanged on the internal line: for example,
for a gluon we have to sum the possible helicity assignments.
The recursive formula (3.8) gives a manifestly gauge-invariant construction of scatter-

ing amplitudes. Thus (3.8) is the general “prototype” of the on-shell recursion relations
for tree-level amplitudes under a valid shift of the external momenta. We did not use
any special properties of 4d spacetime, so the general derivation of the recursion re-
lations is valid in D spacetime dimensions. In the following, we specialize to D = 4
again.
3.2 BCFW recursion relations
We shifted all external momenta democratically in (3.1), but with a parenthetical remark
μ μ
that some of the light-like shift-vectors ri might be trivial, ri = 0. The BCFW shift is one
in which exactly two lines, say i and j, are selected as the only ones with non-vanishing
shift-vectors. In D = 4 spacetime dimension, the shift is implemented on angle and square
spinors of the two chosen momenta:
|î] = |i] + z | j] , | ĵ] = | j] , |î = |i , | ĵ = | j − z|i . (3.9)
No other spinors are shifted. We call this a [i, j-shift. Note that [îk] and ĵk are lin-
ear in z for k = i, j while î ĵ = i j, [î ĵ] = [i j], îk = ik, and [ ĵk] = [ jk] remain
unshifted.
Exercise 3.1
μ μ
Use (2.15) to calculate the shift vectors ri and r j corresponding to the shift (3.9).
Then show that your shift vectors satisfy properties (i)–(iii) of Section 3.1.
With the two momenta i and j shifted according to (3.9), the BCFW recursion relation for
tree amplitudes takes the form
1 î PÎ ^j
An = ÂL (z I ) Â R (z I ) = L R .
diagrams I
PI2 diagrams I
(3.10)
The sum is over all channels I such that the shifted lines i and j are on opposite sides of the
factorization diagram in (3.10). As in the general recursion relations (3.8), there is also an
implicit sum over all possible on-shell particle states that can be exchanged on the internal
line.
Before diving into applications of the BCFW recursion relations (such as proving the
Parke–Taylor formula), let us study the shifts a little further. As an example, consider the
Parke–Taylor amplitude
124
An 1− 2− 3+ . . . n + = . (3.11)
1223 · · · n1
Explore the following properties:
Exercise 3.2
Convince yourself that for large-z the amplitude (3.11) falls off as 1/z under a [−, −-
shift (i.e. choose i and j to be the two negative helicity lines). What happens under the
three other types of shifts? Note the difference between shifting adjacent/non-adjacent
lines.2
Exercise 3.3
Consider the action of a [1, 2-shift of (3.11). Identify the simple pole. Calculate the
residue of Ân (z)/z at this pole. Compare with (3.4). What happens if you try to repeat
this for a [1, 3-shift?
Exercise 3.4
The spin-1 polarization vectors (2.50) have denominators with qp and [qp] (with
q the reference spinor) which may shift under a BCFW shift (3.9) involving p. Why
are there no terms in the on-shell recursion relations (3.8) corresponding to poles at
q p̂ = 0 or [q p̂] = 0?
The validity of the BCFW recursion relations requires that the boundary term Bn in (3.4)
is absent. The typical approach is to show that the shifted amplitude vanishes in the limit
of large z, as in (3.7):
Ân (z) → 0 for z → ∞. (3.12)
In pure Yang–Mills theory, an argument [18] based on the background field method estab-
lishes the following large-z behavior of color-ordered gluon tree amplitudes under a BCFW
shift of adjacent gluon lines i and j of helicity as indicated:
[i, j [−, − [−, + [+, + [+, −

1 1 1 . (3.13)
Ân (z) ∼ z3
z z z
If i and j are non-adjacent, one gains an extra power 1/z in each case. Thus any one of the
three types of shifts [−, −, [−, +, [+, + gives valid recursion relations for gluon tree
amplitudes.
We are now going to use the BCFW recursion relations (3.10) to construct an inductive
proof of the Parke–Taylor formula (3.11). The formula (3.11) is certainly true for n = 3,
as we saw in Section 2.5, and this establishes the base of the induction. For given n,
suppose that (3.11) is true for amplitudes with less than n gluons. Then write down the
recursion relation for An [1− 2− 3+ . . . n + ] based on the valid [1, 2-shift: adapting from
2 Of course, we cannot use the large-z behavior of formula (3.11) itself to justify the method to prove this formula!
A separate argument is needed and will be discussed shortly.
(3.10), we have
1^− 2^−
− − + +

n
+ PÎ
An 1 2 3 . . . n = n
L R 3+
k=4
k+ k−1+

n
1
= Ân−k+3 1̂− , P̂Ih I , k + . . . , n +
k=4 h I =±
PI2

Âk−1 − P̂I−h I , 2̂− , 3+ . . . , (k − 1)+ . (3.14)
The internal momentum is labeled PI , meaning that for a given k = 4, . . . , n we have

PI = p2 + p3 + · · · + pk−1 and P̂I = p̂2 + p3 + · · · + pk−1 . There are no diagrams where
lines 1 and 2 belong to the same subamplitude, because in that case, the internal momentum
would not be shifted and then there is no corresponding residue in (3.4). Only diagrams
that preserve the color-ordering of the external states are included. Note that we are also
explicitly including the sum over the possible helicity assignments for the particle exchanged
on the on-shell internal line: if the exchanged gluon is outgoing from the left subamplitude
and has negative helicity, then it will be a positive helicity outgoing gluon as seen from the
right subamplitude.
Since one-minus amplitudes An [− + · · · +] vanish except for n = 3, (3.14) reduces to
1^− 2^− 1^− 2^−

PÎ + PÎ
An 1− 2− 3+ . . . n + = L + 3+ + n L R
− R − +
n+ n−1+ 4+ 3+
+
1 − − +
= Â3 1̂− , − P̂1n , n+ 2
Ân−1 P̂1n , 2̂ , 3 . . . (n − 1)+
P1n
−
1 +

+ Ân−1 1̂− , P̂23 , 4+ . . . , n + 2
Â3 − P̂23 , 2̂− , 3+ . (3.15)
P23
It is understood here that P̂I is evaluated at the residue value of z = z I such that P̂I2 = 0.
The notation Pi j means Pi j = pi + p j .
The next step is to implement special kinematics for the 3-point subamplitudes. In the
first diagram of (3.15), we have a 3-point anti-MHV amplitude
+
[ P̂1n n]3
Â3 1̂− , − P̂1n , n+ = . (3.16)
[n 1̂][1̂ P̂1n ]
Here we use the following convention for analytic continuation:
| − p = −| p , | − p] = +| p] . (3.17)
μ μ μ
Since P̂1n = p̂1 + pn , the on-shell condition is
0 = P̂1n
2
= 2 p̂1 · pn = 1̂n[1̂n] = 1n[1̂n] . (3.18)
For generic momenta, the only way for the RHS to vanish is for z to take a value such that
[1̂n] = 0. This means that the denominator in (3.16) vanishes! But so does the numerator:
from
| P̂1n [ P̂1n n] = − P̂1n |n] = −( p̂1 + pn )|n] = |1[1̂n] = 0 , (3.19)
we conclude that [ P̂1n n] = 0 since | P̂1n is not zero. Similarly, one can show that [1̂ P̂1n ] = 0.
Thus, in the limit of imposing momentum conservation, all spinor products in (3.16) vanish;
with the three powers in the numerator versus the two in the denominator, we conclude that
+
special 3-point kinematics force Â3 [1̂− , P̂1n , n + ] = 0. So the contribution from the first
diagram in (3.15) vanishes.
In the second diagram of (3.15), the 3-point subamplitude is also anti-MHV, but it does
not vanish, since the shift of line 2 is on the angle spinor, not the square spinor. This way,
the big abstract recursion formula (3.10) reduces – for the case of the [−, − BCFW shift of
an MHV gluon tree amplitude – to an expression with just a single non-vanishing diagram:
1^−
− − + PÎ 2^−
An 1 2 3 . . . n + = n+ L R
− +
4+ 3+
−
1 +

= Ân−1 1̂− , P̂23 , 4+ , . . . , n + 2
Â3 − P̂23 , 2̂− , 3+ . (3.20)
P23
Our inductive assumption is that (3.11) holds for (n − 1)-point amplitudes. That, together
with the result (2.81) for the 3-point anti-MHV amplitude, gives
1̂ P̂23 4 1 [3 P̂23 ]3

An 1− 2− 3+ . . . n + = × × . (3.21)
1̂ P̂23 P̂23 445 . . . n 1̂ 23[23] [ P̂23 2̂][2̂3]
We could now proceed to evaluate the angle and square spinors for the shifted momenta.
But it is more fun to introduce you to a nice little trick. Combine the factors from the
numerator:
1̂ P̂23 [3 P̂23 ] = −1̂ P̂23 [ P̂23 3] = 1̂| P̂23 |3] = 1̂|( p̂2 + p3 )|3] = 1̂| p̂2 |3]
= −1̂2̂[2̂3] = −12[23] . (3.22)
In the last step we used 1̂2̂ = 12 and |2̂] = |2]. Playing the same game with the factors
in the denominator, we find
P̂23 4[ P̂23 2̂] = 4| P̂23 |2̂] = 4|3|2] = −43[32] = −34[23] . (3.23)
Now use (3.22) and (3.23) in (3.21) to find
−123 [23]3
An 1− 2− 3+ . . . n + =

−34[23] 45 · · · n1 23[23] [23]
124
= . (3.24)
12233445 · · · n1

This completes the inductive step. With the 3-point gluon amplitude A3 1− 2− 3+ fixed
completely by little group scaling and locality, we have then proven the Parke–Taylor
formula for all n. This is a lot easier than calculating Feynman diagrams!
You may at this point complain that we have only derived the Parke–Taylor formula
recursively for the case where the negative helicity gluons are adjacent. Try your own hands
on the proof for the non-adjacent case. In Chapter 4 we will use supersymmetry to derive
a more general form of the tree-level gluon amplitudes: it will contain all MHV helicity
arrangements in one compact expression.
We have now graduated from MHV-level and are ready to embark on the study of NMHV
amplitudes. It is worthwhile to consider the 5-point example A5 [1− 2− 3− 4+ 5+ ] even though
this amplitude is anti-MHV: constructing it with a [+, +-shift is a calculation very similar
to the MHV case – and that would by now be boring. So, instead, we are going to use a
[−, −-shift to illustrate some of the manipulations used in BCFW recursion.
Example. Consider the [1, 2-shift recursion relations for A51−2−3−4+5+: there are
two diagrams
1^ – 2^ – 1^ – 2^ –
P^15 P^23
A5 1− 2− 3− 4+ 5+ = − + 3– + 5 +
− + .
+ +
5 4 + 4 3 –
diagram A diagram B
(3.25)
We have indicated the required helicity for the gluon on the internal line. Had we chosen
the opposite helicity option for the internal gluon in diagram A, the righthand subam-
plitude would have helicity structure −− −+, so it would vanish. Diagram B also van-
ishes for the opposite choice of the helicity on the internal line. For the helicity choice
+
shown, the righthand subamplitude of diagram B is MHV, A3 [− P̂23 , 2̂− , 3− ], and since
+
|2 is shifted, the special 3-particle kinematics actually makes A3 [− P̂23 , 2̂− , 3− ] = 0,
just as we saw for the anti-MHV case in the discussion below (3.16). So diagram B
vanishes, and we can focus on diagram A. Using the Parke–Taylor formula for the
MHV subamplitudes, we get
1̂ P̂3 1 2̂34

A5 1− 2− 3− 4+ 5+ = × × . (3.26)
P̂551̂ 15[15] 2̂3344 P̂ P̂ 2̂
Here P̂ stands for P̂15 = p̂1 + p5 . We have three powers of | P̂ in the numerator and
three in the denominator. A good trick to simplify such expressions is to multiply
(3.26) by [ P̂ X ]3 /[ P̂ X ]3 for some useful choice of X such that [ P̂ X ] = 0. In this case,

it is helpful to pick X = 2. Grouping terms conveniently together, we get:
• 1̂ P̂[ P̂2] = −1̂|1̂+5|2] = 1̂5[52] = −15[25] (since |1̂ = |1).

• 5 P̂[ P̂2] = −5|1̂+5|2] = 51[1̂2] = 51[12] .
• 4 P̂[ P̂2] = −4|1̂ + 5|2] = 4|2̂+3+4|2] = 4|3|2] = −43[32] = −34[23].
• 2̂ P̂[ P̂2] = −2 p̂2 · P̂ = 2 p̂2 · ( p̂2 + p3 + p4 ) = ( p̂2 + p3 + p4 )2 − ( p3 + p4 )2
= P̂ 2 − 34[34] = −34[34],
since the amplitude is evaluated at z such that P̂ 2 = 0.
Using these expressions in (3.26) gives

[25]3 2̂33
A5 1− 2− 3− 4+ 5+ = . (3.27)
[12][23][34][15]343
Despite the simplifications, there is some unfinished business for us to deal with: (3.27)
depends on the shifted spinors via 2̂3. This bracket must be evaluated at the residue
value of z = z 15 which is such that P̂15
2
= 0:
[15]
0 = P̂15
2
= 15[1̂5] , i.e. 0 = [1̂5] = [15] + z 15 [25] , i.e. z 15 = − . (3.28)
[25]
Use this and momentum conservation to write
23[25] + 13[15] 34[45]
2̂3 = 23 − z 15 13 = = . (3.29)
[25] [25]
Inserting this result into (3.27) we arrive at the expected anti-MHV Parke–Taylor
expression
[45]4
A5 1− 2− 3− 4+ 5+ = . (3.30)
[12][23][34][45][51]
As noted initially, the purpose of this example was not to torture you with a difficult
way to compute A5 [1− 2− 3− 4+ 5+ ]. The purpose was to illustrate the methods needed
for general cases in a simple context.
You may not be overly impressed with the simplicity of the manipulations needed to reduce
the raw output of BCFW. Admittedly it requires some work. If you are unsatisfied, go ahead
and try the calculations in this section with Feynman diagrams. Good luck.
Now you have seen the basic tricks needed to manipulate the expressions generated by
BCFW. So you should get some exercise.
Exercise 3.5
Let us revisit scalar-QED from the end of Section 2.4. Use little group scaling and
locality to determine A3 (ϕ ϕ ∗ γ ± ) and compare with your result from Exercise 2.15.
Then use a [4, 3-shift to show that (see Exercise 2.16)

1424
A4 ϕ ϕ ∗ γ + γ − = g 2 . (3.31)
1323
[Hint: this is not a color-ordered amplitude. See also Exercise 2.16.]

What is the large-z falloff of this amplitude under a [4, 3-shift?3
Exercise 3.6
Calculate the 4-graviton amplitude M4 (1− 2− 3+ 4+ ): first recall that little group scaling
& locality fix the 3-graviton amplitudes as in Exercise 2.34. Then employ the [1, 2-shift
BCFW recursion relations (they are valid [18, 20]).
Check the little group scaling and Bose-symmetry of your answer for
M4 (1− 2− 3+ 4+ ).
[Hint: your result should match one of the amplitudes in Exercise 2.33.]
Show that M4 (1− 2− 3+ 4+ ) obeys the 4-point “KLT relations” [21]
M4 (1234) = −s12 A4 [1234] A4 [1243] , (3.32)
where A4 is your friend the Parke–Taylor amplitude with negative helicity states 1
and 2, and s12 = −( p1 + p2 )2 is a Mandelstam variable. When you are done, look up
ref. [22] to see how difficult it is to do this calculation with Feynman diagrams.
Let us now take a look at some interesting aspects of BCFW for the split-helicity NMHV
amplitude A6 [1− 2− 3− 4+ 5+ 6+ ]. Let us first look at the recursion relations following from
the [1, 2-shift that we are now quite familiar with. There are two non-vanishing diagrams:
1^ – 2^ – 2^ –
P^16 1^ –
− − − + + + 3– P^156
A6 1 2 3 4 5 6 = + 6+ 3– . (3.33)
+
− + 4+ +
− +
6 5 4+
diagram A 5+ diagram B
Exercise 3.7
Show that the 23-channel diagram does not contribute in (3.33).
The first thing we want to discuss about the 6-gluon amplitude is the 3-particle poles in the
2
expression (3.33). Diagram B involves a propagator 1/P156 , so there is a 3-particle pole at
P156 = 0. By inspection of the ordering of the external states in A6 [1− 2− 3− 4+ 5+ 6+ ] there
2
should be no distinction between the (− + +) 3-particle channels 561 and 345, so we would
expect the amplitude to have a pole also at P3452
= P126
2
= 0. But the [1, 2-shift recursion
relation (3.33) does not involve any 126-channel diagram. How can it then possibly encode
the correct amplitude? The answer is that it does and that the P345 2
= P126
2
= 0 pole is
actually hidden in the denominator factor 2̂ P̂16 of the righthand subamplitude of diagram
A in (3.33). Let us show how.
As in the 5-point example (3.26), we multiply the numerator and denominator both by
three powers of [ P̂16 3]. Then write
2̂ P̂16 [ P̂16 3] = 21[1̂3] + 2̂6[63] . (3.34)
3 Of course, we cannot use the BCFW result for the amplitude to validate the shift, only test self-consistency. The
independent demonstration of the large-z falloff under the shift used here is given in [19].
2
It follows from P̂16 = 0 that z 16 = −[16]/[26], and this is then used to show that 2̂6 =
(16[16] + 26[26])/[26] and [1̂3] = [12][36]/[26]. Plug these values into (3.34) to find
[36]
[36] 2
2̂ P̂16 [ P̂16 3] = − 12[12] + 16[16] + 26[26] = − P . (3.35)
[26] [26] 126
2
So there you have it: the 3-particle pole P126 is indeed encoded in the BCFW result (3.33).
The second thing we want to show you is the actual representation for the 6-gluon NMHV
tree amplitude, as it follows from (3.33):
3|1 + 2|6]3
A6 1− 2− 3− 4+ 5+ 6+ = 2
P126 [21][16]34455|1 + 6|2]
1|5 + 6|4]3
+ . (3.36)
2
P156 [23][34]56615|1 + 6|2]
The expression (3.36) may not look quite as delicious as the Parke–Taylor formula, but
remember that it contains the same information as the sum of 38 Feynman diagrams!
Exercise 3.8
Check the little group scaling of (3.36). Fill in the details for converting the two
diagrams in (3.33) to find (3.36).
The third thing we would like to emphasize is that using the [1, 2-shift recursion relations
is just one way to calculate A6 [1− 2− 3− 4+ 5+ 6+ ]. What happens if we use the [2, 1-shift?
Well, now there are three non-vanishing diagrams:
1^ – 2^ – 1^ – 2^ –
P^16 1^ –
P^156
2^ –
P^23
A6 1− 2− 3− 4+ 5+ 6+ = 3–
4+
+ 6+ 3 – + 6+ .
+ – − + 5+ + − +
6+ 5+ 5+ 4 + 4 3–
diagram A′ diagram B′ diagram C′
anti−MHV × NMHV MHV × MHV MHV × MHV
(3.37)
Special 3-particle kinematics force the diagram A to have helicity structure anti-
MHV×NMHV, as opposed to the similar diagram A in (3.33) which has to be MHV×MHV.
It is the first time we see a lower-point NMHV amplitude show up in the recursion relations.
This is quite generic: the BCFW relations are recursive both in particle number n and in
N K MHV level K .
The two BCFW representations (3.33) and (3.37) look quite different. In order for both
to describe the same amplitude, there has to be a certain identity that ensures that diagrams
A+B = A + B + C . To show that this identity holds requires a nauseating trip through
Schouten identities and momentum conservation relations in order to manipulate the angle
and square brackets into the right form: numerical checks can save you a lot of energy when
dealing with amplitudes with more than five external lines. It turns out that the identities
that guarantee the equivalence of BCFW expressions such as A+B and A + B + C actually
originate from powerful residue theorems [23] related to quite different formulations of the
amplitudes. This has to do with the description of amplitudes in the Grassmannian – we get
to that in Chapters 9 and 10, but wanted to give you a hint of this interesting point here.
Exercise 3.9
Show that the BCFW recursion relations based on the [2, 3-shift give the following
representation of the 6-point “alternating helicity” gluon amplitude:

A6 1+ 2− 3+ 4− 5+ 6− = {M2 } + {M4 } + {M6 } , (3.38)
where
i, i +24 [i +3, i −1]4
{Mi } = ,
P̃i2 i| P̃i |i +3]i +2| P̃i |i −1]i, i +1i +1, i +2[i +3, i −2][i −2, i −1]
(3.39)
and P̃i = Pi,i+1,i+2 . [Hint: {M4 } is the value of the 12-channel diagram.]
In Chapter 9 we discover that each {Mi } can be understood as the residue associated
with a very interesting contour integral (different from the one used in the BCFW
argument).
The fourth thing worth discussing further is the poles of scattering amplitudes. Color-
ordered tree amplitudes can have physical poles only when the sum of momenta of adjacent
external lines go on-shell. We touched on this point already in Section 2.8. There we also
noted that MHV gluon amplitudes do not have multi-particle poles, only 2-particle poles.
Now you have seen that 6-gluon NMHV amplitudes have both 2- and 3-particle poles. But
as you stare intensely at (3.36), you will also note that there is a strange denominator-factor
5|1 + 6|2] in the result from each BCFW diagram. This does not correspond to a physical
pole of the scattering amplitude: it is a spurious pole. The residue of this unphysical pole
better be zero – and it is: the spurious pole cancels in the sum of the two BCFW diagrams
in (3.36). It is typical that BCFW packs the information of the amplitudes into compact
expressions, but the cost is the appearance of spurious poles; this means that in the BCFW
representation the locality of the underlying field theory is not manifest. Elimination of
spurious poles in the representations of amplitudes leads to interesting results [24] that we
discuss in Chapter 10.
Finally, for completeness, note that the color-ordered NMHV amplitudes
A6 [1− 2− 3+ 4− 5+ 6+ ] and A6 [1− 2+ 3− 4+ 5− 6+ ] are inequivalent to the split-helicity am-
plitude A6 [1− 2− 3− 4+ 5+ 6+ ]. More about this in Chapter 4.
Other comments:
1) In our study of the recursion relations, we kept insisting on “generic” momenta. However,
special limits of the external momenta place useful and interesting constraints on the
amplitudes: the behavior of amplitudes under collinear limits and soft limits is described
in Section 2.8.
2) In some cases, the shifted amplitudes have “better than needed” large-z behavior.
For example, this is the case for a BCFW shift of two non-adjacent same-helicity lines
in the color-ordered Yang–Mills amplitudes: Ân (z) → 1/z 2 for large z. The vanishing
of the amplitude at large z implies the validity of a recursion relation for An . Let us

briefly outline the reason. Start with C Ânz(z) = 0 with C a contour that surrounds all the
simple poles. The unshifted amplitude An , which is the residue of the z = 0 pole, is

therefore (minus) the sum of all the other residues. We write An = I d I , where d I is
the factorization diagram (3.8) associated with the residue at z = z I . This summarizes the
derivation of the recursion relations from Section 3.1. Now, an extra power
in the large-z
falloff, Ân (z) ∼ 1/z 2 , means that there is also a bonus relation from C Ân (z) = 0 (with

C as before). It gives I d I z I = 0, because there is no residue at z = 0. Bonus relations
have practical applications, for example they have been used to demonstrate equivalence of
different expressions for graviton amplitudes [25].
3.3 When does it work?
In Section 2.6 we learned that 3-point amplitudes for massless particles are uniquely
determined by little group scaling, locality, and dimensional analysis. As we have just seen,
with the on-shell BCFW recursion relations, we can construct all higher-point gluon tree
amplitudes from the input of just the 3-point gluon amplitudes. That is a lot of information
obtained from very little input! It prompts suspicion: when can we expect on-shell recursion
to work? We will look at some examples now.
Yang–Mills theory and gluon scattering. The Yang–Mills Lagrangian (2.66) has two
types of interaction terms, the cubic A2 ∂ A and the quartic A4 . Given the former, the latter
is needed for gauge invariance, and the quartic must be included along with the cubic
as interaction vertices in the Feynman rules. However, existence of valid BCFW recursion
relations indicates that the cubic term A2 ∂ A fully captures the information needed for all on-
shell gluon amplitudes, at least at tree-level, with no need for A4 . The key distinction is that
the 3-vertex (and hence the Feynman rule cubic vertex) is an off-shell gauge non-invariant
object, while the 3-point on-shell amplitude is gauge invariant. Since A4 is determined from
A2 ∂ A by the requirement of off-shell gauge invariance of the Lagrangian, it contains no
new on-shell information. In a sense, that is why the recursion relations for on-shell gluon
amplitudes even have a chance to work with input only from the on-shell 3-point amplitudes.
We can rephrase the information contents of A2 ∂ A in a more physical way. The actual
input is then this: a 4d local theory with massless spin-1 particles (and no other dynam-
ical states) and a dimensionless coupling constant. With valid recursion relations, this
information is enough to fix the entire gluon tree-level scattering matrix!
Scalar-QED. As a second example, consider scalar-QED. The interaction between the

photons and the scalar particles created/annihilated by a complex scalar field ϕ is encoded
by the covariant derivatives Dμ = ∂μ − ie Aμ in

L ⊃ −|Dϕ|2 = |∂ϕ|2 + ie Aμ (∂μ ϕ ∗ )ϕ − ϕ ∗ ∂μ ϕ − e2 Aμ Aμ ϕ ∗ ϕ . (3.40)
In terms of Feynman diagrams, the tree amplitude A4 (ϕ ϕ ∗ γ γ ) is constructed from the

sum of two pole diagrams and the contact term from the quartic interaction (Exercise 2.17).
3.3 When does it work? 63
We have seen in Exercise 3.5 that this 4-point amplitude is constructible via BCFW. So it is
clear that only the information in the 3-point vertices is needed, and the role of Aμ Aμ ϕ ∗ ϕ
is just to ensure off-shell gauge invariance of the Lagrangian. Thus this case is just like the
Yang–Mills example above.
Thus emboldened, let us try to compute the 4-scalar tree amplitude A4 (ϕ ϕ ∗ ϕ ϕ ∗ ) using
BCFW recursion (3.10). Using a [1, 3-shift, there are two diagrams and their sum simplifies
to
132 242
ABCFW (ϕ ϕ ∗ ϕ ϕ ∗ ) = ẽ2 . (3.41)
4
12233441
If, however, we calculate this amplitude using Feynman rules from the interaction terms in
(3.40), we get

132 242
(ϕ ϕ ∗ ϕ ϕ ∗ ) = ẽ2 1 +
Feynman
A4 . (3.42)
12233441
Ugh! So BCFW did not compute the amplitude we expected. So what did it compute?
Well, let us think about the input that BCFW knows about: 4d local theory with massless
spin-1 particles and charged massless spin-0 particles (and no other dynamical states) and
a dimensionless coupling constant. Note that included in this input is the possibility of
a 4-scalar interaction term λ|ϕ|4 . So more generally, we should consider the scalar-QED
action from (2.64):
1 1
L = − Fμν F μν − |Dϕ|2 − λ|ϕ|4
4 4
1 1
= − Fμν F μν − |∂ϕ|2 + ie Aμ (∂μ ϕ ∗ )ϕ − ϕ ∗ ∂μ ϕ − e2 Aμ Aμ ϕ ∗ ϕ − λ|ϕ|4 . (3.43)
4 4

∗ ∗
In Exercise 2.19 you were asked to calculate A4 ϕ ϕ ϕ ϕ in this model. The answer was
given in (2.65): it is

132 242
A4 ϕ ϕ ∗ ϕ ϕ ∗ = − λ + ẽ2 1 + . (3.44)
12233441
So it is clear now that we have a family of scalar-QED models, labeled by λ, and that
our BCFW calculation produced the very special case of λ = ẽ2 . How can we understand
this? Validity of the recursion relations requires the absence of the boundary term Bn (see
Section 3.1). For the general family of scalar-QED models, there is a boundary term under
the [1, 3-shift, and its value is −λ + ẽ2 (as can be seen from (3.44) by direct computation).
The special choice λ = ẽ2 eliminates the boundary term, and that is then what BCFW
without a boundary term computes.
The lesson is that for general λ, there is no way the 3-point interactions can know the
contents of λ|ϕ|4 : it provides independent gauge-invariant information. That information
needs to be supplied in order for recursion to work, so in this case one can at best expect
recursion to work beyond 4-point amplitudes. The exception is of course if some symmetry,
or other principle, determines the information in λ|ϕ|4 in terms of the 3-field terms. This
is what we find for λ = ẽ2 . Actually, the expression (3.41) is the correct result for certain
4-scalar amplitudes in N = 2 and N = 4 super Yang–Mills theory (see Chapter 4), and in
those cases the coupling of the 4-scalar contact term is fixed by the Yang–Mills coupling
by supersymmetry.
Scalar theory λφ 4 . The previous example makes us wary of λφ 4 interaction in the context
of recursion relations – and rightly so. Suppose we consider λφ 4 theory with no other
interactions. It is clear that one piece of input must be given to start any recursive approach,
namely in this case the 4-scalar amplitude A4 = λ. In principle, one might expect on-shell
recursion to determine all tree-level An amplitudes with n > 4 from A4 = λ – what else
could interfere? After all, this is the only interaction in the Feynman diagrams. However,
given that the 6-scalar amplitude is A6 = λ2 ( s1231
+ . . . ), it is clear that any BCFW shift
0
gives O(z )-behavior for large z and hence there are no BCFW recursion relations without
boundary term for A6 in λφ 4 theory. Inspection of the Feynman diagrams reveals that the
O(z 0 )-contributions are exactly the diagrams in which the two shifted lines belong to the
same vertex. The sum of such diagrams equals the boundary term Bn from (3.4). Thus,
in this case of λφ 4 theory one can reconstruct Bn recursively.4 Hence A4 does suffice to
completely determine all tree amplitudes An for n > 4 in λφ 4 theory; but it is (in more than
one sense) a rather trivial example.
N = 4 super Yang–Mills theory. This is the favorite theory of many amplitugicians. We

will review the theory in more detail in Section 4.4, for now we just comment on a few
relevant aspects. The spectrum consists of 16 massless states: gluons g ± of helicity ±1,
four gluinos λa and λa of helicity ±1/2, and six scalars S ab . The indices a, b = 1, 2, 3, 4
are labels for the global SU (4) symmetry. The Lagrangian contains standard gluon self-
interactions, with standard couplings to the gluinos and the scalars. All fields transform
in the adjoint of the SU (N ) gauge group, so we consider color-ordered tree amplitudes
defined in the same way as the color-ordered gluon amplitudes. The Lagrangian includes
a scalar 4-point interaction term of a schematic form [S, S]2 . It contains, for example,
the interaction S 12 S 23 S 34 S 41 . The result for the corresponding color-ordered amplitude is
(suppressing the gauge coupling constant):

A4 S 12 S 23 S 34 S 41 = 1. (3.45)
Since this amplitude has no poles, it cannot be obtained via direct factorization. Actually, the
amplitude (3.45) and its cousin 4-scalar amplitudes with equivalent SU (4) index structures
are the only tree amplitudes of N = 4 SYM that cannot be obtained from the BCFW
recursion formula (3.10); that may seem surprising, but it is true [27].
When supersymmetry is incorporated into the BCFW recursion relations, all tree ampli-
tudes of N = 4 SYM can be determined by the 3-point gluon vertex alone. The so-called
super-BCFW shift mixes the external states in such a way that even the 4-scalar ampli-
tude (3.45) can be constructed recursively. We will introduce the super-BCFW shift in
Section 4.5.
4 See [16]. Or avoid the term at infinity by using an all-line shift [26], to be defined in Section 3.4.
Gravity. We have already encountered the 4-point MHV amplitude M4 (1− 2− 3+ 4+ ): you
“discovered” it from little group scaling in Exercise 2.33 and constructed it with BCFW
in Exercise 3.6. The validity of the BCFW recursion relations for all tree-level graviton
amplitudes [18, 20] means that the entire on-shell tree-level S-matrix for gravity is deter-
mined completely by the 3-vertex interaction
of three gravitons. In contrast, the expansion
√
of the Einstein–Hilbert action 2κ1 2 d 4 x −g R around the flatspace Minkowski metric
gμν = ημν + κ h μν contains infinitely many interaction terms. It is remarkable that all these
terms are totally irrelevant from the point of view of the on-shell tree-level S-matrix;
their sole purpose is to ensure diffeomorphism invariance of the off-shell Lagrangian. For
on-shell (tree) amplitudes, we do not need them. Much more about gravity amplitudes in
Chapter 12.
Summary. We have discussed when to expect to have recursion relations for tree-level
amplitudes. The main lesson is that we do not get something for nothing: input must be
given and we can only expect to recurse that input with standard BCFW when all other
information in the theory is fixed by our input via gauge invariance. If another principle –
such as supersymmetry – is needed to fix the interactions, then that principle should
be incorporated into the recursion relations for a successful recursive approach. Further
discussion of these ideas can be found in [26], mostly in the context of another recursive
approach known as CSW, which we will discuss briefly next.
3.4 MHV vertex expansion (CSW)
We introduced recursion relations in Section 3.1 in the context of general shifts (3.1)
satisfying the set of conditions (i)–(iii). Then we specialized to the BCFW shifts in
Section 3.2. Now we would like to show you another kind of recursive structure.
Consider a shift that is implemented via a “holomorphic” square-spinor shift:
|î] = |i] + z ci |X ] and |î = |i . (3.46)

n
Here |X ] is an arbitrary reference spinor and the coefficients ci satisfy i=1 ci |i = 0.
Exercise 3.10
Show that the square-spinor shift (3.46) gives shift-vectors ri that fulfill requirements
(i)–(iii) in Section 3.1.
The choice c1 = 23, c2 = 31, c3 = 12, and ci = 0 for i = 4, . . . , n implies that the
shifted momenta satisfy momentum conservation. This particular realization of the square-
spinor shift is called the Risager shift [28].
We consider here a situation where all ci = 0 so that all momentum lines are shifted via
(3.46) – this is an all-line shift. It can be shown [29] that N K MHV gluon tree amplitudes
fall off as 1/z K for large z under all-line shift. So this means that all gluon tree-level
amplitudes can be constructed with the all-line shift recursion relations; except the MHV
amplitudes (K = 0). It turns out that in this formulation of recursion relations, the tower of
MHV amplitudes constitutes the basic building blocks for the N K MHV amplitudes. Let us
see how this works for NMHV. The recursion relations give
^
^ PÎ ^
^.
ANMHV
n = L R (3.47)
diagrams I ^ ^
If you consider the possible assignments of helicity labels on the internal line, you will
see that there are two options: either the diagram is anti-MHV3 ×NMHV or MHV×MHV.
The former option vanishes by special kinematics of the 3-point anti-MHV vertex, just as
in the case of the first diagram in (3.15). So all subamplitudes in (3.47) are MHV. Let us
write down the example of the split-helicity NMHV 6-gluon amplitude:
1^− 1^− 1^− 2^−

2^− 2^−
6^+ ^
An 1− 2− 3− 4+ 5+ 6+ =
^
− +
3−
4^+
+ 6^ +
− +
^
3− + 5^+ − +
6^+ 5^+ 5^+ 4^+ 4+ 3−
^
2^− ^
3− 2^− ^
3− 2^− ^
3−
4^+ 1^−
+ + 1^ −
4^+ +
+ − 5^+ + − 6^+ + −
1^− 6^+ 5^+ 4^+
6^+ 5^+
(3.48)
All six diagrams are non-vanishing and this may look a little daunting, especially compared
with the BCFW version where there were just two diagrams in the simplest version (3.36).
However, the diagrams in (3.48) are easier to evaluate than the BCFW diagrams: the MHV
amplitudes depend only on angle spinors, so the only way they know about the square-spinor
shift is through the internal line angle spinors | P̂I , for example
1^ – 2^ –
1 P̂I 4 1 234
– = .
6^ + − + 3^
1 P̂I P̂I 55661 P156
2
23344 P̂I P̂I 2
(3.49)
5^ + ^
4 –
We can write
[ P̂I X ] 1 1
| P̂I = P̂I |X ] = PI |X ] . (3.50)
[ P̂I X ] [ P̂I X ] [ P̂I X ]
In the last step we can drop the hat, because the shift of ( P̂I )ȧb is proportional to the
reference spinor [X |b of the shift (3.46). Note that the diagrams are necessarily invariant
under little group scaling associated with the internal line. Therefore the factors [ P̂1X ] in
I
(3.50) cancel out of each diagram and we can use the prescription
| P̂I → PI |X ] . (3.51)
This gives
1^ – 2^ – 1|P156 |X ]4 1 234
6^ + 3^ – =
5^ +
− + 1|P156 |X ]5|P156 |X ]5661 P156
2 23344|P |X ]2|P |X ]
156 156
4^ –
(3.52)
and similarly for the other “MHV vertex diagrams” in (3.48). Note that we can drop the
indication ˆ of the shift on the external lines in the MHV vertex diagrams since the square-
spinor shift does not affect the MHV vertices and all that is needed is the prescription (3.51)
for the internal lines.
In general, each diagram depends explicitly on the reference spinor |X ], but of course
the full tree amplitude cannot depend on an arbitrary spinor: the Cauchy theorem argument
of Section 3.1 guarantees that the sum of all the diagrams will be independent of |X ] and
reproduce the correct tree amplitude. Numerically, it is not hard to verify independence
of |X ] and that the expressions (3.48) and (3.36) indeed produce the same scattering
amplitude.
The expansion of the amplitude in terms of MHV vertex diagrams generalizes beyond the
NMHV level. In general, the N K MHV tree amplitude is written as a sum of all tree-level
diagrams with precisely K +1 MHV vertices evaluated via the replacement rule (3.51).
This construction of the amplitude is called the MHV vertex expansion: it can be viewed as
the closed-form solution to the all-line shift recursion relations. However, it was discovered
by Cachazo, Svrcek, and Witten in 2004 [15] before the introduction of recursion relations
from complex shifts. The method is therefore also known as the CSW expansion and the
rule (3.51) is called the CSW prescription. The first recursive derivation of the MHV vertex
expansion was given by Risager [28] using the 3-line Risager shift mentioned above applied
to the three negative helicity line of NMHV amplitudes. The all-line shift formulation was
first presented in [29].
Exercise 3.11
Construct A5 [1− 2− 3− 4+ 5+ ] from the CSW expansion. Make a choice for the reference
spinor |X ] to simplify the calculation and show that the result agrees with the anti-MHV
Parke–Taylor formula (2.115).
The MHV vertex expansion was the first construction of gluon amplitudes from on-shell
building blocks. The method is valid also in other cases, for example in super Yang–
Mills theory [27, 29] or Higgs amplitudes with gluons and partons [30, 31]. There are
also applications of the MHV vertex expansion at loop-level – for a review see [32] and
references therein.
The MHV vertex expansion can also be derived directly from a Lagrangian [33]: a field
redefinition and suitable light-cone gauge choice brings it to a form with an interaction
term for each MHV amplitude. The N K MHV amplitudes are then generated from the
MHV vertex Lagrangian by gluing together the MHV vertices. The reference spinor |X ]
arises from the light-cone gauge choice. There is also a twistor-action formulation of the
MHV vertex expansion [34].
In the case of the BCFW shift, we have applied it to gluon as well as graviton amplitudes.
A version of the MHV vertex expansion was proposed for gravity in [35] based on the
Risager shift. However, the method fails for NMHV amplitudes for n ≥ 12: under the
Risager shift, Ân (z) ∼ z 12−n for large-z, so for n ≥ 12 there is a boundary term obstructing
the recursive formula [36]. An analysis of validity of all-line shift recursion relations in
general 4d QFTs can be found in [26].
At this stage, you may wonder why tree-level gluon scattering amplitudes have so many
different representations: one from the MHV vertex expansion and other forms arising from
BCFW applied to various pairs of external momenta. The CSW and BCFW representations
reflect different aspects of the amplitudes, but they turn out to be closely related. We need
more tools to learn more about this. So read on.
Supersymmetry 4
The focus of this chapter is supersymmetry in the context of scattering amplitudes. We

begin with a very brief introduction to supersymmetry; it serves to provide the minimal
background material we need for our amplitude studies. We then discuss supersymmetry
Ward identities for the amplitudes and introduce on-shell superspace as a tool for organizing
the amplitudes into superamplitudes. This is particularly powerful in N = 4 super Yang–
Mills theory where it allows us to solve a supersymmetric version of the BCFW recursion
relations to find all tree-level superamplitudes in this theory. We had better get started.
4.1 N = 1 supersymmetry: chiral model
Let us begin with a simple example of supersymmetry. Consider the free Lagrangian for a
Weyl fermion ψ and a complex scalar field φ:
L0 = iψ † σ̄ μ ∂μ ψ − ∂μ φ̄ ∂ μ φ . (4.1)
The bar on φ denotes the complex conjugate. In addition to Poincaré symmetry, L0 also
has a symmetry that mixes fermions and bosons:
δ φ = ψ , δ φ̄ = † ψ † ,
μ † μ (4.2)
δ ψa = −iσa ḃ †ḃ ∂μ φ , δ ψȧ = i∂μ φ̄ b σbȧ .
This is an example of a supersymmetry transformation. The anti-commuting constant

spinor is the supersymmetry parameter (a fermionic analogue of the infinitesimal angle
†
θ of a rotation transformation), and ψ = a ψa and † ψ † = ȧ ψ †ȧ are the 2-component
spinor products with spinor indices raised/lowered with the 2-index Levi-Civita symbol,
just as in (2.13).
The conserved Noether charges associated with the supersymmetry transformations are
anticommuting supercharges Q and Q † . They enhance the Poincaré algebra to a super-
Poincaré graded Lie algebra with both bosonic and fermionic generators. Thus, in addition
to the translation and rotation/boost generators Pμ and Mμν whose algebra is
μ

P μ, P ν = 0 , P , M ρσ = i ημρ P σ − ημσ P ρ ,
μν
(4.3)
M , M ρσ = i ημρ M νσ − ηνρ M μσ − ημσ M νρ + ηνσ M μρ ,
70 Supersymmetry
the super-Poincaré algebra has

μ μ
Qa , P μ = 0 , Q a , M μν = 4i ȧ ċ σa ȧ σcνċ − σaνȧ σcċ Q c ,
†
(4.4)
Qa , Qb = 0 , Q a , Q ḃ = −2σaμḃ Pμ ,
and similarly for Q † . Note that the supercharges Q and Q † commute with the spacetime
translation operators, but not with the rotations/boosts, and that the anticommutator of Q
and Q † is a spacetime translation.
If you have not previously seen supersymmetry, you should promptly go ahead and do
these two exercises:
Exercise 4.1
Check that L0 in (4.1) is invariant under the supersymmetry variation (4.2) up to a total
derivative.
Exercise 4.2
Calculate [δ1 , δ2 ] by acting with it on the fields; first try on φ, then afterwards on ψ. You
should find that in each case, the combination of two supersymmetry transformations
is a spacetime translation.
The 2-component Weyl spinors can be combined into 4-component Majorana spinors
ψ
a
M = †ȧ and M = †aȧ , (4.5)
ψ
and the supersymmetry transformations are then defined with suitable L- and R-projections
PL ,R from (A.9). We write the free field expansions as

a− ( p) ei p.x + a+†
φ(x) = dp ( p) e−i p.x

ψa (x) = PL M (x) = bs ( p) PL u s ( p) ei p.x + bs† ( p) PL vs ( p) e−i p.x , (4.6)
dp
s=±
and similarly for φ̄(x) and ψ †ȧ = PR M . The Lorentz-invariant measure dp is defined
below (2.5).
Upon canonical quantization, the coefficients satisfy the algebra of bosonic/fermionic
creation-annihilation operators,
† †
a± ( p), a± ( p ) = (2π )3 2E p δ 3 ( p − p ) , b± ( p), b± ( p ) = (2π )3 2E p δ 3 ( p − p ) ,
(4.7)
with all other (anti)commutators vanishing.
For the fermions, the ±-subscripts on the operators indicate the helicity h = ± 12 of
the associated 1-particle states. As a matter of later convenience, we have also labeled
the two sets of annihilation/creation operators associated with the complex field φ with
4.1 N = 1 supersymmetry: chiral model 71
±-subscripts. The corresponding particles are of course scalars with h = 0, but the label
indicates which spinor helicity state the scalar state is matched to via supersymmetry. Let
us see how that works.
The supersymmetry transformations (4.2) transform the fields ψ and φ into each other,
and therefore the associated annihilation/creation operators are also related. The relationship
is straightforward to extract from the free field expansions (4.6). Recalling from Section
2.2 that PL vs ( p) is equal to | p] for s = + and vanishes for s = − (and similarly for PR ),
one finds
δ a− ( p) = [ p] b− ( p) , δ b− ( p) = p a− ( p) ,
(4.8)
δ a+ ( p) = p b+ ( p) , δ b+ ( p) = [ p] a+ ( p) .
†
We have introduced anti-commuting bra-kets |ȧ = ȧ and |]a = a for the supersym-
metry parameter. Using −| p] p| = pa ḃ , it is easy to see that [δ1 , δ2 ]O( p) = a μ pμ O( p)
for O( p) any one of the creation/annihilation operators. Hence, the commutator of two
supersymmetry transformations is a translation, as in the algebra (4.4). The translation
parameter a μ can be written in terms of Majorana spinors as a μ = M,2 γ μ M,1 .
|Q]
a
The supersymmetry generators Q M = can be deduced from
|Q † ȧ

δ O = ¯M Q M , O = [ Q] + Q, O . (4.9)
Using this, one finds (and you should check it) that

† †
|Q]a = dp | p]a a+ ( p) b+ ( p) − b− ( p) a− ( p) ,
(4.10)

† †
| pȧ a− ( p) b−
|Q † ȧ = dp ( p) − b+ ( p) a+ ( p) ,
reproduces (4.8).
Exercise 4.3
Show that {|Q]a , Q † |ḃ } equals pa ḃ times a sum of number operators.
The action of the supercharges (4.10) on the annihilation operators is

[Q , a− ( p)] = | p] b− ( p) , [Q † , b− ( p)] = | p a− ( p) ,
[Q , b− ( p)] = 0 , [Q † , a− ( p)] = 0 ,
(4.11)
[Q, b+ ( p)] = | p] a+ ( p) , [Q † , a+ ( p)] = | p b+ ( p) ,
[Q, a+ ( p)] = 0 , [Q † , b+ ( p)] = 0 ,
where [ . , . ] is a graded bracket that is an anticommutator when both arguments are
Grassmann and otherwise a commutator. The 2-component spinor-indices are suppressed.
A similar set of relations holds for the creation operators.
The action of the supersymmetry generators on the creation/annihilation operators im-
plies that the associated bosonic and fermionic 1-particle states are related – they are said
72 Supersymmetry
to belong to the same supermultiplet. Note that Q lowers the helicity by 12 and that Q †
raises it by 12 . Because the supersymmetry generators commute with the Hamiltonian, P 0 ,
and the space-translation generators, P i , it follows that states in the same supermultiplet
must have the same mass.
In the chiral model (4.1), the mass of the bosons and fermions is of course zero. It
follows from our discussion that the spectrum of the chiral model splits into a “nega-
tive helicity sector” and a “positive helicity sector”; CPT symmetry requires us to have
both.
In a supermultiplet, the number of on-shell bosonic states equals the number of on-shell
fermionic states. This is manifest in the example of the chiral model: the two real scalar
states encoded in the complex scalar field are supersymmetry partners of the two fermionic
states with h = ±1/2.
Interactions. Next, we would like to introduce interactions – after all, this is all about
scattering amplitudes so we need something to happen! We want to study interactions
that preserve supersymmetry. For our chiral model, one can introduce a “superpotential”
interaction of the form
L I = 12 g φ ψψ + 12 g ∗ φ̄ ψ † ψ † − 14 |g|2 |φ|4 . (4.12)
A small modification of the supersymmetry transformations (4.2) is now needed in the

transformation rule of the fermion field:
δ φ = ψ , δ φ̄ = † ψ † ,
μ † μ † (4.13)
δ ψa = −iσa ḃ †ḃ ∂μ φ + 12 g ∗ φ̄ 2 a , δ ψȧ = i∂μ φ̄ b σbȧ + 12 gφ 2 ȧ .
Note that the coupling of the 4-scalar interaction in (4.12) is fixed in terms of the Yukawa
coupling g; this is required by supersymmetry.
Exercise 4.4
Show that (4.13) is a symmetry of L = L0 + L I .
A linear version of the supersymmetry transformations can be given using an auxiliary

field. Supersymmetric actions can be expressed compactly and conveniently using off-shell
superspace formalism. You can find much more about this in textbooks such as Wess and
Bagger [37].
Extended supersymmetry
So far we have focused on a very simple case of a “chiral supermultiplet” in which a spin-
0 particle is partnered with a spin- 12 particle. One can repeat the analysis for any N = 1
supersymmetric model with particles of spin (s, s + 12 ). The N counts the number of super-
†
symmetry generators Q A and Q A with A = 1, 2, . . . , N . When N > 1, the model is said to
have extended supersymmetry. The super-Poincaré algebra with extended supersymmetry
4.2 Amplitudes and supersymmetry Ward identities 73
is very similar to (4.4), but with

†
Q aA , Q B ḃ = −2σaμḃ Pμ δ A B . (4.14)
In some cases of extended supersymmetry, a “central charge” Z AB can appear in the anti-
commutator {Q aA , Q bB } = Z AB ab . In almost all applications of extended supersymmetry
in this book, there will be no central charge.
The extended supersymmetry algebra (without central charges) has a global symmetry
that rotates the supersymmetry charges: this is called an R-symmetry. An R-symmetry is
characterized by the property that it does not commute with the supersymmetry transfor-
mations; this contrasts “ordinary” flavor-type global symmetries that do commute with
supersymmetry. With N -fold extended supersymmetry, the R-symmetry is SU (N ); this
full R-symmetry may or may not be realized for a given model.
As examples of N = 1 supermultiplets, N = 1 super-QED has a photon and a photino
with helicities ±1 and ± 12 . N = 1 super Yang–Mills theory (SYM) with a gluon (h = ±1)
and its fermionic partner, the gluino (h = ± 12 ), will be discussed in Section 4.3.
In extended supersymmetry, there are 2N states in the massless supermultiplets (and the
same in the CPT conjugates). For example, for N = 2 SYM, one supermultiplet consists of a
helicity −1 gluon, two gluinos with helicity − 12 , and a scalar with helicity 0. Thus the N = 2
gluon supermultiplet has two bosonic states and two fermionic states. The CPT conjugate
multiplet contains the same types of states but with opposite signs for the helicities.
The spectrum of states is generated by acting with the supercharges on the “highest
weight state”; in the case of massless particles in four dimensions, this is the highest helicity
state. To avoid states with spin greater than 1, the maximal amount of supersymmetry in
fourdimensions is N = 4. This large symmetry-requirement places such strong constraints
on the theory that it is unique (up to choice of gauge group): this is N = 4 super Yang–
Mills theory. We are going to study the supersymmetry constraints on the N = 4 SYM
amplitudes in much further detail in Section 4.4. For now, let us content ourselves with
N = 1 supersymmetry and study the consequences it has on the scattering amplitudes in
our chiral scalar model.
4.2 Amplitudes and supersymmetry Ward identities
Consider the simple supersymmetric chiral scalar model whose Lagrangian L is the sum of
the kinetic terms (4.1) and the interaction terms (4.12). The 4-point tree amplitudes were
essentially already presented in Exercise 2.8. To adapt the results (2.46), we just need to
take the coupling of the 4-scalar interaction to be λ = |g|2 . The 4-point tree amplitudes are
then:

− + 24
− − + + 12
Atree φφ φ̄ φ̄ = −|g|2 , Atree φ f f φ̄ = −|g|2 , Atree f f f f = |g|2 .
4 4
34 4
34
(4.15)
74 Supersymmetry
By inspection of (4.15), we observe that

− + 24 tree

Atree φ f f φ̄ = A4 φφ φ̄ φ̄ , (4.16)
4
34

− − + + 12 tree
− +
Atree f f f f =− A φ f f φ̄ . (4.17)
4
24 4
These relations hold not just for the tree-level amplitudes, as we have seen just now; super-
symmetry ensures that (4.16) and (4.17) hold at all orders in the perturbation expansion.
We will now see how that comes about.
We can think of an n-point amplitude with all-outgoing particles as the S-matrix element
0|O1 ( p1 ) . . . On ( pn )|0 in which n annihilation

operators
Oi ( pi ), i = 1, . . . , n, act to the
left on the out-vacuum. For example, A4 φ f − f + φ̄ = 0|a− ( p1 )b− ( p2 )b+ ( p3 )a+ ( p4 )|0.
Suppose the vacuum is supersymmetric: Q|0 = 0 = Q † |0. Then for any set of n annihi-
lation (or creation) operators, we have

0 = 0| Q † , O1 ( p1 ) . . . On ( pn ) |0

n
= (−1) j<i |O j | 0|O1 ( p1 ) · · · Q † , Oi ( pi ) · · · On ( pn )|0 , (4.18)
i=1
and similarly for Q. Here the sign-factor takes into account that a minus sign is picked up
every time Q † passes by a fermionic operator: so |O| is 0 when the operator O is bosonic
and 1 if it is fermionic. Now using the action of the supersymmetry generators (4.11) on
the free asymptotic states, the equation (4.18) describes a linear relation among scattering
amplitudes whose external states are related by supersymmetry. Such relations are called
supersymmetry Ward identities and they are valid at any order in perturbation theory. The
supersymmetry Ward identities are best illustrated in explicit examples. Thus, to start with,
consider:

0 = 0| Q † , a− ( p1 )b− ( p2 )a+ ( p3 )a+ ( p4 ) |0
= |2 a− ( p1 )a− ( p2 )a+ ( p3 )a+ ( p4 ) − |3 a− ( p1 )b− ( p2 )b+ ( p3 )a+ ( p4 )
−|4 a− ( p1 )b− ( p2 )a+ ( p3 )b+ ( p4 ) . (4.19)
We have used (4.11) that Q † annihilates a− ( p). Translating (4.19) to amplitudes, we have
the supersymmetry Ward identity

0 = |2 A4 φφ φ̄ φ̄ − |3 A4 φ f − f + φ̄ − |4 A4 φ f − φ̄ f + . (4.20)
Each identity (4.18) encodes two relations, since Q † (Q) has two components. This is also
visible in our example (4.20). We project out the two independent relations by dotting in a
suitable choice of bra-spinor r |. Picking r | = 4|, we find

0 = 42 A4 φφ φ̄ φ̄ − 43 A4 φ f − f + φ̄ . (4.21)
At tree-level, this is precisely the relation noted in (4.16). We emphasize that supersymmetry
ensures that (4.21) holds true at any loop order.
4.3 N = 1 supersymmetry: gauge theory 75
A second relation is extracted from (4.20) by choosing r | = 2|:

23

A4 φ f − φ̄ f + = − A4 φ f − f + φ̄ . (4.22)
24
Note how the factor 23

24
compensates the different little group scaling of the two amplitudes.
Exercise 4.5
Plug the two supersymmetry Ward identities (4.21) and (4.22) into (4.20) to show that
there is no further independent information available in (4.20).
Exercise 4.6
Derive (4.17) as a supersymmetry Ward identity.
Exercise 4.7

Find a Q-supersymmetry Ward identity that gives A4 (φ f − f + φ̄) = − [13] A φφ φ̄ φ̄ .
[12] 4
Show that this relation is equivalent to (4.16).
Let us take a brief look at the supersymmetry Ward identities at higher points, for example
for amplitudes with 6-particles. Starting with 0 = 0|[Q † , a1− a2− b3− a4+ a5+ a6+ ]|0 (using
a short-hand notation to indicate the momentum with a subscript) we find, after dotting in
r |,

0 = r 3 A6 φφφ φ̄ φ̄ φ̄ − r 4 A6 φφ f − f + φ̄ φ̄

− r 5 A6 φφ f − φ̄ f + φ̄ − r 6 A6 φφ f − φ̄ φ̄ f + . (4.23)
There are two relations, but four “unknowns” (the amplitudes), so this time the super-
symmetry Ward identity does not give simple proportionality relations among pairs of
amplitudes. Instead, one gets a web of linear relations. This is typical for “non-MHV type”
amplitudes; we come back to this point briefly in the next two sections.
To summarize, amplitudes whose external states are related by supersymmetry are linked
to each other through linear relationships called supersymmetry Ward identities. They were
first studied in 1977 by Grisaru, Pendleton, and van Nieuwenhuizen [38] and have since
then had multiple applications.
4.3 N = 1 supersymmetry: gauge theory
We have introduced supersymmetry for the simple chiral scalar model and derived the
consequences of supersymmetry for the amplitudes. Let us now adapt these results to
gauge theory. The Lagrangian for pure N = 1 super Yang–Mills theory (SYM) is
1
L = − Tr Fμν F μν + i Tr λ† σ̄ μ Dμ λ , (4.24)
4
76 Supersymmetry
where all fields are in the adjoint of the SU (N ) gauge group and the covariant derivative is
Dμ λ = ∂μ λ − ig[Aμ , λ]. The on-shell spectrum consists of helicity h = ±1 states of the
gluon g ± and h = ± 12 states of the gluino λ± . (For an abelian gauge group, we would call
the states photon and photino.) We will not need the explicit supersymmetry transformation
of the fields (they can be found for example in [39]), but will simply state the consequences
on the states. The supercharge Q̃ raises the helicity by 12 and Q lowers it by 12 , so if
c± (b± ) are the annihilation operators for gluons of helicity ±1 (gluinos of helicity ± 12 ),
then

Q̃, c+ (i) = 0 , [Q, c+ (i)] = [i| b− (i) ,

Q̃, b+ (i) = |i c+ (i) , Q, b+ (i) = 0 ,
(4.25)
Q̃, b− (i) = 0 , Q, b− (i) = [i| c− (i) ,

Q̃, c− (i) = |i b− (i) , Q, c− (i) = 0 ,
for N = 1 SYM. This is derived as (4.11) for the chiral model in Section 4.1.
Supersymmetry Ward identities for N = 1 SYM are derived from the action of the
supercharges on the states in (4.25). Since Q̃ annihilates c+ (i), the supersymmetry Ward
identity 0|[ Q̃, b1+ c2+ . . . cn+ ]|0 = 0 gives |1An [g + g + g + . . . g + ] = 0. This says that the
all-plus gluon amplitude vanishes at all orders in perturbation theory in N = 1 SYM.
Similarly, one can show that the gluon amplitude with exactly one negative helicity gluon
also vanishes. So
super Yang–Mills: AnL-loop [g + g + g + . . . g + ] = AnL-loop [g − g + g + . . . g + ] = 0 . (4.26)
We used only one supersymmetry generator for this argument, so the statement (4.26) is
true also with extended supersymmetry (as long as supersymmetry is unbroken). Moreover,
as indicated, (4.26) holds true at all orders in the loop-expansion because it relies only on
the existence of supercharges that annihilate the vacuum.
In Section 2.7, we showed that (4.26) holds at tree-level in pure non-supersymmetric
Yang–Mills,
+ + + + tree − + + +
Yang–Mills: n [g g g . . . g ] = An [g g g . . . g ] = 0 .
Atree (4.27)
However, at loop-level (4.26) is not true without supersymmetry: for example in pure
Yang–Mills theory, non-vanishing all-plus amplitudes are indeed generated at the 1-loop
level. The reason the result (4.27) holds at tree-level in pure Yang–Mills theory is that the
superpartners of the gluon couple quadratically to the gluon. So an amplitude whose external
states are all gluons “sees” the superpartner states only via loops. Thus, in pure Yang–Mills
the gluon amplitudes at tree-level have no choice but to obey the same supersymmetry
Ward identity constraints as the gluon amplitudes in super Yang–Mills. The argument is
even stronger: it says that the all-gluon amplitudes must be exactly the same at tree-level in
super Yang–Mills theory as in pure Yang–Mills theory.
Exercise 4.8
The non-vanishing of the n = 3 anti-MHV amplitude A3 [1− 2+ 3+ ] escapes the Ward
identity that forces An [g − g + g + . . . g + ] = 0 for n > 3. Explain how.
Exercise 4.9
Show in N = 1 SYM that supersymmetry Ward identities give
13 − − + +
A n g − λ− λ+ g + . . . g + = An g g g g . . . g + . (4.28)
12
Relate An [λ− λ− λ+ λ+ g + . . . g + ] to the MHV gluon amplitude.

Does N = 1 supersymmetry allow you to relate gluon amplitudes such as for exam-
ple An [1− 2− 3+ . . . 4+ ] and An [1− 2+ 3− . . . 4+ ]? Can it relate gluon MHV and NMHV
amplitudes?
The N = 1 supersymmetry Ward identities for NMHV gluon amplitude are more involved
and do not result in simple proportionality relations among pairs of amplitudes, such as
(4.28). For n = 6, the NMHV N = 1 supersymmetry Ward identities have been solved
explicitly [36, 38].
In our discussion of recursion relations in Section 3.3, we learned that in general we
should not expect it to be possible to produce a 4-scalar amplitude recursively from
3-particle amplitudes because of the possibility of independent input from a 4-scalar contact
term. However, in our supersymmetric chiral scalar model of Section 4.1, supersymmetry
determines the coupling of the 4-scalar interaction in terms of the 3-point interactions,
so one could hope that recursion relations work in that case. In particular, one might an-
ticipate that all 4-point amplitudes can be determined by the 3-point ones. For this to be
true, the recursion algorithm has to “know” which principle fixes the 4-scalar coupling,
so supersymmetry must be built into the recursion relations. This is done most efficiently
in two steps: first one introduces on-shell superspace and groups the amplitudes into su-
peramplitudes. Secondly, one incorporates a shift of a Grassmann super-parameter of the
on-shell superspace into the BCFW shift. Then one gets recursion relations for the su-
peramplitudes. This is best illustrated for N = 4 SYM, so that is what we will turn to
next.
4.4 N = 4 SYM: on-shell superspace and superamplitudes
The action for N = 4 super Yang–Mills theory (SYM) can be written compactly as

1 1 i g g2
S = d 4 x Tr − Fμν F μν − (D I )2 + D / + I [ I , ] + [ I , J ]2 .
4 2 2 2 4
(4.29)
78 Supersymmetry
Here Dμ = ∂μ − ig[Aμ , · ] is the covariant derivative, Aμ is the vector potential field,

and I are six real scalar fields with labels I = 1, . . . , 6 of a global S O(6) ∼ SU (4)
R-symmetry. All fields transform in the adjoint of the gauge group, which we take to be
SU (N ). The commutators in (4.29) are associated with the SU (N ) matrix structure of the
fields. The fermions are encoded in 10-dimensional Majorana–Weyl fields and the I
are gamma-matrices of the 10d Clifford algebra. This description of N = 4 SYM follows
from dimensional reduction of N = 1 SYM in 10d [40].
It is convenient to group the six real scalars I into six complex scalar fields ϕ AB =
−ϕ B A , with A, B = 1, 2, 3, 4, that satisfy the self-duality condition ϕ AB = 12 ABC D ϕ C D .
Here ABC D is the Levi-Civita symbol of SU (4) ∼ S O(6), and the scalars ϕ AB transform
in the fully antisymmetric 2-index representation of SU (4). In this language, the 10d
fermion fields give 4+4 gluino states λ A (h = + 12 ) and λ̄ A (h = − 12 ) that transform in the
(anti-)fundamental of SU (4).
In a supersymmetric model, the value V0 of the scalar potential V at the vacuum is
an order parameter of supersymmetry breaking [37]: in flat space, V0 = 0 is necessary
for preserving supersymmetry while V0 > 0 breaks supersymmetry. In N = 4 SYM, the
2
scalar potential is V = g4 [ I , J ]2 , so the theory has a moduli space of N = 4 super-
symmetric vacua with [ I , J ] = 0. At the origin of moduli space, where all the scalar
vacuum expectation values (vevs) vanish, ϕ AB = 0, all states are massless and the theory
contains no dimensionful parameters. In fact, the theory has conformal symmetry: the
trace of the stress-tensor is zero, up to the trace anomaly. In particular, the beta-function
vanishes at all orders in perturbation theory and there is no running of the coupling. The
theory is invariant under an enlarged spacetime symmetry group, namely the conformal
group S O(2, 4) ∼ SU (2, 2). Supersymmetry enhances the conformal symmetry to super-
conformal symmetry SU (2, 2|4). The bosonic part of this super-group consists of the con-
formal group SU (2, 2) and the R-symmetry group SU (4). The fermionic parts are the four
supersymmetry generators as well as their four superconformal partners. The symmetries
of N = 4 SYM are the subject of Chapter 5.
When the scalars acquire vevs in such a way that full supersymmetry is preserved,
i.e. [ I , J ] = 0, the theory is said to be on the Coulomb branch. The scale introduced by
the scalar vevs breaks the conformal symmetry, so the theory is not superconformal away
from the origin, but it still has N = 4 supersymmetry. We will briefly discuss scattering
amplitudes on the Coulomb branch of N = 4 SYM in Exercise 12.3, but otherwise we
focus entirely on the superconformal theory at the origin of moduli space: henceforth,
when we discuss amplitudes in N = 4 SYM this means the theory at the origin of moduli
space.
Given that the theory is conformal, we should clarify what we mean by the scattering-
matrix in N = 4 SYM. One way to think about this is to consider the theory in 4 −
dimensions: then the conformal symmetry is broken and the S-matrix is well-defined. This
turns out to be a little inconvenient for keeping on-shell symmetries manifest, and it can
therefore be better to consider the theory on the Coulomb branch and define the N = 4
SYM S-matrix as the zero-vev limit of the Coulomb branch S-matrix. These subtleties will
not affect the majority of our discussion and therefore we proceed to discuss the amplitudes
of N = 4 SYM without further hesitation.
Spectrum and supersymmetry Ward identities

The spectrum of the massless N = 4 supermultiplet consists of 16 states:
a
, aA
, a AB ,

ABC
a , 1234
a
1 gluon g + 4 gluinos λ A 6 scalars S AB 4 gluinos λ ABC ∼ λ D 1 gluon g − . (4.30)
helicity = 1 1
2
0 − 1
2
−1
The indices A, B, . . . = 1, 2, 3, 4 are labels of the global SU (4) R-symmetry that rotates the
†
four sets of supersymmetry generators Q A and Q̃ A ≡ Q A . The helicity-h states transform in
fully antisymmetric 2(1 − h)-index representations of SU (4). Note the important property
that the N = 4 supermultiplet is CPT self-conjugate: in particular N = 4 supersymmetry
connects the gluon states of h = ±1.
The action of the supercharges on the annihilation operators is

Q̃ A , a(i) = 0 , [Q A , a(i)] = [i| a A (i) ,

Q̃ A , a B (i) = |i δ AB a(i) , Q A , a B (i) = [i| a AB (i) ,
[B C] A BC
Q̃ A , a BC (i) = |i 2! δ A a (i) , Q , a (i) = [i| a ABC (i) , (4.31)
[B C D]
A BC D
Q̃ A , a BC D (i) = |i 3! δ A a (i) , Q ,a (i) = [i| a ABC D (i) ,
A 1234
[B C D E]
Q̃ A , a BC D E (i) = |i 4! δ A a (i) , Q , a (i) = 0.
Note that Q̃ A raises the helicity of all operators by 12 and removes the index A (if it is not
available to be removed, then the operator is annihilated). Q A does the opposite. Truncating
the spectrum to include just the gluons and gluinos, one finds that (4.31) reduces to the
result (4.25) for N = 1 SYM.
Exercise 4.10
Show that the supersymmetry Ward identities give the following relationships among
the color-ordered amplitudes in N = 4 SYM:
0 = −|1An [λ123 g − λ4 g + . . . g + ] − |2An [g − λ123 λ4 g + . . . g + ]

+ |3An [g − g − g + g + . . . g + ] (4.32)
and extract
13
An [g − λ123 λ4 g + . . . g + ] = An [g − g − g + g + . . . g + ] . (4.33)
12
Then derive
132
An [g − S 12 S 34 g + . . . g + ] = An [g − g − g + g + . . . g + ] . (4.34)
122
80 Supersymmetry
Since supersymmetry connects the entire N = 4 SYM multiplet, there is an important

N = 4 supersymmetry Ward identity that connects color-ordered MHV gluon amplitudes
with the negative helicity gluons in different positions:
i j4
N = 4 SYM: An [g + g + . . . gi− . . . g −j . . . g + ] = An [g − g − g + g + . . . g + ] .
124
(4.35)
Exercise 4.11
Derive (4.35) by extending the strategy you used for (4.34).
Convince yourself that the supersymmetry Ward identities cannot mix gluon ampli-
tudes with different K in the N K MHV classification.
You were prompted in Exercise 4.9 to note that N = 1 supersymmetry would not be
sufficient to derive (4.35). We need to relate the positive and negative helicity gluons and
this is only possible because the N = 4 supermultiplet is CPT self-conjugate.
Gluon tree amplitudes do not see the supersymmetry partners, so they must comply with
the supersymmetry Ward identities, in particular the tree amplitudes in pure Yang–Mills
theory must obey (4.35). Indeed, this relation is automatic in the Parke–Taylor formula
(2.114) for the tree-level MHV gluon amplitudes! Furthermore, the focus on the “split-
helicity” MHV amplitude An [1− 2− 3+ . . . n + ] in our recursive proof of the Parke–Taylor
formula in Section 3.2 is now justified: the supersymmetry Ward identities (4.35) ensure
that Parke–Taylor holds for an MHV tree gluon amplitude with the two negative helicity
gluons in any position.
N K MHV classification – revisited

We introduced N K MHV amplitudes in Section 2.7 as the gluon amplitudes with K +2
negative helicity gluons. Now we can expand this to define the N KMHV sector to be all
amplitudes in N = 4 SYM connected to the N K MHV gluon amplitude via supersymmetry.
For example, all the amplitudes in (4.33), (4.34), and (4.35) are MHV. In fact, all MHV
amplitudes in N = 4 SYM are proportional to An [g − g − g + g + . . . g + ].
The global SU (4) R-symmetry implies that an amplitude vanishes unless the external
states combine to an SU (4) singlet. This requires that the (upper) SU (4) indices appear as a
combination of (K + 2) sets of {1234} for N K MHV amplitudes. Since the supersymmetric
Ward identities relate amplitudes with the same number of SU (4) indices, this then provides
an alternative definition of the N K MHV sector.
On-shell superspace
It is highly convenient to introduce an on-shell1 superspace in order to keep track of the
states and the amplitudes. We introduce four Grassmann variables η A labeled by the SU (4)
1 There is no (known) off-shell superspace formalism for N = 4 SYM.

index A = 1, 2, 3, 4.2 This allows us to collect the 16 states into an N = 4 on-shell chiral
superfield (or “superwavefunction”)
1 1
= g+ + η A λ A − η A η B S AB − η A η B ηC λ ABC + η1 η2 η3 η4 g − . (4.36)
2! 3!
The relative signs are chosen such that the Grassmann differential operators
particle g + λ A S AB λ ABC g − = g 1234

N = 4 SYM: (4.37)
operator 1 ∂iA ∂iA ∂iB ∂iA ∂iB ∂iC ∂i1 ∂i2 ∂i3 ∂i4
select the associated state from ( pi ).

In the on-shell formalism, the supercharges are represented as
∂ †ȧ
q A a ≡ [ p|a , q A ≡ | pȧ η A , (4.38)
∂η A
where | p and | p] are the spinors associated with the null momentum p of the particle.
Exercise 4.12
Show that the supercharges satisfy the standard supersymmetry anticommutation re-
lation (4.14), i.e. {q A a , q̃ Bḃ } = δ A B | pḃ [ p|a = −δ A B p ḃa . The supercharges (4.38) act
on the spectrum by shifting states right or left in . Check that this action on matches
(4.31).
Example. The purpose of this example is to clarify the relation between the on-
shell superspace introduced here and the off-shell superspace formalism described
in textbooks on supersymmetry (e.g. [37]). In an off-shell N = 1 formalism, the
ȧ ȧ
superspace is (x μ , θ a , θ ) with Grassmann variables θ a and θ . The algebra of the
supercharges is {Qa , Qb } = 0, {Qȧ , Qḃ } = 0, and {Qa , Qȧ } = i(σ μ )a ȧ ∂μ . As is often
convenient for studies of anti-chiral superfields, we can realize the superalgebra with
an anti-chiral representation
∂ ∂
Qa = , Qȧ = − + iθ a (σ μ )a ȧ ∂μ . (4.39)
∂θ a ∂θ
ȧ
The algebra (4.4) is represented faithfully, even when we replace Qȧ → iθ a (σ μ )a ȧ ∂μ .

In momentum space, the QQ-anticommutator can then be written
! "
∂
, − θ (σ )a ȧ pμ = −(σ μ )a ȧ pμ .
a μ
(4.40)
∂θ a
2 Originally, Ferber [41] introduced these variables as superpartners of bosonic twistor variables. Twistors are
discussed in Chapter 5.
82 Supersymmetry
Let us now go on-shell by assuming that p μ is light-like; then we can rewrite (4.40) in
spinor helicity formalism as
! "
∂
, θ a
| p]a p|ȧ = | p]a p|ȧ . (4.41)
∂θ a
Introduce a new Grassmann-odd variable η = θ a | p]a . Then ∂θ∂ a = | p]a ∂

∂η
, so that
(4.41) becomes
! " ! "
∂ a ∂
| p]a , η p|ȧ = | p]a p|ȧ =⇒ [ p| , | p η = | pȧ [ p|a .
ȧ
(4.42)
∂η ∂η
∂
The arguments, [ p|a ∂η and | pȧ η, of the anticommutator are recognized as N = 1
versions of our on-shell supersymmetry generators q a and q †ȧ in (4.38).
We note that dotting some arbitrary reference spinors |w̃ and [w| (whose brackets
∂
with the p-spinors are non-vanishing) into (4.42), we find {η , ∂η } = 1.
Consider the consequences of the above analysis. The superspace coordinates θ and
θ have mass-dimension (mass)−1/2 and the angle and square spinors have dimension
(mass)1/2 . So the on-shell superspace variables η are dimensionless. Under little group
scaling, the θ and θ are inert, and therefore we have η → t −1 η. Consequently, the
on-shell superwavefunction (4.36) scales homogeneously as → t −2 under a little
group scaling. The state-operator map (4.37) compensates this scaling when extracting
component states from .
Superamplitudes
We can think of i = ( pi ) as a superwavefunction for the ith external particle of a (color-
ordered) superamplitude An [1 , . . . , n ]. It depends on the on-shell momentum pi and a
set of Grassmann variables ηi A for each particle i = 1, . . . , n. Expanding An [1 , . . . , n ]
in the Grassmann variables, we note that the SU (4)-symmetry requires it to be a sum of
polynomials in ηi A of degree 4(K + 2). An example of a legal combination is ηi1 ηi2 ηi3 ηi4 ;
according to the map (4.37), it corresponds to particle i being a negative helicity gluon.
Using the map (4.37) to extract states from the superwavefunction of each particle
i = 1, 2, . . . , n, one projects out the regular amplitudes (sometimes called “component
amplitudes”) from the superamplitudes. For example:
# #
4
∂
4
∂
An 1+ . . . i − . . . j − . . . n + = An 1 , . . . , n
A=1
∂ηi A B=1
∂η j B ηkC =0

∂ ∂ ∂ ∂
An S 12 S 34 3− 4+ . . . n + =
∂η11 ∂η12 ∂η23 ∂η24
#
4
∂
× An 1 , . . . , n . (4.43)
A=1
∂η 3A ηkC =0
Grassmann differentiations can of course equally well be expressed as Grassmann

integrals.
The order K of the Grassmann polynomial precisely corresponds to the N K MHV sector.
So we can organize the full superamplitude as
2
An = AMHV
n + ANMHV
n + AnN MHV
+ · · · + Aanti-MHV
n , (4.44)
where AMHV
n has Grassmann degree 8, ANMHV n has Grassmann degree 12, etc. For example,
expanding the MHV superamplitude in Grassmann monomials, we have

AMHV
n = An g − g − g + . . . g + (η1 )4 (η2 )4 + An g − λ123 λ4 . . . g + (η1 )4 (η21 η22 η23 )(η34 ) + · · · ,
(4.45)
where (ηi )4 ≡ ηi1 ηi2 ηi3 ηi4 . (Quiz: why does the second term in (4.45) come with a
plus?)
Next, consider the supersymmetry Ward identities. In the language of on-shell su-
perspace, the supersymmetry Ward identities are identical to the statement that the
supercharges

n n
∂
n
†

n
QA ≡ qiA = [i| , and Q̃ A ≡ qA = |i ηi A , A = 1, 2, 3, 4,
i=1 i=1
∂ηiA i=1 i=1
(4.46)
annihilate the superamplitude:
Q A An = 0 and Q̃ A An = 0 . (4.47)
Exercise 4.13
It may not be totally obvious that (4.47) encodes the supersymmetry Ward identities,
so the point of this exercise is to illustrate how it works. Start by writing the (relevant
terms in the) MHV superamplitude as in (4.45). The supersymmetry Ward identity
(4.47) says that the coefficient of each independent Grassmann monomial in Q̃ A AMHV n
has to vanish. Pick A = 4 and act with Q̃ 4 on AMHV n to extract all terms whose Grass-
mann structure is (η1 )4 (η2 )4 (η34 ). Use that to show that the “component amplitude”
supersymmetry Ward identity (4.32) follows from Q̃ 4 AMHV n = 0.
μ
The requirement that {Q A , Q̃ B } ∼ i pi annihilates the superamplitude is equivalent to
n
the statement of momentum conservation. The associated delta function δ 4 ( i=1 pi ) has
been left implicit throughout the early chapters, but we will begin to include it explicitly in
Section 5, where it plays a central role. For example, it makes momentum conservation an
μ
operational statement that the momentum operator P μ = i pi annihilates the amplitude:
it simply acts multiplicatively and the vanishing result follows from the fact that x δ(x) has
support nowhere.
84 Supersymmetry
Note that the action of Q̃ A on the superamplitude is multiplicative. We can therefore

solve it easily using a Grassmann delta function δ (8) ( Q̃) defined as

1 # 1 #
4 4 n
δ (8) Q̃ = 4 Q̃ Aȧ Q̃ ȧA = 4 i jηi A η j A . (4.48)
2 A=1 2 A=1 i, j=1
Exercise 4.14
Show that momentum conservation ensures that Q A annihilates δ (8) ( Q̃).
Which property of angle spinors allows you to demonstrate explicitly that
Q̃ A δ (8) ( Q̃) = 0?
Half the supersymmetry constraints, namely Q̃ A An = 0, are satisfied if we write the

N K MHV superamplitude as
K

ANn
MHV
= δ (8) Q̃ Pn(4K ) , (4.49)
(4K ) (4K )
where Pn is a degree 4K polynomial in the Grassmann variables. If Pn is annihilated
by each Q A , then – by Exercise 4.14 – all the supersymmetry constraints are solved. The
Grassmann delta function δ (8) ( Q̃) can be viewed as conservation of supermomentum.
The Grassmann delta function is a degree 8 polynomial in the ηi A s. This means that for
(0)
an MHV superamplitude, δ (8) ( Q̃) fixes the ηi A -dependence completely and Pn is just a
normalization that depends on the momenta. Thus, the tree-level MHV superamplitude of
N = 4 SYM is

δ (8) Q̃
N = 4 SYM tree-level: AMHV [123 . . . n] = . (4.50)
n
1223 · · · n1
This object will appear repeatedly in our study of amplitudes in N = 4 SYM, both at tree-
and loop-level, so now is a good time to become familiar with it.
It is not hard to see that the MHV superamplitude (4.50) produces the Parke–Taylor
gluon tree amplitudes correctly. Just use the map (4.37) to take four derivatives with respect
to ηi A and four with respect to η j A as in (4.43). Then the delta function produces the
numerator-factor i j4 of the component amplitude An [1+ . . . i − . . . j − . . . n + ]. Note that
one component amplitude and supersymmetry uniquely fix the form of all MHV amplitudes
in N = 4 at each order in perturbation theory.
Exercise 4.15
Reproduce the three supersymmetry Ward identities (4.33), (4.34), and (4.35) from the
MHV superamplitude AMHVn in (4.50).
Exercise 4.16
Use the N = 4 SYM superamplitude AMHV n (4.50) to calculate the 4-scalar amplitude
A4 [S 12 S 34 S 12 S 34 ]. Compare your answer to the 4-scalar amplitude (3.41).
Calculate A4 [S 12 S 23 S 34 S 41 ] and compare with (3.45).
In our discussion so far, we have slipped silently by the anti-MHV 3-point amplitudes
Aanti-MHV
n , whose supersymmetry orbit determines the anti-MHV sector with K = −1.
Thus, the 3-point anti-MHV sector is encoded in degree-4 superamplitudes. We simply
state the answer,
1

Aanti-MHV
3 = δ (4) [12]η3 + [23]η1 + [31]η2 , (4.51)
[12][23][31]
with the delta function defined as

#4

δ [12]η3 + [23]η1 + [31]η2 ≡
(4)
[12]η3A + [23]η1A + [31]η2A , (4.52)
A=1
and leave it as an exercise to show that (4.51) is annihilated by the supercharges.
Exercise 4.17
Show that Q A Aanti-MHV
n = 0 and Q̃ A Aanti-MHV
n = 0.
K
Let us now outline three approaches to determining the superamplitudes AN
n
MHV
beyond
the MHV level.
1. Solution to the supersymmetry Ward identities. The N K MHV superamplitudes in N = 4

SYM must obey the supersymmetry Ward identities (4.47). Writing the L-loop super-
N K MHV (4K ) (4K )
amplitude An,L = δ (8) ( Q̃) Pn;L , supersymmetry requires Q A Pn;L = 0. In addi-
(4K )
tion, we need Pn;L to obey the Ward identities of the global SU (4) R-symmetry. As
we have seen, these constraints are trivially satisfied at the MHV level where one L-
(0)
loop component-amplitude suffices to fix Pn;L and hence the full L-loop MHV n-point
superamplitude.
For non-MHV amplitudes, how many component-amplitudes does one need to fix
the N K MHV superamplitude? This question is answered by analyzing the constraints
(4K ) (4K )
Q A Pn;L = 0 and R-symmetry. It turns out that for general K > 0, Pn;L can be written
as a linear combination of L-loop “basis amplitudes” with coefficients built from the
polynomials m i jk,A ≡ [i j]ηk A + [ jk]ηi A + [ki]η j A . This is the same type of Grassmann
polynomial that appears in Aanti-MHV 3 given in (4.51). The solution reveals that the
K
number of component-amplitudes sufficient to determine AN n,L
MHV
is the dimension
of the irreducible representation of SU (n−4) corresponding to a rectangular Young
diagram with K rows and (N=)4 columns. The independent component amplitudes
are labeled by the semi-standard tableaux of this Young diagram.3 A given particle
content (gluons, fermions, scalars) of a basis amplitude corresponds to an ordered
partition of the number 4K into n integers between 0 and 4. For each such partition, the
corresponding Kostka number counts the number of independent arrangements of the
SU (4) R-symmetry indices. You can read more about the solution to the supersymmetry
and R-symmetry Ward identities in [42, 43].
3 By manipulating the color-structure, one can improve further on this count of “basis amplitudes” [42, 43].
86 Supersymmetry
K
2. The super-MHV vertex expansion. The tree-level superamplitudes AnN MHV of N = 4
SYM can be constructed as an MHV vertex expansion in which the vertices are the
MHV superamplitudes AMHV n of (4.50). The expansion can be derived from an all-line
shift, as we discussed in Section 3.4. The validity of the all-line shift for tree-level
superamplitudes in N = 4 SYM was proven in [29] (see also [44]). A new feature
is that one must sum over the possible intermediate states exchanged on the internal
lines of the MHV vertex diagrams. This is conveniently done by integrating over the
Grassmann variables η PI A associated with internal lines; this automatically carries out
a super-sum [36]. More details about such super-sums will be offered in the next
section.
3. Super-BCFW. Superamplitudes can be constructed with a supersymmetric version of
the BCFW shift. This construction plays a central role in many developments, so we will
treat it in detail in the following section.
4.5 Super-BCFW and all tree-level amplitudes in N = 4 SYM
The BCFW shift introduced in Section 3.2 preserves the on-shell conditions pi2 = 0 and
n
momentum conservation i=1 pi = 0. However, the shift does not preserve the condition
n
for supermomentum conservation, i=1 |iηi A = 0. As a consequence, the shifted com-
ponent amplitudes have large-z falloffs that depend on which types of particles are shifted,
for example we have seen the difference between the [−, + and [+, − BCFW shifts of
gluons in (3.13). This can be remedied by a small modification of the BCFW shift (3.9)
that allows us to conserve supermomentum. We simply accompany the BCFW momentum
shift by a shift in the Grassmann-variables [45–47]: for simplicity let us focus on the [1, 2
super-BCFW shift (or [1, 2 “supershift”), defined as
|1̂] = |1] + z |2] , |2̂ = |2 − z|1 , η̂1A = η1A + z η2A . (4.53)
No other spinors or Grassmann variables shift.
Exercise 4.18

Show that the supermomentum is conserved under the supershift (4.53) so that δ (8) Q̃
is invariant.
It follows directly from (4.50) that the MHV superamplitudes have a 1/z falloff under a
supershift of any adjacent lines (1/z 2 for non-adjacent). In fact, this falloff behavior holds
true for any tree superamplitude in N = 4 SYM, as can be shown by using supersymmetry to
rotate the two shifted lines to be positive helicity gluons and using the non-supersymmetric
result (3.13) for the falloff under a [+, +-shift [19, 47].
The tree-level recursion relations that result from the super-BCFW shift (4.53) involve
diagrams with two superamplitude “vertices” connected by an internal line with on-shell
momentum P̂. As in the non-supersymmetric case, we must sum over all possible states that
can be exchanged on the internal line: in this case, this includes all 16 states of N = 4 SYM.
In terms of component amplitudes, the particle exchanged on the internal line depends on
the external states: if they are all gluons, then the internal line is also a gluon and one must
simply sum over the helicities. The superamplitude version of this helicity sum is
$ % & # '
#
4
∂
4
∂
1 1
ÂL ÂR + ÂL 2 ÂR , (4.54)
∂η P̂ A P 2 P ∂η P̂ A
A=1 A=1 η P̂ A =0
where η P̂ A is the Grassmann variable associated with the internal line. In the first term of
(4.54), a negative helicity gluon is projected out from the left sub-superamplitude leaving
a positive helicity gluon coming out of the right sub-superamplitude. For the second term,
the helicity projection is the opposite.
If a gluino can be exchanged on the internal line,4 then we have to move one of the four
Grassmann derivatives from ÂL to ÂR in the first term of (4.54) – in all four possible ways
to sum over the four different gluinos labeled by the SU (4)-indices. And similarly for the
second term. A scalar exchange means that two Grassmann derivatives act on ÂL and the
two others on ÂR . All in all, the entire sum over states exchanged on the internal line can
be written
#
4 & %
∂ 1 1
ÂL 2 ÂR = d 4 η P̂ ÂL ÂR . (4.55)
∂η P̂ A P P2
A=1 η P̂ A =0
Note how the product rule distributes the Grassmann derivatives ∂/∂η P̂ A on the L and R
superamplitudes in all possible ways to automatically carry out the state super-sum. In
(4.55), we have rewritten the Grassmann differentiation as a Grassmann integral. Similar
super-sums are used in evaluation of unitarity cuts (see Chapter 6) of loop amplitudes
where one includes integration over the Grassmann variable associated with the internal
line [36, 48].
Remarkably, the super-BCFW recursion relations can be solved to give closed-form
expressions for all tree-level superamplitudes in N = 4 SYM [49]. We are now going to
show how this works. As a warm-up, we first verify that the MHV superamplitude formula
(4.50) satisfies the super-BCFW recursion relations. Then we present the most essential
details of the derivation of the tree-level NMHV superamplitude. Finally we comment
briefly on the results for N K MHV.
4.5.1 MHV superamplitude from super-BCFW

Consider the super-BCFW recursion relations for the MHV superamplitude. Just as in the
non-supersymmetric case (3.20), there is just one non-vanishing diagram, but we must now
4 This happens when there is an odd number of external gluinos on each side of the BCFW diagram.
88 Supersymmetry
include the super-sum (4.55) over states exchanged on the internal line:
1^ 2^
n PÎ
AMHV
n [123 . . . n] = L R
4 3
MHV anti-MHV

1
= d 4 η P̂ Ân−1 1̂, P̂ , 4, . . . , n 2
Â3 − P̂ , 2̂, 3
P

δ (8) i∈L |îη̂i 1
= d 4 η P̂
1 P̂ P̂ 445 · · · n1 P 2

δ (4) [ P̂2]η3 + [23]η P̂ + [3 P̂]η2
× , (4.56)
[23][3 P̂][ P̂ 2]
where P = P23 = p2 + p3 and we use (4.50) and (4.51) for the (n−1)-point MHV super-
amplitude and 3-point anti-MHV superamplitude. We also use the analytic continuation
rule (3.17) for | − P] = |P].
The new feature is the Grassmann integral in (4.56). The first delta function is

n
δ (8)
|îη̂i = δ (8)
|1η̂1 + | P̂η P̂ + |iηi . (4.57)
i∈L i=4
On the support of the second delta function, we can set η P̂ = −([ P̂2]η3 + [3 P̂]η2 )/[23] to
find
| P̂

|1η̂1 + | P̂η P̂ = |1η̂1 − [ P̂2]η3 + [3 P̂]η2
[23]
= |1η̂1 + |3η3 + |2̂η2
= |1η1 + |3η3 + |2η2 . (4.58)
We used momentum conservation on the right-hand subamplitude to obtain the second line
in (4.58) and the definition (4.53) of the supershift to get the third line. This way, the first
n
delta function simply becomes δ (8) ( i=1 |iηi ) = δ (8) ( Q̃).
Now the only η P̂ -dependence is in δ (4) (. . . ) and the Grassmann integral is easy to carry
out:

d η P̂ δ
4 (8)
|îη̂i δ (4) [23]η P̂ + · · · = [23]4 δ (8) Q̃ .
i∈L
Coming back to (4.56), we then have

δ (8) Q̃ 1 [23]4
AMHV
n [123 . . . n] = . (4.59)
1 P̂ P̂ 445 · · · n1 P 2 [3 P̂][ P̂ 2][23]
Compare this with (3.21) and you will see that the two expressions are the same, except
that δ (8) ( Q̃) has replaced 1 P̂4 and [23]4 has replaced [3 P̂]4 . Using the identities (3.22)
and (3.23), we promptly recover the desired result

δ (8) Q̃
AMHV [123 . . . n] = . (4.60)
n
1223 · · · n1
Thus we have shown that the N = 4 SYM tree-level MHV superamplitude (4.60) satisfies
the super-BCFW recursion relations. Next, we use super-BCFW to derive an important
result for NMHV superamplitudes in N = 4 SYM.
4.5.2 NMHV superamplitude and beyond

The super-BCFW recursion relation for the NMHV superamplitude involves two types of
diagrams
1^ 2^ 1^

n
n PÎ n PÎ 2^
ANMHV
n [12 . . . n] = L R 3 + L R . (4.61)
k=5 k k–1 4 3
MHV MHV NMHV anti-MHV
diagram A diagram B
The diagrams of type A involve two MHV sub-superamplitudes. Diagram B is present only
for n ≥ 6 because the L subamplitude is NMHV and therefore needs at least five legs;
indeed, back in Section 3.2, we only had an MHV×MHV diagram in the calculation (3.27)
of A5 [1− 2− 3− 4+ 5+ ].
Diagram B provides the setting for an inductive proof of the NMHV superamplitude for-
mula we are seeking, while the diagrams of type A give an “inhomogeneous” contribution.
We begin with a detailed evaluation of the type A diagrams. There will be a lot of detailed
calculations in this example calculation, so if you just want the result, you are free to skip
ahead to the answer in (4.82). Right after the example, we summarize the full result for the
NMHV superamplitude.
Example. Calculation of diagram A. The first step is simply to plug in the MHV
superamplitudes:

δ (8) L 1
Diagram A = d η P̂
4
1 P̂ P̂ kk, k + 1 . . . n1 P 2

δ (8) R
× . (4.62)
P̂ 2̂2̂334 . . . k − 1, P̂
Here P = p2 + p3 + · · · + pk−1 and (with the help of the rule (3.17)) we have

n
δ L =δ
(8) (8)
−| P̂η P̂ + |1η̂1 + |r ηr ,
r =k
(4.63)

k−1
δ R = δ (8) | P̂η P̂ + |2̂η2 +
(8)
|r ηr .
r =3
90 Supersymmetry
n
On the support of δ (8) (R), we can write δ (8) (L) = δ (8) (L + R) = δ (8) ( i=1 |iηi ) =
δ ( Q̃), which is independent of η P̂ and expresses the conservation of supermomentum
(8)

for the n external states. Now only δ (8) (R) remains under the state-sum integral d 4 η P̂ ;
since δ (8) enforces two conditions we can project out two separate δ (4) s:

1
k−1
k−1
δ R =
(8)
δ (4)
1 P̂η P̂ + 12̂η2 + 1r ηr δ (4)
P̂ 2̂η2 + P̂ r ηr .
1 P̂4 r =3 r =3
(4.64)
There is η P̂ -dependence only in the first of these two δ (4) s, so it is straightforward
to perform the Grassmann integral: it produces a factor 1 P̂4 which cancels the
normalization factor included in (4.64). All in all, we have then shown that

(8)

(4)
k−1
d η P̂ δ L δ R = δ Q̃ δ
4 (8) (8)
P̂ 2̂η2 + P̂ r ηr . (4.65)
r =3
The factor δ (8) ( Q̃) can be used to pull out an overall factor AMHV n from diagram A in
(4.62) and we can then write

1223k − 1, k δ (4) P̂ 2̂η2 + rk−1 =3 P̂ r ηr
Diagram A = AMHV n . (4.66)
k P̂k − 1, P̂2̂ P̂2̂31 P̂ P 2
Next, we turn our attention to the brackets in (4.66) that involve the shifted spinors
| P̂ and |2̂. As in the non-supersymmetric examples of Section 3.2, we manipulate
these brackets by multiplying the numerator and denominator of (4.66) by [ P̂ 2]4 214 .
We are going to encounter sums of momenta that appear in the ordering fixed by color-
structure, so for convenience we introduce the shorthand notation
yi j ≡ pi + pi+1 + · · · + p j−1 . (4.67)
Further, we declare that y ji with j > i equals −yi j ; when we think of the pi as cyclically
ordered, this simply expresses momentum conservation. We will use the variables yi j
to write the result for diagram A in a form that may look slightly mysterious for now,
but it has some very important features that are discussed in Chapter 5.
One type of bracket in (4.66) is r P̂ with r = 1, 2̂. We manipulate this as follows:

r P̂[ P̂ 2]21 = −r | 2̂ + 3 + · · · + (k − 1) |2]21 = −r |y3k |2]21
= r |y3k · y23 |1 = r |y3k · y13 |1 = −1|y13 · y3k |r . (4.68)
The double angle-bracket is defined as i| p · q| j ≡ i|ȧ pȧb qbċ | jċ . The result (4.68)
applies to the brackets r P̂, k P̂, and k − 1, P̂ in (4.66).
Exercise 4.19
To keep you actively engaged, involved, and awake, here is an exercise: show that
i|K .K | j = −K 2 i j (4.69)
for any momentum K (light-like or not) and any spinors i| and | j.
The two brackets 1 P̂ and 2̂ P̂ from the denominator of (4.62) are dealt with as
follows:
1 P̂[ P̂ 2] = −1|y3k |2] and 2̂ P̂[ P̂ 2] = −2 p̂2 · P̂ = − P̂ 2 + y3k

2
= y3k
2
.
(4.70)
In the last equality, we used the fact that the BCFW diagram is evaluated on the value
of z such that P̂ 2 = 0. It is useful to note what this value of z is:
2
y2k
0 = P̂ 2 = 2̂| P̂|2] + y3k
2
= y2k
2
− z1|y3k |2] =⇒ z= . (4.71)
1|y3k |2]
We use this z to evaluate 2̂3:

2
y2k 1|y3k |2]23 − y2k 2
13
2̂3 = 23 − 13 =
1|y3k |2] 1|y3k |2]
1|y1k |2]23 + 1|y2k · y2k |3 −1|y1k · p2 |3 + 1|y1k · y2k |3
= =
1|y3k |2] 1|y3k |2]
1|y1k · (− p2 + y2k )|3 1|y1k · y3k |3
= = . (4.72)
1|y3k |2] 1|y3k |2]
In the third equality, we use the result of Exercise 4.19. Combining (4.70) and (4.72),
we can write
2̂31 P̂[ P̂ 2] = 1|y1k · yk3 |3 . (4.73)
We have now evaluated all four angle brackets involving | P̂ and |2̂ in (4.66). There
is, however, another manipulation that we would like to do, namely one involving the
propagator 1/P 2 = 1/y2k
2
. It goes like this:
P 2 12 = y2k
2
12 = −1|y2k · y2k |2 = −1|y1k · y3k |2 = 1|y1k · yk3 |2 . (4.74)
And then it is time to put everything together. We can now write diagram A from (4.66)
as

23k − 1, k [ P̂ 2]4 214 δ (4) P̂ 2̂η2 + rk−1=3 P̂ r ηr
Diagram A = AMHV n .
2
y3k 1|y13 · y3k |k1|y13 · y3k |k − 11|y1k · yk3 |31|y1k · yk3 |2
(4.75)
Let us examine the δ (4) in the numerator. We absorb the factors [ P̂ 2]4 214 into the
delta function whose argument then becomes (suppressing SU (4)-indices)

k−1
k−1
≡ 12[2 P̂] P̂ 2̂η2 + 12[2 P̂] P̂ r ηr = 12y3k
2
η2 + 1|y13 .y3k |r ηr .
r =3 r =3
(4.76)
92 Supersymmetry
We are going to do a little work on this expression in order to introduce another piece
of shorthand notation, namely fermionic companions of the yi j s defined in (4.67):

j−1
|θi j,A ≡ |r ηr A . (4.77)
r =i
Then |θ ji = −|θi j encodes supermomentum conservation.
We start by rewriting each of the terms in (4.76):
12 y3k
2
η2 = −1|y3k · y3k |2 η2 = −1|y1k · y3k |2 η2 + 1|y23 · y3k |2 η2
= −1|y1k · y3k |θ13 + 1|y1k · y3k |1 η1 + 1|y13 · y3k |2 η2 (4.78)
and

k−1
1|y13 · y3k |r ηr = 1|y13 · y3k |θ1k − 1|y13 · y3k |1 η1 − 1|y13 · y3k |2 η2 .
r =3
(4.79)
Adding (4.78) and (4.79) to get , the two extra η2 -terms cancel and the η1 -terms
combine to
1|y1k · y3k |1 η1 − 1|y13 · y3k |1 η1 = 1|(y1k − y13 ) · y3k |1 η1
= 1|y3k · y3k |1 η1 = 0 (4.80)
by (4.69). Thus we have
= −1|y1k · yk3 |θ31 − 1|y13 .y3k |θk1 . (4.81)
Our work brings us to the following form of a diagram of type A:

23k − 1, k δ (4) 1|y1k · yk3 |θ31 + 1|y13 · y3k |θk1
Diagram A = AMHV
n .
2
y3k 1|y13 · y3k |k1|y13 · y3k |k − 11|y1k · yk3 |31|y1k · yk3 |2
(4.82)
This completes our calculation of diagram A in the super-BCFW recursion relation
(4.61). Next, we discuss what the full NMHV superamplitude looks like.
The result (4.82) for diagram A is often written AMHV
n R13k , where the so-called R-
invariants5 are defined as

j − 1, jk − 1, k δ (4) 1 jk
R1 jk = , (4.83)
y 2jk 1|y1 j · y jk |k1|y1 j · y jk |k − 11|y1k · yk j | j1|y1k · yk j | j − 1
where
1 jk,A = 1|y1k · yk j |θ j1,A + 1|y1 j · y jk |θk1,A . (4.84)
5 We discuss in Chapter 5 under which symmetries Ri jk is invariant.

For completeness, we repeat here the definitions (4.67) and (4.77) of the variables yi j = −y ji
and θi j,A = −θ ji,A :

j−1
yi j ≡ pi + pi+1 + · · · + p j−1 and |θi j,A ≡ |r ηr A . (4.85)
r =i
Note the structure of (4.83) and (4.84): neighbor indices match up with each other. This is
an important feature.
For n = 5, diagram B vanishes and diagram A with k = 5 is the complete result:
ANMHV
5 = AMHV
5 R135 . (4.86)
n
For n > 5, the diagrams of type A contribute AMHV n k=5 R13k . Diagram B recurses this
form (we will not present the details, see [49]) and the result is a sum of R-invariants:

n−2
n
Diagram B = AMHV
n R1 jk . (4.87)
j=4 k= j+2
Hence, the entire NMHV superamplitude can be expressed in terms of the R-invariants as

n−2
n
ANMHV
n = AMHV
n R1 jk . (4.88)
j=3 k= j+2
For example,

ANMHV
6 = AMHV
6 R135 + R136 + R146 , (4.89)
and

ANMHV
7 = AMHV
7 R135 + R136 + R137 + R146 + R147 + R157 . (4.90)
A generic Feynman diagram calculation of the 7-gluon NMHV component amplitude

requires 154 diagrams. It is a dramatic simplification that reduces the sum of Feynman
diagrams down to the six terms in (4.90). Moreover, note that it is not just the Feynman
diagrams for gluon scattering that are encoded in the sums of Ri jk s, rather the relatively
simple expression (4.88) encodes all tree-level NMHV scattering amplitudes of the entire
N = 4 SYM spectrum.
The remarkably simple result (4.88) for the all-n tree-level NMHV superamplitudes in
N = 4 SYM was first found in studies of loop amplitudes [50]. Later, the NMHV formula
was constructed as we did here using super-BCFW [49].
One final comment about the R-invariants (4.83) is that we defined them here “anchored”
at momentum line 1; this came about because we used a [1, 2-supershift. However, by
cyclic symmetry all pairs of adjacent lines in the superamplitude are on equal footing, so
an [i, i + 1-supershift would have resulted in an NMHV formula anchored at i, giving
R-invariants Ri jk . These are defined by replacing all momentum labels 1 in (4.83) by i. In
94 Supersymmetry
the literature you will often find the expressions for the NMHV superamplitude given with
n−3 n−1
n as the anchor; i.e. ANMHV
n = AMHV
n j=2 k= j+2 Rn jk .
Exercise 4.20
Write down the NMHV superamplitude formula that results from a [2, 3-supershift.
Then project out the gluon amplitude A6 [1+ 2− 3+ 4− 5+ 6− ]. Can you match your result
to the 3-term expression in Exercise 3.9? Can you guess what expression you get from
a [3, 2-supershift after projecting out A6 [1+ 2− 3+ 4− 5+ 6− ]? The answer is given in
Chapter 10.
Formulas for tree-level N K MHV superamplitudes have also been derived – they take the
form of sums of generalized R-invariants [49]. We refer you to the original paper [49], the
discussions in [46, 47, 51], and the review [52] for further details.
The superamplitudes of N = 4 SYM are invariant under a large symmetry group. In
particular, there are both “ordinary” and “hidden” symmetries, and they are realized for the
tree-level superamplitudes. The structure of the R-invariants is essential for this. We will
discuss these symmetries next in Chapter 5.
Symmetries of N = 4 SYM 5
This chapter is dedicated to a detailed description of the symmetries of superamplitudes in

N = 4 SYM. As mentioned in Section 4.4, N = 4 SYM is a superconformal theory.1 In four
dimensions, the conformal group is S O(4, 2) ∼ SU (2, 2), and with N = 4 supersymmetry,
it is enhanced to the superconformal group SU (2, 2|4). Its bosonic symmetries consist of
the conformal group S O(4, 2) and the R-symmetry group SU (4). Our first task is to show
that the generators of SU (2, 2|4) annihilate the N = 4 SYM superamplitudes; this shows
that the superamplitudes respect SU (2, 2|4) symmetry. In Section 5.1, we demonstrate this
property using the spinor helicity formalism and the on-shell superspace from Section 4.4.
And in Section 5.2 we introduce the – for this purpose – more suitable language of twistors.
The superconformal symmetry SU (2, 2|4) can be called the “obvious” symmetry of
N = 4 SYM. This is said to contrast that there is also a “hidden” symmetry: it turns
out that the tree-level superamplitudes of N = 4 SYM enjoy another symmetry, namely
dual superconformal symmetry SU (2, 2|4).2 We treat this in Section 5.3, where we also
consider the consequences of having both “ordinary” and dual superconformal symmetry.
The natural variables for dual superconformal symmetry are momentum twistors and they
are introduced in Section 5.4 for our enjoyment now and later.
5.1 Superconformal symmetry of N = 4 SYM
All theories studied in this book are Poincaré invariant. In four dimensions, the ten Poincaré
symmetry generators, P μ and M μν , can be written in spinor helicity formalism. This is
achieved by converting P μ and M μν to momentum space and then contracting the Lorentz
indices with (σ̄ μ )ȧb and (σ μ σ̄ ν − σ ν σ̄ μ )a b (and its conjugate). The action of the Poincaré
generators on the scattering amplitudes can then be realized by the operators

P ȧb = − |iȧ [i|b , Mab = |i](a ∂[i|b) , Mȧ ḃ = i|(ȧ ∂|iḃ) , (5.1)
i i i
where the sum is over external particle labels i = 1, 2, . . . , n and (..) indicates symmetriza-
tion of the enclosed indices. The operators (5.1) satisfy the Poincaré commutator algebra
(4.3) without imposing momentum conservation on the n momenta.
1 At the origin of moduli space.

2 Dual conformal symmetry at loop-level is discussed in Sections 6.4 and 6.5.
96 Symmetries of N = 4 SYM
Exercise 5.1
Show that

i|ȧ ∂|iḃ jk = ȧ ḃ jk (5.2)
i
and hence Mȧ ḃ annihilates angle brackets. Mab trivially gives zero on any angle brack-
ets. The equivalent conclusions hold for square brackets.
[Hint: use Aab − Aba = −Ac c ab which is valid for any 2 × 2 matrix.]
In the earlier chapters, we have imposed momentum conservation on the external momenta
“by hand.” For the purpose of studying the symmetries of the amplitudes it is necessary to

include the momentum conservation delta function δ 4 ( i pi ) explicitly in the amplitudes.
For example, for the MHV gluon amplitude
i j4

An 1+ . . . i − . . . j − . . . n + = δ4 P . (5.3)
1223 · · · n1
The translation generator P ȧa acts multiplicatively and annihilates an amplitude in the
distributional sense P ȧa δ 4 (P) = 0. The action of the rotations/boosts follows from the
following useful identity:
Example. Calculate

ȧ ∂ P ċd ∂
∂

|iȧ ∂|iḃ δ 4 P = |i δ 4
P = P ȧd
δ 4
P = −2δ ȧ
δ 4
P ,
∂|iḃ ∂ P ċd
ḃ
i i
∂ P ḃd
(5.4)

where the last equality holds as a distribution
since x f (x) ∂x δ(x) = − f (x) δ(x)
and ∂∂ PP ċd = 2δ ȧ ḃ . It follows that Mȧ ḃ δ 4 P = 0.
ȧd

By Exercise 5.1, we already know that the rotations/boosts annihilate the angle and square
brackets, so now we conclude that the Poincaré generators annihilate the amplitudes; this
is of course not surprising, it is the explicit realization of the statement that the amplitudes
are Lorentz-invariant and conserve momentum.
For a supersymmetric theory, the Poincaré generators (5.1) are supplemented by the
supersymmetry generators Q and Q̃; the supersymmetry generators for N = 4 SYM were
given in (4.46) in on-shell superspace. We have already discussed that the annihilation of the
superamplitudes by the supersymmetry generators encodes the supersymmetry Ward identi-

ties. In particular, the supermomentum conserving Grassmann-delta function δ (8) ( i |iηi )
implies that Q̃ A annihilates the superamplitude. Note that for the tree-level MHV super-
amplitude

δ 4 P δ (8) Q̃
AMHV = n , (5.5)
i=1 i, i + 1
n
5.1 Superconformal symmetry of N = 4 SYM 97
momentum and supermomentum conservation (defined in (4.48)) now appear on an equal

footing. In the denominator of (5.5), we set |n + 1 = |1.
The spacetime symmetry of N = 4 SYM is enlarged to the superconformal group. This
means that in addition to the super-Poincaré generators,
♦ four translations and six boosts and rotations in (5.1),

♦ 16 fermionic supersymmetry generators Q Aa and Q̃ ȧA in (4.46),
the superconformal algebra also has
♦ four conformal boosts K a ȧ ,

♦ one dilatation D,
♦ 15 SU (4) R-symmetry generators R A B , satisfying the traceless condition R C C = 0,
♦ 16 fermionic conformal supersymmetry generators S̃ȧA and Sa A .
Together, these 16 + 16 = 32 fermionic and 4 + 6 + 4 + 1 + 15 = 30 bosonic generators

form the graded Lie algebra su(2, 2|4) of the superconformal group.3 Introducing a collec-
tive index A = (a , ȧ, A) we denote the superconformal generators as G A B .
We are going to realize the generators G A B ∈ su(2, 2|4) in the following form, organized
here according to their mass dimensions:

P ȧb = − i |iȧ [i|b ,

Q̃ ȧA = |iȧ ηi A Qa A = [i|ia ∂ηi A
i i
D= 1
2
|iȧ
∂|iȧ + 12 |i]a ∂|i]a + 1
Mȧ ḃ = i|(ȧ ∂|iḃ)
i Mab = |i](a ∂[i|b)
i RA B
= ηi A ∂ηi B − 14 δ A B ηiC ∂ηiC i
i
S̃ȧA = ∂|iȧ ∂ηi A Sa A = ∂[i|a ηi A
i i
K a ȧ = −∂ |iȧ ∂[i|a .
(5.6)
n
These generators are given as a sum, G A B = i=1 G iA B , of operators G iA B that are
each defined on one external leg i; this reflects the local nature of the symmetry. In
Section 5.3, we will encounter symmetries whose generators are “non-local” in that they
involve products of operators that act on different legs.
Exercise 5.2
Show that the action of the first two terms in the dilatation operator D in (5.6) extracts
the mass dimension from an expression constructed from angle and square brackets.
Then show that D annihilates amplitudes when including δ 4 (P).
3 U (2, 2|4) has 32 fermionic and 32 bosonic generators, i.e. it has two more U (1)s than SU (2, 2|4).
We define the “helicity operator” H as

H= |iȧ ∂|iȧ − |i]a ∂|i]a − ηi A ∂ηi A + 2 . (5.7)
i
The first three terms in H act as follows: |iȧ → |iȧ , |i]a → −|i]a , and η A → −η A , i.e. it
pulls out the correct little group weight of each on-shell variable. Hence, when H acts on a
component amplitude, it extracts the sum of the helicity weights plus 2n from the last term

in (5.7): in the N K MHV sector this is i (−2h i ) + 2n = 4K + 8. But such a component
amplitude appears in the superamplitude multiplied by a factor of 4(K + 2) Grassmann
K
variables ηi s. Therefore H AnN MHV = 0.
Together with the helicity operator H , the generators in (5.6) form a closed algebra. For
example, we encourage you to check that:

1 B b 1 B b 1
{Sa A , Q } = δ A Ma + δa R A + δ A δa D − H .
bB b B
(5.8)
2 2 2
Since H vanishes on the amplitude, the generators in (5.6) close into the superconformal
group when acting on the on-shell amplitude.
We are now ready to study the action of superconformal symmetry on the amplitudes of
N = 4 SYM: the superamplitudes are invariant under the full superconformal symmetry
group, so one should find G A B An = 0 .
We restrict our analysis to the MHV superamplitude (5.5). We have already discussed that
this superamplitude is super-Poincaré invariant. To prove that it enjoys the full SU (2, 2|4)
symmetry, it is sufficient to check that the amplitude vanishes under the conformal su-
persymmetries Sa A and S̃ȧA . For example, the anticommutator of Sa A and S̃ȧA gives the
conformal boost K a ȧ . The anticommutator (5.8) is a another example. The following two
examples show that AMHV n is annihilated by Sa A and S̃ȧA .
Example. Let us show that Sa A = i ∂[i|a ηi A annihilates AMHV

n . Note that

Sa A δ 4 P = − |iȧ ηi A ∂ P ȧa δ 4 P = − Q̃ ȧA ∂ P ȧa δ 4 P (5.9)
i
vanishes on the support of δ (8) ( Q̃). Since AMHV does not have any further dependence

n
on square spinors, we conclude Sa A An MHV
= 0.
Example. To start with, consider the action of S̃ȧA = i ∂|iȧ ∂η iA

on δ (8) ( Q̃): it follows
from direct calculation that

∂ηi A δ (8) Q̃ = |iȧ ∂ Q̃ ȧA δ (8) Q̃ and S̃ȧA δ (8) Q̃ = (n − 1 + 3) ∂ Q̃ ȧA δ (8) Q̃ . (5.10)
To show the second identity, you need Q̃ C ȧ Q̃ Cḃ = 12 δȧ ḃ Q̃ C ċ Q̃ Cċ (no sum on C). We use
this result when we write

δ 4 (P)
ȧ δ 4 (P)
S̃ȧA AMHV = S̃ δ
A (8)
Q̃ n + ∂ ḃ δ (8)
Q̃ |i ∂ n .
i=1 i, i +1 i=1 i, i +1
n ȧ Q̃ A |iḃ
i
(5.11)
5.2 Twistors 99
We evaluate the second term using (5.2) and (5.4), and we then have

δ 4 (P)
S̃ȧA AMHV = (n−1+3) ∂ Q̃ ȧA δ (8) Q̃ + ∂ Q̃ ḃ δ (8) Q̃ (−2−n)δḃ ȧ n = 0.
i=1 i, i +1
n A
(5.12)
This completes the proof that AMHV respects the superconformal symmetry SU (2, 2|4)

n
of N = 4 SYM.4
We end this section with a remark about the operation of inversion. Inversion acts on the
spacetime coordinates as
xμ
I(x μ ) = . (5.13)
x2
The inversion operation generates the (super)conformal symmetry algebra from the (su-
per)Poincaré algebra, for example K μ = I P μ I.
Exercise 5.3
In this exercise, we derive the form of the conformal boost generator in position space;
this will be useful for us in Section 5.4. Write the momentum generator in position
space as P ȧa = −i ∂ x∂aȧ .5 Show that I P ȧa I is equivalent to
∂
Kȧa = −i x ȧc x ċa (5.14)
∂ x ċc
by demonstrating that I P ȧa I and Kȧa give the same result when acting on x ḃb .
[Hint: you will need the same type of identity given in the hint of Exercise 5.1.]
5.2 Twistors
The representation of the superconformal generators given in (5.6) is unusual in that the
generators appear with various degrees of derivatives. For example, the bosonic SU (2, 2)
subgroup of conformal transformations has a 2-derivative generator K a ȧ as well as a
multiplicative 0-derivative generator P ȧa . Since the realization of the generators depends on
the choice of variables – here spinor helicity variables – we can hope to find a set of variables
such that all generators are linearized as l-derivative operators. This is actually simple and
can be achieved by performing a “Fourier transformation” on the angle spinor variables:
∂ ∂
i|ȧ → i , → −iμ̃i |ȧ . (5.15)
∂|μ̃i ȧ ∂|iȧ
4 We have ignored here subtle points of non-generic momenta and anomalies.

5 We include here the “−i” in the translation generator; henceforth we will often drop the is in the symmetry
generators.
For example, the translation and conformal boost generators are linearized after the Fourier
transform:

P ȧb → i [i|b ∂μ̃i |ȧ , K a ȧ = i μ̃i |ȧ ∂[i|a . (5.16)
i i
Exercise 5.4
∂
Show that |iȧ → −i .
∂μ̃i |ȧ
Exercise 5.5
Show that the dilatation generator then becomes

D= 1
2
|iȧ ∂|iȧ + 12 |i]a ∂|i]a + 1 → 1
2
|i]a ∂|i]a − 12 μ̃i |ȧ ∂μ̃i |ȧ . (5.17)
i i
The new variables

WiA = [i|a , |μ̃i ȧ , ηi A , (5.18)
with a collective index A = (ȧ, a, A), are called supertwistors. In terms of these variables,
the generators of the superconformal algebra su(2, 2|4) can be written compactly as

n n
1
GAB = G iA B = WiA ∂WiB − δ A B WiC ∂WiC . (5.19)
i=1 i=1
4
The δ A B -term is necessary for the bosonic subgroups, SU (2, 2) and SU (4), to have traceless
generators. However, the term proportional to δ A B simply counts the degree of WiA , so if
the function one is interested in has vanishing weight in each WiA , the generators reduce to:

n
GAB = WiA ∂WiB . (5.20)
i=1
The bosonic components, WiI = ( [i|a , |μ̃i ȧ ), of the supertwistors (5.18) are simply called
twistors. They were first introduced by Penrose [53] in the context of describing flat
Minkowski spacetime. Later they were supersymmetrized by Ferber [41] and used to form
representations of the superconformal group. Under little group scaling, the supertwistors
scale homogeneously, Wi → ti Wi , and this leaves the symmetry generators (5.19) invariant.
That means that we can define the (super)twistors projectively: the bosonic twistors are
points in CP3 while the supertwistors live in CP3|4 .
Amplitudes in twistor space

Let us consider the n-gluon anti-MHV tree amplitude (2.115) in bosonic twistor space,
WiI = ( [i|a , |μ̃i ȧ ). The only angle-spinor dependence is in δ 4 (P), so a Fourier transfor-
mation of | j gives
# $ n '
n #
i j μ j i j μ j
d | je
2
An
anti-MHV
= d | je
2
δ (P) f (|i]) .
4
(5.21)
j=1 j=1
5.2 Twistors 101
Here f (|i]) is a function that only depends on |i]. To ease the integration, write the delta
function itself as a Fourier integration (ignoring factors of 2π ),

δ 4 (P) = d 4 x e−i xaȧ j | j [ j| .
ȧ a
(5.22)
The integration over | j can now be carried out and we find

# #
n n

d 2 | jei j μ̃ j Aanti-MHV
n = d 4
x δ 2
μ̃ |
j ȧ + [ j|a
x a ȧ f (|i]) . (5.23)
j=1 j=1
Hence, the twistor variables WiI = ( [i|a , |μ̃i ȧ ) are localized in twistor space by delta
functions that enforce, for each i, the “incidence relations”
μ̃i |ȧ + [i|a xa ȧ = 0 . (5.24)
This equation says that, for given xa ȧ , WiI is determined by just the input of [i|. Naively,
this leaves two degrees of freedom, but the projective nature of WiI reduces this to just
one degree of freedom. Thus the solution to (5.24) is parameterized by a 1-dimensional
variable, say the first component of the square spinor, [i|1 , and that means the solution is
described as a degree 1 curve, i.e. a line, in CP3 (it is defined as the zeroes of a degree 1
polynomial in Wi ).
Since each xa ȧ defines a different line, the integration d 4 x in (5.23) can be understood
as integrating over all possible lines. Thus amazingly, the n-gluon anti-MHV tree amplitude
corresponds in twistor space to n twistors Wi living on a line!
A tree-level gluon amplitude with q positive helicity gluons lives on a degree (q −1)-
curve.6 This interesting observation was due to Witten [54], and it provided an important
inspiration for the modern developments of scattering amplitudes. The geometric interpre-
tation is given by a twistor string theory, whose tree-level amplitudes are precisely those
of the 4-dimensional N = 4 SYM. More details can be found in the original paper [54].
We leave the twistor story for now, but it will sneak back into the limelight later on when
supertwistors WiA make another appearance in Chapter 10.
Twistor space and spacetime

We introduced twistors as a formalism for linearizing the action of conformal symme-
try. However, twistors are also naturally introduced from the point of view of conformal
symmetry in spacetime. For this purpose, it is convenient to interpret the conformal group
S O(2, 4) as the Lorentz group of a 6d space with metric signature (−, −, +, +, +, +). This
way conformal symmetry is realized as Lorentz symmetry if we embed the 4d spacetime
into the 6d spacetime. Consider a null-subspace in 6d defined by X · X = 0, where X μ
is a 6d vector. Since the null constraint is invariant if we rescale X → r X , it is natural
to identify X ∼ r X on the null-space. The constraint X · X = 0 and the projective nature
6 In [54], the twistor was defined with WiA = (|iȧ , [μi |a , ηi A ), where [μi |a is the Fourier conjugate to [i|a . Thus,
instead, one had the MHV amplitude to be of degree 1 in twistor space, while N K MHV corresponds to degree
K -curves in twistor space. Note that in Section 5.4 we will introduce a variable [μi |a which is not the same as
the one in Witten’s supertwistor Wi . But it will be part of a different type of supertwistor Zi . Please don’t be
too confused.
leaves 6 − 2 = 4 degrees of freedom, so this subspace is indeed 4-dimensional. This is the

so-called “embedding formalism” that was first introduced by Dirac in 1937 [55]. (See [56]
for a more recent discussion.)
We now spinorize the above discussion. Since the S O(2, 4) ∼ SU (2, 2), we can rewrite
the 6d vector X μ as a bi-spinor X I J . This 4 × 4 tensor is antisymmetric, X I J = −X J I , and
transforms in the 6-dimensional irreducible representation of SU (2, 2).
The null condition now translates to:
1
X 2 = I J K L X I J X K L = 0 . (5.25)
2
This implies that X I J has rank 2, and therefore we can write it as
[I J]
X I J = Wi W j , (5.26)
I
where the 4-component spinors W are the twistors, now introduced in the embedding
formalism. From (5.26) we see that a point X in the 4d subspace is defined in twistor space
by the line formed by two twistor variables (Wi , W j ). Since the subspace of X is defined
projectively, we identify Wi ∼ t Wi and therefore the twistor space is really CP3 .
The SU (2, 2) covariant form of the incidence relation in (5.24) is simply:
K]
X [I J Wi = 0. (5.27)
Since we have constructed our 4d spacetime with manifest conformal symmetry, we cannot
properly define the length; the construction does not specify a metric in 4d. Not surprisingly,
to introduce a metric, we need to break the conformal SU (2, 2) symmetry. This is done by
introducing a reference point I I J in the embedding space. The projectively invariant length
between two points X i and X j in the 4d subspace can then be defined as
(X i · X j )
, (5.28)
(I · X i )(I · X j )
where (X i · X j ) ≡ I J K L X iI J X Kj L . The choice of I corresponds to a choice of metric.
If two points X i and X j share a common twistor W , then it follows from (5.26) and
(5.28) that they are null-separated. Since two points define a line, we conclude that a point
in twistor space corresponds to a null-line in spacetime. We have then observed the general
feature that points in spacetime map to lines in twistor space and that points in twistor space
map to null-lines in spacetime.
5.3 Emergence of dual conformal symmetry
We have experienced the advantage of using spinor helicity formalism for scattering ampli-
tudes, both in terms of the restrictive power of consistency conditions, such as little group
scaling, and in terms of the simplicity of the final results. Many of these properties stem
from two properties of the spinor helicity variables, which:
1. trivialize the on-shell condition p 2 = 0, and
2. transform linearly under Lorentz symmetries, so that we get manifestly Lorentz-invariant
expressions for the amplitudes.
In contrast, the ordinary way of representing the amplitudes using polarization vectors
realizes Lorentz invariance via gauge redundancy; but this makes the amplitudes overly
complicated. Alternatively, one can work with only on-shell degrees of freedom by using
light-cone (or space-cone) gauge [57]. However, as these gauges are not Lorentz invariant,
the symmetry generators will act non-linearly on the kinematic variables. Thus, using the
spinor helicity formalism essentially allows us to “have our cake and eat it”: it allows us to
work with only on-shell degrees of freedom, yet the global symmetries are linearly realized.
At this point you may have noticed that there is a glaring hole in the above story: there is an
important part of the Poincaré symmetry that does not act linearly on the spinor variables:
the translations. In momentum space, translation invariance corresponds to momentum
conservation. This symmetry, as well as its supersymmetric partner, is respected by the
scattering amplitudes in a rather ad hoc fashion, namely by being enforced through the
presence of the delta functions:
n n

δ 4
piȧa
and δ (2N )
|iηi .
A
(5.29)
i=1 i=1
Here we have indicated the N -fold supersymmetric case, though in this section we are going
to study only N = 4 SYM. The point here is that if we follow the spirit of what the spinor
helicity formalism brought us, we should try to find new variables that either simplify, or
at least encode, the information of momentum and supermomentum conservation.
As an inspiration, let us visualize momentum conservation geometrically. The condition
μ
that an ordered set of n momenta pi add to zero implies that the vectors form a closed
contour, e.g. for n = 5:
y2
p1 p2
y1 y3 (5.30)
p5 p3
y5 p4 y4
Now there are two different ways to define the contour: it can be defined by the edges or
by the cusps. The former is just the usual momentum representation. For the latter, we take
μ
the cusps to be located at points yi in a dual-space [58]. They are defined by their relation
to the momentum vectors:
yiȧa − yi+1
ȧa
= piȧa . (5.31)
The dual coordinates yi (sometimes called zone variables or region variables) are not
spacetime coordinates; they are dual momentum variables defined by (5.31). In particular,
they have mass-dimension 1. In dual space, n-point momentum conservation simply corre-
sponds to the periodicity condition that yn+1 = y1 . For massless particles, the edges of the
n-edge polygon are light-like.
Since the ordering of the external momenta is crucial for the definition of the polygon
(5.30), we need a well-defined notion of ordering. For (super) Yang–Mills theory, this is
simply the color-ordering. The ordering also lets us define
yi j ≡ yi − y j = pi + pi+1 + · · · + p j−1 . (5.32)
The resulting variables yi j precisely match the yi j s introduced in (4.85) when we calculated
the NMHV tree-level superamplitudes in N = 4 SYM; this is of course no coincidence.
In (4.85) we also defined fermionic variables θiaj A ; those now arise as differences |θi j,A ≡
|θi A − |θ j A of dual-space fermionic coordinates |θi A defined as
|θi A − |θi+1,A = |i ηi A , (5.33)
where the ηi A s are the on-shell superspace Grassmann variables introduced in Section 4.4
and A = 1, 2, 3, 4 are the SU (4) R-symmetry labels.
In dual coordinates, the n-point tree-level MHV superamplitude of N = 4 SYM is
δ 4 (y1 − yn+1 ) δ (8) (θ1 − θn+1 )
AMHV = n , (5.34)
i=1 i, i + 1
n
and the tree-level NMHV superamplitude takes the form

n−3
n−1
ANMHV
n = AMHV
n Rn jk , (5.35)
j=2 k=i+2
where

j − 1, jk − 1, k δ (4) n jk
Rn jk = ,
y 2jk n|yn j .y jk |kn|yn j .y jk |k − 1n|ynk .yk j | jn|ynk .yk j | j − 1 (5.36)
n jk,A = n|ynk .yk j |θ jn,A + n|yn j .y jk |θkn,A .
The NMHV expression is exactly the same as the super-BCFW result in Section 4.5, where
it was already given in dual coordinates. We chose here to anchor the expressions on line n.
The representations in (5.34) and (5.35) may appear somewhat disappointing since the
amplitudes in dual space are basically identical to the original expressions. However, the dual
space description allows us to study a new symmetry, namely (super)conformal symmetry
in the dual coordinates y. This is called dual (super)conformal symmetry. To describe it,
first note that since the defining relation (5.31) of the yi s is invariant under translations, the
amplitude is guaranteed to be translational invariant in the y-space. Let the corresponding
translation operator in dual y-space be P μ . Next, because the conformal boost generator is
Kμ = I P μ I , the dual superconformal property of the amplitude can be extracted by simply
studying how the amplitude transforms under dual inversion I :
μ
μ y y ḃa y ȧb y ḃa
I yi = i2 , I |θi A ȧ = θi A |ḃ i 2 , I [i|a = i 2 |i]b , I |iȧ = i|ḃ i+1
2
.
yi yi yi yi+1
(5.37)
These rules are well-defined only when we have a notion of ordering. The dual inversion
rules for |i and [i| are defined only up to a relative scaling.
Exercise 5.6
μ
The inversion rules for yi and |θi A are standard. Verify the consistency of the rules
for |i and [i| using (5.31) and yi,i+2
2
= i, i + 1[i, i + 1].
[Hint: Show first that the definition (5.31) implies that i|ȧ yiȧa = i|ȧ yi+1
ȧa
. It follows
that I [i, i + 1] = i, i + 1/yi+2 .]
2
The momentum and supermomentum delta function transform under dual inversion as

I δ 4 (y1 − yn+1 ) = y18 δ 4 (y1 − yn+1 ) , I δ (8) (θ1 − θn+1 ) = y1−8 δ (8) (θ1 − θn+1 ) .
(5.38)
Thus we see that the inversion weight of the bosonic delta function exactly cancels7 that of
the Grassmann delta function. So for the tree-level N = 4 SYM MHV superamplitude we
have
n
MHV #
I An = yi AMHV
2
n . (5.39)
i=1
We conclude that under dual superconformal inversion, the tree MHV superamplitude
transforms covariantly with equal weights on all legs.
At this point, you may wonder if this new dual conformal symmetry is secretly just
another incarnation of the conventional conformal symmetry. It is straightforward to see
that this is not the case. The pure Yang–Mills gluon tree amplitude is conformal invariant
because it takes the same form as in N = 4 SYM. However, under dual inversion, the
split-helicity amplitude transforms as:
n
#
2 4
− − + +
I An [1 2 3 · · · n ] = 2
yi y1 An [1− 2− 3+ . . . n + ] . (5.40)
i=1
Clearly, this amplitude does not have homogeneous inversion properties. The situa-
tion is worse for a gluon amplitude without
the split-helicity arrangement, for example
An [1− 2+ 3− . . . n + ]. The result of I 13 is not proportional to 13, so An [1− 2+ 3− . . . n + ]
does not even transform covariantly under dual inversion. This shows that one can have a
conformal invariant amplitude that is not dual conformal covariant. Hence the two symme-
tries are inequivalent. The example also shows that dual conformal invariance is a symmetry
of the superamplitude, not the individual component amplitudes.
What about the tree-level NMHV superamplitude? Well, remarkably, the somewhat
complicated expression in (5.36) is invariant under dual inversion, i.e. I [Rn jk ] = Rn jk .
7 Note that in D = 4 this cancellation only happens for N = 4. In D = 3, a similar cancellation happens for
N = 6; indeed a theory with N = 6 supersymmetry exists, namely ABJM theory, and its amplitudes also
respect dual superconformal symmetry. We discuss this theory in more detail in Section 11.3.6. A similar result
in D = 6 would require a supersymmetric theory with 24 supercharges. It is not clear if an interacting theory
exists that can realize this symmetry in D = 6.
Thus ANMHV
n in (5.35) has the same homogeneous dual inversion weight as the tree-level
MHV superamplitude. In fact, using super-BCFW recursion relations it can be shown
[46, 49] that all tree superamplitudes of N = 4 SYM transform covariantly under dual
inversion,
n
tree #
I An = yi Atree
2
n . (5.41)
i=1
We prove this statement using recursion relations in Section 7.3.
Exercise 5.7
Use the inversion rules (5.37) to show that Rn jk is invariant. Note that it is crucial that
the spinor-products in Rn jk can be arranged to involve adjacent lines.
[Hint: The identity (yn j y jk + ynk yk j )a b + y 2jk δa b = 0 is useful for calculating I [n jk ].]
Due to the non-trivial weights in (5.41), the dual conformal boost generator does not vanish
on the amplitudes. Rather, it generates an “anomaly” term,
n
μ
μ tree
K An = − n .
yi Atree (5.42)
i=1
n μ
If we bring this term to the LHS and redefine K̃μ ≡ Kμ + i=1 yi , then the new generator
μ
K̃ annihilates the amplitudes.
The dual conformal symmetry can be enlarged into an SU (2, 2|4) dual superconfor-
mal symmetry. Recall that N = 4 SYM is also superconformal invariant, with the same
SU (2, 2|4) group. If we combine the two sets of generators, we obtain an infinite dimen-
sional algebra called a Yangian [59]. The generators of this algebra are organized by levels.
For the SU (2, 2|4) Yangian, level 0 consists of the ordinary superconformal generators
n
G A B = i=1 G iA B , where A = (I, A) with I = (ȧ, a) the index of SU (2, 2) and A the
SU (4) R-symmetry index. At level 1, the generators are bi-local in their particle index:

n
level 0: G iA B
i=1

n
level 1: (−1)|C| [G iA C G Cj B − (i ↔ j)]
i< j
..
. (5.43)
Here, |C| is 0 for C = I and 1 for C = A. It turns out that the shifted dual conformal boost
generator K̃μ (not the unshifted one, Kμ ) belongs to level 1. So the “anomaly” in (5.42) was
not a nuisance, but rather it was needed in order for the tree-level superamplitude of N = 4
SYM to be Yangian invariant! Beyond level 1, the new generators can be obtained simply
by repeated (anti)-commutation of level 1 and level 0 generators. For further information
about Yangian symmetry we refer to the original work [59]. The message here is that the
tree-level superamplitudes of N = 4 SYM are Yangian invariant.
5.4 Momentum twistors
Now that we know the N = 4 SYM tree superamplitudes to have dual superconformal
symmetry (in fact even Yangian symmetry), let us again set out on a path to find new
variables that transform covariantly under the new symmetry. This is especially justified
given that Rn jk is fairly unwieldy in its current form (5.36): we would like to write it as
an expression that is manifestly invariant under dual superconformal symmetry. Also, the
μ
presence of both helicity spinors and the vectors yi in Rn jk is a further redundancy of
variables that we would like to eliminate.
μ
As a first step, we redefine the dual-space coordinate yi in terms of spinor variables.
μ μ
Recall that we introduced the yi coordinates by their relation to the momenta: pi = yi −
μ
yi+1 . This relation implies that i|ȧ yi = i|ȧ yi+1 . Instead of referring to the momentum,
ȧa ȧa
μ
we can take this relation to be the defining relation for the dual coordinates yi : these are
called the incidence relations and take the form
[μi |a ≡ i|ȧ yiȧa = i|ȧ yi+1

ȧa
. (5.44)
The incidence relations define the

new variables [μi |a . The statement of the incidence
relations is that for a given Z iI = |i, [μi | , with I = (ȧ, a) an SU (2, 2) index, any two
μ μ
points, yi and yi+1 , in y-space that satisfies (5.44) must be null-separated by the vector
μ μ
yi − yi+1 = pi . Thus a point Z iI = (|i, [μi |) in Z -space corresponds to the null-line in
ȧa
μ μ
y-space determined by the two points, yi and yi+1 .
On the other hand, any point in y-space is determined by a line in Z -space. To see this, note
that the point yiȧa is involved in two incidence relations: [μi | = i|yi and [μi−1 | = i − 1|yi .
Combining these leads to

|iḃ [μi−1 |a − |i − 1ḃ [μi |a = |iḃ i − 1|ȧ − |i − 1ḃ i|ȧ yiȧa = i − 1, i yiḃa ,
(5.45)
so that
|iȧ [μi−1 |a − |i − 1ȧ [μi |a

yiȧa = . (5.46)
i − 1, i
I
This means that yi is determined by Z i−1 and Z iI : these two points define a line in Z -space.
The relationship between y-space and Z -space is illustrated in Figure 5.1.
Exercise 5.8
Use identities from the Appendix to show that |μi ]a = −(yi )a ḃ |iḃ .
We have translated the dual coordinates yi to Z iI ≡ (|iȧ , [μi |a ). The new four-component
spinor variables Z iI are called momentum twistors [24]. The name stems from the analogy
with spacetime twistors introduced in Section 5.2: a point in position space maps to a line
in twistor space, and vice versa. Writing the first equality in the incidence relation (5.44) in
y Z y Z
yi Zi
Zi yi
yi+1 Zi–1
Fig. 5.1 The map between dual space y μ and momentum twistor space Z iI = (|i, [μi |). The left-hand figure illustrates
the incidence relations (5.44): a null-line in dual space, defined by the two points yi and yi+1 , corresponds to a point
Z iI = (|i, [μi |) in momentum twistor space. The right-hand figure shows how a point yi in dual space maps to a
line in momentum twistor space via the relation (5.46).
the form ([μi |a − i|ȧ yiȧa ) = 0 should remind you of the twistor space incidence relation
(5.24). The second equality in (5.44) geometrizes the Weyl equation pi |i = 0.
The defining incidence relations (5.44) imply that |μi ] → ti |μi ] under little group scaling
(2.95), so Z iI → ti Z iI . Hence the momentum twistors are defined projectively.
So what have we achieved by going from yi to Z iI ≡ (|iȧ , [μi |a )? Well, the new variables
transform linearly under dual conformal transformations. The generators G I J of the dual
conformal group can now be written compactly together with the group algebra as8
∂ I
GI J ≡ Z iI , G J , G K L = δJ K G I L − δ I L G K J , (5.47)
i
∂ ZiJ
with I, J, . . . = (ȧ, a). The operator G I J can be thought of as a 4 × 4 matrix with a block
diagonal 2 × 2 structure. To make the generators more concrete, consider the 2 × 2 block
∂
with I = a and J = ȧ: it is G a ȧ = i [μi |a ∂|i ȧ . Its index structure, and the fact that it has
mass-dimension 1, indicate that this should be the dual conformal boost Ka ȧ . Analogously
with the regular conformal boost, given in Exercise 5.3, the dual conformal boost generator

can be written in dual y-space as Ka ȧ = − i ȧ ċ yiḃa yi ċb ∂ḃb . Comparing this expression
∂ yi
to G a ȧ , it is not obvious that G a ȧ = Ka ȧ , but the following exercises show you how it works.
Exercise 5.9
Show that Ka ȧ and G a ȧ are equivalent by demonstrating that they give the same result
when acting on yiċc in (5.46).
Since the yi s and the momenta pi = −|i[i| are related, the change of variables from (|i, yi )
to Z i = (|i, [μi |) should make it possible to express [i| in terms of |i and [μi |. Indeed,
one finds
i + 1, i[μi−1 | + i, i − 1[μi+1 | + i − 1, i + 1[μi |

[i| = . (5.48)
i − 1, ii, i + 1
Exercise 5.10
Derive (5.48) using the incidence relations (5.44) and Schouten identities.
8 For simplicity we consider U (2, 2). The tracelessness condition can be implemented as in (5.19).
Since the momentum twistors Z I carry the dual conformal SU (2, 2) index I , we can form a
dual conformal invariant by contracting four Z I s with the Levi-Civita symbol of SU (2, 2):
we use a 4-bracket for this invariant (the minus sign is chosen for later convenience),
i, j, k, l ≡ − I J K L Z iI Z Jj Z kK Z lL = i j[μk μl ] + ik[μl μ j ] + il[μ j μk ]

+ kl[μi μ j ] + l j[μi μk ] + jk[μi μl ] .
(5.49)
We have expanded out the product in terms of S L(2, C) invariants [μi μ j ] ≡ [μi |a |μ j ]a .
We can gain intuition for the 4-brackets by evaluating them in special cases. For example
k, j − 1, j, r = j − 1, j k|yk j y jr |r . (5.50)
Exercise 5.11
Prove (5.50) by first using (5.49) to rewrite the LHS as a sum of i j[μk μl ]s. Then
apply (5.44) and the Schouten identity to pull out an overall factor j − 1, j.
A special case of (5.50) is
j − 1, j, k − 1, k = j − 1, j k − 1|yk−1, j y jk |k = j − 1, j k − 1, k y 2jk ,

(5.51)
i.e.
j − 1, j, k − 1, k
y 2jk = . (5.52)
j − 1, j k − 1, k
Since 1/yi2j are propagators, the relation (5.52) will appear repeatedly in our discussions.
Looking at (5.50) and (5.52) makes us realize that these are exactly the type of objects
that appear in the denominators of the R-invariants (5.36) of the NMHV tree amplitudes.
Indeed
Rn jk
j − 1, j4 k − 1, k4 δ (4) (n jk )
= .
n, j −1, j, k −1 j −1, j, k −1, k j, k −1, k, nk −1, k, n, j −1k, n, j − 1, j
(5.53)
Note that it was possible to arrange the five denominator factors such that the input of
the 4-brackets go cyclically through the set of five labels (n, j − 1, j, k, k − 1). Now
the denominator is manifestly dual conformal invariant since it is composed entirely of
SU (2, 2)-invariant 4-brackets. However, the 4-brackets transform under little group scaling
with weight 1 for each line in the argument. Thus the denominator is not really an invariant
since the Z I s are defined only projectively. So let us take a closer look at the numerator;
for this purpose we need Grassmann-companions for the Z I s.
Similarly to the bosonic incidence relations (5.44), we use the spinor-fermionic coordi-
nate |θi A ȧ to introduce a Grassmann-odd (spacetime-)scalar coordinate χi A :
χiA = i θi A = i θi+1,A . (5.54)
With these new fermionic twistor variables, we have extended the SU (2, 2) momentum
twistors Z I to SU (2, 2|4) momentum supertwistors

ZiA ≡ |iȧ , [μi |a χi A , where A = (ȧ, a, A). (5.55)
Under little group scaling, Zi scales uniformly, ZiA → ti ZiA , so the momentum super-
twistors are defined projectively and are elements in CP3|4 .
Exercise 5.12
Derive the fermionic versions of (5.46) and (5.48).
In the R-invariant (5.53), the fermionic delta function argument is n jk,A =

n|ynk .yk j |θ jn,A + n|yn j .y jk |θkn,A . To rewrite it in terms of momentum supertwistors,
we need the identity
k, r − 1, r, j − 1χ j A − k, r − 1, r, jχ j−1,A
k|ykr yr j |θ j A = − . (5.56)
r − 1, r j − 1, j
Exercise 5.13
Derive (5.56) by manipulating the RHS using (5.50) and employing the Schouten
identity.
Exercise 5.14
Use (5.56) to show that

j − 1, j, k − 1, k χn A + cyclic
n jk,A =− . (5.57)
j − 1, jk − 1, k
[Hint: use the hint in Exercise 5.7.]
Plugging (5.57) into (5.53), we then have
Rn jk

δ (4) j −1, j, k −1, k χn + cyclic
= .
n, j −1, j, k −1 j −1, j, k −1, k j, k −1, k, nk −1, k, n, j −1k, n, j −1, j
(5.58)
The “+ cyclic” is the instruction to sum cyclically over the labels (n, j − 1, j, k, k − 1),
similarly to the product structure in the denominator. Now Rn jk is manifestly invariant
under both the little group scaling and dual SU (2, 2). Together with the overall factor of
AMHV
n , we then have the building blocks of the NMHV superamplitude in a form that makes
it manifestly dual superconformal invariant.
The expression (5.58) for Rn jk is cyclic in the five labels (n, j − 1, j, k − 1, k). This
motivates us to introduce a 5-bracket defined as

i, j − 1, j, k − 1, k ≡ Ri jk . (5.59)
The 5-bracket is cyclic in its five arguments, so for example [6, 1, 2, 3, 4] = [1, 2, 3, 4, 6].
In terms of the 5-bracket, the n-point NMHV tree superamplitude is simply given by

n−3
n−1

ANMHV
n = AMHV
n n, j − 1, j, k − 1, k . (5.60)
j=2 k= j+2
Let us review what we have accomplished so far. Starting with the simple observation
that momentum conservation is imposed in a rather ad hoc fashion, we introduced the
auxiliary variables yi such that momentum conservation is encoded in a geometric fashion.
This led us to the realization of a new symmetry of the tree amplitude for N = 4 SYM,
namely superconformal symmetry in the dual space yi . The new symmetry set us on
a journey to search for new variables, the momentum (super)twistors, that linearize the
transformation rules. This culminated in the simple symmetric form of the n-point NMHV
tree superamplitude in (5.60). For N K MHV, equation (5.60) generalizes to a sum involving
products of K 5-brackets (see Section 7.3).
The 5-bracket in (5.60) corresponds to the terms in the super-BCFW expansion of the
superamplitude; specifically we have seen in Section 4.5.2 how each Rn2k arises from an
MHV×MHV BCFW diagram while the remaining Rn jk s with j > 2 appear via recur-
sion from the BCFW diagram with NMHV×anti-MHV subamplitudes. As discussed, this
means that the representation (5.60) is not unique, since there are many equivalent BCFW
expansions for a given amplitude, depending on the choice of lines in the BCFW shift.
This implies that the dual conformal invariants (5.59) must have some linear dependencies.
For example, compare for n = 6 the result of the recursions relations based on the BCFW
supershifts [6, 1 and [1, 2: they have to give the same result, so
[6, 1, 2, 3, 4] + [6, 1, 2, 4, 5] + [6, 2, 3, 4, 5] = [1, 2, 3, 4, 5] + [1, 2, 3, 5, 6]

+ [1, 3, 4, 5, 6] . (5.61)
Using the cyclic property of the 5-bracket, we can write this
[2, 3, 4, 6, 1] + [2, 3, 4, 5, 6] + [2, 4, 5, 6, 1] = [3, 4, 5, 6, 1] + [3, 5, 6, 1, 2]

+ [3, 4, 5, 1, 2] . (5.62)
Now you see that the LHS looks like the result of a [2, 3 supershift, while the RHS comes
from a [3, 4 supershift. Actually, the LHS and, independently, the RHS are invariant under
i → i + 2. We can also reverse the labels in the 5-brackets at no cost to get
[2, 3, 4, 6, 1] + [2, 3, 4, 5, 6] + [2, 4, 5, 6, 1] = [3, 1, 6, 5, 4] + [3, 2, 1, 6, 5]

+ [3, 2, 1, 5, 4] . (5.63)
This states that the “parity conjugate” supershifts [2, 3 and [3, 2 give identical results. In
fact, we can now conclude that any adjacent supershifts are equivalent.
The presence of these equivalence-relations between the dual conformal invariants may
strike you as rather peculiar and you may wonder if it has a deeper meaning. Furthermore,
while the expressions in (5.58) and (5.60) are extremely simple, they lack one key aspect
when compared to the Parke–Taylor superamplitude (5.5): cyclic invariance. The presence
of dual conformal symmetry relies heavily on the cyclic ordering of the amplitude, and
hence it is somewhat surprising that the manifestly dual conformal invariant form of the
superamplitude (5.60) breaks manifest cyclic invariance. One may say that we are asking
too much of the amplitude, but considering the payoff we have reaped from the innocent
chase of manifest momentum conservation, we will boldly push ahead with our pursuit of
“having cakes and eating them” in Chapters 9 and 10.
Exercise 5.15
Use cyclicity of the 5-brackets to show that the tree-level 6-point NMHV superampli-
tude can be written in the manifestly cyclic invariant form
1
ANMHV
6 = AMHV
6 × R146 + cyclic , (5.64)
2
where “+ cyclic” means the sum over advancing the labels cyclically, i.e. R146 + R251 +
four more terms.
Momentum twistors
For our further studies, it is worth making a few observations about the momentum twistors.
We introduced the dual yi s in order to make momentum conservation manifest; but the yi s
could not be chosen freely since they are subject to the constraint (yi − yi+1 )2 = 0 that
ensures the corresponding momenta pi to be on-shell. On the other hand, the momentum
twistors Z i are free variables: they are subject to the scaling equivalence Z i ∼ t Z i , so they
live in projective space CP3 . (The momentum supertwistors Zi are elements of CP3|4 .) We
can choose n points Z i in CP3 , subject to no constraints, then study the n lines defined by
consecutive points (Z i , Z i+1 ), with the understanding that the nth line is (Z n , Z 1 ). Equation
(5.46) maps each line (Z i , Z i+1 ) to yi and the incidence relation (5.44) guarantees that the
points yi and yi+1 are null-separated; thus the corresponding momenta pi = yi − yi+1 are
on-shell. The collection of lines (Z i , Z i+1 ) gives, per definition, a closed contour and that
ensures momentum conservation, yn+1 = y1 . So all in all, the map to momentum twistors
geometrizes the kinematic constraints of momentum conservation and on-shellness by
simply stating these requirements as the intersection of n lines (i, i + 1) ≡ (Z i , Z i+1 ) at the
points (i) ≡ Z i in momentum twistor space CP3 . The momentum supertwistors similarly
make conservation of supermomentum automatic.
Pursuing the geometric picture a little further, we consider intersections of lines and
planes in momentum twistor space. The momentum twistors are points (i) ≡ Z i in CP3 .
Two points in CP3 define a line (i, j) that can be parameterized as Z i + τ Z j for τ ∈ C. Now
any two distinct points on that line can be used to define it, so a reparameterization-invariant
[I J]
definition of a line in CP3 is given by the antisymmetric product Z i Z j . This applies
similarly for planes, which are defined in terms of the antisymmetrization of three points.
So we have:
[I J]
line: (i, j) = Z i Z j ,
[I K]
(5.65)
plane: (i, j, k) = Z i Z Jj Z k .
In CP3 , the point ( p) ≡ Z p that corresponds to the intersection of line (i, j) ≡ (Z i , Z j )

with a plane (k, l, m) ≡ (Z k , Z l , Z m ) is given by
(
( p) = (i, j) (k, l, m) = Z i j, k, l, m − Z j i, k, l, m . (5.66)
)
The symbol indicates the intersection of the two objects. Similarly, the line ( p, q) that
corresponds to the intersection of two planes (Z i , Z j , Z k ) and (Z l , Z m , Z n ) is given by
(
( p, q) = (i, j, k) (l, m, n) = (i, j) k, l, m, n + ( j, k) i, l, m, n + (k, i) j, l, m, n .
(5.67)
A more detailed discussion of twistor geometry can be found in [60].

Propagators 1/yi2j are expressed in terms of momentum twistors via (5.52). Hence,
the on-shell condition yi2j = 0 becomes the requirement i − 1, i, j − 1, j = 0. This is
the statement that the four momentum twistors labeled by i − 1, i, j − 1, j are linearly
dependent. Geometrically, that means they lie in the same plane in CP3 . In Section 7.3 we
will see that in momentum twistor space a pole yi2j = 0 in the BCFW-shifted amplitude is
characterized as the intersection between the line (i − 1, i) and the plane (i − 1, j − 1, j);
this motivates our interest in formulas such as (5.66).
It is now time to venture beyond tree-level and wrestle loops: in the second part of the
book, we discuss various approaches to loop-amplitudes.
PART II
LOOPS
Loop amplitudes and generalized unitarity 6
In Part I of the book, we focused exclusively on tree-level amplitudes, the leading contri-
butions in the perturbative expansion of a scattering process. The loop-corrections are of
course highly relevant, both in particle physics applications and for our understanding of
the mathematical structure of the S-matrix. Here, in Part II of the book, we explore the
structure of loop amplitudes.
Feynman rules require momentum conservation at each vertex of a Feynman diagram.
At tree-level, this fixes all momenta of the internal lines in terms of the external momenta.
At loop-level, momentum conservation leaves one momentum undetermined per loop and
one must integrate over all such unfixed momenta. Thus in D-dimensions, one has a D-
dimensional loop-integral for each loop. For example, in φ 3 theory one of the diagrams that
contributes to the 1-loop 4-point amplitude is
p2 p3

d D 1
=
2

. (6.1)
(2π ) 2 − p1 − p1 − p2 2 + p4 2
D
p1 p4
As another example, consider a diagram of the 1-loop correction to the 4-point all-scalar
amplitude in scalar QED:
p1 p2

d D ( − 2 p1 ) · ( + 2 p2 ) ( − p1 + p4 ) · ( + p2 − p3 )
=e 4

2
2
2 ,
(2π ) D 2 − p1 − p1 − p2 + p4
p4 p3
(6.2)
where e is the electric charge of the scalar. Note that the integral (6.2) has loop-momentum
dependence in the numerator.
In general, an L-loop amplitude can be written schematically as
L
# d D l 1 n j cj
An
L-loop
= D
, (6.3)
j l=1
(2π ) S j α j pα2 j
where the sum runs over all possible L-loop Feynman diagrams j. For each diagram,
the l denote the L loop-momenta, α j label the propagators, and S j is the symmetry
factor associated with the diagram. The kinematic numerator factors n j are polynomials of
Lorentz-invariant contractions of external- and loop-momenta and polarization vectors (or
118 Loop amplitudes and generalized unitarity
other external wavefunctions). The constants c j capture the information about couplings
and gauge group factors.
The color-structure we have worked with for super Yang–Mills amplitudes is rather
simple: each color-ordered amplitude is associated with a single trace of gauge group
generators. At loop-level, more intricate trace-structures arise. In this book, however, we
will be concerned with the planar limit of Yang–Mills amplitudes. In this context, the planar
limit means that the amplitudes have only one trace-structure, just as the tree amplitudes.
They are called planar amplitudes. The color-ordered Feynman rules of Section 2.5 apply
and the diagrams that contribute to an amplitude are precisely those in which no lines cross.
At loop-level, it is useful to distinguish the following three objects:
1. The loop-integrand is the rational function inside the loop-momentum integration.

2. The loop-integral is the combination of the integrand and the loop-momentum inte-
gration measure: this is a formal object, since we have not specified the integration
region of the loop-momentum or addressed divergences.
3. The loop amplitude is the result of carrying out the loop-integrations in the loop-
integral. If we integrate over physical momentum space R1,3 , the integral may have
infrared (IR) and/or ultraviolet (UV) divergences. We need to regulate such diver-
gences in order to make the integrated result, the loop amplitude, well-defined.
Our focus will mainly be on the loop-integrand and the loop-integral. However, it is useful to
understand the analytic structure of the loop amplitude, since this reveals which constraints
the loop-integrand, and hence integral, must satisfy.
It is not surprising that integrals of loop amplitudes can be divergent. They integrate a
rational function – the loop-integrand – over L copies of R1,3 and this includes regions where
the loop-momentum is infinite | | → ∞ as well as regions where integrand-denominator
may vanish; both cases give potential divergences. We discuss the UV and IR divergences
in Section 6.1. After regulating the amplitude and separating out the divergences, the finite
pieces of the amplitude are typically not just rational functions. For massless lower-loop
amplitudes, one sometimes encounters functions that can be expressed in the form of
iterative integrals or products of iterative integrals. For example,
x x y x t
dt dt ds dt du
log(x) = , log(x) log(y) = , Li2 (x) = .
0 t 0 t 0 s 0 t 0 1−u
(6.4)
The last function, Li2 , is a dilogarithm. It is the q = 2 special case of the polylogarithm Liq .
Starting with the familiar logarithm Li1 (x) = − log(1 − x), the polylogarithms are defined
iteratively as
x
Liq−1 (t)
Liq (x) = dt . (6.5)
0 t
Such functions are usually referred to as transcendental functions. They are associated
with a degree of transcendentality, given by the dimension of the integral involved. The
three functions in (6.4) are of transcendentality 1, 2, and 2, while it is r for the polylog Lir .
Note that in this categorization, π has transcendentality 1, since iπ = log(−1).
The analytic structure of loop amplitudes is clearly more complicated than for tree
amplitudes: where tree amplitudes are simple rational functions with only simple poles,
the loop-integrations typically give rise to generalized logarithms, as in (6.4), and other
special functions. These have branch cuts in addition to poles. The well-understood pole
structure of tree amplitudes was instrumental for developing the on-shell recursion relations
(Chapter 3), so at first sight it looks challenging to develop a similar “on-shell” approach for
loop amplitudes. Nonetheless, the analytic structure of the loop-integrands can be exploited
to reconstruct the amplitude from lower-order on-shell data. The purpose of this and the
following two chapters is to show how to do this.
In the present chapter, we begin with a brief description of the physics of loop-
divergences. We then focus on the widely applicable and very successful unitarity method.
Chapter 7 studies BCFW recursion for the loop-integrands and in Chapter 8 we discuss
Leading Singularities and on-shell diagrams.
6.1 UV and IR divergences
If a loop-integral diverges as the loop-momentum becomes large, the loop-diagram is said

to have an ultraviolet (UV) divergence. At large loop-momentum, the denominators of the
propagators behave as 2 , so a given integrand behaves as

N | | N +D−1
d 2 P ∼
D
d| | , (6.6)
( ) | |2P
where D is the number of spacetime dimensions, P is the number of propagators, and N
counts the powers of loop-momenta in the numerator. With D + N − 2P = 0, 1, 2, . . . , the
integral is logarithmically, linearly, quadratically divergent, etc. To regulate the UV behavior
of the integral, one can introduce an explicit UV cutoff , but it is often more convenient
to analytically continue the spacetime to 4 − 2 dimensions, and use as a regularization
parameter. This is dimensional regularization, which is widely used in quantum field
theory due to the appealing feature that it preserves gauge invariance.
If the amplitude has UV divergences, this is telling us that, at short distances, the theory
needs to be modified either by changing the constants in the original action or by adding
new local operators to the action. The way in which the theory is modified depends crucially
on the form of the divergence. Since the loop-integral has a definite mass dimension, the
divergence must be multiplied by a kinematic invariant with the corresponding dimension. If
this invariant is the tree amplitude, then we can simply modify the constants, such as overall
coupling factors, of the tree amplitude to absorb this divergence. This modification can be
translated into the modification of the coupling constants and masses in the Lagrangian
description of the theory. If, on the other hand, the kinematic invariant in front of the
loop-divergence is not proportional to the tree amplitude, then this means that one has to
modify the theory by adding new operators to the action such that the divergences can be
absorbed.
The procedure of removing UV-divergences from loop amplitudes is called renormal-
ization. If the UV-divergences of all loop amplitudes for a given theory can be removed by
adding a finite number of terms to the action (including zero), then the theory is said to be
renormalizable. Renormalization is an extremely important subject in quantum field the-
ory. It is crucial for making physical predictions from gauge theories and the techniques of
renormalization are well-tested in the remarkable and successful comparisons of theoretical
predictions with experimental results.
Another type of singularity is infrared (IR) divergences. They can come from two
sources: soft and collinear divergences. To facilitate the discussion, let us consider an
integral with two adjacent massless legs:
p1
p2
. (6.7)
The relevant part of the integrand is

d D
. (6.8)
( + p1 )2 2 ( − p2 )2
Let us consider light-cone coordinates
p+ ≡ p0 + p1 , p− ≡ p0 − p1 , p⊥ ≡ ( p 2 , p 3 , . . . , p D−1 ) . (6.9)
Then p 2 = − p + p − + p⊥ 2
. We can always choose the coordinates in such a way that p1 is
in the p direction. To simplify the discussion, we choose p2 to be in the p + direction; this
−
is not general but the result of the discussion will be. The loop-integral is now (dropping
overall factors)

d + d − d D−2 ⊥
. (6.10)
( − 2 + p1− ) 2 ( 2 − 2 − p2+ )
2
Let us consider the integration region where the loop-momentum becomes soft, μ → 0.
Practically, we scale the integral to this region by changing integration variable from + to
t as
+ → t , − → t − , ⊥ → t ⊥ . (6.11)
All terms in 2 scale homogeneously as t 2 which is subleading to . pi for small t. Thus,

for t near zero, the integral (6.10) gives
D−1
t dt dt
4
= 5−D
. (6.12)
t t
In D = 4, this integral is logarithmic divergent near t = 0.1 We can regulate it by formally
continuing the spacetime dimensions to 4 − 2, and one now has:
3−2
t dt dt 1

4
= 1+2
= + O 0 . (6.13)
t t 2
This divergence is called a soft divergence, since it occurs when the loop-momentum is
“softer” than the external legs.
Next, consider a limit in which the loop-momentum becomes collinear with p1 : we scale
the components to align with the p − direction,
+ → t 2 , − → − , ⊥ → t ⊥ . (6.14)
Again, 2 scales as t 2 , so for small t we have that ( 2 − 2 + p1− ) scales as t 2 and that
( 2 − 2 − p2+ ) goes as t 0 at leading order. Thus the overall behavior of the integral at small
t is again (6.13), so it goes as 1/ in 4−2 dimensions. This is a collinear divergence,
it corresponds to the limit where the loop-momentum becomes collinear with one of the
external momenta.
The soft divergences can occur simultaneously with collinear divergences, so in dimen-
sional regularization D = 4−2, the -integration of the diagram (6.7) behaves at leading
order as 1/ 2 . For an L-loop amplitude, the leading IR divergences are usually of order
1/ 2L . Soft and collinear limits for amplitudes in massless gauge theories have been studied
since the late 1970s and this is an entire subject on its own; see for example the brief outline
in [61] and references therein.
The appearance of IR divergences does not correspond to a need to modify our original
theory. Rather, it is a consequence of a subtlety in defining scattering amplitudes for
massless particles. To define a scattering amplitude, one must have a well-defined notion
of asymptotic states, i.e. what exactly are the particles that we are scattering? If identical
massless particles become collinear, it is difficult to distinguish them from a single massless
particle: a single particle with definite light-like momentum is indistinguishable from
hundreds of soft particles whose momenta are collinear and near zero. This ambiguity
for asymptotic massless states is reflected in the IR divergences of loop amplitudes as
well as in the singularities of the soft and collinear limits of the tree-level amplitude
(Section 2.8).
The tree and loop soft and collinear singularities are intimately related, as can be seen
from the computation of the scattering cross-section. The cross-section of a 4-point process
for four gluons in pure Yang–Mills theory involves the norm-squared product of gluon
1 It is also divergent for t → ∞. But in that limit, our approximation where the loop-momentum is small compared
to the external momenta is no longer valid.
amplitudes, appropriately summed over states. Organizing the perturbative expansion in

powers of the gauge coupling g, we find
dσ

∝ g1 g2 |S|ii|S|g3 g4 = tree tree + 1−L tree + tree tree + O g8 .
d i
g4 g6 g6
(6.15)

The sum i represents the sum over all on-shell multi-particle states and S is the S-matrix.
The dashed line in each diagram represents a sum over the particle species of the on-shell
internal leg. The 4-gluon tree and 1-loop amplitudes are order g 2 and g 4 , respectively, while
the 5-gluon tree-amplitude is order g 3 . Hence the product of the 4-gluon tree and 1-loop
amplitudes contributes at the same order, namely g 6 , as the product of two 5-gluon tree
amplitudes. Now diagrammatically, one can see that the collinear divergence of the 1-loop
4-gluon amplitude in one diagram can indeed be regarded as the collinear divergence of
the 5-point tree amplitude looked at in another way:
. (6.16)
The collinear divergences of the 5-point tree amplitudes “conspire” to cancel the IR diver-
gence of the 4-point 1-loop amplitudes to yield a finite cross-section. A simple toy model
example of such cancellations can be found in [62, 63].
The preceding discussion illustrates that the IR divergences of loop amplitudes are
intimately tied to the soft and collinear behavior of tree amplitudes. Historically, this
has served as an inspirational starting point for the development of tree-level recursion
relations [13] discussed in Section 3. Note that if a scattering process is absent at tree-
level and is non-vanishing only at loop-level, then the leading-order result cannot have
IR singularities; it would have no tree-level counterpart to cancel its IR divergences
in the cross-section. For example, the all-minus or single-plus gluon amplitudes vanish
at tree-level, as we have seen in (2.112), but in pure Yang–Mills theory they are non-
vanishing at 1-loop order. It can be shown by explicit calculation that the results for these
gluon amplitudes are free of IR divergences, and in fact are simple rational functions e.g.
[64, 65]

− − [i 1 i 2 ]i 2 i 3 [i 3 i 4 ]i 4 i 1
pure Yang–Mills: A1-loop 1 2 . . . n− = , (6.17)
n
[12][23] · · · [n1]
where the sum in the numerator is over 1 ≤ i 1 < i 2 < i 3 < i 4 ≤ n and we have dropped
overall normalization factors.
It is useful for the following discussion to consider a set of basic scalar integrals and
(i)
their integrated results in dimensional regularization. The 1-loop scalar integrals Im are
the m-gon Feynman diagrams that appear in φ n -theory:
. (6.18)

In the scalar integrals, the dependence on loop- and external momenta is contained solely
in the propagators. We categorize 1-loop scalar integrals as bubbles, triangles, boxes,
etc. if they have 2, 3, 4, etc. propagators. These integrals are denoted by I2;n (K i , K j ),
I3;n (K i , K j , K k ), and I4;n (K i , K j , K k , K l ), with the subscript indicating the number of
corners of the diagrams as well as the number of external legs involved. K i is the sum of
external momenta at each corner. The 4-point box, triangle, and bubble scalar integrals are
p2 p3
d D 1
= I4;4 ( p1 , p2 , p3 , p4 ) =
2

,
p1
(2π ) 2 − p1 − p1 − p2 2 + p4 2
D
p4
p2
p3 d D 1
= I3;4 ( p1 , p2 , p3 + p4 ) =
2
, (6.19)

p4 (2π ) 2 − p1 − p1 − p2 2
4
p1

p2 p3 d D 1
= I2;4 ( p1 + p2 , p3 + p4 ) =
.
p1

p4 (2π )4 2 − p1 − p2 2
We met the scalar box integral in (6.1). In D = 4, power counting shows that the box and
triangle integrals I4;n and I3;n cannot have UV divergences, only IR divergences due to the
massless corners. The exceptions are the cases in which all corners are massive, i.e. all
K i2 = 0; such massive triangle and box integrals are IR finite. The bubble integrals I2;n
have UV divergences, but no IR divergences because both corners are massive, K i2 = 0.
The scalar bubble integral with a massless corner vanishes in dimensional regularization.
This is because the integral has mass-dimension 2, yet there is no dimensionful quantity to
give it dimension since the only invariant entering the integral is K i2 = 0.
Setting D = 4 − 2, the integrated results of the 4-point box, triangle, and bubble inte-
grals are

γ 2
−2 −
−2 −
2 s
I4;4 ( p1 , p2 , p3 , p4 ) = −μ s + −μ u − ln − π 2
+ O() ,
su 2 u
γ
I3;4 ( p1 , p2 , p3 + p4 ) = 2 (−μ−2 s)−1− + O() , (6.20)

1
I2;4 ( p1 + p2 , p3 + p4 ) = γ − ln(−μ−2 s) + 2 + O() .

2 2 2
3 2 3
1 3
1 1 3

+ + + … = 4
4
5 4
5 4 5 1 5
Fig. 6.1 The sum of residues from all Feynman diagrams with propagators 2 and ( − p1 − p2 )2 on-shell must give the
product of two tree amplitudes.
Here, γ = (1 + ) 2 (1 − )/ (1 − 2) and μ is the regularization scale. These results
and their n-point generalizations can be found in [66]. The 1-loop box and triangle integrals
are 1/ 2 , as expected from the combination of the soft and collinear divergences. The 1/
divergence in the bubble integral is a UV divergence.
The scalar box, triangle, and bubble integrals provide a basis for 1-loop amplitudes in
D = 4 dimensions. We will see how this works next when we study the unitarity method.
6.2 Unitarity method
We begin with an example. Consider the color-ordered planar n-point 1-loop gluon ampli-
tude in pure Yang–Mills theory. Suppose we identify2 the loop-momentum such that in
each Feynman diagram, is the momentum that flows between legs 1 and n, as indicated for
n = 5 in Figure 6.1. Then we can unambiguously collect all the distinct Feynman diagrams
under one integral,

d D Jj . (6.21)
j
The integrands J j take the form indicated in (6.3). To compute the full amplitude we need
to integrate over R4 (after Wick rotation from R1,3 ), but let us focus on the subplane
where the loop-momentum satisfies the two cut conditions
2 = ( − p1 − p2 )2 = 0 . (6.22)
On this subplane, integrands of the form
1 cjn j
Jj = (6.23)
S j . . . ( 2 ) . . . ( − p1 − p2 )2 . . .
become singular. The singularity corresponds to a kinematic configuration where the two
propagators, 1/ 2 and 1/( − p1 − p2 )2 , go on-shell. The sum of the residues from all such
integrands must be equivalent to the product of two on-shell tree amplitudes, as shown
schematically in Figure 6.1.
2 More about this choice in Section 7.1.

6.2 Unitarity method 125
This observation is very useful for constraining the loop-integrand: if the enemy gives
us a rational function and claims that it is the 1-loop integrand of some (unitary) theory,
we can apply cuts, such as (6.22), and test the claim by checking if the function factorizes
correctly into products of tree amplitudes. This way, our knowledge of tree amplitudes can
be recycled into information about the loop-integrand!
The operation of taking loop propagators on-shell is called a unitarity cut. It originates
from the unitary constraint of the S-matrix. To see how, recall that unitarity requires
S † S = 1. Writing S = 1 + i T , where T represents the interacting part of the S-matrix,
unitarity requires −i(T − T † ) = T † T . If we examine this constraint order by order in
perturbation theory, it tells us that the imaginary part of the T -matrix at a given order is
related to the product of lower-order results. In particular, the imaginary part of the 1-loop
amplitude is given by a product of two tree amplitudes. This is illustrated by the diagram
p2 p3
–p1–p2
2 Im 1−Loop = Tree Tree . (6.24)
p1 p4
The product of two tree amplitudes on the RHS involves a sum over all possible on-shell
states that can “cross” the cut. Only states from the physical spectrum of the theory are
included in this sum. In gauge theory, Feynman diagram calculations of loop amplitudes
require ghosts in the loops: the purpose of the ghosts is to cancel unphysical modes in the
loops. In contrast, in a unitarity cut (6.24) we restrict the loop-momenta to be on-shell and
only physical modes are included in the two on-shell amplitudes on the RHS of (6.24).
The cutting rules also include integrals of any remaining freedom in the loop-momentum
after imposing the cut constraints, such as (6.22), and momentum conservation. The integral
over all allowed kinematic configurations, with respect to the diagram (6.24), can be written
as

d D δ+ 2 δ+ ( − p1 − p2 )2 . (6.25)
The subscript + means that we are choosing the solution to the on-shell condition that has
a positive time component, 0 > 0, i.e. it is associated with a particle (as opposed to an
anti-particle) crossing the cut. Note that (6.25) just replaces the two cut propagators with
their on-shell conditions, exactly as we did in Figure 6.1.
The imaginary part of the amplitude probes its branch cut structure, hence the unitarity
cut allows us to relate the “pole structure” of the integrand with the “branch cut structure” of
the loop-integral. One can reconstruct the integrand by analyzing different sets of unitarity
cuts. The unitarity cuts can also involve more than two “cut” lines, i.e. several internal
lines taken on-shell, such that the loop amplitude factorizes into multiple lower-loop on-
shell amplitudes. An N -line cut simply means that N internal lines are taken on-shell.
Reconstruction of the full loop amplitude from systematic application of unitarity cuts is
called the generalized unitarity method [67]. It has been applied to a wide range of scattering
problems, from next-to-leading order precision QCD predictions to the ultraviolet behavior
of perturbative supergravity theories. The generalized unitarity method [67] deserves much
more attention than we are able to offer here. Our introduction to the unitarity method
covers just the minimum needed for you to see the idea and appreciate its power. The
method has been reviewed extensively and you can learn more about it and its applications
to supersymmetric as well as non-supersymmetric theories in the reviews [68–71]. Next,
we discuss its implementation at 1-loop level.
6.3 1-loop amplitudes from unitarity
The information from unitarity cuts can be utilized most efficiently if we know, a priori, a
complete basis of integrals that can appear in the scattering amplitudes. It can be shown
[72–76] that in D-dimensions all 1-loop amplitudes can be rewritten as a sum of m-gon
1-loop scalar integrals Im for m = 2, 3, . . . , D:
(i) (i)
( j) ( j)
(k) (k)
A1-loop = CD ID + C D−1 I D−1 + · · · + C2 I2 + rational terms . (6.26)
i j k
(i) (i)
Here, Cm are kinematics-dependent coefficients for the m-gon scalar integrals Im that
were introduced towards the end of Section 6.1. The 4-point box, triangle, and bubble
(i)
scalar integrals (6.19) were evaluated in (6.20). The n-point box integral I4 takes the form
(henceforth dropping the subscript n)
(i) (i)
K1 K2

d D 1
(i)
= =
2
,
(i) 2
(i) 2
I4 D (i) (i)
(2π ) 2 − K − K1 − K2 + K4
1
(i) (i)
K4 K3
(6.27)
(i) (i) (i) (i)
where (K 1 , K2 , K3 , K4 )
are sums of the external momenta at each corner. The label
i indicates a particular choice of distributing the external lines on the four subamplitudes,
i.e. different 4-line cuts.
It may seem surprising that all 1-loop integrals can be written in terms of a basis of scalar
1-loop integrals as in (6.26), since integrals obtained from Feynman diagrams generally
involve numerators that depend on loop-momentum. The simplification is that one can use
the external momenta to form a basis for any vectors in the integrals. For example, a 4-point
1-loop “vector integral” can be expanded as

μ μ μ μ
d D
2
2
2 = a1 p1 + a2 p2 + a3 p3 . (6.28)
− p1 − p1 − p2 + p4
2
Here we are simply using the fact that the integral must be a Lorentz vector, and since
by momentum conservation there are only three linearly independent momentum vectors
in the 4-particle input, the answer must be a linear combination thereof. The coefficients
in (6.28) are solved using the Passarino–Veltman reduction procedure [76]. The idea is
simply to dot in various combinations of external momenta and simplify the result. For
example, dotting p1 into (6.28), the numerator of the integral becomes
1
2
( · p1 ) = + p1 − 2 . (6.29)
2
The two terms of the RHS each cancel a loop-propagator, leaving behind two scalar integrals.
Then (6.28) gives
$ '
1 1
d
D

2
2 −
2
2
2
2 2 + p1 + p2 − p4 2 + p1 + p1 + p2 − p4
1

= − a2 s + a3 t . (6.30)
2
Repeating the above procedure, one can obtain three linear equations for a1 , a2 , and a3 .
Solving this linear system is straightforward and gives the vector integral (6.28) as a linear
combination of scalar integrals.
Exercise 6.1
Only scalar triangle integrals appear on the LHS of (6.30). Show that scalar box
integrals appear if we contract (6.28) with p2 .
In general, since there are only D independent vectors in D-dimensions, the set of needed
scalar integrals can be reduced to the set of scalar integrals shown in (6.26). Tensor-integrals
can be simplified similarly. A more detailed discussion of integral reductions can be found
in Section 4.2 of [77], also see [3].
The basis expansion (6.26) makes the task of computing the 1-loop amplitude a matter
(i)
of determining the coefficients Cm . Since the scalar integrals have distinct propagator
structures, only a subset contributes to a given unitarity cut. By applying multiple unitarity
(i)
cuts, one obtains a set of linear equations that relate the Cm s to the results of the cuts,
each computed as a product of tree amplitudes. Solving these linear equations gives us the
(i)
coefficients Cm s as a combination of products of tree amplitudes. By (6.26), this determines
the 1-loop amplitude, up to the possibility of rational terms that we discuss below.
(i)
The integral coefficients Cm can be found systematically by organizing the calculation
(i)
according to the number of propagators present in the scalar integrals Im . In D-dimensions,
the loop-momentum has D components, so one can find isolated solutions for such that
(i)
all propagators in the scalar integral I D are on-shell: this corresponds to a D-line cut
and this is the maximally possible cut in D dimensions. It is called a maximal cut and is
discussed in much more detail in Chapter 8.
Since the cut constraints are quadratic in loop-momentum, there are two isolated so-
lutions, denoted ∗(1) and ∗(2) , to a maximal cut. The corresponding coefficient C D is
(i)
i i+1
2
1 j
1 3
n j+1
4
k+1 k
(i) (i) (i)
Fig. 6.2 1-loop box diagram withK1 = p1 + · · · + = pi+1 + · · · + p j , K 3 = p j+1 + · · · + pk
pi , K 2
(i) (i)
and K 4 = pk+1 + · · · + pn . The corresponding box coefficient C4 in (6.32) is the product of the four tree
amplitudes at each corner.
completely determined by the product of D tree amplitudes as
(i) 1
CD = n 1 · · · An D .
Atree tree
(6.31)
2
= ∗(1) , ∗(2)
Note that one averages over the two solutions, ∗(1) and ∗(2) . At 1-loop, the relative weight
between the two solutions can be determined by considering special integrands that integrate
to zero. The associated maximal cut must also vanish and this fixes the above prescription.
See [78] for a concise discussion.
Example. Let us make (6.31) concrete. For an n-point 1-loop amplitude in D = 4, the
coefficient of the box integral shown in Figure 6.2 is given by
&
(Fig. 6.2) 1
C4 = An 1 − 1 , 1, . . . , i, 2 × An 2 − 2 , i + 1, . . . , j, 3
2 ∗(1) ∗(2) states %
= ,

×An 3 − 3 , j + 1, . . . , k, 4 × An 4 − 4 , k + 1, . . . , n, 1 ,
(6.32)

where states indicates a state sum for each internal line 1 = , 2 = − ( p1 + · · · +
pi ), 3 = − ( p1 + · · · + p j ), and 4 = + ( pk+1 + · · · + pn ). The vectors ∗(1) and
∗(2) solve the on-shell conditions 21 = 22 = 23 = 24 = 0 of the 4-line cut.
(i)
Coefficients Cm with m < D are not quite as simple to calculate, but they can be obtained
(i)
systematically. After determining all D-gon coefficients, we treat (D−1)-cuts. Both I D
( j)
and I D−1 integrals can potentially contribute to (D−1)-cuts, but since we have already
(i) ( j)
determined all the C D s, we can unambiguously determine all C D−1 s. Similarly, all integral
coefficients can be determined iteratively. This way, the generalized unitarity method offers
a systematic way to determine the 1-loop amplitude in terms of tree amplitudes. Detailed
discussions of extracting 1-loop integral coefficients in D = 4 can be found in [47, 79].
Example. As a concrete example, the integral coefficients for the fermion-loop correc-
tion of the 1-loop 4-gluon amplitude A4 [1+ , 2− , 3+ , 4− ] are
−u 2 s 2 (u 2 + s 2 )
C4 ( p1 , p2 , p3 , p4 ) = Atree
4 ,
t4
us(u 3 + us 2 )
C3 ( p1 , p2 + p3 , p4 ) = Atree
4 , (6.33)
t4
2u(5s 2 + stu + 2u 2 )
C2 ( p1 + p4 , p2 + p3 ) = Atree
4 ,
6t 3
where the integral coefficients are labeled in the same way as the scalar inte-
grals in (6.19). All other integral coefficients are related to the above by cyclic
permutation.
Exercise 6.2
Use the integrated results in (6.20) for the 4-point box, triangle, and bubble integrals,
and the above integral coefficients to compute the final result for the fermion loop-
correction to the 1-loop 4-gluon amplitude A4 [1+ , 2− , 3+ , 4− ]. Show that there are
only UV and no IR divergences. Explain why. What does the UV divergence indicate
about renormalization of the theory at 1-loop level?
Of course, there is a big elephant in the room – you already met it in (6.26): it is the
rational terms. Rational terms are rational functions that do not possess branch cuts, so
they are undetectable by unitarity cuts. The rational terms arise from the need to regularize
the loop integrals. In dimensional regularization, the loop-momentum is really (D−2)-
dimensional. If we separate the loop-momentum into a D-dimensional part, (D) , and a
(−2)-dimensional part, μ−2 , there can also be contributions from the μ-integrals. An
example of an integrand that gives a branch-cut-free contribution is the (D/2 + 1)-gon
scalar integral with μ2 numerator: it integrates to a finite value

d (D) dμ−2 1 1
D−2
I D/2+1 [μ2 ] = − D/2
+ O() . (6.34)
(2π ) (4π ) (D/2)!
One cannot capture this from the ordinary unitarity cut since it is just a rational func-
tion (here, a constant). The unitarity cut forces the loop-momentum to be on-shell in
D-dimensions and this implies μ2 = 0, so the contribution from the above integrand van-
ishes. On the other hand, if one considers unitarity cuts where the internal lines become
massless in D − 2 dimensions ( (D) )2 + μ2 = 0, or equivalently massive in D-dimensions
with mass m 2 = μ2 , such terms are detectable. Thus rational terms can be reconstructed
from unitarity cuts if we allow the states crossing the cut to be massive.
Rational terms are absent for supersymmetric Yang–Mills theories because supersym-
metry cancellations ensure that the powers of loop-momentum in the integrals do not lead
to rational terms after integral reduction. For non-supersymmetric theories – for example
λφ 4 , QED, QCD – rational terms are present and they are the most time-consuming ones to
compute. We will not discuss this important issue in further detail, but simply refer you to
[80, 81] and references therein. Instead, we illustrate the unitarity method by working out
an explicit example.
Example. In this example, we calculate the 4-point 1-loop superamplitude in N = 4

SYM using the generalized unitarity method. We do so by first considering a 2-line cut
and then infer from it which terms contribute in the integral basis expansion (6.26).
The s-channel unitarity cut is
2
2 3

Cuts = L R = A4 [− 1 , 1, 2, 2 ] × A4 [− 2 , 3, 4, 1 ] .
states
1 4
1
(6.35)
Using the analytic continuation (3.17), the tree-level MHV superamplitudes are
δ (8) (L)
A4 [− 1 , 1, 2, 2 ] = ,
1 1122 2 2 1
(6.36)
δ (8) (R)
A4 [− 2 , 3, 4, 1 ] = .
2 3344 1 1 2
The arguments of the Grassmann delta functions are L = −| 1 η 1 + |1η1 + |2η2 +
| 2 η 2 , and R = −| 2 η 2 + |3η3 + |4η4 + | 1 η 1 . As in the case of tree-level re-
cursion (see the discussion around (4.55)), the intermediate state sum is performed
as an integration of the on-shell Grassmann variables η i associated with each in-
ternal line [36]. These integrals are easy to perform when we use δ (8) (L) δ (8) (R) =
δ (8) (L + R) δ (8) (R) = δ (8) ( Q̃) δ (8) (R). We find

δ (8) Q̃ δ (8) (R)
Cuts = d 4 η 1 d 4 η 2
1234 1 12 2 2 1 2 34 1 1 2

δ (8) Q̃ 1 2 2
=− . (6.37)
1234 1 12 2 2 34 1
On the unitarity cut, one can convert the loop-momentum part of the denominator in
the above expression into propagators:

−su
Cuts = A4 [1234] ×
tree . (6.38)
( 2 + p2 ) ( 1 + p4 ) 21 = 22 =0
2 2
Exercise 6.3
Show that (6.38) follows from (6.37).
Let us now consider the possible integrals from (6.26) that contribute to the s-
channel cut: the integrals that contain the propagators with 21 = 2 and 22 = ( − p1 −
p2 )2 are the box-integral I4 ( p1 , p2 , p3 , p4 ), the triangle-integrals I3 ( p1 , p2 , p3 + p4 ),
and I3 ( p3 , p4 , p1 + p2 ), and the bubble-integrals I2 ( p1 + p2 , p3 + p4 ) and I2 ( p2 +
p3 , p4 + p1 ). In each case, we have indicated the distribution of the external lines. The
result (6.38) for Cuts shows that there are two uncut propagators left after cutting 1 and
2 , so this excludes the triangle- and bubble-integrals. Thus we conclude that only the
1-loop
box-integral is present, i.e. A4 [1234] = C4 I4 ( p1 , p2 , p3 , p4 ). The box-coefficient
C4 is readily determined from (6.38), giving3
1-loop
A4 [1234] = su Atree
4 [1234] I4 ( p1 , p2 , p3 , p4 ) . (6.39)
To make sure that (6.39) is the correct result for the amplitude, we examine other
distinct cuts to see if there could be terms that vanish in the s-channel cut and were
therefore not captured in our analysis. The only other available cut is the u-channel cut.
(Color-ordering excludes the t-channel cut.) But since the RHS of (6.39) is invariant
under cyclic permutations, it is guaranteed to produce the correct u-channel cut. Hence
(6.39) is indeed the correct 1-loop 4-point amplitude for N = 4 SYM. We discuss the
evaluation of I4 in dimensional regularization in Section 6.4.
Working through the details in the example above, you will notice that the unitarity method
does take some work and you may wonder how it compares with a brute-force 1-loop
Feynman diagram calculation. The answer is that the unitarity method is superior, since it
heavily reduces the number of diagrams needed and it avoids gauge obscurities. For the
unitarity method, the input is gauge-invariant on-shell amplitudes. You might find it curious
1-loop
that the first computation of the 1-loop 4-gluon amplitude A4 in N = 4 SYM was not
1-loop
done in QFT, but in string theory: in 1982, Green, Schwarz, and Brink [82] obtained A4
as the low-energy limit of the superstring scattering amplitude with four external gluon
states.
We close this section with some general comments on the 1-loop integral expansion
(6.26). The representation of 1-loop amplitudes in terms of scalar integrals provides an
interesting categorization scheme in terms of whether or not particular classes of integrals
– in 4d: boxes, triangles, bubbles, and rationals – appear or not. For example, we mentioned
earlier that rational terms are absent in N > 0 super Yang–Mills theory. One can ask
which theories involve only box-integrals, i.e. no triangle- or bubble-integrals and no
rational terms. In 4d, such so-called “no-triangle” theories include N = 4 SYM [67],
N = 8 supergravity [83, 84], and N = 2 SYM coupled to specific tensor matter fields
[85]. For pure N = 6 supergravity, only box- and triangle-integrals appear [86], while for
pure N ≤ 4 supergravity all integrals in (6.26), including rational terms, appear.
(i)
One relevant aspect of the above analysis is that in 4d only the bubble integrals I2
contain ultraviolet divergences. In dimensional regularization, all bubble integrals have a
(i)
common leading 1/-term, and hence contribute 1 i C2 to the amplitude. As a result,
(i)
the beta function for a given theory vanishes at 1-loop order precisely when i C2 = 0.
(i)
In fact, in a renormalizable theory, one must have i C2 ∼ Atree n and the proportionality
constant is related to the 1-loop beta function [47, 85, 87, 88]. Note that even though bubble
(i)
coefficients are non-trivial for pure N ≤ 4 supergravity theories, their sum i C2 must
vanish since the theory is known to be free of ultraviolet divergences at 1-loop order.
Next, we offer a quick survey of results for loop amplitudes in planar N = 4 SYM.
3 The minus sign compared with (6.38) comes from the (−i)2 in the cut propagators.
6.4 1-loop amplitudes in planar N = 4 SYM
When we introduced the N = 4 SYM theory in Section 4.4, we mentioned that this is a
conformal theory, there is no running of the coupling. This means that there are no UV
divergences in the on-shell scattering amplitudes at any loop-order. At 1-loop order, this
can be seen from the fact that in N = 4 SYM only the box-integrals contribute in the scalar
integral basis expansion (6.26), i.e.
(i) (i)
N = 4 SYM in 4d: A1-loop
n = Cbox Ibox . (6.40)
i
Only bubble-integrals give UV divergences at 1-loop in 4d, so it is clear that the 1-loop
amplitudes in N = 4 SYM are UV finite. Loop amplitudes in N = 4 SYM theory do have
IR divergences and we will see more of them in the following.
We used the unitarity method in Section 6.3 to construct the 1-loop 4-point superampli-
tude in planar N = 4 SYM. We found that the answer (6.39) could be written in terms of
a single scalar box-integral:
1-loop
4 [1234] I4 ( p1 , p2 , p3 , p4 ) . (6.41)
The scalar box integral I4 , evaluated in dimensional regularization D = 4 − 2, was given
in (6.20). It follows that
!
1-loop 2

−2 2 −

−2 2 −

A4 [1234] = Atree
4 [1234] − μ y + − μ y
2 13 24
y2 "
− ln2 132
− π 2
+ O() , (6.42)
y24
where μ is the regularization scale and yi j = yi − y j are the zone-variables defined in

(5.31). In terms of Mandelstam variables, we have s = −y132
and u = −y24
2
.
Let us now use the notation
K
AN
n;L
MHV
= n-point L-loop N K MHV superamplitude of planar N = 4 SYM , (6.43)
with the color-ordering 12 . . . n of external particles implicit. Since supersymmetry and

SU (4) R-symmetry4 Ward identities hold at each loop-order, this decomposition of the
N K MHV
loop amplitude is sensible and An;L has Grassmann degree 4(K + 2).
It is convenient to factor out an MHV tree-level superamplitude and write the loop-
expansion as
K
K
N K MHV
AN
n;L
MHV
() = AMHV
n;0 Pn;0
N MHV
+ λ Pn;1 () + · · · , (6.44)
4 The SU (4) R-symmetry is non-anomalous [89, 90].

6.4 1-loop amplitudes in planar N = 4 SYM 133
where λ ∼ g 2 N is the t’Hooft coupling written in terms of the gauge coupling g and the
rank of the gauge group SU (N ). We include the dependence on the -regulator explicitly
in the loop amplitudes. At tree-level, we have

n−2
n
Pn;0
MHV
= 1 and Pn;0
NMHV
= R1 jk . (6.45)
j=3 k= j+2
The NMHV result is given in terms of the dual superconformal invariants R1 jk defined in
(4.83) and discussed further in Section 5.4.
Exercise 6.4
Why is it possible to factor out AMHV
n;0 even at loop-level?
The -regulator explicitly breaks the conformal and dual conformal symmetry. You can see
that in the expression (6.42) for the 1-loop 4-point superamplitude: not even the finite part
O( 0 ) respects dual conformal inversion (5.37):
yi2j
I (yi2j ) = . (6.46)
yi2 y 2j
So the raw output of the loop amplitudes does not entertain the ordinary or dual conformal
symmetries of the N = 4 SYM theory. However, the IR divergences take a universal form
that facilitates construction of IR-finite quantities that turn out to respect the symmetries.
K
At 1-loop order, the IR divergent part of AN n;1
MHV
is captured entirely by the MHV
superamplitude in the sense that
N K MHV N K MHV
MHV
An;1 () = An;0 × IRdiv Pn;1 () + O( 0 ) , (6.47)
where
MHV 1
n
IRdiv Pn;1 () = 2 (−μ−2 yi,i+2
2
)− . (6.48)
i=1
Note that for n = 4, this reproduces the IR divergent terms in (6.42). The finite part of the
4-point MHV superamplitude is
2
MHV y13
F4;1
MHV
() ≡ P4;1
MHV
() − IRdiv P4;1 () = −ln2 2
− π 2 + O() . (6.49)
y24
The universal form (6.47) of the 1-loop IR divergences implies that the ratio functions5
K K K
RN
n;1
MHV
() ≡ Pn;1
N MHV
() − Pn;0 Pn;1 ()
N MHV MHV
(6.50)
K MHV
5 The RHS of (6.50) can be viewed as the O(λ) term in the small λ expansion of the ratio AN
n /AMHV
n .
K
are IR finite. Moreover, it has been proposed [50, 91] that Rn;1N MHV
(0)s are actually dual
conformal invariant. This was shown at NMHV level for n ≤ 9 in [50, 91] and for general
n in [92, 93] using generalized unitarity. To give you a sense of the expressions, we present
the result [91] for the ratio function for the 6-point 1-loop NMHV superamplitude. It is
1
RNMHV
6;1 (0) = R146 V146 + cyclic , (6.51)
2
where
1 π2
3
V146 = − log u 1 log u 2 + log u k log u k+1 + Li2 (1 − u k ) − . (6.52)
2 k=1 6
The dilogarithm Li2 was defined in (6.4). The u i s are dual conformal cross-ratios,
2 2 2 2 2 2
y13 y46 y24 y51 y35 y62
u1 = 2 2
, u2 = 2 2
, u3 = 2 2
, (6.53)
y14 y36 y25 y41 y36 y52
so each Vi jk is a dual conformal invariant, as you can see by applying dual inversion (6.46).
The “+ cyclic” in (6.51) is the instruction to sum over the cyclic sum of the external state
labels; note that V251 is just V146 with u 1 → u 2 → u 3 → u 1 .
Recalling that the BCFW recursion relations for the tree-level 6-point NMHV su-
peramplitude only have three terms, you might be surprised to see six terms in the 1-
loop result (6.51). However, in Exercise 5.15 we used the cyclically invariant 5-brackets
[i, j − 1, j, k − 1, k] = Ri jk to rewrite the tree-level superamplitude as ANMHV
6;0 /AMHV
6;0 =
P6;0
NMHV
= 2 (R146 + cyclic). This was done in anticipation of the 1-loop ratio function
1
(6.51), and now you see that (6.51) is just like the tree-level result but with each Ri jk
dressed with a dual conformal invariant Vi jk . Adding loop-orders 0 and 1, we can therefore
write 6-point ratio function
1

RNMHV
6 (0) = R146 1 + λ V146 + cyclic + O(λ2 ) . (6.54)
2
There are two properties worth noting about the 1-loop ratio function RNMHV
6;1 (0):
• It is dual conformal invariant, but not dual superconformal invariant. For a discussion of
this, see [94].
• V146 – and hence RNMHV
6;1 (0) – has uniform transcendentality 2. This can be extended to
the -dependent terms if is assigned transcendentality −1.
Both of these properties carry over to all RNMHV

n;1 (0). At higher-loop order in planar N = 4
SYM, the degree of transcendentality is expected to be uniformly 2L.
At higher-point, there are more dual conformal invariant cross-ratios available than just
the three u i s for n = 6. Consequently, the NMHV 1-loop ratio functions RNMHV n;1 (0) are
more involved; however, they are all known explicitly and they take a similar form to
(6.51). You can find the results for RNMHV
n;1 (0) in [93].
2 3 2 3 3
3
I 41-loop = y 13
2 2
y 24 × 2 a 4, I 42-loop = ( y 13 ) y 24 ×
2 2 2
2 a b 4 + cyclic ,
1 1 4 1 1 4
2 3 3
2 3 3
b
I 43-loop = ( y 13 ) y 24 × 2
2 3 2
a b c 4 2 2 2
+ ( y 13 2
) y 24 y a4 × 2 a 4 + cyclic .
c
1 1 4
1 1 4
The integrands of N = 4 SYM 4-point amplitude to 3-loop order. These are the unique scalar integrands that are dual Fig. 6.3
conformal invariant.
6.5 Higher-loop amplitudes in planar N = 4 SYM
The generalized unitarity method can be applied successfully to higher-loop amplitudes,

both at the planar and non-planar level; for a recent review see [68]. The application of
unitarity is most efficient when a complete integral basis is available; for example, for
1-loop amplitudes in 4d, the basis consists of the scalar box-, triangle-, and bubble-integrals
in (6.26).
Beyond 1-loop, there is not a complete understanding of the basis integrals for amplitudes
in generic quantum field theories, although partial results have been achieved at 2-loops in
the planar limit, see [78, 95–99]. One thing worth noting is that the integral basis is finite
[100].
Without a given basis of integrals, one strategy is to construct the most general integral
Ansatz that satisfies certain criteria, such as dimension-counting, and then use various inte-
gral identities to recast the Ansatz into a basis of independent integrals. Further symmetries,
such as dual conformal invariance in planar N = 4 SYM, can be a strong handle on finding
a complete integral basis. As an example, the diagrams in Figure 6.3 correspond to the
only dual conformal invariant scalar integrals for the 4-point 1-, 2- and 3-loop integrands
of planar N = 4 SYM. The coefficients of each integral are fixed by applying unitarity
cuts [101, 102], so that the LHSs of the equations in Figure 6.3 are the full integrands for
the 4-point 1-, 2- and 3-loop amplitudes in planar N = 4 SYM. The evaluation of these
integrals leads to interesting results that we discuss next.
The analytical result [103] for the 2-loop 4-point amplitude in planar N = 4 SYM was
shown by Anastasiou, Bern, Dixon, and Kosower (ABDK) [101] to be expressible in terms
of the 1-loop amplitude as
1 MHV 2
P4;2
MHV
() = P () + P4;1
MHV
(2) f (2) () + C (2) + O() , (6.55)
2 4;1
where the MHV tree factor is stripped off as in (6.44), f (2) () = −ζ2 − ζ3 − ζ4 2 and

C (2) = −ζ22 /2. Here ζs = ∞ −s
is the Riemann zeta function; note ζ2 = π6 , ζ3 ≈ 1.202,
2
k=1 k
and ζ4 = π90 .
4
It is interesting that the 2-loop 4-point amplitude in planar N = 4 SYM can be written
in terms of the 1-loop result. But the plot thickens! By explicit calculation of the 3-loop
4-point amplitude in planar N = 4 SYM, Bern, Dixon, and Smirnov (BDS) [102] found
that the iterative structure continues:
1 MHV 3
P4;3
MHV
() = − P () + P4;1
MHV
() P4;2
MHV
() + f (3) () P4;1
MHV
(3) + C (3) + O() .
3 4;1
(6.56)
Here f (3) () = 11 ζ + O() and C (3) is a constant.

2 4
The 2- and 3-loop results indicate an exponentiation structure. This motivates the
ABDK/BDS Ansatz for the full MHV superamplitude in N = 4 SYM:
$ '
∞

PnMHV(BDS) () = exp λL f (L) () Pn;1
MHV
(L) + C (L) + O() . (6.57)
L=1
This Ansatz is almost correct: keep reading! In the ABDK/BDS Ansatz, the functions f (L)
(L) (L) (L) (L)
are of the form f (L) () = f 0 + f 1 + 2 f 2 , and the constants C (L) and f 0,1,2 () are
independent of the number of external legs n. In particular, at 1-loop order f (1) () = 1 and
C (1) = 0, and at 2-loops the results for f (2) () and C (2) were given below (6.55).
Exercise 6.5
Show that (6.57) reproduces the 4-point 2- and 3-loop expressions (6.55) and (6.56).
Of course, the way one would go about testing the ABDK/BDS exponentiation Ansatz (6.57)
is by direct calculation of the n-point L-loop amplitudes at L = 2, 3, . . . . But how many 2-
loop amplitudes have you ever calculated? Yeah, it is not an easy task, nonetheless progress
has been made. It has been shown numerically in [104, 105] that the exponentiation Ansatz
correctly produces the 5-point 2-loop amplitude. It is very interesting that something
new happens at 6- and higher-point: while the ABDK/BDS Ansatz matches the IR divergent
structure, it does not fully produce the correct finite part. The ABDK/BDS Ansatz determines
the finite part of the amplitude only up to a function of dual conformal cross-ratios of the
external momenta. This function is called the remainder function and it is defined as
MHV(BDS)
rn;L () ≡ Pn;L
MHV
() − Pn;L () , (6.58)
MHV(BDS)
where Pn;L
MHV
() is the actual MHV L-loop amplitude and Pn;L () is the O(λ L ) terms
in the expansion of the exponential Ansatz (6.57). The remainder function does not show
up for n = 4, 5 because in those cases there are no available conformal cross-ratios.
The first indication of the remainder function came from a strong coupling calculation
by Alday and Maldacena [106] who proposed [107] to use the AdS/CFT correspondence
to calculate PnMHV . Subsequently, it was verified numerically that a remainder function is

needed for the parity-even part of the 6-point 2-loop MHV amplitude [108], whereas
ABDK/BDS successfully determines the parity-odd part [109]. The analytic form of the
remainder function r6;2 for the 6-point 2-loop MHV amplitude was calculated (as a hexag-
onal Wilson-loop) by Del Duca, Duhr, and Smirnov [110, 111]. The result, written in terms
of the three dual conformal cross-ratios u 1,2,3 in (6.53), is a respect-inducing 17-page long
sum of generalized polylogarithms; all terms have transcendentality 4. In an impressive
application of a mathematical tool known as the Symbol, Goncharov, Spradlin, Vergu, and
Volovich [112] managed to simplify this complicated result for r6;2 to an expression that
involves only regular polylogs – Lis and log – and fits in just a few lines of LATEX.
The simple answer [112] for r6;2 is an important step towards a better understanding
of loop amplitudes in planar N = 4 SYM. The Symbol is now being used to understand
higher-loop amplitudes, however, there will be amplitudes in planar N = 4 SYM involving
integrals that the Symbol does not help with. Thus techniques are eventually needed beyond
the Symbol.
We have reviewed the unitarity method and shown you how it allows us to construct
L-loop amplitudes from on-shell lower-loop input. The approach explores the analytic
structure of the loop-integrands in a somewhat different way from the recursive techniques
you know from tree-level amplitudes. BCFW is available at the level of loop-integrands,
and that is the subject of the next chapter.
7 BCFW recursion for loops
The unitarity cuts discussed in Chapter 6 involve cuts of at least two propagators. In contrast,
the tree-level recursion relation of Chapter 3 looks more like a single-line cut construction,
since it uses the factorization-structure of a single propagator going on-shell. It is tempting
to ask if loop amplitudes can similarly be reconstructed from the singularities associated
with taking a single propagator on-shell. In this chapter, we study a recursive approach that
generalizes the tree-level BCFW construction to loop-integrands.
We begin in Section 7.1 with a discussion of what is meant by the loop-integrand of an
amplitude and how and when it is well-defined. The rest of the chapter focuses on super-
BCFW recursion relations for loop-integrands of planar superamplitudes in N = 4 SYM.
In this setting it is quite natural to work with the momentum twistor variables introduced
in Section 5.4; we will in particular make use of the description given there of points, lines,
and planes in momentum twistor space. It is interesting how the description of the BCFW
shift becomes geometric in the momentum twistor description, for example, the on-shell
conditions for the internal line in a diagram become a statement of intersection of lines
and planes. We demonstrate the geometrization in Section 7.2 and apply it in Section 7.3 to
derive the super-BCFW result for the Nk MHV tree-level superamplitude in N = 4 SYM.
Thence, suitably warmed up and prepared, we introduce loop-recursion in Section 7.4. In
Section 7.5, we give a fully worked-through example of its application to the 4-point 1-loop
superamplitude of N = 4 SYM. We end with brief comments about higher loop amplitudes
in Section 7.6.
7.1 Loop-integrands
As seen in Chapter 6, loop amplitudes have a complicated analytic structure, so we focus

on the loop-integrand which is just a rational function with poles at the location of the
propagators, much similar to the tree amplitudes. Suppose we do a BCFW shift on the
μ μ μ μ
external legs, for example p1 → p1 + z q μ and pn → pn − z q μ , with q 2 = 0 as usually.
We can deduce from the Feynman diagrams that the shifted loop-integrands possess two
types of poles in z:
1. poles in loop-independent propagators;

2. poles in propagators involving loop-momentum.
7.1 Loop-integrands 139
^1 4^
^1
4^
2 3
^1 r r
4^
2 3
2 3
Representation of the two types of poles occurring in a BCFW shift of a 4-point 3-loop integrand. Type 1 are poles asso- Fig. 7.1
ciated with loop-momentum independent propagators and type 2 poles are loop-momentum dependent propagators.
The former factorizes the 3-loop integrand into a product of a 2-loop and a 1-loop integrand while the latter gives a
forward limit of a 2-loop amplitude with two extra legs.
The residue of a type 1 pole corresponds to factorization of the integrand into a product
of two lower-loop integrands. The residue of a type 2 pole in an L-loop n-point inte-
grand is an (n+2)-point (L −1)-integrand with two adjacent legs evaluated in the forward
limit
μ μ
pi → r μ , pi+1 → −r μ , with r 2 = 0 . (7.1)
This is illustrated for the example of a 4-point 3-loop amplitude in Figure 7.1. The poles of
type 2 are what we would call single-line cuts in the unitarity method [113, 114].
Thus – provided that the large-z behavior is well-understood – it appears that one
can straightforwardly set up a recursion relation for loop-integrands. However, there are
subtleties we have to resolve:
• The first issue has to do with the identification of the loop-momenta in the loop-integrand.
In the amplitude, we have to integrate the loop-momenta, so the i are just dummy
variables that can be redefined while still giving the same integrated answer. But the
integrand itself can have different pole structures depending on how the i are identified.
As an example, consider the 1-loop 4-point box-integral (6.1) and compare the equivalent
(a) (b) (a)
parameterizations I4 and I4 = I4 ( → + p1 ). BCFW shifting legs 1 and 2 yields
−r r −r
r
p
(a) (b)
Fig. 7.2 Examples of diagrams problematic for the forward limit. In the lefthand diagram of (a), the propagator between the
forward legs diverges due to momentum conservation. The righthand part of diagram (a) illustrates that such a diagram
corresponds to the single-cut of a bubble on an external leg. Similar remarks apply to the diagrams in (b), where the
limit corresponds to cutting a tadpole.
two distinct analytic functions in z:
(a) 1
I4 (1̂, 2̂, 3, 4) = ,
2 (
− p1 − − p1 − p2 )2 ( + p4 )2
z r )2 (
(7.2)
(b) 1
I4 (1̂, 2̂, 3, 4) = .
( + p1 + z r ) ( − p2 + z r )2 ( − p2 − p3 + z r )2
2 2
In general there is no canonical way to identify how the loop-momentum is parameterized,

so that is the first subtlety that needs to be resolved. It basically comes down to the
definition of what we mean by the “un-integrated integrand.”
• The second subtlety has to do with the forward limit (7.1). When loop-momentum-
dependent propagators go on-shell, there is a residue corresponding to a lower-loop
(n+2)-point integrand in the forward limit, but that limit suffers from singularities. For
example, from the explicit Feynman diagrams one sees that if the forward legs are attached
to the same external line, then due to momentum conservation there is a 1/ p 2 singularity
as p 2 → 0. Such diagrams can be identified with cuts of bubbles on external legs or
tadpole diagrams. This is illustrated in Figure 7.2. In massless theories, these integrate to
zero in dimensional regularization and do not contribute to the loop amplitude. However,
prior to integration, they are part of the integrand and will contribute to the single cuts.
Therefore an important, but difficult, task is to identify these contributions in the forward
limit such that one can consistently remove them.
Resolution of the above subtleties has been partially achieved for non-supersymmetric
theories [113] and completely resolved in supersymmetric theories in the planar limit
[114, 115]. In particular, it was shown that for supersymmetric theories the problematic
terms associated with the tadpole as well as for external bubbles cancel in the state sum
over the supermultiplet, and thus one has a perfectly well-defined residue. Furthermore, in
the planar limit, the loop-momenta in the integrand can be defined unambiguously. This is
done by defining the i s with a specific relation to the ordering of the external momenta.
For example at 1-loop, one choice is to declare that is the momentum associated with
the internal line immediately before line 1. For the 4-point 1-loop box integral, this selects
(a)
integrand I4 in (7.2).
The identification of the loop-momentum is naturally done in dual space y, that we
defined in (5.31) in order to make momentum conservation manifest. We noted there that
the yi s are also sometimes called zone variables; that is because we can think of them as
labeling the “zones,” or regions, that the external lines of the amplitude separate the plane
into. This assumes a well-defined ordering of the external lines based on the color-ordering –
and at loop-level it further requires the graphs to be planar. Let us illustrate this for a 6-point
tree-graph and the 4-point 1-loop box diagram:
2 1 2
1 y2 y2
y3
y1 3
tree y1 y0 y3 (7.3)
6 y6 y4
y4
y5
5 4 4 3
At loop-level there are “internal” zones, one for each loop. This offers the opportunity
to switch the integration variable from i to the new internal zone variables. For example
in the 4-point 1-loop box graph, we can use y0 as the loop-parameter instead of . They
are related by = y1 − y0 , similarly to the relationship pi = yi − yi+1 . The dual variables
therefore give an unambiguous definition of the loop-momentum and this facilitates the
loop-level recursion relations for planar integrands.
As a useful example, let us express the 4-point scalar box-integral from (7.3) in dual
variables. Following the rules for identifying the momentum on each internal line in terms
of the zone-variables of the two adjacent zones, we have i2 = (y0 − yi )2 = y0i2 . The box-
integral is therefore simply

d 4 1
I 4 ( p1 , p2 , p3 , p4 ) =
2

(2π ) 2 − p1 − p1 − p2 2 + p4 2
4

d 4 y0
→ 2 2 2 2
. (7.4)
y01 y02 y03 y04
To reiterate, for planar amplitudes in supersymmetric theories, all subtleties in defining

the loop-integrand have been overcome. In the following, we introduce a BCFW recursion
relation that generates the planar loop-integrands for N = 4 SYM; it was developed in
[115] and also considered in [116].
7.2 BCFW shift in momentum twistor space
The planar integrand is well-defined in the dual coordinates yi , so it would seem natural
to formulate the BCFW shifts in the dual representation. However, it is actually more
interesting to use the momentum supertwistors ZiA that we introduced in Section 5.4. The
ZiA s can be chosen freely in CP3|4 and make momentum conservation, supermomentum
conservation, and on-shell conditions automatic. We can therefore set up the BCFW shift
without worrying about these constraints. The simplest possibility is to shift just a single
momentum twistor, say Zi :
Ẑi = Zi + wZi+1 . (7.5)
All other Z j s are unshifted. The shift parameter w is a complex variable, w ∈ C. Geomet-
rically, (7.5) is the statement that the point Ẑi lies on the line (i, i + 1) = (Zi , Zi+1 ).1
Let us translate (7.5) back to the spinor helicity formalism. In components (7.5) says
|î = |i + w |i + 1 , |μ̂i ] = |μi ] + w |μi+1 ] , χ̂i A = χi A + w χi+1,A . (7.6)
Using (5.46) and the incidence relations (5.44), one finds that
ŷi = yi + z |i − 1[i| , (7.7)
where
wi, i + 1
z= . (7.8)
i − 1, i + wi − 1, i + 1
All other y j s are unshifted.
Exercise 7.1
Use (5.46) and (5.44) to show that yi+1 and yi−1 are unshifted. Then derive (7.7).
The shift in y-space makes sense geometrically, because by Figure 5.1 the point ŷi+1 is
determined by the line (î, i + 1) which is equivalent to the line (i, i + 1) since (7.5) exactly
tells us that the point Ẑi lies on (i, i + 1). So we conclude ŷi+1 = yi+1 . On the other hand, ŷi
is determined by the line (i − 1, î) which is different from (i − 1, i), so ŷi = yi for w = 0.
Translating from dual y-space to momentum space, we have

p̂i = ŷi − yi+1 = − |i − z|i − 1 [i| , (7.9)

p̂i−1 = yi−1 − ŷi = −|i − 1 [i − 1| + z[i| . (7.10)
No other momenta shift. We immediately read off that this is a [i − 1, i BCFW shift,
|î = |i − z|i − 1 , |i
− 1] = |i − 1] + z|i] . (7.11)
Since there is also a shift of the Grassmann-components of the momentum supertwistors,
(7.5) actually induces a BCFW [i − 1, i-supershift.
Exercise 7.2
Show that the Grassmann-part of the shift in (7.5) is equivalent to η̂i−1 = ηi−1 + z ηi .
1 As in Section 5.4, we denote the line in momentum twistor space defined by two points (Z j , Z k ) as ( j, k), and
the plane defined by three points (Z j , Z k , Z l ) as ( j, k, l).
Exercise 7.3
It may seem surprising that the shift Ẑi = Zi + wZi+1 is equivalent to a [i − 1, i-
supershift; one might have expected a shift involving lines i and i + 1 instead. Actually,
the shift (7.5) is also equivalent to a [i + 1, i shift: this is because the momentum
twistors are defined projectively, so we could supplement (7.5) with an overall scaling.
For example, one finds that the angle spinor shift in (7.11) is equivalent to
i − 1, i

|î = |i + w|i + 1 . (7.12)
i − 1, i + wi − 1, i + 1
Manipulate |î in (7.11) to find (7.12).
Our next task is to describe the kinematics associated with the internal lines in the BCFW
diagrams – it turns out to have a nice geometric description in momentum twistor space.
Consider a typical BCFW diagram associated with a factorization channel PI :
î
i–1
^y
i yi+1
I i+1
. (7.13)
j+1 yj
j j–1
For simplicity, let us for now suppose that there are no loop-momenta in PI . The shifted
momentum on the internal line is
i − 1, î, j − 1, j
P̂I2 = ( p̂i + pi+1 + · · · + p j−1 )2 = ( ŷi − y j )2 = ŷi2j = . (7.14)
i − 1, i j − 1, j
In the last step, we have written the distance in dual y-space in terms of the 4-bracket, as in
(5.52), and we have also used that i − 1, î = i − 1, i. The shift Ẑ i = Z i + w Z i+1 says
that the point Ẑ i lies on the line (i, i + 1) and its position on that line is parameterized by w.
The condition P̂I2 = 0 is the statement that w is chosen such that i − 1, î, j − 1, j = 0,
so this value w∗ is such that the point Ẑ i lies in the plane (i − 1, j − 1, j). In other words,
Ẑ i can be characterized as the point of intersection between the line (i, i + 1) and the plane
(i − 1, j − 1, j), viz.
(
( î ) = (i, i + 1) (i − 1, j − 1, j) . (7.15)
The intersection formula was given in (5.66) in terms of the 4-brackets. The geometry is
illustrated in Figure 7.3(a).
We now determine the momentum twistor Z I associated with the internal line P̂I . Take
a look at the BCFW diagram in (7.13). The point y j in dual space can be determined by
the line ( j − 1, j) in momentum twistor space. But by inspection of (7.13), y j can also
be determined by the line (I, j). This means that the three points Z I , Z j−1 , and Z j lie on
the same line. Similarly, the point ŷi can be determined by the line (i − 1, î) or by the line
Zi+1 Zi+1
Zi Zj–1
Zi
Zj–1
Zi–1 Zi–1 ZI
Zj ^ Zj
Z i
^
Z i
(a) (b)
Fig. 7.3 The geometry of (7.15) and (7.16): (a) shows that the on-shell condition = 0 fixes the shifted momentum twistor
P̂I2
Ẑ i to be at the intersection of the line (i, i + 1) with the plane (i − 1, j, j − 1). In (b) the momentum twistor
Z I is located at the intersection of line ( j, j − 1) and the plane (i − 1, i, i + 1).
(I, î), so Z I , Z i−1 , and Ẑ i lie on the same line. Since Z I lies on both lines, we conclude
that Z I can be characterized as the intersection point of the lines (i − 1, î) and ( j − 1, j).
We previously learned that Ẑ i lies in the plane (i − 1, j − 1, j), and therefore the plane
contains the line (i − 1, î). Thus we conclude that Z I is the point where the line ( j − 1, j)
intersects the plane (i − 1, i, i + 1):
(
( I ) = ( j, j − 1) (i − 1, i, i + 1) . (7.16)
The geometry is shown in Figure 7.3(b). These results will be useful in the following.
Now we are ready to study the BCFW recursion relations in momentum twistor space.
Consider the BCFW shift (7.5) of an n-point L-loop integrand InL (for tree-level you can
translate “integrand” to “superamplitude”
in your head). The recursion relations are based
on the contour argument for dw w
Î L
n (w), just as in Section 3.1. Poles at finite values of w
are equivalent, via (7.8), to poles at finite z: they arise from propagators with momentum
ŷi j going on-shell and the corresponding BCFW diagrams are those in (7.13). However, in
completing the contour integral argument we also need to consider the large-w limit. It is
clear from the relation (7.8) that z goes to a finite value z ∗ as w → ∞. Specifically,
w→∞ i, i + 1
z −−−→ z ∗ ≡ . (7.17)
i − 1, i + 1
Thus the pole at infinity in the w-plane maps to a finite point in the z-plane and we will
have to consider this contribution too. At w = ∞, the shifted angle spinors are
i − 1, i

|î = |i + 1 . (7.18)
w→∞ i − 1, i + 1
Thus, in the limit w → ∞, the spinors |î and |i + 1 become proportional, and that
implies that P̂i,i+1 = p̂i + pi+1 is on-shell: P̂i,i+1
2
= î, i + 1[i, i + 1] → 0 for w → ∞.
Or equivalently, we may note that z ∗ is exactly the solution to î, i + 1 = 0. Hence the pole
at w = ∞ corresponds to a factorization channel of an integrand into a 3-point anti-MHV
part – the only possibility that can support the special kinematics |î ∝ |i + 1 ∝ | P̂i,i+1 –
and the remainder L-loop integrand:
î i−1
i−1
î
w ∞ ^y . (7.19)
i
MHV yi+1
L L
An(w) An–1(w)
yi+2 i+1
1
1 n n i+2
Note that the lines in this figure are labeled by momentum supertwistors, so we place a hat
just on line i since that is the only shifted momentum twistor.
Exercise 7.4
Show that for w → ∞, the kinematics give
i, i + 1
|i
− 1 = |i − 1 , |i
− 1] = |i − 1] − |i] ,
i + 1, i − 1

i, i − 1
| P̂I = |i + 1 , | P̂I ] = − |i] + |i + 1] . (7.20)
i + 1, i − 1
To summarize, BCFW recursion relations in momentum twistor space express the integrand
(superamplitude) as a boundary contribution B∞ from w = ∞ plus a sum of residues at
finite w. Schematically, we have

Î L (w)
InL = ÎnL (w = 0) = B∞
L
− Residues of n at finite w∗ . (7.21)
w∗ =0
w
The boundary term B∞ is computable and is given by the diagram (7.19). The rest of the
residues come from diagrams such as (7.13).
To become familiar with how this works in practice, we first apply the recursion relations
to tree-level superamplitudes before moving on to loop-integrands in Section 7.4.
7.3 Momentum twistor BCFW at tree-level
In Section 4.5.2, we used the super-BCFW recursion relations to show that the NMHV
tree-level superamplitude of N = 4 SYM can be written as the MHV superamplitude times
a sum of the dual superconformal invariants Ri jk ; see (4.88). We then rewrote the NMHV
formula in terms of momentum twistors in Section 5.4 and found (5.60),

n−3
n−1

ANMHV
n = AMHV
n n, j − 1, j, k − 1, k , (7.22)
j=2 k= j+2
where the 5-brackets [n, j − 1, j, k − 1, k] ≡ Rn jk are invariant under cyclic permutations

of the five labels. It was claimed then that the N K MHV tree superamplitudes took a similar
form but with the sum involving products of K 5-brackets. We now prove this statement
using the momentum twistor formulation of super-BCFW. This also serves to prove that
the tree superamplitudes of N = 4 SYM are dual superconformal covariant.
Adapted to tree-level, the Ẑ i -shift BCFW relation (7.21) reads

i+2
Ân (w)
An = B∞ − Residues of at ŷi j = 0 .
2
(7.23)
j=i−3
w
We begin with a detailed evaluation of the boundary term.
The boundary term B∞ . Per definition, the boundary contribution is the residue of the
pole at infinity,2
*
dw
B∞ = − Ân (w) . (7.24)
C∞ w
Here C∞ is a contour that surrounds w = ∞ counterclockwise. Since we are more familiar

with the shift in momentum space, let us change variables from w to z. With the help of
(7.8) we find
*
z∗
B∞ = dz Ân (z) , (7.25)
Cz ∗ z(z − z ∗ )
i,i+1
where z ∗ = i−1,i+1 is the value of z at w = ∞. Now we need to find out how the shifted
n-point amplitude behaves for z near z ∗ . We have already established in (7.19) that the
N K MHV superamplitude factorizes as N K MHVn−1 ×anti-MHV3 at z = z ∗ . Let us focus
on the MHV case (K = 0) first to see explicitly how this comes about. Under the [i −1, i-
supershift, the Grassmann delta function in the MHV superamplitude is inert, and the only
part of the amplitude affected by the shift is the denominator factor î, i + 1. This exactly
is the factorization pole for z → z ∗ . Therefore, near z ∗ we can write
z→z ∗ 1 anti-MHV PI2 MHV 1 anti-MHV
ÂMHV
n (z) −−−→ ÂMHVn−1 (z ∗ ) Â 3 (z ∗ ) = Â n−1 (z ∗ ) Â 3 (z ∗ ) .
P̂I2 P̂I2 PI2
(7.26)
We know from super-BCFWing the MHV superamplitude in Section 4.5.1 that the
factor [. . .] in (7.26) equals AMHV
n (remember, for MHV only one diagram contributed
in the recursion relations based on a BCFW shift of adjacent lines). The prefactor is
PI2
P̂ 2
= −z ∗ /(z − z ∗ ). Thus
I
*
z ∗2
B∞
MHV
= −AMHV dz = AMHV . (7.27)
n
Cz ∗ z(z − z ∗ )2 n
2 We ignore the 2πi of the Cauchy theorem since all such factors drop out at the end.
In the second equality, we evaluated the double pole integral using

* *
f (z) d f (z)
dz = dz = f (z ∗ ) (7.28)
(z − z ∗ ) 2 dz ∗ (z − z ∗ )
for the case f (z) = 1/z.
What we have achieved for the MHV case here is a verification of the simple statement
that the MHV tree-level superamplitude satisfies the super-BCFW recursion relations,
which for MHV include only one term, namely the MHVn−1 ×anti-MHV3 diagram. We
knew that already more than 50 pages ago (Section 4.5.1), but the point is that here we have
set up the calculation in a way that facilitates the generalization to N K MHV level. And that
is what we do next.
N K MHV case. Assume inductively that the (n − 1)-point tree-level N K MHV superampli-
(K )
tude can be written as an MHV prefactor times a function we will call Yn−1 . We know that
(0) (1) (0) (1)
Yn = 1 and that Yn is the sum of 5-brackets given in (7.22). In particular, Yn and Yn
are dual superconformal invariants. The aim here will be to use BCFW to derive a recursive
(K ) (K )
expression for Yn in terms of Y s with smaller K and n. This will prove that Yn is a dual
superconformal invariant for all n and k.
The calculation of the contribution from w = ∞ follows the same steps as the MHV
case, except that the factorization (7.26) is now replaced by
z→z ∗
+(K ) 1 anti-MHV
ÂMHV
n (z) −−−→ ÂMHV
n−1 (z ∗ ) Yn−1 (z ∗ ) Â3 (z ∗ ) . (7.29)
P̂I2
Let us take a closer look at the Y -factor. It is naturally a function of momentum supertwistors
Y +(K ) (. . . , Zi−1 , Z I , Zi+2 , . . .) .
+(K ) (z ∗ ) = Y (7.30)
n−1 n−1
Now in our analysis of the kinematics, we learned that the momentum twistor Z I is
characterized as the intersection (7.16) between the line ( j − 1, j) and the plane (i −
1, i, i + 1). In our case here, we have j = i + 2, so (7.16) says that Z I is the point of
intersection between the line (i + 1, i + 2) and the plane (i − 1, i, i + 1). Obviously this
intersection point is Zi+1 , i.e.
(
For w = ∞ case: ( I ) = (i + 1, i + 2) (i − 1, i, i + 1) = (i + 1) . (7.31)
So we can freely substitute Z I → Zi+1 to find
+(K ) (z ∗ ) = Y (K ) (. . . , Zi−1 , Zi+1 , Zi+2 , . . .) .
Y (7.32)
n−1 n−1
This factor is independent of z and we can therefore repeat our argument from the MHV
case to find
*
N K MHV dw N K MHV (K )
B∞ = Ân (w) = AMHV Yn−1 (Z1 , . . . , Zi−1 , Zi+1 , . . . Zn ) . (7.33)
C∞ w n
This completes the calculation of the boundary term.
Residues at finite w. Now we extract the residues of the finite poles in the w-plane. They
arise from propagators 1/ ŷi2j going on-shell. Writing the shifted propagator in terms of
momentum twistor, we find
1 î, i − 1 j, j − 1 i, i − 1 j, j − 1

= =
2
ŷi j î, i − 1, j, j − 1 i, i − 1, j, j − 1 + wi + 1, i − 1, j, j − 1
w∗ 1
= 2 , (7.34)
yi j w − w∗
i,i−1, j, j−1
where w∗ = − i+1,i−1, j, j−1
. This means that

dw 1 dw w∗ 1 1
− f (w) = − f (w) = 2 f (w∗ ) . (7.35)
C(w∗ ) w ŷi j C(w∗ ) w yi j w − w∗
2 2
yi j
So the contribution from the shifted propagator is simply the unshifted propagator, exactly
the same as the usual BCFW rules. In a given factorization channel, we can write the left
and the right subamplitudes as:
An L = AMHV
nL Yn(KL L ) (Z I , Z j−1 , . . . , Zi+1 , Zî ),
(7.36)
An R = AMHV
nR Yn(KR R ) (Z I , Zi−1 , . . . , Z j+1 , Z j ) .
Here the Grassmann degrees obey K R + K L + 1 = K ; in particular this is why there were
no such diagrams for the MHV case. The contribution of the BCFW channel is simply

ÂMHV ÂMHV
nL nR
Y +n(K R ) (Z I , Zi−1 , . . . , Z j+1 , Z j ) .
+n(K L ) (Z I , Z j−1 , . . . , Zi+1 , Ẑi ) Y
2 L R
states
PI
(7.37)
The Z I appearing in the dual superconformal invariants YL,R can be written in terms of the
external line supermomentum twistors using the characterization of Z I as an intersection
point (7.16) and the formula (5.66), similarly for Ẑi , via the intersection rule (7.15).
As a consequence, the state sum – which is an integration over the η I variables – acts
solely on the MHV prefactors. Furthermore, the factor in parenthesis is in (7.37) simply the
BCFW term of the NMHV amplitude arising from the PI factorization channel. We have
calculated this in Section 4.5.2, and later learned that in the momentum twistor language
the answer is written in terms of the 5-bracket:

ÂMHV
nL ÂMHV
nR
2
= AMHV
n i − 1, i, i + 1, j − 1, j . (7.38)
states
PI
Thus we have finally arrived at the BCFW recursion relation for tree-level amplitudes in
N = 4 SYM, written in momentum twistor space:
!
K (K )
i−2

AN
n
MHV
= AMHV
n Yn−1 (. . . , Zi−1 , Zi+1 , Zi+2 , . . .) + i − 1, i, i + 1, j − 1, j
j=i+3 "
+ +
× Yn L (Z I , Z j , Z j+1 , . . . , Zi−1 ) × Yn R (Z I , Ẑi , Zi+1 , . . . , Z j−1 ) .
(K L ) (K R )
(7.39)
The above relation corresponds to the shift defined in (7.5), and the momentum twistors
Ẑi and Z I are given by (7.15) and (7.16) respectively. Also, there are implicit sums over
K R + K L + 1 = K and n L + n R = n + 2.
Exercise 7.5
Show that (7.39) gives the known NMHV expression (7.22) for K = 1.
The result (7.39) verifies the claim that all tree-level amplitudes of N = 4 SYM can be
written as an MHV prefactor times polynomials of 5-brackets: given that this is true for
the NMHV amplitudes, (7.39) ensures that the 5-brackets are recycled into the higher-
K results. Since the 5-brackets are manifestly dual superconformal invariant, so are all
tree-level superamplitudes of N = 4 SYM.
7.4 Momentum twistor BCFW for planar loop-integrands
To initiate the discussion of BCFW recursion for planar loop-integrands of N = 4 SYM,

let us examine a specific example to get some intuition for the good looks and behaviors of
integrands. In other words, we start with the answer and let that guide our discussion.
In Section 6.2, we used the generalized unitarity method to construct the 1-loop N = 4
SYM superamplitude. We found
1-loop
4 [1234] I4 ( p1 , p2 , p3 , p4 ) , (7.40)
where I4 is the 1-loop box-integral which we wrote in dual y-space in (7.4) as

d 4 y0
I 4 ( p1 , p2 , p3 , p4 ) = 2 2 2 2
. (7.41)
y01 y02 y03 y04
Here the propagator-terms y0i2 = (y0 − yi )2 involve the zone-variable y0 associated with
the loop-momentum, as indicated in (7.3).
The expressions (7.40), (7.41) determine the loop-integrand for the 4-point 1-loop N = 4
SYM superamplitude to be (using −s = y13 2
and −u = y242
)
2 2
1-loop y13 y24
I4 [1234] = Atree
4 [1234] 2 2 2 2
. (7.42)
y01 y02 y03 y04
Now, to translate this to momentum twistor space, recall that a point y in dual space maps
to a line in momentum twistor space. So let us take y0 to be mapped
4 to some line (A, B)
determined by two points Z A and Z B ; then the loop-integral d y0 maps to an integral
over all inequivalent lines ( A, B). There is a story here of how to define the integration
measure appropriately – we postpone this until later in this section in order to first discuss
the structure of the loop-integrands.
Using (5.52) to rewrite all the dual variables y in the integrand in terms of 4-brackets, in
A,B,i−1,i
particular y0i2 = ABi−1,i , we arrive at the expression
1-loop 12342 AB4

I4 [1234] = −Atree
4 [1234] . (7.43)
AB12AB23AB34AB41
Note that all i − 1, is dropped out. The factor AB4 will eventually be absorbed in the
integration measure and what remains is manifestly dual conformal invariant.
The expression (7.43) is an example of what a 1-loop integrand looks like in momentum
twistor space. Under a super-BCFW shift Ẑ4 = Z4 + wZ3 , the integrand has a pole that
involves the loop-momentum: it comes from AB 4̂1 = 0. The residue of such a pole is the
new input we need for the loop-level recursion relations.
Exercise 7.6
What type of super-BCFW shift is induced in momentum space by the momentum
twistor shift Ẑ4 = Z4 + wZ3 ?
Next, we outline the form of the recursion relations for general L-loop integrands.
Structure of the BCFW recursion for loop-integrands

Without loss of generality, we consider the recursion relations derived from a shift of the
nth momentum twistor,
Ẑn = Zn + wZn−1 . (7.44)
For an n-point L-loop integrand, there will be three distinct types of contributions to the
recursion relations:
1. The boundary contribution from w → ∞. This contribution is calculated just as in the

tree-level case of the previous section, so we simply state the result (suppressing the K
of the N K MHV classification)

term at w → ∞: Atree n,MHV Yn−1 Z1 , . . . , Zn−1 .
L
(7.45)
L
Here Yn−1 is, by the inductive assumption, an L-loop dual superconformal invariant.
2. Residues of factorization channels from propagators that do not involve loop-momenta
correspond to poles in ŷ12 j ∝ n̂, 1, j − 1, j = 0. The results for these also follow the
same steps as the tree-level case, and one finds

n−2

Atree
n,MHV j − 1, j, n − 1, n, 1 YLL 1 Z I j , Z j , Z j+1 , . . . , Ẑn j YRL 2
j=3

× Z I j , Z1 , Z2 , . . . , Z j−1 , (7.46)
) )
where Ẑn j = (n − 1, n) (1, j − 1, j) and Z I j = ( j, j − 1) (n − 1, n, 1). This in-
cludes an implicit sum over loop-orders L 1 and L 2 in the sub-integrands such that
L 1 + L 2 = L. Also, the N K MHV level was suppressed, and one must sum over the
Grassmann degrees associated with the sub-integrands such that K L + K R = K − 1.
y1 y1
y5
p6 p1 p1
p5 p6
y5 p5
y6
y2 p4 y6 y2
p4
p2 p2
y4 y4
p3 p3
y3 y3
The forward limit illustrated in dual coordinates. The limit corresponds to y5 and y1 approaching a point while satisfying Fig. 7.4
2
y14 = 0.
3. Residues of factorization channels from propagators that do involve loop-momenta; they

correspond to
A B n̂ 1 = 0 . (7.47)
These are the new forward limit contributions at loop-level, so we will take a closer look
at them now.
Forward limit contributions

For an L-loop n-point integrand, the residue of the pole (7.47) is an (L −1)-loop (n+2)-
point integrand whose two extra legs are evaluated in the forward limit (7.1), as shown in
Figure 7.1. The following example of n = 4 will illustrate the idea of how to do this.
Example. Start with a 6-point (L − 1)-loop integrand. Translated to dual coordinates,

the forward limit of p5 and p6 approaching p5 = − p6 = r is the limit of taking
y1 → y5 while y6 remains fixed. This is illustrated in Figure 7.4. In momentum twistor
space, the point y1 is determined by the line (1, 6) and y5 by (4, 5), so y1 and y5 can
be identified only when (Z 1 , Z 6 , Z 5 , Z 4 ) lie on the same line. It is easy to achieve this
configuration if the line (1, 4) intersects line (6, 5), because then we can send Z 5 and
)
Z 6 to the intersection point (6, 5) (1, 4). Note that this does not change y6 , but the
result is y1 → y5 .
However, momentum twistors live in CP3 where two lines generically do not in-
tersect. So we cannot take the limit as naively as above. Instead, we modify the
momentum twistor Z 4 → Ẑ 4 = Z 4 + w Z 3 , and tune w such that the new line (1, 4̂)
intersects (5, 6): let Z B̂ be the point of intersection. Since Ẑ 4 per construction lies on
the line (3, 4), we can characterize Z B̂ as the intersection point between the line (5, 6)
and the plane (3, 4, 1) (see Figure 7.5):
(
Z B̂ = (5, 6) (3, 4, 1) . (7.48)
Because the lines (1, 4̂) and (5, 6) are arranged to intersect, it follows that Ẑ 4 lies
in the plane (5, 6, 1), see Figure 7.5. But Ẑ 4 is also on the line (3, 4), so the shifted
(5,6)
Z4^ = (3,4) ^
(1,B)
Z5
ZB
^ = (5,6) (3,4,1)
Z4 Z1 Z6
Z3
(3,4,1)
Fig. 7.5 The geometry of the forward limit Figure 7.4 illustrated here in momentum twistor space. The point Z B̂ is defined as
the intersection of line (Z 5 , Z 6 ) with plane (Z 1 , Z 3 , Z 4 ). The BCFW-deformed Ẑ 4 is fixed at the interaction of
line (Z 3 , Z 4 ) and (Z B̂ , Z 1 ). The gray points lie in the same plane.
momentum twistor can be identified in terms of the unshifted lines as

(
Ẑ 4 = (3, 4) (5, 6, 1) . (7.49)
The setup with (7.48) and (7.49) allows us to take the forward limit directly by sending
Z 5 , Z 6 to the intersection point Z B̂ . We can summarize the deformation and forward
limit p5 = − p6 = r as

(Z 1 , Z 2 , Z 3 , Z 4 , Z 5 , Z 6 ) → (Z 1 , Z 2 , Z 3 , Ẑ 4 , Z 5 , Z 6 ) . (7.50)
Z 5 ,Z 6 →Z B̂
It is important to note that Ẑ 4 satisfies
5, 6, 4̂, 1 = 0 . (7.51)
Comparing (7.51) with (7.47), we recognize the single-cut condition (or equivalently,
momentum dependent pole in BCFW) provided that Z 5 and Z 6 are identified as the
loop-momentum twistors Z A and Z B . This is also the statement that y6 has been
identified as the Lth loop integration region, as Figure 7.4 indicates that it should
be.
Let us return to the general case. The forward limit is taken for (L − 1)-loop (n + 2)-point
integrands by sending the momentum twistors ( Ẑ n , Z A , Z B ) to the forward configuration
discussed in the example. One must multiply by an overall MHV factor as well as the result
for the cut propagator. The result (which we discuss further below) is

Atree × f (A, B, n − 1, n, 1) × Y L−1
[Z 1 , Z 2 , . . . , Ẑ n AB , Z A , Z ]
B
, (7.52)
n,MHV n+2
A,B→ B̂
where adapting (7.48) and (7.49) to the n-point case with 5 → A and 6 → B identifies
( (
Ẑn AB = (n − 1, n) (A, B, 1) and B̂ = (A, B) (n − 1, n, 1) . (7.53)
Here, f (A, B, n − 1, n, 1) represents the kinematic function which includes the cut prop-
agator as well as possible Jacobian factors that arise from solving the single cut constraint,
A, B, n̂, 1 = 0. We will determine this function shortly, but first we address one impor-
tant missing piece: what to do about the Z A and Z B momentum twistors and how the extra
loop-momenta integral emerges from the forward limit. That is the next step.
The integration measure

The forward limit is taken of a higher point amplitude/integrand, so we need to devise a
way to remove the information of the two extra external legs. The most naive proposal is to
apply the following integration:

d 4|4 Z A d 4|4 Z B . (7.54)
Surprisingly, this is essentially the correct answer! The reason it is correct is rather non-
trivial. Let us first consider the bosonic part of the integration. The integration over Z A
and Z B can be decomposed into two pieces, one is the integration over all possible lines
(A, B), and the other is the movement of Z A and Z B along a particular line ( A, B). To aid
this separation, consider the following G L(2) transformation on (Z A , Z B ),
A
Z A c c A B ZA
= AA B . (7.55)
Z B cB cB ZB
The above 2 × 2 matrix exactly parameterizes the movement along the line ( A, B), because
the new pair (A , B ) defines the same line as (A, B). In the forward limit, we are sending
Z A and Z B on a given line to the intersection of (A, B) with the plane (n − 1, n, 1), so this
limit corresponds to a particular solution for the G L(2) matrix. In light of this, it will be
convenient to separate the bosonic integral as
4
d Z Ad4 Z B
d4 Z Ad4 Z B = . (7.56)
Vol[G L(2)] G L(2)
The Vol[G L(2)] in the denominator indicates that as one integrates over the 4 × 2 di-
mensional space of Z A and Z B , one needs to mod out the 2 × 2 c-matrix in (7.55) that
parameterizes an arbitrary G L(2) transformation. The explicit integration measure for
G L(2) can be fixed by requiring it to be S L(2) invariant and having G L(1) weight 4 in both
A and B.3 This fixes the form to be

= c A dc A c B dc B c A c B 2 , (7.57)
G L(2)
where c A c B = c A c B − c A B c B A . After one has separated out the G L(2) integral, the

A
B
remaining integration measure can be naturally related to the y0 measure. To see this, note
that after stripping off the G L(2) part, the remaining measure is purely integrating over
all distinct lines ( A, B). Recall that distinct lines in twistor space define distinct points in
dual space, so this tells us that this measure is precisely proportional to d 4 y0 . The precise
3 This follows from the fact that there are four-components in Z A and Z B .
momentum twistor integral that is equivalent to the loop-integral over y0 is

d4 Z Ad4 Z B
d 4 y0 = , (7.58)
Vol[G L(2)]AB4
where the four extra factors of AB in the denominator are necessary for the measure to
be projective. The angle bracket AB breaks the S L(4) invariance because it picks only
the angle spinor piece of the momentum twistors. This breaks dual conformal invariance –
but that is expected because the d 4 y0 inverts non-trivially under dual conformal inversion.
From our example (7.43), we see that the AB4 -factor in the measure is exactly canceled
by the AB4 -factor that appeared when we rewrote the box-integral in momentum twistor
space. This is a general feature which follows from (or, if you prefer, is necessary for) the
dual conformal invariance of the loop-integrand. Henceforth, we simply implicitly assume
the cancellation of the AB4 s. Let us for later reference write what the 1-loop 4-point
superamplitude looks like when dressed in full momentum twistor regalia:

1-loop d4 Z Ad4 Z B 12342
A4 [1234] = −Atree
4 [1234] . (7.59)
Vol[G L(2)] AB12AB23AB34AB41
Back to the forward-limit discussion. To integrate over all possible configurations of the
forward limit, one should only integrate over all distinct lines ( A, B). This requires us to
remove the G L(2)-part of the integration in (7.54). Thus we have two problems to solve,
how to put the higher-point amplitude on the forward limit and how to remove the G L(2)
redundancy. Fortunately, we can scare two birds with one stone!4 We begin by simply
presenting the resolution: the correct prescription for the computation of the forward limit
is
4|4
d Z A d 4|4 Z B
Atree
n,MHV [A, B, n − 1, n, 1] × Yn+2
L−1
[Z1 , Z2 , . . . , Ẑn AB , Z A , Z B̂ ] ,
Vol[G L(2)] GL(2)
(7.60)
)
where Z B is sent to the intersection B̂ = (A, B) (n − 1, n, 1). Note the appearance of the
factor [A, B, n − 1, n, 1]. The role this factor plays is two-fold:
• It contains the invariants A, n − 1, n, 1 and B, n − 1, n, 1 in the denominator. The

vanishing of these invariants is precisely the forward limit, and therefore these poles can
be used to localize the G L(2) integral on to the forward limit. (The two birds fly away.)
• It also contains the factor 1/A, B, n, 1, which is precisely the single-cut propagator.
Hence the G L(2) integration is understood to encircle poles that correspond to the forward
limit. One may ask if [A, B, n − 1, n, 1] is the unique function that satisfies the above two
points. The answer is no, however, it can be easily justified by dual conformal invariance.
L−1
The recursion better preserves this symmetry. With Yn+2 already an invariant, [A, B,
n − 1, n, 1] is the unique invariant that satisfies the above two properties. Thus, using
4 No need for excessive aggression.

symmetry arguments we did not need to know a priori what the function f (A, B, n −
1, n, 1) in (7.52) should be; it is whatever [A, B, n − 1, n, 1] evaluates to once the G L(2)
integral is localized. This is admittedly rather abstract, but we are going to realize the
contents of the discussion here explicitly when we compute the 4-point 1-loop amplitude in
Section 7.5.
Finally, we need to sum over all N = 4 SYM states that can run in the forward limit
loop. In (7.60) this is naturally achieved in a way that preserves the dual superconformal
symmetry by simply extending the bosonic momentum twistor integration to include the
Grassmann-components, χ A and χ B . We are now ready to put everything together.
Result of the BCFW recursion for L-loop integrands

Summarizing the preceding discussion, the loop-level BCFW recursion relation is given
by
AnL-loop
!

= Atree
n,MHV
L
Yn−1 Z1 , . . . , Zn−1

n−2

+ j − 1, j, n − 1, n, 1 YLL 1 Z I j , Z j , Z j+1 , . . ., Ẑn j YRL 2 Z I j , Z1 , Z2 , . . ., Z j−1
j=3
"
d 4|4 Z A d 4|4 Z B
+ [A, B, n − 1, n, 1] L−1
Yn+2 [Z1 , Z2 , . . . , Ẑn AB , Z A , Z B̂ ] .
Vol[G L(2)] G L(2)
(7.61)
In the second line, L 1 and L 2 are summed over subject to L 1 + L 2 = L, as are the Grassmann
degrees K 1 + K 2 = K − 1, and we have
( (
Ẑn j = (n − 1, n) (1, j − 1, j) , Z I j = ( j, j − 1) (n − 1, n, 1) ,
(
Ẑn AB = (n − 1, n) (A, B, 1) . (7.62)
In the last line of (7.61), the Grassmann degree of Yn+2 L−1

is 4(K + 1), i.e. one level higher
than the amplitude on the LHS of the equation.
Before moving on to an explicit application of the loop-integrand recursion relations, it is
important to note that (7.61) provides us with the tool to prove dual conformal properties of
loop-integrals, just as we did for the tree-level recursion. Assuming the n-point L < L-loop
as well as the (n+2)-point (L −1)-loop integral is given by an MHV tree amplitude times
a dual conformal invariant function, then – through (7.61) – the n-point L-loop amplitude
will have the same property.
We will now apply the recursion relations developed in this section to show how the
4-point 1-loop integrand (7.43), can be derived recursively from the recursion relation with
the input of just a tree amplitude. Sharpen your pencils and keep your eraser close at
hand.
7.5 Example: 4-point 1-loop amplitude from recursion
The 4-point 1-loop amplitude is the simplest example that can illustrate all the novelties of
the loop-recursion. Let us examine the potential terms in the recursion formula (7.61):
The first term with Y3L=1 is absent. This is because there are no 3-point 1-loop amplitudes.
Another way to understand this is that this contribution came from the pole at w → ∞. If
we sneak-peak at the answer for the 4-point 1-loop amplitude (7.43), we realize that while
the shifted MHV prefactor does have a w → ∞ pole (as we saw and used in Section 7.3),
its residue is actually zero for the 1-loop integrand because the 1/AB 4̂1 → 0 as w → ∞.
In other words, we have a consistent picture for the absence of the first term Y3L=1 in (7.61).
The second term in (7.61) is absent, again because Y3L=1 = 0. This is consistent with
(7.43) not having any momentum-independent poles at finite w.
The third term in (7.61) is
4|4
1-loop d Z A d 4|4 Z B
A4 = Atree
4,MHV [A, B, 3, 4, 1] × Y6 [Z1 , Z2 , Z3 , Ẑ4 AB , Z A , Z B̂ ] .
Vol[G L(2)] G L(2)
(7.63)
Unfortunately, we now have to evaluate this thing!!
Y6 is the tree-level NMHV dual conformal invariant for n = 6, discussed previously in
(5.60):
Y6 [Z1 , Z2 , Z3 , Ẑ4 AB , Z A , Z B̂ ] = [ B̂, 1, 2, 3, 4̂] + [ B̂, 1, 2, 4̂, A] + [ B̂, 2, 3, 4̂, A] .

(7.64)
The hatted momentum twistors can be found explicitly using the intersection formulas
(5.66). Since the twistors are defined projectively, one can freely include a scaling-factor:
( 1

(4̂) = (3, 4) (A, B, 1) =⇒ Z4̂ = Z4 3AB1 − Z3 4AB1 , (7.65)
3AB1
( 1

( B̂) = (A, B) (3, 4, 1) =⇒ Z B̂ = − Z A B341 + Z B A341 . (7.66)
A341
For convenience, we pick overall factors such that the “hatted” twistors have the same
projective weights as the un-hatted ones. Note that some 4-brackets remain unshifted:
3, 4̂, . , . = 3, 4, . , . and A, B̂, . , . = A, B, . , ..
Let us first do the fermionic integrals in d 4|4 Z A d 4|4 Z B , i.e. d 4 χ A d 4 χ B . We have to
saturate the Grassmann integrals, so it is only relevant to look at the χ A - and χ B -terms in
the Grassmann delta functions. Begin with the 5-bracket [ A, B, 3, 4, 1] that multiplies each
of the three terms in Y6 . Its Grassmann delta function involves

[A, B, 3, 4, 1] ∝ δ (4) χ A B341 − χ B A341 + · · · . (7.67)
It follows from (7.66) that χ B̂ ∝ χ A B341 − χ B A341, so any appearance of χ B̂ in the

three 5-brackets in (7.64) vanishes on the support of the δ (4) in (7.67) under the d 4 χ A d 4 χ B -
integral. In particular, the only contribution from χ A , χ B in [ B̂, 1, 2, 3, 4̂] is through χ B̂ , so
7.5 Example: 4-point 1-loop amplitude from recursion 157
we immediately conclude that

d 4 χ A d 4 χ B [A, B, 3, 4, 1] × [ B̂, 1, 2, 3, 4̂] = 0 . (7.68)
In the next case, [ B̂, 1, 2, 4̂, A], we have δ (4) (χ B̂ 124̂ A + χ A B̂124̂ + · · · ). As before the
χ B̂ -term can be dropped. Moreover, one can show that B̂124̂ vanishes (see Exercise 7.7
below), so we conclude

d 4 χ A d 4 χ B [A, B, 3, 4, 1] × [ B̂, 1, 2, 4̂, A] = 0 . (7.69)
Exercise 7.7
The 3-term Schouten identity (2.48) for angle and square spinors is the statement that
three vectors in a plane are linearly dependent. As 4-component objects, the momentum
twistors Z I , I = (ȧ, a), similarly satisfy a 5-term Schouten identity,
i, j, k, lZ m + j, k, l, mZ i + k, l, m, iZ j + l, m, i, jZ k + m, i, j, kZ l = 0 .

(7.70)
Use (7.70) to derive the two identities
123434AB
B̂124̂ = 0 , 234 B̂ = − . (7.71)
A341
Hint: AB 4̂1 = 0; this identity follows from (7.65).
With the help of the second identity in (7.71), one finds that the result of integrating the
δ (4) s in [A, B, 3, 4, 1] × [ B̂, 2, 3, 4̂, A] gives 12344 34AB4 .
In conclusion, after Grassmann integration, only the third 5-bracket in (7.64) contributes.
After some simplifications one finds

1234AB34
d 4 χ A d 4 χ B [A, B, 3, 4, 1] × [ B̂, 2, 3, 4̂, A] = × I4 (A, B) , (7.72)
A234B341
where
12342
I4 (A, B) = . (7.73)
AB12AB23AB34AB41
Exercise 7.8
Derive (7.72).
I4 (A, B) is the answer we expect, cf. (7.59), so now we have to massage its prefactor in
(7.72). The recursion relations (7.63) instruct us to perform the forward limit by evaluating
the G L(2) integral:
4
1-loop d Z Ad4 Z B 1234AB34
A4 = A4,MHV
tree
I4 (A, B) × . (7.74)
Vol[G L(2)] G L(2) A234B341
This can be done by first making a G L(2) rotation (7.55) of Z A , Z B and then integrating
over the G L(2) parameters. Since the integral is G L(1) invariant, we can fix the scale in
the G L(2) matrix and set c A A = c B B = 1. With this “gauge fixing,” we have

ZA 1 c A ZA
→ . (7.75)
ZB cB 1 ZB
The result of this transformation on the 4-brackets is
Ai jk → Ai jk + c A Bi jk ,
ABi j → ABi jc A c B , (7.76)
Bi jk → c B Ai jk + Bi jk ,
and the gauge fixing means that c A c B = 1 − c A c B . Also, c A dc A = dc A and
c B dc B = dc B . So including the appropriate G L(2) measure (7.57), we have
4
1-loop d Z Ad4 Z B
A4 = A4,MHV
tree
I4 (A, B) 1234AB34
Vol[G L(2)]

dc A dc B
× . (7.77)
1 − c A c B A234 + c A B234 c B A341 + B341
Now the plan all along was that the G L(2) integration was supposed to localize us on the
forward limit. So consider the denominator factor (c B A341 + B341). The vanishing
of this expression is the statement that Ẑ B = Z B + c B Z A is sent to the intersection point
of the line (A, B) and the plane (3, 4, 1): but this is exactly part of the forward limit
)
Ẑ B → Z B̂ = (A, B) (3, 4, 1). So to realize this, we take the contour (7.77) in the c B -
plane to surround the pole c B ∗ = −B341/A341. Now we also want to send A to the
intersection point B̂, but the integral (7.77) appears not to have a pole that achieves this.
However, when we evaluate the c B -integral to localize B → B̂, the factor (1 − c A c B )
actually develops the desired pole, namely c A ∗ = −A341/B341. Just do it:

1
dc A dc B
C(c A ∗ ) C(c B ∗ ) 1 − c A c B A234 + c A B234 c B A341 + B341

dc A
=
C(c A ∗ ) A341 + c A B341 A234 + c A B234
1
=
A234B341 + A341 B234
1
=− . (7.78)
AB341234
In the last line we used the 5-term Schouten identity (7.70). Plugging this result into (7.77),
the factors AB341234 cancel and we are left with
4
1-loop d Z Ad4 Z B
A4 = −Atree
4,MHV I4 (A, B) . (7.79)
Vol[G L(2)]
This is the correct result, as we discussed in Section 7.4.
7.6 Higher loops 159
Note that in this derivation, the G L(2) integral ended up localizing the integrand on the
forward limit, where the poles that were used in this localization were given by the extra
[A, B, 3, 4, 1]. This precisely realizes the idea we described around (7.60). Now you might
be a little concerned that we could have chosen to localize on “non-forward” poles in the
G L(2) integral instead, but the answer would have been the same, as guaranteed by the
large-c A ,B falloff of the integrand.
Finally, you may find it discouraging that it takes much more work and sophistication to
calculate even the simplest of all N = 4 SYM amplitudes with BCFW than it did with the
generalized unitarity method, as we showed in Section 6.2. However, while it is not directly
practical, it is morally encouraging – and perhaps even fascinating – that all information
about the 4-point MHV 1-loop amplitude is encoded already in the 6-point NMHV tree-
level amplitude. This is a realization of an interesting connection between amplitudes with
different numbers of particles n, different N K MHV levels, and different loop-orders L.
7.6 Higher loops
The planar loop-integrand recursion relations studied above can also be directly applied to
higher-loop order in the planar limit of N = 4 SYM. This was already shown in (7.61).
However, in contrast to the unitarity approach, the integrands obtained from recursion
generally contain spurious poles. Local poles (non-spurious) refer to propagator-like poles,
these take the form of 1/i, i − 1, j, j − 1 or 1/A, B, i − 1, i in momentum twistor
space. Spurious poles, on the other hand, could take the form 1/A, B, 4, 2; this is non-
local in that it does not arise from a propagator in the loop-diagrams. At 4-point, since
there is only one term in the recursion, such spurious poles must vanish by themselves, and
indeed the final result is free of spurious poles. However, at higher-points the spurious poles
cancel between various terms in the recursion relations, but their presence in the individual
BCFW terms makes it difficult to carry out the integration of the loop-integrand to obtain
the actual amplitude. Spurious poles are a hallmark of BCFW recursion relations – we
discussed this for tree-level BCFW at the end of Section 3.2. BCFW builds in unitarity and
gauge-invariance at the expense of manifest locality. While it provides us with a method to
compute loop-integrands, it leads to complications as one eventually has to integrate these
non-local functions in momentum space.
Given the large amount of symmetry enjoyed by N = 4 SYM – superconformal sym-
metry and dual conformal symmetry as well as their enhancement to the Yangian – it
is tempting to be ambitious and ask if it is possible to manifest both locality and dual
conformal invariance at the same time. Certainly the unitary method discussed previously
would suffice for this purpose, since the dual conformal invariant scalar integrals are local.
However, when extended beyond 4-point, the number of dual conformal invariant integrals
becomes large and not all of them may contribute to a given amplitude.
At 1-loop level, the 1-loop box integral in (7.59) is the only available dual conformal
invariant local 4-point integral. In other words, dual conformal symmetry forces the triangle
and bubble-contributions to be absent in N = 4 SYM. Could it be that all amplitudes

N = 4 SYM are fixed by similar considerations? To study this involves maximal cuts and
Leading Singularities – and some of the principles involved also extend beyond the planar
limit and to SYM with less supersymmetry. We introduce the basic setup in the following
chapter.
Leading Singularities and on-shell diagrams 8
Chapter 6 introduced unitarity cuts. In D-dimensions, we can at most cut D propagators

per loop since each loop-momentum only has D components. When the maximum number
of propagators, D × L, is cut, the unitarity cut is called a maximal cut [117, 118]. The
maximal cuts are very useful for determining the integrand, in particular in 4d planar N = 4
SYM.
Example. As an example of a maximal cut, consider the quadruple cut of the 1-loop
n-point amplitude in 4d:
i i+1
1 j
Four propagators are put on-shell:
(8.1)

2
2
2
n j+1 2 = − K 1 = − K 2 = − K 3 = 0 .
k+1 k
Here K 1 = p1 + · · · + pi , K 2 = pi+1 + · · · + p j , K 3 = p j+1 + · · · + pk , and K 4 =
pk+1 + · · · + pn . No further propagators can be put on-shell in 4d since the loop-
momentum only has four components.
The result of a maximal cut is a product of on-shell tree amplitudes, Atree n 1 · · · An j , appro-
tree
priately summed over all possible intermediate states, with the loop-momenta evaluated on
the solutions to the cut constraints. the 1-loop box in 4d N = 4 SYM, the
For example,treefor tree
value of the maximal cut (8.1) is d 4 η I Atree
n1 A n2 A n 3 An 4 evaluated on the two solutions
tree
to the quadratic loop-momentum constraints. At L > 1, the cut constraints generically have
2 L distinct solutions, however, there are situations where there are not enough propagators
to cut; a simple example in 4d is the following 2-loop double-box integral
. (8.2)
It has only 7 propagators, but we need to take 2 × 4 = 8 propagators on-shell for a maximal
cut. For such cases, the solution space for the loop-momenta is not a set of isolated points but
rather a continuous multi-dimensional manifold. However, if we impose the cut constraints
162 Leading Singularities and on-shell diagrams
on one loop-momentum at a time, new poles appear and they can be used to fix the remaining
degrees of freedom, again leaving us with a set of isolated solutions for the loop-momenta.
We demonstrate this explicitly for the double-box (8.2) in Section 8.2.
The method of generalized unitarity is to find an integrand that reproduces all the
unitarity cuts, including of course all the maximal cuts. But how exactly do we treat the
distinct solutions to the maximal cut constraints? There are two ways to proceed:1
1. Appropriate sampling over all solutions. In one approach, we require that the correct
integrand matches the maximal cut evaluated on a sampling of all 2 L solutions, with
each solution given an appropriate weight. At 1-loop there are just two solutions and
the proper weight is 1/2 for both, thus in effect averaging over the two solutions, as in
(6.31) and (6.32). For higher-loops, one starts with a set of integrals that integrate to
zero. The appropriate weight for each solution is determined by the requirement that
their contributions to the vanishing integrals need to sum to zero. Explicit examples and
further discussions at 2-loop order can be found in [96].
2. Match each solution. In the other approach, we require the integrand to reproduce each
of the cut solutions individually. The individual cut solutions are treated as independent
entities and the resulting value for the maximal cut evaluated on each solution is called
a Leading Singularity (LS). The name reflects that these objects are the residues of
the most singular configuration of the loop-integrand (for generic external kinematics).
Note that, despite the name, these contributions are finite.
We focus here on the second approach. A major motivation is that all planar loop-integrands
of N = 4 SYM can be written as a linear combination of dual conformal invariant “unit
Leading Singularity integrands” (to be introduced below) [119]. The characterization of
Leading Singularities turns out to be a quite interesting mathematical problem [120].
The Leading Singularities offer insight into the structure of planar N = 4 SYM loop-
integrands at all loop orders, but note that to obtain the actual amplitudes, one still needs to
perform the challenging loop-integrations. While the notion of dual conformal invariance
is only applicable in the planar limit, it is a well-defined (and open) question whether the
full non-planar loop-integrands of N = 4 SYM can also be determined by the Leading
Singularities.
We begin our study of the Leading Singularities at 1-loop order. All amplitudes in this
chapter are in 4d planar N = 4 SYM.
8.1 1-loop Leading Singularities
To build intuition for the Leading Singularities at 1-loop order, we start with the simplest
case, namely the 4-point integrand, then consider the new features at 5-point, and finally
generalize to n-point.
1 The maximal cut was formulated in [118] for 1-loop amplitudes of N = 4 SYM and generalized to multi-loop
amplitudes in [117]. A more detailed review is offered in [68].
4-point
For n = 4, the maximal cut conditions (8.1) are simply
1 2
2 = ( − p1 )2 = ( − p1 − p2 )2 = ( − p1 − p2 − p3 )2 = 0 . (8.3)
4 3
2
In dual variables, this is simply y01 = y02
2
= y03
2
= y04
2
= 0 (see Section 7.1). And translat-
ing that to momentum twistor space (as in the early part of Section 7.4), we have
AB12 = AB23 = AB34 = AB41 = 0 . (8.4)
The geometric interpretation of the cut constraints (8.4) is that ( A, B) is a line in CP3 that
intersects each of the four lines (1, 2), (2, 3), (3, 4), and (4, 1). As anticipated from the
quadratic nature of the constraint (8.3), there are two independent solutions. This is rather
obvious geometrically:
1 2 1 2
(a) (A, B) = (1, 3) (A,B) (b) (A, B) = (2, 4) .

(A,B)
3 3
4 4
(8.5)
Now, on the quadruple cut, the 1-loop integrand factorizes into a product of four tree-level
3-point amplitudes, which can be either MHV or anti-MHV. Recall that special kinematics,
(2.58) and (2.59), apply to the 3-point amplitudes – we summarize it here:
i i
MHV |i − 1] ∝ |i] ∝ |i + 1] , MHV |i − 1 ∝ |i ∝ |i + 1 .
i−1 i+1 i−1 i+1

(8.6)
We use a black blob to indicate an MHV 3-point subamplitude (or vertex), and a white blob
for 3-point MHV = anti-MHV.
A configuration where two MHV subamplitudes with external legs are adjacent
vanishes:
1 2
= 0. (8.7)
This is because special kinematics (8.6) implies |1 ∝ |2, so that s12 = −( p1 + p2 )2 =
12[12] = 0. Of course this does not hold true for generic momenta p1 and p2 . Hence,
for generic external momenta, we are not allowed to have helicity configurations, such as
(8.7), where two MHV or two anti-MHV subamplitudes with external legs are adjacent.
The only helicity options for the 4-point quadruple cut are therefore
1 2 1 2
. (8.8)
4 3 4 3
Now these two diagrams have to be evaluated on the kinematic solutions (8.5). For the
solution (a) where the line (A, B) is (1, 3), we can simply pick the momentum twistor
of the loop line to be Z 1 = ( |1, [μ1 | ). This means that | ∝ |1 and that selects the
kinematics where the vertex that line 1 attaches to is anti-MHV, i.e. this picks the first helicity
configuration in (8.8). Likewise, solution (b) with (A, B) = (2, 4) selects the opposite
helicity configuration. So we conclude that the maximal cuts have the two solutions:
1 2 1 2
(a) (A, B) = (1, 3) (b) (A, B) = (2, 4) . (8.9)
4 3 4 3
These two diagrams encode the Leading Singularities for the 4-point 1-loop amplitude.
The Leading Singularities, LS(a) and LS(b) , are calculated as the product of the four subam-
plitudes summed over all intermediate states and evaluated on the respective solutions (a)
and (b), times a Jacobian factor. The Jacobian factor 1/J comes from a change of variables
that converts the associated loop-integral over R4 to a contour integration with four contours
encircling each of the four propagator poles in the quadruple cut.2 The conversion of the
2 Since the solutions to the cut conditions may be complex-valued, we should really consider the loop-integral as
an integral over the real section in C4 ; that makes it more natural to convert to a contour integral encircling the
poles corresponding to the on-shell propagators.
integral can be done via a change of variables u i = y0i2 , for i = 1, 2, 3, 4, giving

d 4 y0 du 1 du 2 du 3 du 4
2 2 2 2
= J, (8.10)
y01 y02 y03 y04 u1 u2 u3 u4
μ
where J = det(∂ y0 /∂u i ) is the Jacobian. As we show explicitly in the example below, the
Jacobian is
1 12233441
J= =− . (8.11)
2 2
y13 y24 12342
The Leading Singularity for the 4-point 1-loop amplitude is then
$# 4
'
LS(a) = J d 4 η i AMHV
3 (− 4 , p1 , 1 ) AMHV
3 (− 1 , p2 , 2 )
i=1

× AMHV
3 (− 2 , p3 , 3 ) AMHV
3 (− 3 , p4 , 4 ) . (8.12)
= (a)
A similar expression is found for LS(b) . Evaluating the LS(a) and LS(b) , one finds
LS(a) = LS(b) = Atree

4 . (8.13)
Exercise 8.1
Evaluate the RHS of (8.12) to show that LS(a) = Atree
4 .
Now before exploring the Leading Singularities further, let us illustrate how the Jacobian
is obtained. It can of course be calculated brute-force, but in the example below we carry
out the calculation via a tour to momentum twistor:
Example. We calculate J in (8.10) via the momentum twistor formulation. From

(7.40)–(7.43), we read off
4
d 4 y0 d Z Ad4 Z B 12233441
= . (8.14)
2 2 2 2
y01 y02 y03 y04 vol(G L(2)) AB12AB23AB34AB41
The loop momentum twistors Z A and Z B can be expanded on the basis of four external
line momentum twistors as
Z A = a1 Z 1 + a2 Z 2 + a3 Z 3 + a4 Z 4 , Z B = b1 Z 1 + b2 Z 2 + b3 Z 3 + b4 Z 4 . (8.15)
This linear transformation gives d 4 Z A d 4 Z B = 12342 d 4 ai d 4 bi . The 4-brackets eval-

uate to A, B, i − 1, i = 1234Mi+1 , where M j is the jth minor of the 2 × 4 matrix

a1 a2 a3 a4
. (8.16)
b1 b2 b3 b4
For example, AB34 = 1234M1 = 1234(a1 b2 − a2 b1 ). Now, consider a G L(2)-
transformation (7.55) of Z A , Z B . We can use it to set a4 = b2 = 0 and a2 = b4 = 1.
Exercise 8.2
Convince yourself that a G L(2) rotation of Z A , Z B allows you to make the above
choice of parameters, but that setting a4 = b4 = 0 would be illegal.
In this gauge, we have AB12 = 1234 a3 , etc., and the integrand then has no depen-
dence (obviously) on a4 , b2 , a2 , b4 . This means that the G L(2)-volume factor cancels
and we are then left with

d 4 y0 da1 db1 da3 db3 12233441
= . (8.17)
2 2 2 2
y01 y02 y03 y04 a1 b1 a3 b3 −12342
This way we have brought the loop-integral to the form on the LHS of (8.10) and we
see that the Jacobian is indeed (8.11).
Now our integration variables ai and bi in (8.17) are not exactly the u i = y0i2 that
41AB 1234
we introduced above (8.10): for example u 1 = y01 2
= 41AB = − 41AB b3 . So the
u i s are proportional to the a1,3 and b1,3 , but the factors of proportionality drop out of
the du i /u i measure.
Recall that we are studying the Leading Singularities in order to find an integrand that
faithfully reproduces both LS(a) and LS(b) . The integrand that we already know for the
4-point 1-loop amplitude does the job – let us see how. We have previously found (see
(6.39)) that

1-loop d 4 y0
A4 = Atree
4 y 2 2
y
13 24 2 2 2 2
. (8.18)
y01 y02 y03 y04
2 2
When we convert this to the contour integral, the prefactor y13 y24 exactly cancels the
Jacobian (8.11), so we are left with

1-loop da1 db1 da3 db3
A4 = Atree
4 . (8.19)
a1 b1 a3 b3
From this, we can directly read off the residue at the propagator poles a1 = b1 = a3 =
b3 = 0. The result is independent of which of the two solutions (a) or (b) we use to localize
the loop-integral, so the quadruple cuts of the integral match exactly with the Leading
Singularities, LS(a) = LS(b) = Atree
4 . This may not shock you, but once we venture beyond
4-point amplitudes, things are not so simple.
The result LS(a) = LS(b) is special for the 4-point case. It can be represented diagram-
matically as
1 2 1 2
. (8.20)
4 3 4 3
This identity is called the square move and it will show up later in our discussions of
higher-loop Leading Singularities and on-shell diagrams.
5-point
At this stage, we have constructed the 4-point 1-loop amplitude of N = 4 SYM in three
different ways: generalized unitarity, loop-level BCFW, and Leading Singularities. It is time
to move ahead.
We consider a specific maximal cut of the 5-point 1-loop amplitude:
5 y1
y5 1
4
y4 y0 y2 2
y01 = y02
2
= y03
2
= y04
2
= 0. (8.21)
3 y3 2
The cut constraints for this maximal cut can be written in momentum twistor space as
AB12 = AB23 = AB34 = AB51 = 0 . (8.22)
This says that (A, B) is a line that intersects the four lines (1, 2), (2, 3), (3, 4), and (5, 1).
There are two solutions:
1 1
2 2
5 (A,B) 5
(A,B)
3 3
4 4
(a) (A, B ) = (1, 3) (b) (A, B ) = (5, 1, 2) (2, 3, 4) . (8.23)
It is straightforward to see that (a) is a solution to (8.22). As for (b), note that generically
two planes in CP3 intersect in a line. Any points A and B on the intersection of the two
planes in (b) will be linearly dependent with any two points in either plane. This establishes
that (b) is a solution to (8.22).
Since intersections of planes in CP3 may not feel as natural to you as brushing your teeth
(hopefully), let us make the solutions (8.23) explicit in momentum space. With = y10 ,
the constraints are 2 = ( − p1 )2 = ( − p1 − p2 )2 = ( − p1 − p2 − p3 )2 = 0 and it is
not hard to verify that the two solutions for the loop-momentum can be written

23 [23]
(1) = −|1 [1| + [2| , (2) = − |1 + |2 [1| . (8.24)
13 [13]
Note that even though (1) is formally the complex conjugate of (2) , their geometric
interpretations in momentum twistor space are quite different. This is because momentum
twistors are chiral objects (only |i appears, not |i]).
Exercise 8.3
Show that (1) and (2) in (8.24) solve the cut constraints. Check little group scaling.
Then show that two solutions, (1) and (2) , correspond to the two geometric solutions
(a) and (b) of (8.23), respectively, in momentum twistor space.
The solution (1) has | ∝ |1 and by momentum conservation these are also proportional
to the angle spinor of ( − p1 ). This means that the special 3-point kinematics forces the
vertex where line 1 attaches to be anti-MHV. Likewise, the solution (2) forces the same
vertex to be anti-MHV. By (8.7), the rest of the helicity structure is then fixed, and we see
that the two solutions (a) and (b) correspond to the two options
5 5
1 1
4 4
(1) (2)
. (8.25)
3 2 3 2
(a) (b)
The gray blob for the 4-point vertex does not have specific helicity designation because
4-point on-shell amplitude is simultaneously MHV and anti-MHV.
Now, let us count the number of Grassmann ηs of these blob-diagrams. MHV has 8
ηs and 3-point anti-MHV has 4 ηs. For each of the four internal lines we have to do a
d 4 η-integral. For diagram (a), this then gives 8 + 4 + 4 + 8 − 4 × 4 = 8 corresponding to
the 5-point MHV sector. For diagram (b): 8 + 8 + 8 + 4 − 4 × 4 = 12 which identifies it
as belonging to the NMHV sector. Including the appropriate Jacobians, the results for the
two diagrams (8.25) are the respective MHV and NMHV 5-point tree-level amplitudes,
respectively.
The above discussion tells us that for a given MHV or NMHV 1-loop 5-point amplitude,
only one of these solutions to the maximal cut conditions (8.21) is relevant. If we focus
on the MHV sector, only diagram (a) in (8.25) matters, and it equals Atree 5 for solution
(1)
and is zero when evaluated on solution (2) . This leads us to the crux of the problem we
mentioned in the beginning of this chapter: whether the integral basis Ansatz we write for
the integrand faithfully reproduces all Leading Singularities. Let us illustrate this explicitly.
Consider the scalar box-integral whose propagators are those considered in the maximal
cut (8.21):
51231234
I5,box (1, 2, 3, 4) = . (8.26)
AB51AB12AB23AB34
We use the labels on the n-point box-integral In,box (i, j, k, l) to specify the first external leg
on each of the vertices. For example, for the arrangement in (8.1), the corresponding scalar
box-integral would be labeled In (1, i + 1, j + 1, k + 1).
When evaluating the quadruple cut for the integral I5,box (1, 2, 3, 4), the Jacobian cancels
the numerator factor 51231234, and since there is no other dependence on the loop-
momenta than the four propagators we are cutting, this integral produces the same answer,
namely 1, no matter if we evaluate it on solution (1) or (2) : i.e.

I5,box (1) = I5,box (2) = 1 . (8.27)
On the other hand, we now know that for the MHV sector the corresponding Leading
Singularities of diagram (a) in (8.25) are
LS(1) = Atree
5,MHV , LS(2) = 0 . (8.28)
This means that the Ansatz

1-loop
A5,MHV = Atree
5,MHV × I5,box (1, 2, 3, 4) + other box-integrals (8.29)
does not produce the Leading Singularities faithfully. However, it does produce the average
of the two maximal cuts correctly because, when multiplied by a factor of Atree
5,MHV ,
1 1
I5,box (1) + I5,box (2) = (1 + 1) = 1 (8.30)
2 2
equals the sum of the Leading Singularities LS(1) + LS(2) . The message is that the inte-
grand Ansatz (8.29) can produce the correct maximal cut when one averages over the two
constraints (as is usually done in applications of the generalized unitarity method), but it
does not produce each Leading Singularity honestly. If you just want an answer for the
amplitude, you don’t have to care. But let us try to be caring people and see where it takes
us.
We have learned now that we need something else in the Ansatz (8.29) in order to match
the Leading Singularities. That something else turns out to be the pentagon integral
)
A, B|(1, 2, 3) (3, 4, 5)2451
I5,pentagon = . (8.31)
AB12AB23AB34AB45AB51
)
The numerator includes the bi-twistor (1, 2, 3) (3, 4, 5) that characterizes the line of
intersection between the planes (1, 2, 3) and (3, 4, 5); the intersection formula was given in
(5.67). When we evaluate the maximal cut (8.21) of the pentagon, the residue depends on
the loop-momentum and hence on which solution (8.24) we choose to evaluate it. Including
the Jacobian J = 51231234, one finds

I5,pentagon (a) = 0, I5,pentagon (b) = −1 . (8.32)
This is good news, because now the improved Ansatz

1-loop
A5,MHV = Atree
5,MHV × I5,box (1, 2, 3, 4) + I5,pentagon + other (8.33)
(where “other” is assumed not to contribute to our quadruple cut) has the following maximal
cut (8.21): on two solutions, (1) and (2) , it gives

Atree
5,MHV I5,box (1) + I5,pentagon (1) = A5,MHV × (1 + 0) = A5,MHV = LS(1) ,
tree tree

(8.34)
Atree
5,MHV I5,box (2) + I5,pentagon (2) = A5,MHV × (1 − 1) = 0 = LS(2) .
tree
So it produces the correct Leading Singularities for the cut (8.21)!

Now, unfortunately we are not done yet, because we have to worry about all the other
cuts: there are a total of 2 × 5 = 10 Leading Singularities for the 5-point 1-loop amplitude.
With just the box diagram and the pentagon diagram in (8.33), there is no chance that
this can be the full answer: the reason is simply that the sum of those two integrals is not
cyclically invariant. It takes just one more integral to achieve cyclic invariance, namely the
box integral I5,box (3, 4, 5, 1). Diagrammatically, we can express the final answer as

1-loop
A5,MHV = Atree
5,MHV × I5,box (1, 2, 3, 4) + I5,pentagon + I5,box (3, 4, 5, 1)
3
5 1 2 3
2 4
4 1
= Atree
5,MHV
+ + . (8.35)
3 2 1 5 5 4
The diagrammatic notation for the pentagon integral has a wavy line indicating that the bi-
)
twistor (1, 2, 3) (3, 4, 5) goes in the numerator in (8.31). Our previous results plus cyclic
invariance then guarantee that (8.35) produces all ten Leading Singularities correctly.
Exercise 8.4
Show that (8.32) is true.
Exercise 8.5
Show that (8.35) is invariant under cyclic permutations of the external lines.
We have introduced here box and pentagon 1-loop integrands, (8.26) and (8.31), whose
quadruple cuts evaluate to either +1, −1 or 0. Such integrands are called unit Leading
Singularity integrands.
It has been shown [119] that all planar loop amplitudes of N = 4 SYM can be obtained
as a linear combination of unit Leading Singularity integrands (times a tree amplitude).
The coefficients in front of the unit Leading Singularity integrands are determined by the
Leading Singularity, thus knowing them is sufficient to determine the entire amplitude.
For planar amplitudes, we need unit Leading Singularity integrands that are also dual
conformal invariant and local, and this is a rather restrictive class of integrands. We have
seen the Leading Singularity method at work for 4- and 5-point 1-loop amplitudes. The
structure generalizes to higher points, as we now outline.
6-point and beyond

The 1-loop n-point MHV amplitude is given by a simple generalization of the 5-point
result:
i+1 j−1
1-loop
An,MHV = Atree i j
n,MHV
1<i<j<n
i−1 j+1
1 n
)
A,B|(i − 1, i, i + 1) ( j − 1, j, j + 1)i, j, n, 1
= .
A,B 1<i< j<n A,Bi, i −1AB, i, i +1A,B, j −1, jA,B j, j +1A,B,n,1
(8.36)
In the sum, there are two boundary cases: i = 2, j = 3 and i = n − 2, j = n − 1. These

correspond to box integrals whose numerators are simply the Jacobian coming from cutting
all four propagators. More precisely, we have
2 3

n1231234
= (8.37)
A,B AB12AB23AB34ABn1
4
1 n
and
n−2 n−1

n−3, n−2, n−1, nn−2, n−1, n, 1
= .
n−3 A,B A,B,n−3,n−2A,B,n−2,n−1A,B,n−1,nA,B,n,1
1 n
(8.38)
In conclusion, the two Leading Singularities of arbitrary 1-loop MHV amplitudes can be
reproduced by including the simple combination of tensorial (due to the loop-momentum
dependence in the numerator) pentagon integrals. These are all local unit Leading Singu-
larity integrands. Note that these integrands can be used as part of the basis for 1-loop
amplitudes in any massless quantum field theory. The special situation for N = 4 SYM
is that these integrals provide the entire answer, whereas for a generic QFT, one needs in
addition the various lower-gon integrals that are not captured by the maximal cuts.
You may (and should) be puzzled that in the beginning of Section 6.2, we stated that
the 1-loop amplitudes in a unitary 4d quantum field theory can be expanded on a basis
of scalar box-, triangle-, and bubble-integrals with possible additional input from rational
terms. This was summarized in equation (6.26), and we noted that in N = 4 SYM, the
only non-vanishing contributions were the box-integrals. There were no pentagons in that
story! So what is the deal? The point of the pentagon integrals in the present section is
that they allow us to write the N = 4 SYM 1-loop integrand in a form that reproduces
each Leading Singularity faithfully. On the other hand, (6.26) determines the 1-loop N = 4
SYM amplitudes as a sum of box-integrals whose coefficients are evaluated by quadruple
cuts, evaluated as the average of the two loop-constraint solutions. More box-diagrams
contribute in (6.26) than in (8.36). So what is the deal? Well, the two representations
of the integrand must yield the same answer for the amplitude. The integrals have to be
regulated, and if one uses dimensional regularization 4 − 2, the difference between the
two representations is only in the O()-terms. Specifically, the pentagon integrals contain
the information about the “missing” boxes plus O() [121]. Thus the two procedures yield
the same integrated answer.
8.2 2-loop Leading Singularities
At the start of this chapter we noted that not all loop-diagrams appear to have enough
propagators available for a maximal cut of 4L-lines. A representative example is the
double-box diagram of the 2-loop 4-point amplitude:
2 −p −p 3
1 1 2
. (8.39)
1 2
1 4
With seven propagators, we can only localize seven of the eight components of the two
loop-momenta, leaving behind a 1-dimensional loop-integral. However, when the seven
propagators are on-shell, the 4-point 1-loop analysis tells us that the lefthand box in (8.39) is
4 [ 2 , p1 , p2 , 2 − p1 − p2 ],
a Leading Singularity that equals the 4-point tree amplitude, Atree
where 2 parameterizes the loop-momentum in the righthand box. But this tree amplitude
has a propagator 1/( 2 − p1 )2 that can now be used to localize the final component of
the loop-momenta, thus providing a maximal cut. Moreover, on this pole, the 4-point tree
amplitude factorizes into two 3-point amplitudes, and therefore the result is simply an
on-shell 4-point 1-loop box diagram. The procedure is illustrated here:
2 2 3 2 3
3 2 3
A4
tree
A4
tree .
1 4
1 4 1 4 1 4
(8.40)
The last step uses the fact that the 1-loop 4-point Leading Singularity is equal to the tree
amplitude Atree
4 [1234]. In conclusion:
(i) the Leading Singularities are well-defined for the 2-loop 4-point amplitude, even
though the double-box only has seven loop propagators, and
(ii) the 4-point “double-box Leading Singularity” equals the 4-point tree amplitude.
A Leading Singularity that involves a “pole-under-a-pole" is called a composite Leading

Singularity. A prototype of such a composite object is the 3-variable contour integral
*
1
d x d y dz . (8.41)
x(x + yz)
This integrand appears to have only two poles, insufficient to localize the 3d integral.
However, if the x-contour encircles the pole at x = 0, then an additional pole emerges in
the form 1/(yz) and this can then localize the remaining two integrals, giving the residue
1 (ignoring 2πis).
The idea of composite Leading Singularities resolves the subtlety about defining maximal
cuts and Leading Singularities for higher-loop amplitudes. Henceforth we work with the
understanding that the Leading Singularities of multi-loop amplitudes are always well-
defined.
8.3 On-shell diagrams
We have found in Section 8.1 that the 4-point 1-loop Leading Singularity is equal to the
4-point tree amplitude:
1 2
1 2
= tree . (8.42)
A4
4 3
4 3
This looks rather peculiar since the LHS is a 1-loop diagram while the RHS is a tree
amplitude. It actually turns out that the LHS can be interpreted as a super-BCFW diagram!
We now show how.
Consider the top two vertices in the Leading Singularity diagram
1 2
c
|a ∝ |c ∝ |1 ,
a b (8.43)
|b] ∝ |c] ∝ |2] .
d
4 3
The MHV and anti-MHV designations imply the indicated special 3-particle kinematics.
Up to an overall factor z, this determines pc to be pc = −z|1[2|. Momentum conservation
then fixes pa and pb to be

pa = −|1 [1| + z[2| and pb = − |2 − z|1 [2| . (8.44)
We recognize pa and pb as BCFW [1, 2-shifted momenta p̂1 and p̂2 !

What about the Grassmann variables? Let us carry out the ηc -integral in the product of
the Grassmann delta functions of the first two vertices,

d 4 ηa d 4 ηb d 4 ηc δ (4) [1c]ηa + [ca]η1 + [a1]ηc δ (8) |2η2 − |bηb − |cηc

(8.45)

∝ d ηa d ηb δ |1 ηa − (η1 − zηb ) − |2 ηb − η2 .
4 4 (8)
The last integral localizes ηa and ηb to be
ηa = η1 − zη2 and ηb = η2 . (8.46)
This is exactly the shift of the Grassmann variables associated with the supersymmetrization
of the BCFW shift (8.44).
Finally, let us see how the internal line d in (8.43) fixes z. The on-shell condition is
0 = pd2 = ( p3 + pb )2 = (23 − z13)[23], i.e. z = 23/13. This corresponds exactly
to the pole where the propagator 1/ p̂23 in the [1, 2-shifted 4-point tree amplitude goes on-
shell. As we know from the super-BCFW calculation (4.56), this is exactly the factorization
pole that allows us to determine the full 4-point tree amplitude in N = 4 SYM from the
MHV3 × anti-MHV3 super-BCFW diagram.
We have established the connection between the Leading Singularity diagram on the LHS
of (8.42) and the 4-point super-BCFW diagram (4.56), and this allows us to understand
why the 4-point 1-loop Leading Singularity is just the 4-point tree superamplitude. The
connection is summarized diagrammatically as the BCFW bridge
1 2 1 2
−z |1>[2|
c
a b ^p ^p , (8.47)
1 2
4 3 4 3
where the upper two vertices, surrounded by the shaded region, are the “bridge.” The bridge
provides the BCFW super-shift.
Exchanging black and white dots in the BCFW bridge simply corresponds to the conju-
gate BCFW shift. This also gives another meaning to the square move (8.20):
1 2 1 2
= . (8.48)
4 3 4 3
It says that the two BCFW super-shifts [1, 2 and [2, 1 give the same 4-point amplitude.
With the square move and the BCFW shift, it becomes fun to calculate Leading Singu-
larities. Starting from the fundamental 3-point vertices, we can build on-shell diagrams
that contain information about the higher-loop amplitudes. Each 3-point vertex represents
the MHV or anti-MHV 3-point amplitude, along with the implication that the square or
angle spinors of its legs are proportional. The vertices are glued together by “on-shell
propagators” whose rules can be written

I d 2 |I d 2 |I ]d 4 η I
= . (8.49)
U (1)
The η I integral is the usual state sum. The integration over the momentum variables will be
localized by the momentum conservation delta function on both sides of the propagator.
In addition to the square move (8.48), there are two rules that help us simplify complicated
on-shell diagrams. The first rule follows from the observation that each MHV 3-vertex
imposes that the square spinors of the associated lines are proportional, so two consecutive
MHV vertices imply that all four square spinors are proportional. This gives the rule
= = . (8.50)
There is of course an equivalent rule for anti-MHV. The black 4-vertex blob imposes, per
definition, that the four lines have proportional square spinors. This blob does not represent
a 4-point MHV tree amplitude; it is just a short-hand notation for the double-blob diagrams.
The second rule is
= . (8.51)
The 3-particle kinematics forces the internal lines in the bubble to be collinear, and this
collapses the bubble. This formally eliminates a loop-integral.
Combining the two rules (8.50) and (8.51) gives
= (8.52)
Exercise 8.6
Show that the internal lines in the bubble (8.51) are collinear and that (8.52) follows
from (8.50) and (8.51).
The point of these rules is to simplify the evaluation of on-shell diagrams. For the on-shell
diagram of the 2-loop 4-point Leading Singularity, we first apply the square move (8.48)
and then the collapse-moves (8.50) and (8.52) to get
2 3 2 3
2 3 2 3
= = = .
1 4 1 4 1 4 1 4
(8.53)
This shows that the 4-point 2-loop Leading Singularity equals the 4-point 1-loop Leading
Singularity, which in turn is just the 4-point tree amplitude. We had already found this result
in Section 8.2 by evaluating the composite Leading Singularity. The rules for the on-shell
diagrams offer a simpler diagrammatical derivation.
Exercise 8.7
Write down 3-loop on-shell diagrams for the 4-point MHV amplitude and show that
they reduce to the 1-loop result.
Using the BCFW bridge, we can begin to build up more complicated on-shell diagrams.
For example, we can use the BCFW bridge to interpret the on-shell diagram:
. (8.54)
The circles highlight the 4-point MHV tree amplitude and 3-point anti-MHV tree subam-
plitudes. The BCFW bridge, indicated with the shaded area, induces the BCFW shift on the
two affected lines, so this on-shell diagram represents the BCFW diagram for the 5-point
MHV tree-level superamplitude.
The super-BCFW recursion relations for the 6-point NMHV tree amplitude can be
represented with on-shell diagrams as
(8.55)
The left-hand diagram is the MHV3 × anti-MHV5 BCFW diagram, the middle diagram is
the BCFW diagram with two 4-point MHV tree subamplitudes and the right-hand diagram
is the MHV5 × anti-MHV3 BCFW diagram.
Exercise 8.8
Do the η-counting to show that (8.54) represents an on-shell diagram for an MHV
amplitude and (8.55) an NMHV amplitude.
Exercise 8.9
Interpret the effect of the rule (8.50) on the on-shell diagram (8.54). Show that the
3-loop diagram
(8.56)
is equivalent to (8.54). What does that tell you about Leading Singularities?
The concept of the Leading Singularity is well-defined for lower-N SYM and for non-super-
symmetric theories, and so are the on-shell diagrams. The only distinction is that the edges
need arrows because the on-shell states split into two CPT conjugate multiplets, one with
the positive helicity gluon and the other with the negative helicity gluon. (N = 4 SYM is
special in that its supermultiplet is CPT self-conjugate.) Opposite helicity multiplets must
sit at different ends of the propagators, so they are dressed with arrows to indicate the
assignment.
In the next section, we discuss a formula that reproduces all Leading Singularities in
planar N = 4 SYM. This formula will tell us that the number of distinct Leading Singu-
larities for given n and given N K MHV-level is fixed. For example, for K = 0 (MHV) there
is only one Leading Singularity: that is why both the 1- and 2-loop Leading Singularities
for the 4-point amplitude evaluate to the same value, namely the 4-point tree amplitude.
That this pattern continues is rather remarkable, since it says that on-shell diagrams with
four external lines when evaluated all give the 4-point tree amplitude, no matter how many
hundreds of loops we add in.
The problem of determining and classifying all distinct on-shell diagrams, under the
equivalence-moves, turns out to be an interesting mathematical problem that has become
an exciting research topic [120].
PART III
TOPICS
Grassmannia 9
The on-shell BCFW recursion formulas (7.39) and (7.61) have taught us that tree superam-
plitudes and loop-integrands of planar N = 4 SYM can be written as
AnL-loop = Atree
n,MHV × Yn
L-loop
, (9.1)
L-loop
where Yn is a dual conformal invariant. In the N K MHV sector, Yntree is a sum of K
products of 5-brackets (or R-invariants in Section 4.5), as found in Sections 5.3 and 5.4. At
L-loop
loop-level, Yn is a linear combination of dual conformal invariant integrands.
The MHV tree amplitude prefactor in (9.1) serves an important purpose: since it is dual
conformal covariant with homogeneous dual conformal inversion-weight of the external
particles – as in (5.39) – it generates the necessary dual conformal “anomaly” that modifies
the dual conformal generators in such a way that they become part of the level 1 generators
of a Yangian symmetry. Hence, as described in Section 4.5, the MHV factor in (9.1) is
essential for Yangian symmetry.
There are several interesting points to consider:
1. The color-ordered planar N = 4 SYM superamplitudes have cyclic symmetry in the

labels of the external states. However, in the dual conformal BCFW representations
(7.39) and (7.61), the cyclic symmetry is completely obscured. This is not surprising,
because these recursion formulas are based on shifts of two adjacent external lines:
making two lines special breaks the cyclic symmetry. In the pursuit of happiness and
manifest symmetries, we may ask if there is a formalism for the planar N = 4 SYM
superamplitudes in which both the (dual) conformal symmetry and the cyclic symmetry
are manifest? This suggestion will guide us in Section 9.1.
2. In Section 8.3, we gave examples of how individual Leading Singularity diagrams can
be understood as BCFW terms of tree amplitudes. Since each BCFW term is Yangian
invariant, as indicated in the recursion formula (7.39), this implies that the Leading
Singularities are also Yangian invariant. Thus, understanding the most general Yangian
invariants is a step towards gaining control of the planar superamplitudes in N = 4
SYM at any loop order. Of course, one still needs to understand how to put the Yangian
invariants together to obtain a given superamplitude; read on.
3. Super-BCFW recursion for N = 4 SYM amplitudes can be based on any choice of two
shifted external momenta. Different choices can give very different representations of
the same amplitude, in particular with distinct spurious poles. For the amplitude to be
local, i.e. free of spurious poles, the residues of the spurious poles must cancel in the
sum of BCFW diagrams. Thus, the equivalence of two different BCFW representa-
tions is intimately related to locality. Each BCFW diagram is Yangian invariant, so by
182 Grassmannia
understanding how to enforce locality in a Yangian invariant way, it turns out that the
equivalence between the different BCFW representations can be trivialized.
Interestingly, the above three points can be addressed jointly. The strategy is to find a way
to generate the most general Yangian invariant rational function from a formula in which
cyclic symmetry is manifest. That’s our job now, so let’s get to work.
9.1 Yangian invariance and cyclic symmetry
The level 0 Yangian generators are the superconformal generators studied in Section 5.1. In
Section 5.2, we introduced the supertwistors WiA = ([i|a , |μ̃i ȧ , ηi A ) in order to linearize
the action of the superconformal generators. The Grassmann components of the supertwistor
are simply the on-shell superspace coordinates ηi A , and |μ̃i ȧ is the Fourier conjugate
coordinate of |iȧ . A function f of on-shell momentum space spinor-helicity variables is
Fourier transformed to the (super)twistor space as
&# %
n

n
d 2 |i f [i|, |i, ηi ei j=1 j μ̃ j ≡ f˜(WiA ) . (9.2)
i=1
In supertwistor space, the level-0 superconformal generators are

n
n
∂
G AB = G iA B = WiA , (9.3)
i=1 i=1
∂WiB
and the level-1 generators can be written in bi-local form as (Section 5.3)

n

(−1)|C| G iA C G Cj B − (i ↔ j) . (9.4)
i< j
Our aim is a cyclic invariant formula that generates Yangian invariant rational functions;
these are the building blocks for N K MHV superamplitudes of planar N = 4 SYM. Let us
try to motivate the construction, step by step. To start with, note that the level 0 generators
(9.3) act on the supertwistor variables as S L(2, 2|4) linear transformations. Any delta
function δ 4|4 whose argument is a linear combination of the supertwistors is invariant under
the linear S L(2, 2|4) transformation: for example

n n n n
δ 4|4
Ci WiA
≡ δ 2
Ci [i| δ
a 2
|i Ci δ
ȧ (4)
Ci ηi A , (9.5)
i=1 i=1 i=1 i=1
with some arbitrary auxiliary coefficients Ci ∈ C. For a level 0 generator G A B with A = B,

this is because it transforms the argument of one of the delta functions to that of another
delta function; schematically
∂
x δ(x)δ(y) = x δ(x) δ (y) = 0 . (9.6)
∂y
9.1 Yangian invariance and cyclic symmetry 183
For “diagonal” generators, such as G 1 1 , invariance follows only after using the tracelessness
condition, i.e. writing the generators in the form analogous to (5.19).
For an N K MHV superamplitude, we need Yangian invariants that are Grassmann poly-
nomials of degree 4(K + 2); so it is natural to take
k ≡ K +2 (9.7)
products of (9.5). Note that k counts the number of negative helicity gluons in the pure gluon
amplitude. To avoid having k identical delta functions, we introduce k sets of the auxiliary
variables Cai labeled by a new index a = 1, 2, . . . , k. So now we have a Grassmann degree
4(K + 2) object
#
k
n
δ 4|4 Cai WiA (9.8)
a=1 i=1
that is S L(2, 2|4) invariant. The parameters Cai are sometimes called link variables [23].
The n × k parameters Cai are arbitrary, so to remove the dependence on them let us
integrate (9.8) over all Cai . This has the further benefit of making the integrated result
cyclically invariant: a permutation of the Wi s is compensated by a permutation of the
Cai s. This is a change of integration variable with unit Jacobian. However, it is not clear
what measure we should use when integrating over the Cai s. So let us allow for a general
cyclically invariant function f (C) and write our candidate “generating function” as
#
k
n
d k×n C f (C) δ 4|4 Cai WiA . (9.9)
a=1 i=1
The integral of the k × n complex parameters is intended to be carried out as a contour

integral. The choice of contour is a very important and physically relevant aspect that will
be discussed in Section 9.3.
When the level 1 generators are considered, it turns out that there is a unique choice of
f (C) such that (9.9) is Yangian invariant. We will not repeat the argument here, but refer
you to [122, 123]. The unique function that gives (9.9) full Yangian symmetry is
1
f (C) = , (9.10)
M1 M2 · · · Mn
where Mi is the ith ordered minor of the k × n matrix Cai : this is the determinant of the
k × k submatrix whose first column is the ith column of Cai , specifically
Mi ≡ a1 a2 ... ak Ca1 i Ca2 ,i+1 · · · Cak ,i+k−1 , (9.11)
with i = 1, 2, . . . , n. One goes around cyclically when reaching the end of the C-matrix.
For example, the n = 5 and k = 2 matrix

C11 C12 C13 C14 C15
C= (9.12)
C21 C22 C23 C24 C25
gives M1 = C11 C22 − C12 C21 and M5 = C15 C21 − C11 C25 .
184 Grassmannia
At this stage, we have learned that the integral

#k n
d k×n C
δ 4|4 Cal WlA (9.13)
M1 M2 · · · Mn a=1 l=1
is Yangian invariant and has cyclic symmetry.

Before declaring victory, there are some loose ends to tie. First, a minor issue (yes, that’s
a pun) is that if Mi contains columns that are not strictly increasing due to cyclicity (for
example Mn−1 = · · · Can Ca1 · · · ) then the proof of Yangian invariance goes through only
on the support of the bosonic delta functions in (9.13). Second, a major issue is that the
integral we so proudly wrote down in (9.13) is not at all well-defined – it is divergent.
To see this, note that the product of delta functions is invariant under a G L(k) rotation of
the k a-indices. The minors only respect S L(k) transformations: G L(1) takes Cai → tCai ,
hence Mi → t k Mi , but this excess weight is canceled by the Jacobian of d k×n C. Thus the
integral has G L(k) symmetry. To define a proper integral we need to “gauge fix” the G L(k)
redundancy. We indicate the need to gauge fix G L(k) by writing
#

d n×k Cai
k n
L̃n,k Wi = n δ 4|4 Cal WlA . (9.14)
G L(k) j=1 M j a=1 l=1
It turns out [122, 123] that for given n and k, Ln,k is the unique cyclically invariant
integral-expression that generates all Yangian invariants! We are going to give exam-
ples in the following sections. The formula (9.14) was first introduced by Arkani-Hamed,
Cachazo, Cheung, and Kaplan [23], who at the time conjectured that it produces all Lead-
ing Singularities of planar N = 4 SYM. A similar integral formula was presented by
Mason and Skinner [60] based on momentum supertwistors Z, as opposed to the “regu-
lar” supertwistors W, thus interchanging the role of the ordinary superconformal and dual
superconformal symmetries. The twistor and momentum twistor versions of the Grassman-
nian integral are directly related (Arkani-Hamed et al. [60]). For a streamlined proof and
more recent review of the Grassmannian formulation, see Elvang et al. [60].
We used supertwistors WiA = ([i|a , |μ̃i ȧ | ηi A ) to emphasize superconformal and Yan-
gian symmetry in the construction above. However, since we are more familiar with scat-
tering amplitudes in momentum space ([i|a , |iȧ | ηi A ), we are going to inverse-Fourier
transform all |μ̃i in (9.14) back to |i. This is conveniently done in a gauge-fixing of the
G L(k) symmetry where the first k × k block of Cai is the unit matrix. For example for
n = 7 and k = 3, we have
⎛ ⎞
1 0 0 c14 c15 c16 c17
C = ⎝0 1 0 c24 c25 c26 c27 ⎠ . (9.15)
0 0 1 c34 c35 c36 c37
In this gauge, (9.14) becomes

d (n−k)×k c # 2
k n n n
n δ [a| + cal [l| δ |μ̃a +
2
|μ̃l cal δ (4)
ηa + cal ηl .
j=1 M j a=1 l=k+1 l=k+1 l=k+1
(9.16)

Performing the inverse-Fourier transform d 2 |μ̃ j e−i j μ̃ j for each j = 1, . . . , n gives
$ '

d (n−k)×k c #
k
n
n
Ln,k [i|, |i, ηi = n δ [a| +
2
cal [l| δ (4)
ηa + cal ηl
j=1 M j a=1 l=k+1 l=k+1
$ '
#
n
k
× δ |i −
2
|acai . (9.17)
i=k+1 a=1
Exercise 9.1
Fill out the details of the inverse Fourier transformation to derive (9.17).
The representation (9.17) is central in the next section where we study the geometric
interpretation of Ln,k . In Section 9.3, we show that familiar amplitude expressions can be
derived from Ln,k .
9.2 The Grassmannian
It is very convenient to view the n × k matrices Cal in (9.14) as k n-component vectors

that define a k-plane in Cn . The space of all k-planes in an n-dimensional space is called
the Grassmannian Gr(k, n). The formula (9.14) for L̃n,k is therefore naturally viewed as a
cyclic invariant integral over all k-planes in the Grassmannian. Since any non-degenerate
linear transformation of the k n-vectors gives the same plane, there is a natural G L(k)
invariance. It is precisely the same G L(k) redundancy we encountered previously in the
discussion of the integral (9.14): the Grassmannian integral (9.14) is well-defined only when
“gauge fixing” the G L(k) redundancy. Because of the G L(k) redundancy, the dimensions
of the Grassmannian Gr(k, n) are k × n − k 2 = k(n − k).
With this geometric picture in mind, let us now examine the bosonic delta functions in
the gauge-fixed expression (9.17) for Ln,k . They enforce the constraints

n
n
Cai [i|a = 0 , C̃a i i| = 0 , (9.18)
i=1 i=1
where C and C̃ are k × n and (n − k) × n matrices respectively, so a = k + 1, . . . , n. They

are explicitly given as
⎛ ⎞ ⎛ ⎞
1 0 · · · 0 c1,k+1 ··· c1n −c1,k+1 · · · −ck,k+1 1 0 ··· 0
⎜0 1 · · · 0 c2,k+1 ··· c2n ⎟ ⎜−c1,k+2 · · · −ck,k+2 0 1 · · · 0⎟
⎜ ⎟ ⎜ ⎟
C = ⎜. .. .. .. .. .. .. ⎟ , C̃ = ⎜ .. .. .. .. .. .. .. ⎟ .
⎝ .. . . . . . . ⎠ ⎝ . . . . . . .⎠
0 ··· 0 1 ckk+1 · · · ckn −c1n ··· −ckn 0 0 ··· 1
(9.19)
186 Grassmannia
An important feature is that

n
C C̃ T = Cai C̃a i = 0 . (9.20)
i=1
Exercise 9.2
Construct C̃ associated with (9.15) and check that (9.20) holds.
We can view C̃ as (n − k) n-vectors spanning an (n − k)-plane in n dimensions. The

condition (9.20) states that the (n − k)-plane defined by C̃ is the orthogonal complement
of the k-plane defined by C.
In this notation, we can reinstate the G L(k) redundancy and write our momentum space
Grassmannian integral (9.17) as
$ '
d n×k C #
k
Ln,k = δ2 i C ai [i| δ
(4)
i C ai ηi A
G L(k) nj=1 M j a=1
$ '
#
n
× δ 2
i C̃ a i i| , (9.21)
a =k+1
with the understanding that C̃ is defined as the complement to C in the sense of (9.20).
Now a geometric picture is emerging of the meaning of the constraints (9.18). The
collection of the n |is defines a 2-plane in an n-dimensional space,

|11̇ |21̇ · · · |n1̇
. (9.22)
|12̇ |22̇ · · · |n2̇
Similarly the [i|s define a 2-plane in an n-dimensional space.

The constraints (9.18) say that the 2-plane spanned by the [i|s is orthogonal to the k-
plane C and that the 2-plane defined by the |is is orthogonal to the (n − k)-plane C̃. This
is illustrated in Figure 9.1. Since C̃ and C are orthogonal complements, we immediately
conclude that C̃ must contain the 2-plane [i| while C must contain |i. This in turn tells us
that the 2-plane of [i| must be orthogonal to the 2-plane of |i, i.e.

n
|i[i| = 0 . (9.23)
i=1
This is just the statement that the external momenta satisfy momentum conservation.
Thus, seemingly out of nowhere, the cyclic- and Yangian-invariant generating function Ln,k
“knows” about momentum conservation!
As we have just shown, the bosonic delta functions in (9.21) give non-vanishing results
only on the support of momentum conservation, δ 4 (P). With this in mind, let us count
the number of “free” integration variables in (9.21), i.e. the number of cai s not localized
by the bosonic delta functions. After gauge fixing G L(k), we have a total of k × (n − k)
2−plane [i|
2−plane |i〉
k−plane C
∼
(n−k)−plane C
The geometry of planes in the Grassmannian. The k-plane C and (n − k)-plane C̃ are orthogonal complements, Fig. 9.1

so the constraint i Cai [i| = 0 in (9.18) means that the 2-plane spanned by the n [i|s is orthogonal to C and
hence must be contained in C̃. Similarly, C contains the 2-plane spanned by the |is. It follows from the geometry
that the 2-planes spanned by [i| and |i, respectively, are orthogonal, but that is exactly the statement of momentum
n
conservation i=1 |i[i| = 0.
cai -variables. There are [2k + 2(n − k)] = 2n bosonic delta functions, but this includes
momentum conservation δ 4 (P), so four of the delta functions do not localize any cai -
variables. Therefore, a total of
#(integration variables) = [k × (n − k) − (2n − 4)] = (k − 2)(n − k − 2) (9.24)
cai -variables are left to be integrated.
In the MHV sector, k = 2 so we learn from (9.24) that the integral (9.21) is fully localized
by the bosonic delta functions. For generic k and n, Ln,k is a multi-dimensional integral that
localizes on the poles in the minors; we work out an explicit example in the next section. For
k = 0, the integral vanishes since it is proportional to δ 2 (|i) which does not have support for
generic momenta. This is simply the statement that the “all-plus” gluon amplitude vanishes
in N = 4 SYM. For k = 1, the last delta function in (9.21) forces all of the |is to be
proportional to each other, but this lacks support for generic momenta, with the exception
of special kinematics for n = 3. Not surprisingly, this says that the all-plus-and-one-minus
gluon amplitudes vanish in N = 4 SYM for n > 3.
The counting of integration variables in (9.24) is invariant under k → (n − k − 2). This
corresponds to a flip of what we identify as positive and negative helicity, i.e. which
states are associated with highest/lowest Grassmann weight. Indeed, for k = n − 2, the
(super)amplitude is anti-MHV, so it makes sense that the Grassmannian integral is localized
completely by the bosonic delta functions, just as it is for the MHV sector k = 2.
Perhaps you have noticed that the dimension (k − 2)(n − k − 2) of the Ln,k -integral is
also the dimension of the Grassmannian Gr(k − 2, n − 4). This is not a coincidence. The
bosonic delta functions enforce that Cai contains the 2-plane |i, so using G L(k) redundancy
we can choose the first two rows of Cai to be |i. The 2k − 4 delta functions (where the
−4 comes from momentum conservation) impose that the remaining (k − 2)-plane in
188 Grassmannia
n-dimensions is orthogonal to | j]. This places 2(k − 2) constraints on Cai . Taking the
G L(k − 2) redundancy into account, the number of free variables in Cai is therefore n(k −
2) − (k − 2)2 − 2(k − 2) = (k − 2)(n − k − 4). This is the dimension of the Grassmannian
Gr(k − 2, n − 4).
So far we have constructed a cyclic and Yangian invariant integral Ln,k in the Grass-
mannian, but we have not really done anything with it. In fact, other than showing that it
captures momentum conservation, we have given you little reason to believe that there is
any connection to scattering amplitudes in planar N = 4 SYM. So now we had better show
how it works.
9.3 Yangian invariants as residues in the Grassmannian
We carry out the Grassmannian integral (9.21) in the simplest cases to illustrate how the
familiar MHV and NMHV superamplitudes appear.
9.3.1 MHV amplitudes

By the counting (9.24), the bosonic delta functions completely localize the integral (9.21) for
the MHV sector (k = 2). Furthermore, the geometric description in the previous section tells
us that for MHV, the bosonic delta functions exactly encode conservation of 4-momentum
on the n external states: δ 4 (P). What about the Grassmann delta function in (9.21)? Well,
when k = 2, C defines a 2-plane, and since the 2-plane |i must be contained in C, we can
simply identify the two 2-planes; up to a G L(2) transformation we therefore have

C11 C12 · · · C1n |11̇ |21̇ ··· |n1̇
= . (9.25)
C21 C22 · · · C2n |12̇ |22̇ ··· |n2̇
Using this explicit representation of the Cai s, the Grassmann delta function becomes the
familiar statement of supermomentum conservation,
#
2 #2

δ (4) C η
i ai i = δ (4) i |i ηi = δ
ȧ (8)
Q̃ . (9.26)
a=1 ȧ=1
Finally, consider the minors in (9.21):
Mi = ab Cai Cb,i+1 = −ȧ ḃ |iȧ |i + 1ḃ = −i, i + 1 . (9.27)
Putting everything together, we find
δ (8) ( Q̃) δ 4 (P)

Ln,2 = (−1)n n = (−1)n AMHV
n,tree . (9.28)
i=1 i, i + 1
So for k = 2, the cyclic invariant integral Ln,2 nicely produces the MHV tree amplitude
(up to an overall convention-dependent sign).
Example. Did the argument above go a little fast? Fair enough, let us evaluate Ln,2 in
full detail, starting with the gauged-fixed expression (9.17), which for k = 2 gives
$ 2 '
d (n−2)×2 c # 2
n n
Ln,2 = δ [a| + cal [l| δ (4)
ηa + cal ηl
M1 · · · Mn a=1 l=3 l=3
$ n ' (9.29)
#
× δ 2 |i − |1c1i − |2c2i .
i=3
The last set of delta functions localize the 2(n − 2) components c1i and c2i :

1 i2 i1
δ 2 |i − |1c1i − |2c2i = δ c1i − δ c2i − . (9.30)
12 12 21
Thus, on the support of these delta functions, the first four bosonic delta functions in
(9.29) give
n n
δ [1| +
2
c1l [l| δ [2| +
2
c2l [l| = 122 δ 4 (P) . (9.31)
l=3 l=3
This is how the momentum conservation delta function appears.

Likewise for the Grassmann delta function: on the support of (9.30) it gives
#
2 n
1

δ (4) ηa + cal ηl = δ (8) Q̃ . (9.32)
a=1 l=3
124
Finally, we evaluate the minors Mi . With the help of the Schouten identity we find
23 34 45 n1
M1 = 1 , M2 = , M3 = − , M4 = − , . . . , Mn = − ,
12 12 12 12
(9.33)
and hence
n
#
n
(−1)n #
Mi = i, i + 1 . (9.34)
i=1
12n i=1
Inserting everything into (9.29) we indeed obtain (−1)n times the MHV tree superam-
plitude, just as in (9.28).
Exercise 9.3
Derive (9.30)–(9.34). Then plug the results into (9.29) to verify that all powers of 12
cancel.
The Grassmannian integral Ln,k has given a unique result, AMHV

n,tree , for k = 2. Given that
the Grassmannian integral produces all possible Yangian invariants [122, 123], it must be
190 Grassmannia
that all MHV superamplitudes have the same Leading Singularities in planar N = 4 SYM,
up to a sign, to all loop-orders. We have already seen a non-trivial manifestation of this
in Section 8.2 (and again in Section 8.3), where the Leading Singularities of the 1-loop
and 2-loop 4-point superamplitudes were found to be the MHV tree superamplitude. We
now know that even if we were to take the 234-loop MHV superamplitude and solve the
on-shell constraints that localize the 234 × 4 = 936 loop-momenta, the result of the Leading
Singularities will again be MHV tree superamplitudes! No other Yangian invariants are
available at MHV order.
9.3.2 6-point NMHV amplitudes

Let us now move on to a slightly more complicated – hence more exciting – example, the
6-point NMHV amplitude.1 With k = 3 and n = 6, the counting formula (9.24) reveals that
the Grassmannian integral L6,3 involves just one non-trivial integration. To evaluate it, we
choose the gauge
⎛ ⎞
c21 1 c23 0 c25 0
⎝c41 0 c43 1 c45 0⎠ . (9.35)
c61 0 c63 0 c65 1
The c-variables are labeled such that the bosonic delta functions in (9.21) can be written
as:
⎛ ⎞ ⎛ ⎞

δ 2 ⎝[ ī | + c ī j [ j|⎠ , δ 2 ⎝| j − | ī c ī j ⎠ , (9.36)
j ī
where ī = 2, 4, 6 and j = 1, 3, 5. Since the integral is 1-dimensional, there must be a 1-

parameter family of solutions that solve the delta function constraints (9.36). Indeed, if c∗ī j
is a solution, then
ĉ ī j (τ ) = c∗ī j + 14 τ ī j̄ k̄ j̄ k̄ jkl [kl] , (9.37)
is also a solution for any τ . Here ī j̄ k̄ is a Levi-Civita symbol for the indices ī = 2, 4, 6,
and similarly for jkl . There are implicit sums over repeated labels in (9.37). That ĉ ī j (τ ) is
a solution can be seen from the result that the τ dependence drops out from the constraints
in (9.36) due to the Schouten identity:

[ ī |a + ĉ ī j (τ )[ j|a = 14 τ ī j̄ k̄ j̄ k̄ jkl [ j|a [kl]
j j

= 1
2
τ ī j̄ k̄ j̄ k̄ [1|a [35] + [3|a [51] + [5|a [13] = 0 . (9.38)
We can now remove the bosonic delta functions by localizing the integral on the solution to
the constraints (9.36) such that the remaining integral is over the 1-dimensional parameter
1 For a much simpler evaluation of the NMHV residues in momentum twistor space, see Elvang et al. [60].
τ . That gives
&# %
d 9 c ī j
L6,3 = δ | j −
2
| ī c ī j
M1 · · · Mn j
ī
&# %
× δ 2 [ ī | + c ī j [ j| δ (4) η ī + c ī j η j
ī j j
9 #
d c ī j dτ 9
= δ 4 (P) δ c ī j − ĉ ī j (τ ) δ (4) η ī + c ī j η j
M1 · · · Mn j
ī
#
dτ
= δ 4 (P) δ (4) η ī + ĉ ī j η j . (9.39)
M̂1 · · · M̂n ī j
The “hat” indicates dependence on τ via (9.37). In the gauge (9.35), the minors are
M̂1 = ĉ43 ĉ61 − ĉ41 ĉ63 , M̂3 = ĉ23 ĉ65 − ĉ25 ĉ63 , M̂5 = ĉ21 ĉ45 − ĉ25 ĉ41 ,
(9.40)
M̂2 = −ĉ63 , M̂4 = −ĉ25 , M̂6 = −ĉ41 .
At this stage, there appears to be no a priori prescription of which contour to pick in the
τ -plane. Each minor M̂i has a simple pole in τ , so there are six different residues that we
denote {Mi }. Let us focus on the pole in M̂4 . This means that τ is evaluated at τ∗ such that
ĉ25 (τ∗ ) = 0. We can make the calculation simpler by choosing the origin for τ such that
∗
M̂4 = 0 for τ = 0; in other words, we choose ĉ25 = 0. Let us use the constraints (9.36) to
∗
solve for the eight other ĉ ī j s. From
∗ ∗ ∗ ∗
|5 − |4c45 − |6c65 = 0, [2| + c21 [1| + c23 [3| = 0 , (9.41)
we deduce
∗ 56 ∗ 45 ∗ [23] ∗ [12]
c45 = , c65 = , c21 =− , c23 =− . (9.42)
46 46 [13] [13]
And this in turn allows us to solve
∗ ∗ ∗ ∗ ∗ ∗
[4| + c41 [1| + c43 [3| + c45 [5| = 0, [6| + c61 [1| + c63 [3| + c65 [5| = 0 , (9.43)
to find
∗ 6|4+5|3] ∗ 6|4+5|1] ∗ 4|5+6|3] ∗ 4|5+6|1]
c41 =− , c43 = , c61 = , c63 =− .
46[13] 46[13] 46[13] 46[13]
(9.44)
Exercise 9.4
Use the above results for c∗ī j to show that the unused constraints in (9.36) give δ 4 (P).
We can now substitute the solutions cî∗j into the minors (9.40) and the Grassmann delta
functions to obtain the residue of the integral (9.39) of the pole 1/M4 , denoted by {M4 }.
For simplicity, consider a particular component amplitude, namely the gluon amplitude
with helicity assignments (+, −, +, −, +, −). For this amplitude, the coefficient from the
192 Grassmannia
Grassmann delta functions is just 1. Taking into account the extra factor of 46[13] coming
from ĉ25 = −τ 46[13], we find
464 [13]4
{M4 } = . (9.45)
4|5+6|1]6|4+5|3][21][23]5456P456
2
Exercise 9.5

Show that M̂1 τ =0 = 46[13]
2
P456
. Evaluate the other minors at τ = 0 and use them to derive
the result (9.45) for the residue at τ = 0.
We could calculate the residues associated with each of the other minors similarly. {M6 }
and {M2 } are just cyclic permutations of {M4 } by two sites, so we have
624 [35]4
{M6 } = ,
6|1+2|3]2|6+1|5][43][45]1612P612
2
(9.46)
244 [51]4
{M2 } = .
2|3+4|5]4|2+3|1][65][61]3234P234
2
For the residues {M1 }, {M3 }, {M5 }, it is convenient to choose the gauge
⎛ ⎞
1 c12 0 c14 0 c16
⎝0 c32 1 c34 0 c36 ⎠ . (9.47)
0 c52 0 c54 1 c56
Then following the same steps as before we find that the {M1 } residue for the
(+, −, +, −, +, −) amplitude is
−6|2+4|3]4
{M1 } = . (9.48)
1|5+6|4]5|6+1|2][23][34]5661P561
2
The other residues, {M3 } and {M5 }, are obtained by relabeling the external states in (9.48).
We have now extracted six residues {Mi } from (9.39) for a projection that corresponds to
the helicity configuration (+, −, +, −, +, −) of a gluon amplitude. But it is not yet clear
what the residues have to do with the amplitude. Each of the {Mi }s contains spurious poles,
such as 4|5+6|1] in {M2 } and {M4 }. However, in the sum {M2 } + {M4 }, this spurious
pole cancels. In fact, in the sum {M2 } + {M4 } + {M6 } all three spurious poles – 4|5+6|1],
6|4+5|3], and 2|6+1|5] – cancel, so this is a local object. Your brain may even be
tingling with the sensation that you have seen this combination before. Go back to look at
Exercise 3.9: there we calculated the 6-point tree amplitude A6 [1+ 2− 3+ 4− 5+ 6− ] from a
[2, 3-BCFW shift and found that it was exactly
A6 [1+ 2− 3+ 4− 5+ 6− ] = {M2 } + {M4 } + {M6 } . (9.49)
In the BCFW construction, each of the three terms in (9.49) corresponds exactly to a BCFW
diagram. Now we have also seen that each term can be understood as the residue of a pole
τ τ
M2 M2
M1 M1
= −
M6 M6
M4 M4
M3 M3
M5 M5
The “tree contour” in the Grassmannian. It circles the residues of the poles {M2 }, {M4 }, {M6 }. Through contour Fig. 9.2
deformation, the result is equivalent to minus the sum of {M1 }, {M3 }, {M5 }.
associated with the minor Mi in the cyclically invariant Grassmannian integral. So for
N = 4 SYM, individual BCFW terms are in one-to-one correspondence with the residues
of the Grassmannian integral. Since the Grassmannian integral was constructed to produce
Yangian invariants, we now understand that each super-BCFW term is a Yangian invariant.
Consider the contour that encircles the minors {M2 }, {M4 }, {M6 }. It is this contour that
gives a Yangian invariant rational function that is local and free of spurious singularities.
The statement of locality has become a choice of contour.
Through contour deformation, illustrated schematically in Figure 9.2, we have
{M2 } + {M4 } + {M6 } = −{M1 } − {M3 } − {M5 } . (9.50)
This means that the tree amplitude A6 [1+ 2− 3+ 4− 5+ 6− ] can also be represented by (minus)
the sum of {M1 }, {M3 }, and {M5 }. Indeed, this is the representation that one obtains
from the BCFW shift [3, 2, the “parity conjugate” of the shift [2, 3 that produced the
{M2 }, {M4 }, {M6 } representation. Actually, this is a little too quick, because the shift [3, 2
would be an illegal [+, −-shift for the component-amplitude A6 [1+ 2− 3+ 4− 5+ 6− ]. The
correct statement is that the {M1 }, {M3 }, {M5 } representation is the result of a [3, 2 BCFW
super-shift recursion relation with a subsequent projection to the (+, −, +, −, +, −) gluon
helicity states.
The insight gained here is that the mysterious six-term identity (9.50) that arises from
the equivalence of the two conjugate BCFW super-shifts [2, 3 and [3, 2 is simply a
consequence of the residue theorem of the Grassmannian integral Ln,k ! Actually, the identity
(9.50) is the 5-bracket six-term identity (5.63) projected to the (+, −, +, −, +, −) gluon
helicity states. In Chapter 10 we expose an underlying geometric interpretation of such
identities.
The Grassmannian, the tree contour, and the twistor string

In Witten’s twistor string [54], mentioned briefly in Section 5.2, the N K MHV superampli-
tudes in N = 4 SYM are calculated as open string current algebra correlators integrated
over the moduli space of degree (K +1) curves in supertwistor space. It turns out that
194 Grassmannia
the so-called RSV connected prescription [124] for the twistor string has a direct relation
to the BCFW recursion relations. Moreover, different BCFW representations are related
via (higher-dimensional versions of) Cauchy’s theorem [125]. Sounds similar to the prop-
erties of the Grassmannian integral, does it not? In fact, it can be shown that the tree
contour in the Grassmannian precisely gives the RSV connected prescription for the twistor
string.
Let us be more explicit about how this works. Recall the procedure for carrying out the
Grassmann integral Ln,k in (9.21). Gauge fixing G L(k) leaves nk − k 2 integration variables
of which the bosonic delta functions fix 2k + 2(n − k) − 4 = 2n − 4. (The −4 is momen-
tum conservation.) So this leaves nk − k 2 − (2n − 4) = (k − 2)(n − k − 2) variables and
the strategy is to fix those by localizing all of them on zeroes of the minors in the denomi-
nator of Ln,k . Reversing the procedure, we ignore the delta functions to begin with and first
localize on the minors: this leaves an integral with nk − k 2 − (k − 2)(n − k − 2) = 2n − 4
variables. These will of course be fixed by the bosonic delta functions, but note that 2n − 4
is the dimension of the Grassmannian Gr(2, n). Thus, after localizing on the minors, the
integral Ln,k naturally lives in Gr(2, n). Now the punchline is that precisely with the tree
contour, reducing the Gr(k, n)-integral to a Gr(2, n)-integral results in the twistor string
formula after a suitable Fourier transform [126]. Thus the underlying property that makes
the tree contour special is that it localizes the Gr(k, n) Grassmannian integral to Gr(2, n)
in a particular fashion that is intimately tied to locality and the twistor string. We will
see this story repeat itself when we consider the Grassmannian formula for the 3-
dimensional ABJM theory in Section 11.3. There are several papers in the litera-
ture on the relationship between amplitudes and the twistor string, see for example
[54, 124–129].
The Grassmannian picture is interesting and has given us insight about locality, but there
is perhaps a small stone in our shoe: we have not yet been able to see how the cancellation
of spurious poles takes place within the tree superamplitude in a manifest fashion. There is
a geometric story about how that happens and we draw it in Chapter 10.
9.4 From on-shell diagrams to the Grassmannian
The terms appearing in the BCFW expansion of the 6-point NMHV amplitude have now
appeared in two distinct guises. We have seen in the previous section that they are given as
residues of an integral over a Gr(3,6) Grassmannian manifold. In Section 8.3, they were the
result of gluing on-shell cubic vertices together into on-shell diagrams. So can we make a
connection between the Grassmannian and the on-shell diagrams? Yes, we can!
In Section 8.3, we did not explicitly compute any of the on-shell diagrams beyond the
simplest box-diagram. The reason is simple: explicitly solving all momentum conservation
constraints at each vertex is a complicated task because it is quadratic in spinor variables.
In this section, we have seen that momentum conservation can be converted into a linear
constraint with the aid of the Grassmannian. This means that if we convert all 3-point vertices
in an on-shell diagram into Grassmannian integrals, then the momentum conservation
constraints are just a set of linear equations. Let us see how this is done in practice. For the
MHV 3-point amplitude, the Grassmannian integral is simply Gr(2,3):

d 2×3 C

AMHV
3 = δ 2×2 Ci [i| δ (4×2) Ci ηi δ 2×1 C̃i i| , (9.51)
M1 M2 M3
where we have used the momentum space representation (9.21): C is a 2 × 3 matrix, and
C̃ is its 1 × 3-dimensional orthogonal complement. For the anti-MHV 3-point amplitude,
the analogue Grassmannian integral in Gr(1,3) is

d 1×3 C

Aanti-MHV
3 = δ 2×1 Ci [i| δ (4×1) Ci ηi δ 2×2 C̃i i| , (9.52)
M1 M2 M3
where now C is a 1 × 3-dimensional matrix and C̃ is its 2 × 3-dimensional orthogonal

complement. To see that this indeed gives the correct 3-point amplitude, note that the first
bosonic delta function requires that C is orthogonal to the 2-plane [i|, so this localizes the
integral (up to an irrelevant overall rescaling) to

C = [23] , [31] , [12] . (9.53)
Substituting this into (9.52), one indeed recovers the anti-MHV 3-point amplitude.
Exercise 9.6
Given (9.53), determine a representation of C̃ai . Substitute the result into (9.52) to
recover the 3-point amplitude. Note that all potential Jacobian factors can be fixed by
dimension-counting and symmetry analysis.
In summary, the MHV and anti-MHV 3-point amplitudes can be viewed as providing 2 × 2
and 1 × 2 linear constraints respectively for the [i|s. For later convenience, we parameterize
the Gr(2, 3) and Gr(1, 3) Grassmannians as follows:
b c a b

1 0 αb
a , (9.54)
0 1 αc
c
a b c b

1 βb βc a . (9.55)
c
The arrows represent the particular gauge that we have chosen for each Grassmannian
integral. Incoming lines on the 3-point vertex indicate that the corresponding columns
in the Grassmannian are G L(k)-gauge-fixed to be the identity matrix. Outgoing lines
correspond to unfixed columns. In this gauge, the [i| parts of the bosonic delta functions
196 Grassmannia
are:

MHV: δ 2 [b| + αb [a| δ 2 [c| + αc [a| , anti-MHV: δ 2 [a| + βb [b| + βc [c| .
(9.56)
For each vertex, the spinors of the incoming lines are expressed as a linear combination
of those of the outgoing lines. One can perform a similar analysis for the fermionic delta
functions and the bosonic delta functions of i|. The analyses are exactly parallel, so we
leave them implicit.
Exercise 9.7
What does the bosonic delta function for i| look like? What constraints does it
impose?
We are now ready to start gluing 3-point on-shell vertices together. Recall from (8.49) that
each internal line in the on-shell diagram corresponds to an integral over the set of internal
variables,
2
d |I d 2 |I ] d 4 η I
. (9.57)
U (1)

Since the spinors |I , |I ] also appear in the vertices on each end of the line, the bosonic
delta functions of these vertices can be used to localize the integral (9.57). This can be
made manifest in a graphical way. For each on-shell diagram, we decorate the lines with
arrows following the rule that for each black vertex, there should be two incoming lines and
one outgoing line, while for each white vertex there should be one incoming line and two
outgoing lines, just as in (9.54), (9.55). One might wonder if it is always possible to find such
decoration consistent throughout the on-shell diagram. For diagrams of physical relevance
the answer is yes, since one can interpret the outgoing lines as + helicity, incoming lines
as − helicity, and a consistent decoration is equivalent to consistent helicity assignments.
For example, consider gluing six vertices together to form a double box diagram. We can
have a consistent decoration if there is at least one different color vertex, but not if they are
all the same:2
. (9.58)
The lines of the second diagram cannot be consistently oriented. It follows from the previous
discussion that the spinors of the internal line are completely determined by the outgoing
2 You may find it a little peculiar that we show examples corresponding to on-shell diagrams for all-plus amplitudes;
the point is just that they give simple illustrations of the consistency conditions.
lines of one of the vertices. For example, decorations of the diagram

2 3
I
. (9.59)
1 4
dictate the bosonic delta functions for the square spinors to be

δ 2 [2| + α2 [1| , δ 2 [I | + α I [1| , δ 2 [3| + β4 [4| + β I [I | . (9.60)
2
The second bosonic delta function localizes the d |I ] integral, while the remaining two
delta functions become

δ 2 [2| + α2 [1| , δ 2 [3| + β4 [4| − β I α I [1| . (9.61)

The delta functions in (9.61) can be combined to the form 2a=1 δ 2 (Cai [i|) with the Gr(2,4)
Grassmannian given as
1 2 3 4

α2 1 0 0
Cai = . (9.62)
−β I α I 0 1 β4
So gluing the two 3-point vertices together now gives a new Grassmannian integral

dα2 dα I dβ4 dβ I 1

δ 2×2 Ci [i| δ (4×2) Ci ηi δ 2×2 C̃i i| , (9.63)
α2 α I β4 β I U (1)
where the Cai is identified in (9.62).
Note the leftover 1/U (1) in (9.63). We have been treating (|I , |I ]) as independent
variables, each being fixed by the bosonic delta functions. However, there remains a gauge-
fixing functional that is present to remove the little-group redundancy. This functional is
represented by this 1/U (1) factor. We do not need its explicit form, just remember that it
can be used to fix an additional degree of freedom.
The above simple example generalizes to an arbitrary decorated on-shell diagram. A
diagram with n b black vertices and n w white vertices contains 2 × (2n b + n w ) constraints
on the |i]s. If it has n I internal lines, then the 2 × n I integrations over the internal |I ]s can
be localized by these bosonic delta functions. Finally, there will be 2 × (2n b + n w − n I )
constraints left and they can be conveniently grouped into a degree 2 × k delta func-

tion ka=1 δ 2 (Cai [i|) where k = (2n b + n w − n I ) and n = 3(n w + n b ) − 2n I . Note that by
counting the Grassmann degrees of the black (2 × 4) and white (1 × 4) blobs minus the n I
internal Grassmann integrations (n I × 4), this k is exactly the same k as in the Nk+2 MHV
classification. Thus each on-shell diagram corresponds to the following Grassmannian
integral:
# nb
n
#
n
#

dαi1 dαi2 w
dβi1 dβi2 I
1

δ 2×k Ci [i| δ (4×k) Ci ηi δ 2×(n−k) C̃i i| .
i=1
αi1 αi2 i=1
βi1 βi2 i=1
U (1)i
(9.64)
198 Grassmannia
As we have seen, this Gr(k, n) Grassmannian integral is parameterized by an on-shell

diagram decorated with arrows consistently throughout the diagram. The Grassmannian
integral (9.64) is in a G L(k)-gauge-fixed form. Each vertex contains two degrees of freedom,
but the n I internal lines each leave a 1/U (1) gauge-fixing function, so the dimension of
this Grassmannian integral is
dim(C) = 2 × n v − n I , (9.65)
where the number of vertices is n v = n b + n w . Using Euler’s formula for a planar diagram
(n f − n) − n I + n v = 1, where n f is the number of faces in a diagram, we find that the
dimension of the Grassmannian integral corresponding to a particular on-shell diagram is
dim(C) = n f − 1 . (9.66)
The total number of bosonic delta functions is 2 × (k + n − k) − 4 = 2n − 4. If dim(C) =
2n − 4, then all the degrees of freedom in the integral are completely localized by the
bosonic delta functions. This is the case for the on-shell diagrams that correspond to the
BCFW terms in Section 8.3. We saw this in Section 9.3.2 for Gr(3,6): prior to solving the
bosonic delta functions, each BCFW term is obtained by localizing on the zeroes of one of
the minors. This is precisely the 9 − 1 = 8-dimensional Grassmannian manifold indicated
by the on-shell diagrams.
If dim(C) < 2n − 4, then the bosonic delta functions over-constrain the external data and
can only be satisfied in special kinematics. This is precisely the scenario for the example
that led to (9.62), where dim(C) = 3 < 4. From (9.62) one can readily read off what the
special kinematics is: [1| ∼ [2|.
What happens if dim(C) > 2n − 4? It corresponds to a term in the BCFW representation
of a loop amplitude, where the remaining integrations can be translated into the loop-
momentum integration! For example, consider attaching a BCFW bridge on the forward
limit of (8.55). In our calculation of the loop-recursion in Section 7.5, only the BCFW term
that corresponds to the middle on-shell diagram of (8.55) has a non-vanishing forward limit
contribution. If we attach a BCFW bridge to this middle diagram and perform a series of
equivalence moves, we find:
= .
I
I
1 4
1 4
(9.67)
The final diagram is precisely the 1-loop 4-point amplitude. One can count that with
n = 4, dim(C) = 9 − 1 = 4 + (2n − 4), i.e. four integrals remain after solving all bosonic
delta functions. These four extra integrals correspond to the integration over the four
components of the loop-momentum. We can readily identify these extra components in the
on-shell diagram: since the original tree-diagram contains no extra integration variable, the
new degrees of freedom must arise from the procedure of taking the forward limit. This
introduces a factor of d |I d 2 |I ] d 4 η I /U (1) in (9.67); that is three integrations because
2
thedzU (1) is mod’ed out. The presence of the BCFW bridge introduces the fourth integration
z
. The loop-momentum can then be written
= |I [I | + z|1[4| . (9.68)
It is quite remarkable that starting out with a fully on-shell construction of “on-shell
diagrams” leads to a loop-integrand construction in which the loop-momentum is off-shell.
Exercise 9.8
Use equivalence moves to prove the last two diagrams of (9.67) are equivalent to each
other.
There is much more information in the connection between the Grassmannian and the
on-shell diagrams, but we also have other fish to fry and birds to scare. If we have awoken
your appetite for blob diagrams and Grassmannians and you are interested in learning
more about their relations to permutations, stratifications, amalgamation, dimers, bipartite
graphs, and quivers, you should take a look at reference [120] for more details.
10 Polytopes
In Chapter 9, we learned that the individual terms in a BCFW expansion of an Nk−2 MHV
n-point superamplitude are residues of a cyclically invariant integral-formula in the Grass-
mannian Gr(k, n). In this language, we found that the different BCFW representations of the
6-point NMHV amplitude A6 [1+ 2− 3+ 4− 5+ 6− ] are related by a simple contour deforma-
tion. This was encoded in the 6-term identity (9.50). In momentum supertwistor space [60],
the 6-term identity is promoted to the 5-bracket relation
[2, 3, 4, 6, 1] + [2, 3, 4, 5, 6] + [2, 4, 5, 6, 1] = [3, 1, 6, 5, 4] + [3, 2, 1, 6, 5]

+ [3, 2, 1, 5, 4] . (10.1)
We have already encountered this version of the 6-term identity in (5.63) when we discussed
the equivalence of the BCFW recursion relations based on [2, 3 and [3, 2 super-shifts.
The 5-brackets were defined in Section 5.4 as

δ 4 χi A jklm + cyclic
[i, j, k, l, m] ≡ , (10.2)
i jkl jklmklmilmi jmi jk
with 4-brackets i jkl ≡ I J K L Z iI Z Jj Z kK Z lL involving the bosonic components Z iI =
(|i, [μi |) of the SU (2, 2|4) momentum supertwistors ZiA ≡ (|iȧ , [μi |a | χi A ), A =
(ȧ, a, A). The [μi | are defined by the incidence relations (5.44).
Recall that momentum conservation is automatic for momentum twistors, so the 6-term
identity (10.1) must hold for any six momentum twistors; it is not specific to the NMHV
6-point amplitude but is an intrinsic property of the 5-brackets (10.2). Hence it seems
worthwhile to try to understand the structure of (10.1) better. Using cyclic and reflection
symmetry of the 5-brackets, we can rewrite (10.1) as
[1, 2, 3, 4, 5] − [2, 3, 4, 5, 6] + [3, 4, 5, 6, 1] − [4, 5, 6, 1, 2]

+ [5, 6, 1, 2, 3] − [6, 1, 2, 3, 4] = 0 . (10.3)
Pretend for a moment that we do not know what the 5-brackets are. Consider a fully anti-
symmetric 5-bracket i, j, k, l, m defined as the contraction of five 5-component vectors
ZiI with a 5-index Levi-Civita tensor. Such an object would satisfy the 6-term Schouten
identity
1, 2, 3, 4, 5 ZI6 − 2, 3, 4, 5, 6 ZI1 + 3, 4, 5, 6, 1 ZI2

− 4, 5, 6, 1, 2 ZI3 + 5, 6, 1, 2, 3 ZI4 − 6, 1, 2, 3, 4 ZI5 = 0 . (10.4)
Since this looks quite similar to (10.3), including relative signs, we might be tempted
to think that (10.3) somehow arises as a Schouten identity. This is of course too
10 Polytopes 201
speculative: the 5-brackets [i, j, k, l, m] really represent rational functions of the 4-

component momentum twistors Z i , not some 5-index objects contracted with a 5-index
Levi-Civita tensor. However, we can entertain the idea a little further. Could the 5-bracket
(10.2) be written in terms of some new 5-vectors? The fermionic variables χi A appear in
(10.2) on a different footing than their bosonic counterparts. In the name of democracy, let
us define the following purely bosonic 5-component vector

I Z iI
Zi = , I = 1, . . . , 5 , (10.5)
χi · ψ
where χi · ψ = χiA ψ A is SU (4)-invariant and ψ A is an auxiliary Grassmann variable com-
mon for all external particles i = 1, 2, . . . , n. If we define i, j, k, l, m as the contraction
of five of these 5-vectors with a 5-indexed Levi-Civita tensor, then they will satisfy the
Schouten identity (10.4) – but that is not what we are after, so read on.
To write the 5-bracket [i, j, k, l, m] in terms of the 5-vectors (10.5), we must remove the
auxiliary variable ψ A . Since it is fermionic, this can be done via a Grassmann-integration:
one finds that the 5-bracket (10.2) can be written as
[i, j, k, l, m]

1 i, j, k, l, m4
= d 4ψ , (10.6)
4! 0, i, j, k, l0, j, k, l, m0, k, l, m, i0, l, m, i, j0, m, i, j, k
where i jklm ≡ I J K L M Z iI Z Jj Z kK Z lL Z mM and we have introduced the auxiliary reference
5-vector
⎛ ⎞
0
⎜0⎟
⎜ ⎟
⎜ ⎟
ZI0 = ⎜ 0 ⎟ . (10.7)
⎜ ⎟
⎝0⎠
1
The representation (10.6) is certainly not just contractions of five 5-vectors with a Levi-
Civita tensor, so the origin of the identity in (10.3) is not a Schouten identity. But let us not
give up just yet, for itwill be worthwhile to examine (10.6) further.
Since the integral d 4 ψ is universal for all 5-brackets, we ignore it for the time being
and focus on the integrand of (10.6). Each ZiI appears an equal number of times in the
numerator and the denominator, so the integrand is invariant under ZiI → ti ZiI for each
i = 1, 2, . . . , n. In other words, the 5-vectors ZiI appear projectively in (10.6), and therefore
we can think of the ZiI as homogeneous coordinates of points in projective space CP4 . The
presence of the reference vector ZI0 in the denominator breaks projective invariance, but
only at this particular point.
There is an analogous case where we have encountered something similar. The momen-
tum twistors Z iI in Section 5.4 are defined projectively and are elements in CP3 . The map
in Figure 5.1 shows how to relate momentum twistors Z iI with points yi in dual space.
Specifically, the distance between two points yi and y j in dual space is
i − 1, i, j − 1, j i − 1, i, j − 1, j
yi2j = = . (10.8)
i − 1, i j − 1, j I0 , i − 1, iI0 , j − 1, j
202 Polytopes
The first equality is simply (5.52): it has a momentum twistor 4-bracket in the numerator
and regular angle brackets in the denominator. In the second equality we have rewritten the
denominator in a more suggestive form involving only 4-brackets, at the cost of introducing
a reference bi-twistor I0I J defined as

0 0
I0 =
IJ
. (10.9)
0 ȧ ḃ
In the literature, I0 is often referred to as the infinity twistor,1 and its role is to break S L(4)
conformal invariance and provide a preferred metric for the definition of distance.
The expression (10.8) is similar to the integrand of (10.6): both are projectively defined,
except for the reference bi-twistor/vector. The reference bi-twistor appears twice in the
denominator of (10.8), reflecting the fact that this expression gives the distance between
the two points i and j. An analogous expression involving three points and a reference
vector appearing thrice defines the area of a triangle. And so on. The appearance of the
reference vector ZI0 five times in the denominator of (10.6) gives us a hint that the rational
integrand in (10.6) is the volume of a geometric figure defined by five points in CP4 ! In
the following, we pursue the interpretation of 5-brackets as volumes of simplices and their
sum – the superamplitudes – as volumes of polytopes. Definitions and explanations follow
next.
10.1 Volume of an n-simplex in CPn
Let us begin with the concepts of polytopes and simplices before reconnecting with the
motivation above.
Polytopes and simplices: definitions and examples

We are all familiar with polygons: triangles, squares (or more generally quadrilaterals),
pentagons, hexagons, chiliagons, star-shapes, etc. These are figures in the plane bounded
by a finite number of straight line-segments. Their 3-dimensional analogues – tetrahedrons,
cubes, prisms, dodecahedrons, etc – are solids whose faces are polygons. The n-dimensional
versions of polygons and polyhedrons are called polytopes or n-polytopes. A 2-polytope is
a polygon and a 3-polytope is a polyhedron.
A simplex is in a sense the simplest example of an polytope. To define it, recall first that
a convex set C (in, for example, Rn or CPn ) has the property that the line segment between
any two points in C lies entirely in C. In the plane, triangles are convex, but star-shaped
polygons are not. Given a set of points S, the convex hull of S is the intersection of all
convex sets containing S. Examples from the plane: 1) the convex hull of a circle is the
closed disk bounded by the circle; 2) the convex hull of three points is a triangle. Adding

ab 0
1 The infinity twistor in (10.9) corresponds to a flat space metric. For AdS4 it is given as I0I J = ,
0 ȧ ḃ
where is the cosmological constant.
a fourth point that lies inside the triangle, the convex hull of the four points is the same
triangle. A fourth point outside the triangle (but in the same plane) gives a convex hull that
is a convex quadrilateral, or in pictures:
2
2
1 3 . (10.10)
4
1 3
4
An n-simplex is the convex hull of a set of n + 1 points. Examples:
0-simplex = a point,
1-simplex = line segment,
(10.11)
2-simplex = triangle,
3-simplex = tetrahedron.

An n-simplex is bounded by n+1 (n−1)-simplices which intersect each other in n+1 2
(n−2)-simplices. For n+1 generic points in Rn , an n-simplex has an n-dimensional volume.
(For CPn it will be n-complex dimensional.) The volume of a polytope can be calculated
by “tessellating” it into simplices, whose volumes are easier to calculate.
Now that we know what simplices and polytopes are, let us progress towards understand-
ing how the integrand in (10.6) represents the volume of a 4-simplex in CP4 , as claimed.
As a warm-up, we begin in 2-dimensions with a 2-simplex (a triangle).
Area of a 2-simplex in CP2

The area of a triangle in a 2-dimensional plane can be computed as
x,y
x x2 x3
2 1 1
Area = y1 y2 y3 , (10.12)
2
1 3 1 1 1
where the (xi , yi ) are the coordinates of the three vertices.
Exercise 10.1
If the area formula (10.12) is not familiar, you should derive it by showing that it is
equivalent to the “ 12 × base × height” formula that was imprinted on your brain in
elementary school.
The 1s in the last row of (10.12) are redundant as we can write the same formula as a sum
of the 2 × 2 minors. In physics, when faced with a redundancy we can choose to eliminate
it or promote it to a feature. Choosing the latter, we define three 3-vectors along with a
reference vector:
⎛ ⎞ ⎛ ⎞
xi 0
Wi I = ⎝ yi ⎠ , Z0I = ⎝ 0 ⎠ , I = 1, 2, 3 . (10.13)
1 1
204 Polytopes
The area can now be written

2 x,y
1 1, 2, 3
Area = , (10.14)
2 (Z0 · W1 )(Z0 · W2 )(Z0 · W3 )
1 3
where the 3-bracket is the contraction of a 3-index Levi-Civita tensor with the three
Wi vectors: 1, 2, 3 = I J K W1I W2J W3K . Using (10.13), the 1, 2, 3-numerator exactly
equals the 3 × 3-determinant in (10.12), so you might consider the trivial dot-products
Z0 · Wi = Z0I Wi I = 1 in the denominator a provocation of your sense of humor. However,
written in the form (10.14), the redundancy has been promoted to projective symmetry: the
new area-formula (10.14) is invariant under scalings Wi → ti Wi . When the 3-vectors Wi
are “gauge fixed” to the canonical form in (10.13), we immediately recover the original area
formula. Since the triangle vertices are specified in terms of projectively defined 3-vectors,
we can think of the triangle as an object in CP2 and the Wi s as the homogeneous coordinates
of the vertices.
The area formula in (10.14) involves an antisymmetric 3-bracket as well as the inner
product of 3-vectors. To make contact with (10.6) and (10.8), we would like to have a
representation that is given solely in terms of 3-brackets. To achieve this, it is useful to
characterize the triangle by its edges instead of its vertices. We define a “dual space” whose
points ZaI are associated with lines in W-space: a given line is defined as the set of points
W I satisfying the incidence relations
ZI WI = 0 . (10.15)
Since Z I is a vector in the 2-dimensional space CP2 , the constraint indeed defines a 1-
dimensional subspace, i.e. a line.
Now, to define the triangle in terms of three lines in dual space, note that each Wi I is
characterized by lying simultaneously on two lines. Labeling the three edges of the triangle
as a, b, and c, the vertex W1I is the intersection of lines a and c, so that ZaI W1I = ZcI W1I = 0.
These two constraints are easily solved and we have
2
W1 = ∗, Zc , Za
a b → W2 = ∗, Za , Zb (10.16)
1 3 W3 = ∗, Zb , Zc ,
c
where the ∗ indicates the free index, e.g. W1I = ∗, Zc , Za = I J K ZcJ ZaK .
Plugging the map (10.16) into (10.14), we find that the area is now given as
2
1 a, b, c2
Area a b = ≡ a, b, c , (10.17)
1 3 2 0, b, c0, a, b0, c, a
c
where a, b, c = I J K ZaI ZbJ ZcK and “0” indicates the reference vector Z 0I introduced in
(10.13).
Exercise 10.2
Show that (10.17) follows from (10.14).
As advertised earlier, we now see that the “volume” (i.e. area) of a 2-simplex is given by a
rational function whose denominator is the product of all 3-brackets involving two of the
edge variables ZiI , i = a, b, c, and a reference vector Z0I . Requiring projective invariance
for each ZiI with i = a, b, c uniquely fixes the numerator. For later convenience, we have
introduced the notation [a, b, c] to denote the volume (10.17). Note that the area-formula
(10.17) comes with an “orientation” in the sense that [a, b, c] is fully antisymmetric in
a, b, c.
Volume of an n-simplex in CPn

The expression in (10.17) can be generalized to the volume of an n-simplex in CPn : we
denote it by an antisymmetric (n+1)-bracket
1 i 1 , i 2 , . . . , i n+1 n
Zi1 , . . . Zin+1 = , (10.18)
n! 0, i 1 , . . . , i n 0, i 2 , . . . , i n+1 · · · 0, i n+1 , i 1 , . . . , i n−1
where the angle-brackets are the contractions of n+1 CPn -vectors Zi with an (n+1)-index
Levi-Civita. The n+1 variables Zi1 , . . . , Zin+1 ∈ CPn carry the information about the n+1
boundaries of the n-simplex as follows. For a given vector ZiI , the set of W I s satisfying the
incidence relation ZiI W I = 0 span an (n−1)-dimensional subspace of CPn . These subspaces
contain the n + 1 faces of the n-simplex; each face is an (n−1)-simplex.
In the example of the 2-simplex in CP2 , the 3 Zi s label the 1-dimensional lines a, b, c
bounding the triangle. Each pair of lines intersect in a point that is a vertex of the triangle:
we can label the vertices (a, b), (b, c), and (c, a). They are defined in terms of the Zi s in
(10.16) and the denominator of the volume formula (10.17) is the dot-product of all vertex
point vectors with the reference vector.
As a second example, consider a 3-simplex (tetrahedron) in CP3 . The four 4-component
homogeneous coordinates of (dual) CP3 – ZaI , ZbI , ZcI , ZdI – have 2-dimensional orthogonal
complements spanned by the W I s satisfying ZiI W I = 0. Pairwise, these generic 2-planes
intersect in a line: this gives six lines, (a, b), (b, c), etc., that define the 1-simplex edges
of the tetrahedron. Three generic 2-planes in CP3 intersect in a point: this defines the four
vertices of the tetrahedron and we label them (a, b, c), (b, c, d), (c, d, a), and (d, a, b), as
illustrated here:
(b,c,d)
c
d
(10.19)
b
(a,b,c) (c,d,a)
a (d,a,b)
206 Polytopes
Just as in the case of the triangle, the denominator of the volume formula (10.18) is
the product of each vertex coordinate dotted into the reference vector Z0 . The numerator
compensates the scaling of each “face”-variable ZaI , ZbI , ZcI , ZdI to make the volume formula
projective.
Now, here comes the highlight of your day: for n = 4 the 5-bracket volume expression
(10.18) for a 4-simplex is identical to the integrand in the amplitude 5-bracket expression
(10.6)! This verifies our statement at the beginning of the chapter that the rational function
in (10.6) is indeed the volume of a 4-simplex in CP4 . The denominator factors in (10.18)
involve the five vertices of the 4-simplex. Since an NMHV tree superamplitude is a sum of
5-brackets, we are led to view the amplitude as a volume of a polytope in CP4 . We realize
this expectation in the next section and discuss its consequences.
10.2 NMHV tree superamplitude as the

volume of a polytope
The simplest NMHV case is the 5-point (anti-MHV)superamplitude

ANMHV
5 [1, 2, 3, 4, 5] = AMHV
5 × 1, 2, 3, 4, 5 . (10.20)
Thus, up to the MHV factor, ANMHV

5 is the volume of a 4-simplex in CP4 .
Next, for the NMHV 6-point superamplitude, consider the [2, 3 super-BCFW represen-
tation on the LHS of (10.1):

ANMHV
6 [1, 2, 3, 4, 5, 6] ∝ 2, 3, 4, 6, 1 + 2, 3, 4, 5, 6 + 2, 4, 5, 6, 1 . (10.21)

4,6,1,2, 2,3,4,6 6,2,3,4, 4,5,6,2 2,4,5,6, 6,1,2,4
Apart from the overall MHV factor, the 6-point NMHV superamplitude is the sum of the
volumes of three 4-simplices in CP4 ; we expect this to be the volume of a polytope obtained
by somehow gluing the three simplices together. But how exactly does this work? To address
this question, it is useful to examine the poles in the 5-brackets.
Recall from (5.50) and (5.52) that momentum twistor 4-brackets i − 1, i, j − 1, j in the
denominator give local poles, whereas other 4-brackets, such as 1, 2, 4, 6, give spurious
“non-local” poles. Examining the denominator terms of the 5-brackets in (10.21), we find
that each of them has two spurious poles; they are listed under each 5-bracket. The spurious
poles come in pairs – for example 4, 6, 1, 2 and 6, 1, 2, 4 in the first and third 5-brackets –
and cancel in the sum (10.21), as required by locality of the physical amplitude. In the
geometric description of a 5-bracket as a 4-simplex in CP4 , each of the five factors in
the denominator of the volume-expression (i.e. the 5-bracket) is determined by a vertex
of the associated 4-simplex. In particular, spurious poles must be associated with vertices
in CP4 that somehow “disappear” from the polytope whose volume equals the sum of the
simplex-volumes in (10.21). We now discuss how the “spurious” vertices disappear in the
sum of simplices. Let us start in CP2 where the polytopes are easier to draw.
10.2 NMHV tree superamplitude as the volume of a polytope 207
Polytopes in CP2
In 2-dimensions, consider the 4-edge polytope
(2,3) 3 (3,4)
2
(1,2)
4 . (10.22)
1
(4,1)
All vertices for this CP2 “amplitude” are defined by adjacent edges and are in this sense
local. We would like to compute the area of the 2-polytope (10.22) using 2-simplex volumes
[a, b, c]. There are several different ways to do this, corresponding to different triangulations
of the polytope. As an example, introduce a “non-local” point (1, 3) as the intersection of
lines 1 and 3. The resulting triangulation is
(1,3) (2,3) 3 (3,4) (1,3) (3,4)

3
2
(1,3) (2,3)
(1,2) 3
= 1 4 −
4 1 2
1
(1,2)
(4,1) (4,1)

= 4, 1, 3 − 2, 1, 3

= 4, 1, 3 + 1, 2, 3 . (10.23)
The area of the 4-edge polytope is given by the difference of two triangular areas. The non-
local vertex (1, 3) appears in both triangles. Comparing the last two lines, the sign of the
3-bracket indicates the orientation of the triangle with respect to a particular predetermined
ordering of all edges (or, in higher dimensions, boundaries).
It is useful to also consider another triangulation, so introduce the point (2, 4):
(2,4) (2,4)
(2,3)
3 2
2 (3,4) (2,4)
2
(1,2) = (1,2) − 4
(1,2)
4 4 3 (4,1)
1 1
(4,1) (4,1)

= 1, 2, 4 − 2, 4, 3

= 1, 2, 4 + 2, 3, 4 . (10.24)
208 Polytopes
The two triangulations (10.23) and (10.24) compute the same area (“amplitude”), so we
have a CP2 version of the identity (10.1), namely [4, 1, 3] + [1, 2, 3] = [1, 2, 4] + [2, 3, 4]
which can also be written

2, 3, 4 − 1, 3, 4 + 1, 2, 4 − 1, 2, 3 = 0 . (10.25)
Exercise 10.3
Suppose the 4-vertex polytope in the example above was not convex as drawn in
(10.24): show that the volume of a non-convex 4-vertex polytope can also be written
[1, 2, 4] + [2, 3, 4].
Polytopes in CP4
Extending the simple CP2 example to CP4 , one finds that the BCFW representation of a 6-
point NMHV tree superamplitude corresponds to a triangulation of the associated polytope
by introduction of three new auxiliary vertices. This allows one to use the given external
data, the boundaries of the polytope, to efficiently construct the corresponding triangulation.
Efficiency here means using a minimum number of 4-simplices; we come back to this point
in Section 10.4. Different BCFW constructions simply correspond to different choices of
auxiliary vertices. As an example, the auxiliary vertices for the BCFW representation in
(10.21) are (2, 4, 5, 6), (6, 1, 2, 4) and (2, 3, 4, 6). Note that these exactly label the spurious
poles in (10.21).
Let us now see how the removal of an auxiliary vertex works in CP4 . Since it can be
slightly challenging to draw a 4-dimensional object on paper, we go to the 3d boundary of
the 4-dimensional polytope. Specifically, at the 3d boundary defined by Z1 · W = 0, only the
two simplices [2, 3, 4, 6, 1] and [2, 4, 5, 6, 1] in (10.21) contribute, and their projections to
the boundary are the tetrahedrons defined by the faces Z2 , Z3 , Z4 , and Z6 and, respectively,
by Z2 , Z4 , Z5 , and Z6 . The two boundary tetrahedrons share the non-local vertex (1, 2, 4, 6),
which we simply label (2, 4, 6) on the boundary. Since this vertex does not appear in
other terms of (10.21), we should be able to visualize its cancellation geometrically on the
boundary defined by Z1 . On the 3d subspace, the superamplitude contains the combination
Z1 bdr
2, 4, 5, 6, 1 + 2, 3, 4, 6, 1 −−−→ 6, 2, 4, 5 − 6, 2, 4, 3 = vol(bdr polytope) .
(10.26)
On the RHS we have arranged the common faces, Z6 , Z2 , and Z4 , to appear in the same
order to facilitate the geometrical interpretation. The pictorial representation of (10.26) is
(2,4,6)
(2,4,6) (6,2,3)
(4,3,2) 3
4 6 (3,4,6)
2
− 2 = 2 .
6 (6,2,3) (4,5,2) (6,2,5)
4 (4,3,2) 4 6
(4,5,2) (6,2,5) 3
5
(3,4,6) (5,4,6)
5 (5,4,6)
(10.27)
The “non-local” auxiliary vertex (2, 4, 6) indeed “cancels” in the sum and leaves behind
the volume of the 3-dimensional polytope with five faces and six local vertices! To see that
the remaining vertices are local, remember that we are in the subspace Z1 · W = 0, so each
vertex is really represented as (1, ∗, ∗, ∗) in CP4 . Thus each of the six vertices in (10.27),
namely
(1, 2, 3, 4), (1, 2, 3, 6), (1, 3, 4, 6), (1, 4, 5, 2), (1, 4, 5, 6), (1, 2, 5, 6) , (10.28)
involves two pairs of adjacent labels, and therefore – by (5.52) – they correspond to local
poles. Thus we conclude that on the subspace Z1 · W = 0, which involves only two of
the simplices in (10.21), the amplitude is free of non-local vertices. One can similarly
understand the cancellation of the two other spurious poles in (10.21).
We have found that the 6-point NMHV tree superamplitude is given by the volume
of a polytope in CP4 . It is defined as the sum of the three 4-simplices in (10.21) and
its six boundaries are in 1-to-1 correspondence with the momentum supertwistors ZiI ,
i = 1, . . . , 6. Different BCFW representations correspond to the different tessellations
of the polytope into 4-simplices; each representation requires introduction of “spurious”
vertices and the associated spurious poles cancel because they are absent in the orig-
inal polytope. The vertices of the polytopes are all local and can be characterized as
the nine quadruple intersections (i, i + 1, j, j + 1) of the six boundaries determined by
ZiI .
The polytope interpretation of the amplitudes was first presented by Hodges [24] with
the goal of geometrizing the cancellation of spurious poles in the BCFW expansion. Build-
ing on Hodges’ work, the authors of [130] constructed the representation of the NMHV
superamplitude where both dual superconformal symmetry and locality are manifest.
10.3 The boundary of simplices and polytopes
We have studied the volumes of simplices and polytopes, but their boundaries also have
interesting properties. Let us again start with a simple triangle in CP2 ,
2
1 . (10.29)
3
The subspace defined by Z1 · W = 0 contains part of the boundary of the triangle, namely
the line segment bounded by the intersections of lines 2 and 3 with line 1. The length of the
line segment is just the projection to the subspace defined by Z1 , namely [2, 3]. Note that
since the “volumes” (i.e. lengths) of the line segments are defined with a choice of sign, we
have to pick an orientation for each: here and in the following, we pick the orientation of
the faces to point into the volume of the polytope that they are bounding. With this choice
of orientation, the circumference is [12] + [23] + [31].
210 Polytopes
To see a little more structure, consider the tetrahedron in CP3 ,

(2,3,4)
3 4
. (10.30)
2
(1,2,3) (3,4,1)
1 (4,1,2)
The volume is [1, 2, 3, 4]. The 2-plane defined by Z1 contains the face bounded by the
intersections of the plane 1 with the planes 2, 3, and 4. Therefore the area of this face is
[2, 3, 4]. Keeping careful track of the orientations of the faces, we find that the total area of
boundary of the tetrahedron is [2, 3, 4] + [1, 4, 3] + [2, 4, 1] + [2, 1, 3].
We can summarize the results for the boundary “volumes” so far as

bdr of CP2 triangle : ∂ 1, 2, 3 = 2, 3 − 1, 3 + 1, 2 ,

bdr of CP3 tetrahedron : ∂ 1, 2, 3, 4 = 2, 3, 4 − 1, 3, 4 + 1, 2, 4 − 1, 2, 3 .
(10.31)
This motivates us to define the boundary operation for any (n−1)-simplex:
n

∂ 123 . . . n = (−1)i+1 1, 2, . . . , i − 1, i + 1, . . . , n . (10.32)
i=1
Exercise 10.4
The boundary of a boundary vanishes, so our definition (10.32) had better yield a
nilpotent operator, ∂ 2 = 0. Show that the action of ∂ 2 on any n-simplex is zero.
At this stage you might have noticed the similarity between the RHS of the tetrahedron
boundary identity (10.31) and the CP2 vanishing identity (10.25). This is easy to understand:
in CP2 , we cannot construct a 3-simplex [1, 2, 3, 4] with a 3d volume, so in particular the
boundary of such a formal object must vanish:

CP2 : 0 = ∂ 1, 2, 3, 4 = 2, 3, 4 − 1, 3, 4 + 1, 2, 4 − 1, 2, 3 . (10.33)
This gives another geometric
interpretation
of the CP2 BCFW identity (10.25).
Similarly, a 5-simplex 1, 2, 3, 4, 5, 6 in CP has vanishing boundary:
4

CP4 : 0 = ∂ 1, 2, 3, 4, 5, 6 ≡ 2, 3, 4, 5, 6 − 3, 4, 5, 6, 1 + 4, 5, 6, 1, 2

− 5, 6, 1, 2, 3 + 6, 1, 2, 3, 4 − 1, 2, 3, 4, 5 .
(10.34)
The RHS is exactly the 6-term identity (10.3) which originated from the equivalence of
different super-BCFW shifts (10.1). This was the identity that motivated our study at the
beginning of the chapter: we understand now that it is not a Schouten identity, but that it
has an interpretation as the vanishing boundary of a formal 5-simplex in CP4 .
Let us now take a look at the action of the boundary operation on a superamplitude. As
per usual, we start with CP2 to get intuition for the problem. Consider the triangulation
used in (10.23) to calculate the volume of a 4-sided polygon in CP2 ,
(1,3) (2,3) 3 (3,4) (1,3) 3 (3,4)
2
(1,3) 3 (2,3)
(1,2)
= 1 −
4 4 1
1 2
(1,2)
(4,1) (4,1)

= 4, 1, 3 − 2, 1, 3 . (10.35)
We apply the boundary operator to each 2-simplex and find:

∂ 4, 1, 3 = 1, 3 − 4, 3 + 4, 1 ,
(10.36)
∂ 2, 1, 3 = 1, 3 − 2, 3 + 2, 1 .
The RHS of
each equation is the circumference

of the respective
triangles. Let us try to
interpret ∂ [4, 1, 3] − [2, 1, 3] . It contains [4, 1] − [2, 1] : this is a difference of lengths
of two line segments in the subspace defined by Z3 and thus it
is the length of the side labeled
3 in the 4-sided polygon on the LHS of (10.35). Similarly, [4, 3] − [2, 3] is the length of
side 1 of the polygon. Now we are left with two terms that are both labeled [1, 3] in (10.36).
It is tempting to cancel these two terms, but this is not quite correct: [1, 3] in ∂[4, 1, 3] lives
in the subspace defined by Z4 , while ∂[2, 1, 3] is in the Z2 -subspace. So the two [1, 3]s are
the lengths

of the sides 2 and 4 in the polygon (10.35). Why does their difference show up
in ∂ [4, 1, 3] − [2, 1, 3] instead of their sum? Easy: that is because they have the opposite
orientations: in our conventions, side 4 in the big triangle in (10.35) is oriented to point
into the polygon, but side 2 in the small triangle points out of the 4-sided polygon. Flipping
the orientation and labeling the 2-brackets by the subspace Zi they live on, we see that the
difference of the two terms in (10.36) exactly calculates the circumference of the 4-sided
polygon on the LHS of (10.35):

∂ [4, 1, 3] − [2, 1, 3] = [4, 3] − [2, 3] Z1 + 3, 1 Z2 + [4, 1] − [2, 1] Z3 + 1, 3 Z4 .
(10.37)
The boundary operator ∂ was introduced [130] as a formal operation useful for studying
the cancellation of spurious poles in the BCFW expansion, without emphasis on the inter-
pretation as the “boundary volume” we have presented here. So let us now comment on
the application of ∂ in [130]. Note that for a simplex, there is a unique point “opposite”
each face: in particular in the triangles in (10.35) the point labeled (1, 3) is the non-local
“spurious” point that sits across from the line segments [1, 3]Z2 and [3, 1]Z4 , respectively.
So since the two [1, 3]s define the same point (1, 3), one can in a vertex-interpretation of
the boundary operation cancel them in ∂([4, 1, 3] − [2, 1, 3]): one can think of this as the
cancellation of the spurious point in the (1, 3) in this particular triangulation. To distinguish
212 Polytopes
the vertex-interpretation from the boundary volume, we include a V (for vertex) with each
term; then we write

∂ [4, 1, 3] − [2, 1, 3] = V 3, 4 + V 4, 1 + V 2, 3 + V 1, 2 . (10.38)
Note how each term on the RHS is of the form V [i, i + 1] indicating that the polytope has
only local vertices; the non-local vertex V [1, 3] canceled. This is the interpretation of the
boundary operation given in [130].
Exercise 10.5
As an example in CP3 , consider the dissection (10.27) of the 5-faced polytope into
two tetrahedrons. Keep careful track of the orientations of the boundaries to show
that ∂([6, 2, 4, 5] − [6, 2, 4, 3]) calculates the surface area of the 5-faced 3-polytope
on the RHS of (10.27). Next use the vertex-interpretation discussed above to show that
the spurious poles cancel. Lift the example back to CP4 (remember that (10.27) was
the projection on the subspace defined by Z1 ) to see that each boundary vertex term is
of the form V [i, i + 1, j, j + 1] as in (10.28).
Enough of toy-examples! Let us compute the boundary of the NMHV 6-point tree super-
amplitude in the BCFW representation

ANMHV
6 [1, 2, 3, 4, 5, 6] = A MHV
6 × 1, 3, 4, 5, 6 + 3, 5, 6, 1, 2 + 5, 1, 2, 3, 4 .
(10.39)
Using the vertex-interpretation, ∂ acts on the first two 4-simplices to give

∂ 1, 3, 4, 5, 6 + 3, 5, 6, 1, 2 = V 3, 4, 5, 6 + V 4, 5, 6, 1 + V 6, 1, 3, 4

+ V 5, 6, 1, 2 + V 2, 3, 6, 1 + V 2, 3, 5, 6

+ V 1, 3, 4, 5 + V 1, 2, 3, 5 . (10.40)
All vertices on the RHS are local except the last two. Including the boundary of the third
5-bracket in (10.39), the non-local vertices cancel completely and we have

∂ 1, 3, 4, 5, 6 + 3, 5, 6, 1, 2 + 5, 1, 2, 3, 4

6
3

= V i, i + 1, i + 2, i + 3 + V i, i + 1, i + 3, i + 4 , (10.41)
i=1 i=1
where, as usually, the arguments are understood cyclically. This shows that, indeed, the
boundary of the polytope that corresponds to our tree superamplitude ANMHV
6 contains only
local vertices. The power of the boundary operation is that it makes the cancellation of
spurious points clear without a need to draw any polyhedrals.
Exercise 10.6
How many local vertices are there in the polytope corresponding to the 7-point NMHV
tree superamplitude?
10.4 Geometric aftermath
The BCFW triangulation is an efficient representation of the tree-level NMHV superampli-

tudes in the sense that it involves only relatively few terms. However, we can imagine other
triangulations. For example, consider the 4-sided polygon in CP2 . Introducing an auxiliary
point W∗ inside the polygon, we can triangulate it as
(2,3) 3 (3,4)
2
(1,2) 4
1 W∗ , W(i−1,i) , W(i,i+1)
W* = 2 (Z0 · W∗ )0, i − 1, i0, i, i + 1
. (10.42)
4 i=1
1
(4,1)
This gives a 4-term expression for the volume of the polygon, as opposed to the 2-term
BCFW triangulations in (10.23) or (10.24). In this sense, BCFW is more efficient. The
representation (10.42) may remind you of another representation of scattering amplitudes,
namely the CSW expansion (or MHV vertex expansion) of Section 3.4. It is actually
not quite the same; CSW in momentum twistor space involves a reference supertwistor
Z ∗ = (0, |X ], 0) instead of W∗ ; see [131] for details.
As yet another way to calculate the amplitudes, we might ask if there is a triangulation
that does not give spurious poles? A prescription for such a representation was given in
[130] for the tree-level NMHV superamplitudes. To give a hint of how it works, consider the
boundary Z1 · W = 0 that we also analyzed in (10.27). Instead of the 2-term triangulation
applied in (10.27) at the cost of a non-local vertex, we can triangulate the 5-sided polytope
as
(6,2,3) (6,2,3)
(4,3,2) (4,3,2) (6,2,3)
3 (4,3,2)
2 (3,4,6) = + (3,4,6)
4
(4,5,2) (6,2,5) (4,5,2) (6,2,5)
6
(5,4,6) (5,4,6)
5 (5,4,6)
(6,2,3)
(6,2,3) (6,2,3)
(4,3,2) (4,3,2)
(3,4,6)
= + (4,5,2) (6,2,5) + .
(4,5,2)
(5,4,6) (5,4,6)
(5,4,6)
(10.43)
214 Polytopes
The manifestly local tessellation of the polytope gives more terms than the BCFW repre-
sentation. You can find the general expression in [130].
We have argued that each n-point tree-level NMHV superamplitude of N = 4 SYM
can be interpreted as the volume of a polytope in CP4 . It should be rather obvious by
now that the reverse is not true: not all polytopes in CP4 correspond to tree-level su-
peramplitudes in N = 4 SYM. An example is the polytope obtained by gluing together
two of the three simplices in the BCFW representation (10.39): that is a perfectly fine
polytope, but it has non-local poles so does not correspond to a physical amplitude. The
color-ordering plays a key role in interpreting the superamplitudes polytopes. It is of
course an interesting question if this geometric picture can be extended beyond the lead-
ing color level – or if other polytopes might have interpretations in terms of scattering
processes.
The current discussion of tree superamplitudes utilizes the dual superconformal invari-
ance of planar N = 4 SYM, focusing on the 5-brackets invariants. Since the tree amplitudes
of pure Yang–Mills theory can be projected out from N = 4 SYM superamplitudes, a sim-
ilar analysis can be applied to pure YM as well. In fact, it was in pure Yang–Mills theory
that Hodges first realized the polytope picture [24].
The polytope interpretation described here is valid for NMHV n-point tree superampli-
tudes as well as 1-loop n-point MHV integrands in planar N = 4 SYM [130]. The gen-
eralization is not obvious. Tree-level N K MHV superamplitudes involve sums of products
of K 5-brackets, so a geometric interpretation in terms of simplex-volumes is not straight-
forward. There is nonetheless a geometrization of all N K MHV tree superamplitudes and
loop-integrands in planar N = 4 SYM: it goes under the name of the amplituhedron [132].
This is a polytope defined in a space whose coordinates are a union of momentum super-
twistors and Grassmannian coordinates that extend the χi · ψ-construction in (10.5). For
tree-level NMHV, the amplituhedron reduces to a dual of the polytope discussed here, with
vertices and faces interchanged. The amplituhedron description makes locality manifest,
while unitarity is an emergent property. BCFW arises as a particular triangulation. Loops
appear from integrating out pairs of “hidden” points, in a somewhat similar way to the
description of the loop-integrands in Chapter 7.
It is curious that in connection to amplitudes, polytopes can appear in different guises.
An example, different from our discussion so far, is the observation [133] that a 1-loop
box integral can be interpreted as the volume of a tetrahedron in AdS5 . The vertices of
the tetrahedron are the four dual region variables yi (in the embedding formalism) and
the edges are geodesics in AdS5 . Since all N = 4 SYM 1-loop amplitudes are given by
an expansion in box integrals, the amplitudes can be interpreted as sums of volumes of
such AdS5 tetrahedrons, weighted by the appropriate box-coefficients. Non-planar 1-loop
amplitudes can be given as linear combinations of planar ones, so the same conclusion
extends to non-planar amplitude as well [134].
We have encountered many different representations of the tree amplitude and have
now seen a unifying geometric interpretation. But there is yet another representation of
amplitudes that we had a glimpse of in Section 2.5, namely the “BCJ representation” in
which the color- and kinematic-structures enter on a dual basis. For such representations
the amplitude can be manifestly local. We discuss BCJ further in Chapter 13, but note
here that the color-ordering is crucial for the relation between polytopes and amplitudes: it
allows us to relate the polytope to the momentum space representation of an amplitude and
this is key for the statements about locality. Polytope or amplituhedron interpretations for
the planar BCJ representation, non-planar, or non-color-ordered amplitudes are still open
problems.
11 Amplitudes in dimensions D = 4
Just in case it slipped your mind, our discussion up to now has focused on scattering
amplitudes in D = 4 spacetime dimensions. There is a good reason for this: for one, this
review was written in 3+1 dimensions (as far as we know) and this is where our particle
physics experiments take place. And secondly, the power of the D = 4 spinor helicity
formalism and its extensions to twistors and momentum twistors allows us to explore the
rich and exciting mathematical structure of 4d scattering amplitudes, especially those in
planar N = 4 SYM. However, there are also interesting quantum field theories in other
dimensions; in this chapter we take a look at their scattering amplitudes. We discuss
D = 6 briefly, but otherwise our eyes are on D = 3, particularly on the N = 8 and N = 6
superconformal theories known as BLG and ABJM.
11.1 Helicity formalism in D = 4
We have often emphasized that the modern on-shell approach relies heavily on having a
“good” set of variables to parameterize the on-shell degrees of freedom: “good” typically
means that the variables trivialize the kinematic constraints and transform linearly under
the global symmetries of the theory. This is realized strikingly by the supertwistors and
momentum supertwistors of planar N = 4 SYM, but the trivialization of the massless
on-shell condition pi2 = 0 in the spinor helicity formalism with |i and |i] was our starting
point. So this is also where we begin for D = 4.
To parameterize massless kinematics in D-dimensions, consider bosonic spinors that
carry a spinor index A of the Lorentz group Spin(1, D − 1) and a fundamental index a of
the little group S O(D − 2):
A ← Lorentz
λia ← little grp . (11.1)
As per usual, i = 1, 2, . . . , n is a particle label. The spinor-type (Weyl, Majorana, etc.)

will be specified when we specialize to a given dimension D. If the spinors are complex,
there will be a conjugate spinor λ̃ whose A and a indices are in the appropriate conjugate
representations.
The 4d Lorentz group is Spin(1,3) = S L(2, C) and the little group is S O(2) = U (1). The
spinors
D = 4: λ̃i−
ȧ
= |iȧ , λi+
a
= [i|a , (11.2)
11.1 Helicity formalism in D = 4 217
are Weyl-spinors, so the index A is the familiar S L(2, C) indices a, ȧ. The little group
index a is + or − depending on how the spinors transform under the S O(2) = U (1) little
group transformations. The D = 4 light-like momentum is written as the bi-spinor as the
familiar relation
D = 4: piȧa = −λ̃i−
a
λi+
ȧ
= −|iȧ [i|a . (11.3)
μ
As discussed in Section 2.2, pi is real when the spinors (11.2) are complex conjugate.
The relation (11.3) implies that the 2×2 matrix piȧa has rank 1 and therefore solves the
D = 4 massless constraint pi2 = − det( pi ) = 0. To see if a similar construction could be
available in D dimensions, we count degrees of freedom. A real light-like vector has D−1
degrees of freedom with the −1 from the condition pi2 = 0.1 So the strategy is to find a
Spin(1,D−1) spinor representation that allows forming a little group invariant bi-spinor
with D−1 degrees of freedom. Here is how the counting works in D = 4. The complex
2-component spinor λi+ a
= [i|a has four real degrees of freedom, and when combined with
its complex conjugate λ̃i− = |iȧ , the resulting bi-spinor (11.3) is invariant under the U (1)
ȧ
little group rotation. Thus subtracting out the U (1) redundancy, we indeed have 4 − 1 = 3
degrees of freedom, matching that of a real light-like vector in 4d. Now let us look at how
the counting works in other dimensions.
Exercise 11.1
Do the above counting for the case where the spacetime signature is (− − ++).
For D = 3, the Lorentz group is Spin(1,2)=S L(2, R) and the minimal spinor representation
is a 2-component Majorana spinor λia , where a = 1, 2 is an S L(2, R) index. The null
momentum is given by
D = 3: piab = λia λib . (11.4)
For pi real, the spinors λia may be either real or purely imaginary. Either way, they encode
two real degrees of freedom. So the RHS of (11.4) has two degrees of freedom, the correct
count for a 3d light-like vector. Note that no little group index was included on the spinors λia
because the little group Z2 is discrete. It acts as λia → −λia , indeed leaving the momentum
(11.4) invariant.
For D = 6, we have Spin(1,5) =SU ∗ (4) and the little group is S O(4) = SU (2) × SU (2).
The ∗ on the SU ∗ (4) indicates that it is pseudo-real.2 We pick a chiral spinor λia A
in the
∗
fundamental of SU (4), so A = 1, 2, 3, 4. The spinor is chiral, as opposed to anti-chiral,
because it is in the fundamental, not anti-fundamental, representation of SU ∗ (4). The two
SU (2)-factors of the little group belong to the chiral and anti-chiral spinors, respectively,
so λia
A
carries a little group index a = 1, 2 of the chiral SU (2) factor. A candidate for the
1 An on-shell massive momentum has pi2 = −m i2 , but we view the constraint as imposed on D + 1 degrees of
μ
freedom, pi and m i .
2 Pseudo-real means that for each group element g, the complex conjugate g ∗ is related to g via a similarity
transformation g = g ∗ −1 , where is an antisymmetric matrix. (If is symmetric, then the representation
is a real.)
light-like momentum can now be formed as the little group invariant bi-spinor
D = 6: piAB = λiAa λia
B
. (11.5)
This works to give the right number of degrees of freedom, namely five, for a massless
momentum in 6d: the spinor λiAa has 4 × 2 degrees of freedom, but we have to mod out by
the little group SU (2)-factor, giving 4 × 2 − 3 = 5.
The results for D = 3, 4, 6 can be summarized as follows:
Spin(1, D − 1) little group p2 = 0
D=3 S L(2, R) Z2 piab = λia λib
(11.6)
D=4 S L(2, C) S O(2) = U (1) piȧa = −λia λ̃iȧ
D=6 SU ∗ (4) S O(4) = SU (2) × SU (2) piAB = λiAa λiaB
How about general D dimensions? The strategy is to introduce a bosonic spinor λia A
(where
a transforms under the little group, or a subgroup as in the 6d example) and use it (and
possibly its conjugate spinor) to form a (real) light-like vector as a little group invariant
bi-spinor, e.g. λia λ̃i . However, for this to encode a null momentum, the number of real
A aB
degrees of freedom of the bi-spinor, modulo the number of little group generators, has to
match that of a light-like vector:
A aB
DOF λia λ̃i − #(little group generators) = D − 1 . (11.7)
This is a non-trivial constraint because the bi-spinor typically has more than D − 1 degrees
of freedom. One has to find a minimal spinor representation with maximal little group
redundancy; this was particularly clear in the 6d example above. Indeed, we know solutions
to these constraints only for D = 3, 4, 6.3
Exercise 11.2
What is the smallest possible number of degrees of freedom for a little group invari-
ant bi-spinor in D = 10? [Hint: In D = 10, the minimum spinor representation is a
Majorana–Weyl spinor; it has 16 real components.]
It is possible to reduce the number of independent spinor degrees of freedom further

by imposing the equations of motion, i.e. the zero-mass Dirac equation. Now you may
be puzzled, because back in Chapter 2 we set up the D = 4 spinor helicity formalism
by requiring at the starting point that the spinors |i and |i] satisfied the Dirac–Weyl
equation. For D = 3, 4, 6, this approach is equivalent: the Lorentz contraction of the (D−
1)-component bi-spinor with one of its spinors gives zero, so the momentum space form of
the massless Dirac equation is automatic.
For D = 3, 4, 6, setting up a spinor helicity formalism is possible but the resulting
spinors are constrained in the sense that the Dirac equation is imposed as a non-trivial
condition [135–137]. Constrained spinors are more difficult to work with, especially if one
wants to construct symmetry generators in order to study symmetries of the D-dimensional
3 D = 3, 4, 6 are precisely the dimensions in which twistor constructions that describe conformal symmetry are
known; see Section II.C.5 of Siegel’s “Fields” [57].
amplitudes. For this reason, we focus on D = 3, 6 in this chapter: we describe D = 6

briefly, then offer more details about the interesting structure of D = 3 amplitudes.
11.2 Scattering amplitudes in D = 6
Oh, who cares about 6d scattering amplitudes!!? Don’t we live in 4d? Well, the 6d massless
condition
− p02 + p12 + p22 + p32 + p42 + p53 = 0 (11.8)
can be viewed from 4d spacetime as the on-shell condition for a massive 4d momentum
vector: take p42 + p52 = m 2 (or = m m̃ if you are willing to accept complex masses). Then
2
p4d = −m 2 follows from (11.8) with p4d denoting the first four components of the 6d
momentum. This makes the 6d formalism very useful for studies of 4d amplitudes with
massive particles. An example of an explicit application to Higgs production processes can
be found in [138]. See [87, 139, 140] for other examples. So let’s care about 6d scattering
amplitudes, at least for one section.
The 6d spinor helicity formalism was first developed by Cheung and O’Connell [141] and
its supersymmetrization was carried out in [142]. It has been applied to tree- and loop-level
scattering amplitudes in maximal super Yang–Mills theory in 6d [139, 143, 144] and also
used in other 6d theories [145–148].
In 4d, we used (σ μ )a ḃ and (σ̄ μ )ȧb to define the 2 × 2 matrices pa ḃ = pμ (σ μ )a ḃ and
p = pμ (σ̄ μ )ȧb . Similarly, the 6d Lorentz group S O(1, 5) ∼ SU ∗ (4) has antisymmetric
ȧb
4×4 matrices (σ μ ) AB and (σ̃ μ ) AB , A, B = 1, 2, 3, 4, that allow us to define

p AB = pμ (σ μ ) AB , p AB = pμ (σ̃ μ ) AB . (11.9)
The explicit form of the (σ, σ̃ ) matrices as well as their relation to the 6d 8 × 8 γ -matrices
can be found in Appendix A of [141].
The pseudo-real property of SU ∗ (4), implies that p AB and p AB are related as
1 ABC D

p AB = pC D . (11.10)
2
In this notation, the S O(1, 5) invariant product p μ pμ can be written as the manifestly
SU ∗ (4)-invariant contraction
1 1
p μ pμ = − p AB p AB = − ABC D p AB pC D . (11.11)
4 8
Now, in momentum space, the 6d Dirac equation for massless spinors is
p AB λiBa = 0 , p AB λ̃i B ȧ = 0 , (11.12)
where λ Ba and λ̃ B ȧ are chiral and anti-chiral spinors, and a = 1, 2 and ȧ = 1, 2 are fun-
damental indices of the two SU (2)s of the little group S O(4) = SU (2) × SU (2). The two
pairs of Weyl spinors
λiAa = i a | A = A
|i a and λ̃i B ȧ = [i ȧ | B = B |i ȧ ] (11.13)
are the building blocks of the 6d spinor helicity formalism. There is no distinction between
bras and kets because there is no raising or lowering of the SU ∗ (4) indices.
The little group indices can be raised/lowered using the SU (2) Levi-Civita symbol as
λa = ab λb and λ̃ȧ = ȧḃ λ̃ḃ . This allows us to form little group invariants, as in the bi-spinor
construction (11.5). Indeed, the massless momentum is given as
piAB = λiAa λia

B
, pi AB = λ̃i Aȧ λ̃i B ȧ . (11.14)
Due to the antisymmetric contraction of the SU (2) indices, the bi-spinors in (11.14) are
automatically antisymmetric in the SU ∗ (4) indices A and B. By (11.14), the 4 × 4 matrix
piAB has rank 2, so pi2 ∼ ABC D piAB piC D is zero. Hence the massless on-shell condition
pi2 = 0 is satisfied. Thus this realizes the construction (11.5).
Reverting the momentum in (11.14) from matrix form to vector form, we have
μ 1 1
pi = − i a |σ μ |i a = − [i ȧ |σ̃ μ |i ȧ ] . (11.15)
4 4
These expressions are the 6d versions of the 4d relation k μ = 12 k|γ μ |k] you derived in
Exercise 2.4.
The Dirac equation (11.12) implies that λiAa λ̃i Aȧ = 0, so the chiral and anti-chiral spinors
are related. Construction of symmetry generators using these variables must take such
constraints into account. However, if only chiral spinors are needed, we can still work with
unconstrained variables.
To get a better feeling for the 6d 4 × 2 spinors – and to facilitate reduction to 4d – consider
the embedding of our good old 4d spinors in the new 6d spinors. Choosing μ = 0, 1, 2, 3
to be the 4d subspace and setting p4 = p5 = 0, the 4d spinors appear in the 6d ones as

0 i|ȧ 0 |iȧ
λia =
A
, λ̃i Aȧ = . (11.16)
[i|a 0 −|i]a 0
Thus the constraint λiAa λ̃i Aȧ = 0 becomes nothing but the familiar i i = [i i] = 0.
In 6d massless kinematics, the basic Lorentz invariant spinor products are:
• i a | jḃ ] ≡ λiAa λ̃ j Aḃ = [ jḃ |i a ,

• i a j b k cl d ≡ ABC D λiAa λ Bb j λk λl ,
Cc Dd
• [i ȧ jḃ kċlḋ ] ≡ ABC D

λ̃i Aȧ λ̃ j B ḃ λ̃kC ċ λ̃l Dḋ .
In particular, the Mandelstam variable si j = −( pi + p j )2 is

1
si j = − ab ȧḃ i a | jȧ ]i b | jḃ ] = − deti a | jȧ ] . (11.17)
2
We have outlined the 6d spinor helicity formalism, so now it is time to apply it to amplitudes.
Let us begin with 3-point amplitudes; this involves special 3-particle kinematics because
all si j vanish. In 4d, we got around this by working with complex kinematics such that
ji = ([i j])∗ and that allowed us to choose either all the angle- or the square-brackets to
be non-vanishing, but not both. In 6d 3-particle kinematics, the only Lorentz invariants
available for si j are the brackets i a | jȧ ]. But since 0 = si j = − deti a | jȧ ], the 2 × 2 matrix
i a | jȧ ] must be rank 1. We have encountered 2 × 2 matrices of rank 1 before, namely the
4d massless pa ȧ , and by now it should be a simple reflex to introduce two 2-component
spinors, u ia and ũ j ȧ such that i a | jȧ ] = u ia ũ j ȧ [141], so the 3-point amplitudes in 6d are
written in terms of these “auxiliary” 2-component spinors.
Just as in 4d, the 6d 3-point amplitudes are highly constrained by the little group and
Lorentz invariance. For example, one finds that the 3-vector amplitudes only come in two
types, one is generated by the A A∂ A vertex of the Yang–Mills action while the other is
generated by the operator Fμ ν Fν ρ Fρ μ [141]. A wide class of possible 3-point interactions
was categorized in [146]. In particular, for 6d self-dual antisymmetric tensors – which are
part of the (2,0) supermultiplet that describes the degrees of freedom of M5-membranes
in M-theory – one can demonstrate [146] that a 3-point amplitude cannot be both Lorentz
invariant and carry the correct little group indices to describe scattering of three self-dual
tensors; so it does not exist.
The 6d 4-point Yang–Mills amplitude is given by:
1a 2b 3c 4d [1ȧ 2ḃ 3ċ 4ḋ ]

A4 (1, 2, 3, 4) = . (11.18)
su
The 6d gluons are not labeled by the 4d concept of helicity: instead, a 6d massless spin-1
particle has four physical states labeled by the little group indices a ȧ .
Exercise 11.3
Use the map in (11.16) to reduce the 6d amplitude (11.18) to 4d. You should find
the usual suspect, the MHV gluon amplitude. But that is not all: identify the other
possibilities and describe their origin.
In maximal SYM in 6d, the 4-point superamplitude takes the simple form

1
A4 (1, 2, 3, 4) = δ 6 P δ (4) Q δ (4) Q̃ 2 2
. (11.19)
y13 y24
We have used dual space to write s = −y13 2

and u = −y24 2
. The supermomentum
delta functions are defined in [142]. If we write the n-point superamplitude as An =
δ 6 (P) δ (4) (Q) δ (4) ( Q̃) f n , we note from (11.19) that I [ f 4 ] = y12 y22 y32 y42 f 4 under dual con-
formal inversion (5.37), i.e. f 4 inverts in exactly the same way as the 4d 4-point superam-
plitude of N = 4 SYM. Using a 6d version super-BCFW recursion, it was proven [144]
for all n that
$ n '
#
I [ fn ] = yi f n .
2
(11.20)
i=1
In 4d, it was essential for dual superconformal symmetry of planar superamplitudes in N =

4 SYM that the inversion weights of the bosonic and fermionic delta functions canceled,
as shown in (5.38). This, however, does not happen in 6d maximal SYM: δ 6 (P) inverts
with weight 6, while the Grassmann delta functions have weight −(4 + 4)/2. Therefore,
the planar superamplitudes of 6d maximal SYM do not have uniform inversion weight.
Nonetheless, as is often the case with scattering amplitudes, even if a symmetry is not
exact, it is still useful if it is broken in a predetermined fashion, as is the case here.
Remarkably, using generalized unitarity methods it has been shown [144] that the planar L-
loop integrands of 6d maximal SYM theory have the same dual conformal inversion weight
as in 4d. (A similar result was found for 10d SYM [136].) While the origin of this form
of dual superconformal symmetry is not clear (and the 6d and 10d SYM theories are not
(super)conformal), it has non-trivial implications in 4d for the structure of (super)amplitudes
on the Coulomb branch of N = 4 SYM [140, 149].
11.3 Scattering amplitudes in D = 3
Scattering amplitudes in D = 3 turn out to have very interesting properties. After introduc-
ing the necessary kinematic tools and basic examples of amplitudes, we focus on scattering
processes in the 3d N = 8 and N = 6 superconformal theories called BLG and ABJM.
11.3.1 D = 3 kinematics
We construct 3d kinematics by reduction from 4d using the fact that the 4d massless con-
dition, − p02 + p12 + p22 + p32 = 0, is equivalent to a 3d massive constraint. It is convenient
to identify the p2 -component with the 3d mass as p22 = m 2 so that the 3d momentum p μ
with μ = 0, 1, 3 satisfies p μ pμ = −m 2 .
Recall that in 4d, the momentum can be given as
0
− p + p3 p1 − i p2
D = 4: pa ḃ = . (11.21)
p1 + i p2 − p0 − p3
We restrict this to 3d by removing p2 and writing
0
− p + p3 p1
D = 3: pab = . (11.22)
p1 − p0 − p3
Then det pab = −(− p02 + p12 + p32 ) = m 2 .

The 2 × 2 matrix pab is symmetric. If the 3d momentum p μ , μ = 0, 1, 3, is real, pab is
also real.4 A generic real symmetric 2 × 2 matrix can be written as [150]5
m = 0: pab = λa λ̄b + λb λ̄a , (11.23)
where λ̄a = (λa )∗ when p μ is real.
4 This contrasts with the 4d case, where pa ḃ is complex valued, and it reflects the different Lorentz groups,
S O(1, 2) = S L(2, R) in 3d and S L(2, C) in 4d.
5 We could also have written pab = λa λb + μa μb , where μa is some real 2-component spinor not parallel to λa ,
but this is equivalent to (11.23) by a linear redefinition.
Exercise 11.4
If p μ is complex, we take λa and λ̄a to be independent. For each case, p μ real or
complex, count the number of degrees of freedom on each side of (11.23).
By direct calculation of the determinant of (11.23), we find that m 2 = det pab = −λλ̄2 ,
where λλ̄ = λa λ̄a and spinor indices are raised and lowered with the 2-index Levi-Civita
of the S L(2, R) Lorentz group. For 3d massless kinematics, m = 0, we must therefore have
λλ̄ = 0, implying that λ̄ ∝ λ. Thus, we can write
m = 0: pab = λa λb = p|a p|b , (11.24)
where λ = p| was rescaled such that the prefactor is just 1. Note that p| must be either
purely real or purely imaginary for pab to be real.
It follows from (11.24) that in 3d massless kinematics, all Lorentz invariants are built out
of one kind of angle brackets, namely i j = λia λ j a . For example, since 2 pi . p j = −i j2 ,
the Mandelstams si j are
D = 3: si j = −( pi + p j )2 = i j2 . (11.25)
n μ
Momentum conservation i=1 pi = 0 can be written

n
D = 3: |ii| = 0 . (11.26)
i=1
Our 3d kinematics is ready, so let us see some amplitudes. As it is our style, we start
with 3-particle amplitudes. These are particularly easy in 3d, because 3-particle kinematics
requires all si j = i j2 = 0 and hence there are no Lorentz invariants available for a massless
3-point amplitude. Thus for massless kinematics, there are no 3-point on-shell amplitudes
in 3d.
The little group for massless kinematics in 3d is the discrete group Z2 ; it acts on the
spinor variables as |i → −|i. The homogeneous scaling of the scattering amplitudes
distinguishes only two types of particles in 3d: scalar particles scale with +1 and fermions
scale with −1. And spin-1 vector particles? A massless vector in D-dimensions has D − 2
degrees of freedom, so in D = 3 this is just 1, the same as a scalar.
Tree-level scattering amplitudes of 3d super Yang–Mills theory can be obtained directly
from 4d ones using dimensional reduction. For example, the dimensional reduction of the
4-point gluon amplitude A4 [1− 2+ 3− 4+ ] of 4d Yang–Mills theory gives
134 4d → 3d 134
A4 [1− 2+ 3− 4+ ] = −−−−→ − . (11.27)
12233441 122 232
We have used 3d momentum conservation 3441 = −3221 to simplify the result.

The two helicity states of the 4d gluon become two degrees of freedom in 3d that we can
organize as a 3d “gauge boson” and a scalar.
11.3.2 3d SYM and Chern–Simons theory

The 3-dimensional Yang–Mills action

1
LYM = 2 d 3 x Tr Fμν F μν (11.28)
g
has a coupling g 2 of mass-dimension (mass)1 . We are particularly interested in theories

with extra symmetry (after all, we keep getting marathonic mileage out of planar N = 4
SYM), but a superconformal theory needs dimensionless couplings.
In 3d, the gauge field can be introduced with a dimensionless coupling via the Chern–
Simons Lagrangian

κCS μνρ 2i
LCS = Tr Aμ ∂ν Aρ + Aμ Aν Aρ . (11.29)
4π 3
The coupling κCS is an integer and is called the Chern–Simons level.

The equation of motion derived from varying LCS with respect to the gauge field is
∂[μ Aν] + i[Aμ , Aν ] = Fμν = 0 . (11.30)
The solution to this equation is simply Aμ = g∂μ g −1 , where g is an arbitrary element in

the gauge group. This means that the gauge field is pure gauge, or a flat connection. For
us, the relevant implication is that the Chern–Simons gauge field does not carry any
physical degrees of freedom, since one can always choose a gauge such that Aμ = 0. This
is an important difference between a gauge field whose dynamics is governed by LCS versus
the usual Yang–Mills Lagrangian LYM : the gauge boson scattering amplitudes of 3d Yang–
Mills theory are non-trivial, but for a theory with just a Chern–Simons term the scattering
amplitudes are trivially zero because there are no physical states to scatter.
There can be non-trivial scattering amplitudes for Chern–Simons theory provided matter
fields are introduced. The Chern–Simons matter Lagrangian is typically written
L = LCS + Lφψ , (11.31)
where the matter Lagrangian Lφψ encodes the interactions of scalars φ and fermions ψ
with the Chern–Simons gauge field as well as their mutual interactions. In 3d, the (complex)
scalar- and fermion-interactions with dimensionless couplings are of the form φ 3 φ̄ 3 and
ψ̄ψ φ̄φ. Thus for superconformal theories, Lφψ takes the form
Lφψ = −D μ φ̄ Dμ φ + i ψ̄ D ψ + Vψ ψ̄φ φ̄ + Vφ 3 φ̄ 3 , (11.32)

where Vψ ψ̄φ φ̄ and Vφ 3 φ̄ 3 are quartic and sextic interaction terms. The explicit form of these
terms depends on the theory; we will show you two examples, namely the N = 8 and
N = 6 superconformal 3d theories (Sections 11.3.5 and 11.3.6). But let us first explore the
properties of amplitudes in 3d a little further.
11.3.3 Special kinematics and poles in amplitudes

The are 3-particle interaction terms in the Lagrangians discussed in Section 11.3.2, but we
have learned in Section 11.3.1 that all on-shell 3-point amplitudes vanish in 3d. Nonetheless,
the 3-particle vertices still make their presence felt by hiding in special kinematic limits
of higher-point amplitudes. As an example of this, consider the limit s12 = 122 → 0 of a
4-point amplitude. In this limit, |1 becomes proportional to |2, so |1 = α|2 for some α.
Further, we must have (1 + α 2 )s23 = 0, since
s12 →0
0 = p42 = ( p1 + p2 + p3 )2 −−−→ 0 = s13 + s23 = (1 + α 2 )s23 . (11.33)
There are two types of solutions to this constraint. For generic α, s23 must be zero and one
can conclude that all Lorentz invariants vanish. This is in line with our previous discussion
that there are no Lorentz invariants for on-shell 3-point kinematics. However, the constraint
(1 + α 2 )s23 = 0 also admits a solution that allows non-trivial Lorentz invariants: α = ±i.
For α = ±i, we have p1 = − p2 and similarly p3 = − p4 . Thus this corresponds to the
kinematic configuration where two particles are traveling in straight lines:
2 3
k1 = k2 k3 = k4 . (11.34)
1 4
From momentum conservation, any exchange between the two particle lines must have
zero momentum, so when we approach the s12 → 0 limit the amplitude should develop a
singularity associated with the propagator of an exchanged soft particle.
To see this in an explicit example, consider the 3d YM gluon amplitude (11.27)
134
A4 [1234] = − . (11.35)
122 232
Taking the limit |1 → i |2, the amplitude indeed develops a non-trivial 1/s12 singularity
with a non-vanishing residue:

s23
A4 [1− 2+ 3− 4+ ] =− . (11.36)
|1 → i |2 s12
The 1/s12 singularity reflects the 1/ p 2 behavior of the gluon propagator.

Thus, although there are no massless 3-point amplitudes in 3d, the 4-point amplitude still
develops a non-trivial “soft” pole. The origin of this singularity comes from the exchange
of a soft particle between two particles going in straight lines. Note that the exchanged
particle has momentum p μ → 0, so it is not strictly going on-shell. This is also reflected
in the observation that amplitude (11.36) does not factorize into two 3-point amplitudes.
Importantly, the precise behavior of the singularity is dictated by the propagator of the
intermediate particle. If the exchanged particle is an ordinary Yang–Mills gluon, then we
should observe a 1/ p 2 singularity. That is what happened in the example (11.36). However,
if it is a fermion or a Chern–Simons gauge boson, one should find a 1/ p 2 singularity.6 In
the case of a Chern–Simons boson, it comes from the propagator of the gauge field in the
Lagrangian LCS of (11.29); in Landau gauge it is
μρν pρ
Aμ ( p)Aν (− p) = . (11.37)
p2
We are going to use information about poles in this special kinematic limit to constrain the
possible 4-point amplitudes in Section 11.3.5.
11.3.4 D = 3 superconformal algebra

We stated in Section 11.1 that the minimal spinors in 3d are 2-component Majorana spinors.
They satisfy the Majorana reality condition, and so do the supersymmetry charges. Thus
for N -fold supersymmetry in 3d, we have N real supercharges and the R-symmetry group
is S O(N ). Since the 3d theories discussed in this chapter have N = 8 or N = 6 supersym-
metry, we focus on N =even in the following. When N = 2M, the real supercharges can
be grouped into M = N /2 complex spinors Q a A and their complex conjugates Q 2aA . Here
A = 1, . . . , M is the index of the reduced SU (M) R-symmetry.
We introduce M on-shell superspace coordinates ηiA for each external leg. The super-
charge can now be written as
∂
2aA =
Q |ia ηi A , Q a A = |ia . (11.38)
i
∂ηi A
2aA , Q bB } = δ A B P ab .
You can quickly see that { Q
The generators (11.38) are part of a larger symmetry group: the O Sp(N |4) superconfor-
mal group. The notation O Sp(N |4) means that the bosonic generators include the S O(N )
6 Our reasoning here is valid only for n = 4. For example, the n-point Parke–Taylor amplitude with n > 4 has
only 1/i, i + 1 poles that do not exhibit the 1/ p 2 of the Yang–Mills gluon propagator. A careful inspection
μ
of the 3-point gluon vertex reveals that the only non-vanishing term is proportional to k2 (1 · 2 ) in the limit
where legs 1, 2 are the two gluons propagating in a straight line. This is dotted into the remaining Feynman
diagram which for n = 4 is simply another 3-point vertex that in this limit contributes just one term proportional
μ
to k3 (3 · 4 ). Hence, on this soft pole, the residue is simply given by the product of the two 3-point vertices.
For n > 4, the remaining Feynman diagrams give multiple contributions, and thus the residue of this soft pole
contains several terms that can cancel to leave behind a milder singularity. Therefore, our discussion of soft-pole
structure is only valid for n = 4.
R-symmetry as well as the Sp(4) conformal symmetry generators. More precisely, the
generators are:

P ab = |ia |ib
i
2aA =
Q i |i ηi A
a
Q a A = i |ia ∂ηi A

Mab = i i|(a ∂|ib) D = i 12 |ia ∂|ia + 12

(11.39)
R AB = i ηi A ηi B R A B = i ηi A ∂ηi B − 12 δ A B R AB = i ∂ηi A ∂ηi B

S̃a A = i ∂|ia ηi A SaA = i ∂|ia ∂ηi A

K ab = i ∂|ia ∂|ib .
The S O(N ) R-symmetry generators are separated into U (N /2) generators R A B and coset
generators R AB and R AB of S O(N )/U (N /2).
As an important application for these generators, let us explore what kind of constraint the
U (1) piece of the U (N /2) R-symmetry imposes on the superamplitudes in a 3d N = 2M
superconformal theory. The U (1) piece is given by
M
RC C = ηiC ∂ηiC − . (11.40)
i
2
The RC C generator annihilates the superamplitude, RC C An = 0, if

M
ηiC ∂ηiC An = n An . (11.41)
i
2
The LHS simply counts the Grassmann degree of η A s in An . Since one cannot have
fractional degrees of η in An , the equation (11.41) can only hold for odd M if n = even.
So we learn that only even-multiplicity scattering amplitudes can be non-vanishing for a
superconformal theory with M odd.
The same is also true for M even. This is because superconformal theories generally
require the presence of a gauge field whose self-interaction is described by the Chern–
Simons Lagrangian (11.29). The form of the Chern–Simons matter Lagrangian (11.31),
(11.32) implies that any odd-multiplicity Feynman diagram has at least one external leg
associated with a gauge field. As discussed previously, the Chern–Simons gauge field does
not carry any physical degrees of freedom, so this means that the odd-multiplicity scattering
amplitudes must vanish.
So we learn that in a 3d superconformal theory, the Grassmann degree of the superam-
plitudes, i.e. the N K MHV-level, is rigidly tied to the number of external particles, contrary
to its freer life in 4d. For example for N = 8, an MHV superamplitude has Grassmann
degree 8 and by (11.41) it exists only for n = 4 external particles in a 3d superconformal
theory; a 6-point superamplitude on the other hand must have Grassmann degree 12, so it
has to be NMHV. Thus in a 3d N = 8 superconformal theory, there is no tower of MHV
superamplitudes, no equivalent of the n-gluon Parke–Taylor amplitude. Similarly, in a 3d
N = 6 superconformal theory, the 4-point superamplitude must have Grassmann degree 6.
11.3.5 N = 8 superconformal theory: BLG

A 3d superconformal theory with N = 8-fold supersymmetry has an on-shell spectrum with
eight scalars (φ, φ AB , φ̄) and eight fermions (ψ A , ψ̄ A ). Just as in N = 4 SYM in 4d, it is
convenient to encode the degrees of freedom in an on-shell superfield,
1 1
= φ + η A ψ A − η A η B φ AB − ABC D η A η B ηC ψ̄ D + η1 η2 η3 η4 φ̄ . (11.42)
2 3!
We have η A → −η A under little group transformations, so the superfield is inert. This

means that the superamplitudes are also invariant under little group transformations.
Since there are no massless 3-point amplitudes, let us consider the most general 4-
point tree superamplitude that enjoys N = 8 3d superconformal symmetry. To start with,
invariance under N = 8 supersymmetry implies that an n-point superamplitude takes the
form

2 f n |i, ηi ,
An = δ 3 P δ (8) Q (11.43)

2 = 4A=1 ( 1 Q
where δ (8) ( Q) 2a Q 2 ). The function f n is constrained further by the supercon-
2 A aA
formal generators (11.39).
As noted at the end of Section 11.3.4, the U (1) generator (11.41) requires the 4-point
superamplitude to have degree 8. Since the supermomentum delta function is already degree
8 in the ηi s, we infer that f 4 can only depend on the bosonic variables |i.
Next, annihilation of the superamplitude by the dilatation operator D in (11.39)
implies
n
DAn = 0 → 1
2
|ia ∂|ia An = − An . (11.44)
i
2

As in 4d (see Exercise 5.2), the operator i ( 12 |ia ∂|ia ) counts the mass dimension when
acting on a function of spinor brackets. It also acts on the delta functions in An , giving a
2 Thus, by (11.44), dilatation invariance requires the
factor of −3 on δ 3 (P) and 4 on δ (8) ( Q).
mass-dimension of f 4 to be − 2 − (−3 + 4) = −3.
4
Exercise 11.5

i ( 2 |i ∂|ia ) δ (P) = −3δ 3 (P).
1 a 3
Use the example around equation (5.4) to show that
To summarize, from invariance under U (1) R-symmetry and dilatation, we conclude that
f 4 is a purely bosonic function of mass-dimension −3. Finally, taking into account that
the superamplitude must be little group invariant, we can write the 4-point superamplitude
[148] as

1
2
A4 (1 2 3 4 = δ 3 P δ (8) Q . (11.45)
122331
We can now project out the 4-scalar amplitude A4 (φφ φ̄ φ̄) using (11.42): the Grass-
mann delta function produces a factor of 344 , so we get (with the help of momentum
conservation)
344 343
A4 (φφ φ̄ φ̄) = =− , (11.46)
122313 2423
The astute reader should object: multiplying the solution (11.45) by an arbitrary function of
1324
1423
still satisfies all previous criteria. This is indeed a valid objection; however, such a
function would change the pole structure of component amplitudes, such as (11.46), gener-
ated by A4 . We have imposed in (11.45) that the amplitudes only have 1/ p 2 poles. Why?
Well, since the only scalar-fermion interactions are of the form φ 3 φ̄ 3 and ψ̄ψ φ̄φ, poles in
the tree-level amplitude A4 (φφ φ̄ φ̄) cannot arise from scalar or fermion propagators. Hence
the only option is that they come from gauge boson exchanges. Since we are considering
a 3d superconformal theory, the gauge boson self-coupling must be dimensionless; this
rules out 3d Yang–Mills theory and rules in Chern–Simons gauge theory. Hence all poles
in A4 (φφ φ̄ φ̄) must be 1/ p and this fixes the 4-point tree superamplitude in an N = 8
2
superconformal 3d theory to be (11.45).

The result (11.45) for the superamplitude has a very important property: it is antisym-
metric under the exchange of any two external particles. This property is inherited by the
component amplitude A4 (φφ φ̄ φ̄) in (11.46) and that contradicts the expected Bose sym-
metry. We encountered something similar in Section 2.6 when we wrote down the 3-point
gluon amplitudes in 4d from just little group scaling and dimensional analysis. The reso-
lution was to include the antisymmetric structure constants f abc of the Yang–Mills gauge
group.
At the superamplitude level, the same issue arises: the physical degrees of freedom are
contained in the bosonic superfield , so A4 (1 2 3 4 ) should be Bose symmetric under
the exchange of any two external legs. But – as you see from (11.45) – it is fully antisymmet-
ric. We could avoid this contradiction if the amplitudes were color-ordered. However, the
presence of the 1/24 pole in for example (11.46) invalidates this interpretation. Instead,
the contradiction can be resolved if we include more than one supermultiplet, each one
with a label ai . Then we can introduce a new 4-index “coupling constant” f a1 a2 a3 a4 that is
completely antisymmetric in all four indices. Using this we write

f a1 a2 a3 a4
2
A4 a11 a22 a33 a44 = δ 3 P δ (8) Q . (11.47)
122331
Now Bose symmetry is respected. Thus by requiring N = 8 superconformal symmetry in

3d, the 4-point superamplitude forces us to introduce a completely antisymmetric 4-index
coupling constant.
This new coupling constant looks similar to the totally antisymmetric 3-index struc-
ture constant of Yang–Mills theory f a1 a2 a3 . This resemblance is not a coincidence. In the
search for an N = 8 super Chern–Simons matter theory, Bagger, Lambert, and Gustavsson
(BLG) [151, 152] found a Lagrangian whose gauge symmetry is built on a Lie 3-algebra.
This algebra is defined through a triple product
[T a , T b , T c ] = f abc d T d . (11.48)
The gauge indices are raised/lowered with h ab = Tr T a T b and its inverse. The structure
constants f abcd = f abc e h ed are totally antisymmetric. Just as the structure constants of
the usual gauge Lie 2-algebra satisfy the Jacobi identity (2.69), the 3-algebra structure
constants are required to satisfy a four-term “fundamental identity”:
f f gd
e f abce − f f ga
e f bcde + f f gb
e f cdae − f f gc
e f dabe = 0 . (11.49)
The fields in BLG theory [151, 152] consist of eight scalars X aIv with Iv = 1, . . . , 8 trans-
forming as a vector of S O(8), eight real spinors aIc with Ic = 1, . . . , 8 transforming as
a chiral spinor of S O(8), and a Chern–Simons gauge field Aab μ . The BLG Lagrangian is
[153]

1 1 μνρ 1 abcd 1
LBLG = f Aμab ∂ν Aρcd + f cda g f e f gb Aμab Aνcd Aρe f
κCS 48 2 3
1 i ¯ Ic
− D μ X aIv Dμ X aIv + D aIc + i3 f abcd ¯ a Iv Jv b X cIv X dJv
2 2 a
− 12 f abcd f a e f g (X bIv X cJv X dK v )(X eIv X Jf v X gK v ) . (11.50)
In the Lagrangian construction [151, 152], the need for the antisymmetric 4-index structure
constant comes from the requirement that the supersymmetry transformations on the fields
close into the correct algebra. Linear combinations of the eight scalars and fermions can be
identified as the (φ, φ AB , φ̄) and (ψ A , ψ̄ A ) components of our superfield (11.42). Indeed,
the 4-point amplitudes computed from the Lagrangian (11.50) match [148] the component
amplitudes of the 4-point superamplitude (11.47).
It is quite non-trivial for an antisymmetric f abcd to satisfy (11.49) and currently the only
known example is if a is an index of S O(4) and f abcd ∼ abcd . In search for other examples,
there were many attempts to relax the symmetry properties of the 4-index structure constant
[154–156]. However, as we have shown from the on-shell analysis, N = 8 superconformal
symmetry only allows for a totally antisymmetric structure constant. Indeed, all known
examples of Lie 3-algebras with f abcd not totally antisymmetric correspond to Chern–
Simons matter theories with N < 8 supersymmetries.
11.3.6 N = 6 superconformal theory: ABJM

Let us now consider a 3d superconformal theory with N = 6 supersymmetry. The R-
symmetry is S O(6) = SU (4) and the physical degrees of freedom are four complex scalars
X A and four complex fermions ψ Aa as well as their complex conjugates X̄ A and ψ̄Aa . They
transform in the fundamental or anti-fundamental of SU (4) and A = 1, 2, 3, 4. To arrange
these states in on-shell superspace, we introduce three anticommuting variables η A and

write
1
= X 4 + η A ψ A − ABC η A η B X C − η1 η2 η3 ψ 4 ,
2 (11.51)
¯ = ψ̄4 + η A X̄ A − 1 ABC η A η B ψ̄C − η1 η2 η3 X̄ 4 .

2
We have split the fields as X A → (X 4 , X A ) and ψ A → (ψ 4 , ψ A ), and similarly for X̄ A

and ψ̄A . So only an SU (3) subgroup of the SU (4) is manifest in this on-shell superspace
formalism.
The on-shell superspace representation (11.51) involves a bosonic superfield and a
fermionic superfield .
¯ Having two superfields is standard for superamplitudes in theories
with less-than-maximal supersymmetry. For example in 4d N < 4 SYM, the spectrum is
not CPT self-conjugate and therefore a superfield is needed for each of the CPT conjugate
supermultiplets; details and applications of the formalism can be found in [87]. In 3d, the
need for two superfields comes from R-symmetry. Just as in 4d N < 4 SYM, where
the two superfields have states that are parity-conjugate with respect to each other, in 3d
the two superfields contain states that are conjugate to each other under R-symmetry.
Since fermions transform with a minus under 3d little group transformations, the super-
amplitude must be odd under |i → −|i and ηi → −ηi if i is a ¯ state. Following the
same steps as for the N = 8 BLG theory in Section 11.3.5, we then find that the 4-point su-
peramplitude in a 3d N = 6 superconformal theory is fixed up to a multiplicative constant
to be [157]

1
A4 ¯ 1 2 2
¯ 3 4 = δ 3 P δ (6) Q . (11.52)
1443
The 4-point superamplitude (11.52) precisely encodes the color-ordered 4-point amplitudes
of an N = 6 Chern–Simons matter theory that was constructed by Aharony, Bergman,
Jafferis, and Maldacena (ABJM) [158]. The theory, known as the ABJM theory, contains
two gauge fields Aa b and Âȧ ḃ with gauge group U (N ) × U (N ). The matter fields are in
the bi-fundamental, meaning that they transform in the fundamental of one U (N ) gauge
group and the anti-fundamental of the other U (N ). Hence, the index structures of the matter
fields are (X A )ȧ a , ( X̄ A )a ȧ , (ψ A )ȧ a , and (ψ̄A )a ȧ . The Lagrangian is [159, 160]
&
k 1 μνρ 2i 2i
LABJM = Tr Aμ ∂ν Aρ + Aμ Aν Aρ − Âμ ∂ν Âρ − Âμ Âν Âρ
2π 2 3 3
%
− (D μ X A )† Dμ X A + i ψ̄A D ψ A + L4 + L6 , (11.53)
where the covariant derivatives for the bi-fundamental fields are
D μ X A ≡ ∂μ X A + i Âμ X A − i X A Aμ ,
(D μ X A )† ≡ ∂μ X̄ A + i Aμ X̄ A − i X̄ A Âμ , (11.54)
with the same definitions for ψ A and ψ̄A .

The quartic and sextic interaction terms in (11.53) are

L4 = i Tr X̄ B X B ψ̄A ψ A − X B X̄ B ψ A ψ̄A + 2X A X̄ B ψ A ψ̄B − 2 X̄ A X B ψ̄A ψ B

− ABCD X̄ A ψ B X̄ C ψ D + ABCD X A ψ̄B X C ψ̄D , (11.55)
1
L6 = Tr X A X̄ A X B X̄ B X C X̄ C + X̄ A X A X̄ B X B X̄ C X C + 4 X̄ A X B X̄ C X A X̄ B X C
3

− 6X A X̄ B X B X̄ A X C X̄ C . (11.56)
For theories whose external states are bi-fundamental matter fields, the color structure of
the amplitude is given in terms of a product of Kronecker deltas. In particular, with n = 2m
the full color-dressed amplitude is [157]
ȧσ ȧ aσ̄ a
An (1̄, σ1 , σ̄1 , . . . , σ̄m−1 , σm ) δȧ1̄ 1 · · · δȧσ̄σm δaσ11 · · · δaσ1̄m , (11.57)
m−1
σ ∈Sm , σ̄ ∈ S̄m−1
where the sums are over all distinct permutations of m even sites and m − 1 odd sites.
Each partial amplitude An is multiplied by a product of Kronecker deltas, and this naturally
defines an ordering, very similar to Yang–Mills amplitudes. However, since the on-shell
degrees of freedom are contained in two distinct supermultiplets, the color-ordered super-
amplitude is not cyclically invariant, but invariant up to a sign under cyclic rotation of two
sites:

An=2m ¯ 1 2 . . . 2m = (−1)m−1 An=2m ¯ 3 4 . . . 2m
¯ 1 2 ; (11.58)
the minus signs come from the exchanges of s. ¯ For the superamplitude (11.52), the 2-site
cyclic property (11.58) is ensured by momentum conservation.
After having seen a Lie 3-algebra appear in the N = 8 superconformal BLG theory in
Section 11.3.5, you may wonder if the above Lagrangian can also be rewritten in terms of
a 3-algebra. Indeed it can! In fact, we can read off the properties of the 4-index structure
constants from the 4-point superamplitude (11.52). It is symmetric under the exchange of
the legs that correspond to the fermionic supermultiplet , ¯ while it is antisymmetric under
the exchange of the bosonic multiplets . This is opposite from the expected symmetry
properties of A4 ( ¯ 1 2
¯ 3 4 ), and therefore one can consider dressing the superamplitude
with a 4-index structure constant f a2 a4 ā1 ā3 that is antisymmetric with respect to the exchange
of barred or unbarred indices, respectively. The color-dressed superamplitude is then7

ā1 a2 ā3 a4

f a2 a4 ā1 ā3
A4
¯ 2
¯ = δ 3 P δ (6) Q . (11.59)
1 2 3 4
1443
7 You might wonder why this issue did not come up when we stated that the amplitude in (11.52) matched that
derived from the Lagrangian (11.53). The reason is that it matched in the context of a color-ordered amplitude
where the exchange of external lines is not a symmetry. In contrast, here we are considering a fully color-dressed
amplitude. In other words, we are asking what properties the color factor should have such that the amplitude
can be considered as a color-dressed amplitude.
It has been shown [161] that the Lagrangian (11.53) is completely equivalent to an alterna-
tive one where the matter fields carry the 3-algebra indices indicated in (11.59).
11.3.7 BCFW recursion in 3d

We argued in Section 11.3.4 that only even-point amplitudes are non-vanishing in 3d
superconformal theories. This means that the 4-point superamplitudes are the building
blocks of higher-point amplitudes in these theories. Conveniently, we found that the 4-point
tree-level superamplitudes in N = 8 and N = 6 theories in Sections 11.3.5 and 11.3.6 are
completely determined by the requirements of symmetries and pole structure. Now is the
time to go to higher-point, and of course our favorite tool is BCFW recursion.
To get started, we have to set up a BCFW recursion relation in 3d. And 3d is different
from all other D > 3 in terms of defining a BCFW deformation. To see this, recall from
Chapter 3 that we shift two external momenta i and j linearly,
pi → pi + zq, p j → p j − zq , (11.60)
with a vector q that satisfies
q · pi = q · p j = q 2 = 0 . (11.61)
This ensures that the shifted momenta remain on-shell and that invariants P̂i2j...k are linear
in z, so that each propagator going on-shell corresponds to a unique simple pole in the
z-plane.
Unfortunately (or, very interestingly, if that is how you like it), in 3d the only q that
satisfies these constraints is q = 0. The reason is this. A 3d vector q with q 2 = 0 can be
written as a bi-spinor q = |qa |qb . The |q is a 2-component spinor so it cannot be linearly
independent from the |i and | j of the two light-like momenta we are shifting. Hence
|q = α|i + β| j (11.62)
for some numbers α and β. Solving for α and β subject to the constraints q · pi = q · p j = 0
in (11.61) gives α = β = 0 and hence q = 0.
So in order to make progress, we need to relax some of the constraints imposed on the
shifted momenta. We cannot give up on momentum conservation and on-shellness for the
shifted momenta. Instead, we can either shift three or more external momenta or give up
on the property that the momenta shift linearly in z. The former is similar to the shift
associated with CSW (Section 3.4) and comes at the price of involving many diagrams for
the superamplitudes. Opting for the solution with fewer diagrams, we choose to give up on
linearity in z and consider the following general 2-line “deformation” [162]

|î |i
= R(z) . (11.63)
| ĵ | j
R(z) is a 2 × 2 matrix that depends on z. Since we want the shift to respect momentum
conservation, the matrix R must satisfy:
R(z)T R(z) = I . (11.64)
Thus R(z) is an orthogonal matrix and we can parameterize it as

z+z −1 −1
cos θ − sin θ 2
− z−z2i
R(z) = = . (11.65)
sin θ cos θ z−z −1 z+z −1
2i 2
If we define the deformation on the fermionic variables ηi and η j in the same fashion,
supermomentum conservation is also preserved by the shift:

η̂i A
T ηi A
ˆq̃i A + q̃ˆ j A = |î , | ĵ = |i , | j R (z)R(z) = q̃i A + q̃ j A .
η̂ j A ηjA
(11.66)
The deformation matrix (11.65) becomes the identity when z = 1, so unshifted kinematics
corresponds to z = 1 and not z = 0. This leads to the following contour integral represen-
tation of the unshifted tree-level amplitude:
*
1 Ân (z)
An = , (11.67)
2πi z=1 z − 1
where the contour wraps just the pole at z = 1. If the deformed superamplitude Ân (z)
vanishes as z → ∞,8 one can perform a contour-deformation and evaluate the amplitude
as a sum of the residues at finite z = 0, 1.
Just as in 4d, the poles at finite z = 0, 1 correspond to propagators going on-shell. Let us
take a closer look at what the singularities look like. Without loss of generality, we choose
1 and n as the deformed momenta:
1 ab
p̂1ab = ( p + pnab ) + z 2 q ab + z −2 q̃ ab ,
2 1
1
p̂nab = ( p1ab + pnab ) − z 2 q ab − z −2 q̃ ab . (11.68)
2
Here q and q̃ are given by
1 1
q ab = (|1 + i|n)a (|1 + i|n)b , q̃ ab = (|1 − i|n)a (|1 − i|n)b . (11.69)
4 4
ab
Defining P12...i = p1ab + p2ab + · · · + piab , the on-shell condition for the shifted propagator
2
P̂12...i takes the form
2
P̂12...i = q̃|P23...i |q̃z −2 + q|P23...i |qz 2 − (P23...i · Pi+1...n−1 ) = 0 , (11.70)
μ
where ( pi · p j ) = pi p jμ and i|P| j ≡ λia Pa b λ jb . One can explicitly write down the
values of z that correspond to the propagator going on-shell

∗ 2 ∗ 2 (P2...i · Pi+1...n−1 ) ± (P2...i )2 (Pi+1...n−1 )2
(z 1,i ) , (z 2,i ) = . (11.71)
2q|P2...i |q
8 One should also make sure that there are no poles at z = 0. Exchanging 1/z ↔ z in (11.65) can be compensated
by extra sign factors in the kinematics of the shifted legs, so if An (z) vanishes as z → ∞ for generic kinematics,
then it also vanishes at z = 0.
Exercise 11.6
Prove the following useful identity

∗ 2
∗ 2 2
P12...i
(z 1,i ) − 1 (z 2,i ) −1 = . (11.72)
q|P2...i |q
As the propagator goes on-shell, the amplitude factorizes into two lower-point amplitudes.
This allows us to write the sum of residues at z = 1 as a sum over distinct single propa-
gator diagrams with legs 1 and n on opposite sides of the propagator. For Chern–Simons
matter theories, we also require that only even multiplicity subamplitudes appear on each
side of the propagator. For each propagator, one needs to sum over the four solutions,
∗ ∗ ∗ ∗
(z 1,i , −z 1,i , z 2,i , −z 2,i ) to the on-shell constraint (11.70). The final result is then [162]
∗ ∗

∗
H (z 1, f , z 2, f )
∗

∗ ∗

An = d ηI
3
Â L z 1, f; ηI 2
Â R z 1, f; iη I + z 1, f ↔ z 2, f ,
i
P12...i
(11.73)
where the function H (a, b) is

a(b2 − 1)
H (a, b) ≡ (11.74)
a 2 − b2
and the Grassmann integral takes care of the intermediate state sum. Did you notice the i
in Â R ? That comes from the analytic continuation of the incoming → outgoing internal
line. In 3d massless kinematics, we only have one type of spinor, namely | p, so with
pab = −| pa | pb we must have
| − p = i | p . (11.75)
Hence we must also have η− p = i η p , since – as you can check – this ensures that the
arguments of the L and R Grassmann delta functions add up to the overall supermomentum
2
Q.
Exercise 11.7
Show that the contour deformation of (11.67) gives the representation (11.73).
[Hints: The identity in (11.72) will be useful. Furthermore, since one of the shifted legs,
1 or n, necessarily corresponds to the fermionic multiplet, we have A L (−z)A R (−z) =
−A L (z)A R (z).]
The validity of (11.73) relies on whether or not the super-shifted superamplitude vanishes
as z → ∞. It was shown in [162] that this criterion is satisfied for the ABJM and BLG
theories.
Exercise 11.8
Recall that in Section 3.3 we discussed when a BCFW recursion is valid: we showed
that the presence of contact terms, for example φ 4 , in the action tends to spoil the
recursion since such terms go to a constant as z → ∞. This issue can be avoided in

supersymmetric theories since amplitudes where such terms are present are related via
supersymmetry to those from which it is absent. This is accomplished via the super-
BCFW shifts. One can illustrate the idea by carefully choosing the external states such
that contact terms do not contribute to a particular component amplitude; then (loosely
speaking) supersymmetry ensures that the superamplitude, which contains this well-
behaved component amplitude, also goes to zero for z → ∞. Let us test whether such
a component amplitude can be found for ABJM theory at 6-point. Consider the 6-point
contact term in (11.56). Show that if we choose all R-symmetry indices to be the same,
say 1, then the sextic interaction terms vanish. Thus the 6-point scalar amplitude with
all scalars having the same SU (4) indices has good large-z behavior.
One thing is deriving the recursion relations, another thing is using them! So let us now apply
the 3d recursion relations to compute the 6-point amplitude A6 ( X̄ 4 X 4 X̄ 4 X 4 ψ̄4 ψ 4 ) in ABJM
theory. For simplicity, we drop the SU (4) indices on the component-fields, i.e. X̄ 4 → X̄ .
Choosing lines 1 and 6 for the shift, the only factorization channel is (123|456), so there is
only one diagram, namely
6^ P^123 1^
5 2. (11.76)
4 3
The recursion relation (11.73) then reads

A6 ¯ ¯ ¯
& %

H (z 1∗ , z 2∗ )

= d 3 η Â4 ¯ 2
¯ 3 A ¯ ¯ 5
+ (z ∗
↔ z ∗
) .
1̂ P̂123 2 4 − P̂123 4 6̂ 1 2
z=z 1∗ P123 z=z 1∗
(11.77)
To project out the amplitude A6 ( X̄ X X̄ X ψ̄ψ) from the superamplitude (11.77) we need
the coefficient of the (η1 )3 (η3 )3 (η6 )3 monomial, where (ηi )3 = ηi1 ηi2 ηi3 . This follows from
(11.51). After manipulation of the Grassmann delta functions and using η̂1 (z)η̂6 (z) = η1 η6 ,
we find

H (z 1∗ , z 2∗ )

A6 X̄ X X̄ X ψ̄ψ = Â4 X̄ 1̂ X 2 X̄ 3 X P̂123 ∗ 2
Â 4 X̄ X ψ̄ ψ
P̂123 4 5 6̂ + (z 1∗ ↔ z 2∗ ) ,
z=z 1 P123 z=z 1∗
(11.78)
where the 4-point amplitudes, obtained from the superamplitude (11.52), are

1̂33
P̂123 6̂2
Â4 X̄ 1̂ X 2 X̄ 3 X P̂123 = − and Â4 X̄ P̂123 X 4 ψ̄5 ψ6̂ = . (11.79)
1̂ P̂123 P̂123 3 6̂5
By (11.71), the poles in the z-plane are located at

2
2
∗2 162 − 23 − 45 ∗2 162 − 23 + 45
z1 =

, z2 =

. (11.80)
1| + i6| P45 |1 + i|6 1| + i6| P45 |1 + i|6
After repeated use of momentum conservation and the Schouten identity, we find that the
6-point amplitude is
$
3

1 2|P123 |6 + i3145
A6 X̄ X X̄ X ψ̄ψ = − 2

2P123 1|P123 |4 + i2356 3|P123 |6 + i1245

3 '
2|P123 |6 − i3145
−

.
1|P123 |4 − i2356 3|P123 |6 − i1245
(11.81)
Here the first term is the result of evaluating the first term in (11.77) while the second term
above is the (z 1∗ ↔ z 2∗ ) contribution.
Exercise 11.9
Let us derive (11.81) from (11.78). First prove

∗2
−i1|P23 |6 + 2345
(z 1 ) + 1 (z 2∗ )2 − 1 = . (11.82)
q|P2...i |q
Next, use (11.82) to show that

i23 2|P123 |6 + i3145
1̂3 z 1∗ (z 2∗ )2 − 1 = . (11.83)
2q|P23 |q
Now continue to manipulate the tree amplitudes (11.79) to derive the first line in
(11.81).
You may worry about the apparently spurious poles in the expression (11.81), since each
only appears in one term and not the other and thus cannot cancel. But have no fear, these
are really local poles in disguise! To see this, we rewrite them as (see Exercise 11.10)
1 1|P123 |4 + i2356 1|P123 |4 + i2356

= = . (11.84)
1|P123 |4 − i2356 1|P123 |42 + 232 562 2
P123 2
P234
Thus each spurious-looking pole in (11.81) is really a product of local poles. Note that
this tells us that the two terms in the BCFW result (11.81) are individually local and free
of spurious poles! The reason behind this will be discussed further in Sections 11.3.8 and
11.3.10.
Exercise 11.10
The final manipulation in (11.84) made use of the identity
i| p j + pk |l2 − ( pi + p j + pk + pl )2 jk2 = ( pi + p j + pk )2 ( p j + pk + pl )2 ,
which holds for any four massless vectors pi , p j , pk , pl in 3d. Prove it.
Exercise 11.11
Although the two terms in (11.81) are individually local, they actually need to come
in the combination in (11.81): show that the relative minus sign is necessary for the
amplitude to have the correct little group properties.
11.3.8 ABJM and dual conformal symmetry

Let us dive straight into the deep end and define 3d dual variables yiab and θiaA such that
yiab − yi+1
ab
= piab and θiaA − θi+1,A
a
= q̃iaA . Momentum and supermomentum delta functions
for a 4-point superamplitude are then

δ 3 P δ (N ) Q̃ → δ 3 y1 − y5 δ (N ) θ1a − θ5a . (11.85)
We define dual conformal inversion on the variables yi and θi the same way in any
spacetime dimension, namely as in (5.37). It then follows from (11.85) that the inversion
weights of the momentum and supermomentum delta function exactly cancel for N = 6
supersymmetry. Using
2
yi,i+2 = si,i+1 = i, i + 12 , (11.86)
we deduce (as in Exercise 5.6) the dual inversion rule for a 3d angle bracket
i, i + 1
I [i, i + 1] = 3 . (11.87)
yi2 yi+2
2
For the 4-point superamplitude (11.52) of N = 6 ABJM theory, this then implies
3
I A4 (
¯ 1 2
¯ 3 4 ) = y 2 y 2 y 2 y 2 A 4 (
1 2 3 4
¯ 1 2
¯ 3 4 ) . (11.88)
It can be shown [162], using the N = 6 super-BCFW recursion relations, that dual inversion
on the n-point tree-level superamplitude gives
n 3
#
I [An ] = yi2 An . (11.89)
i=1
Thus the 3d ABJM tree-level superamplitudes are dual conformal covariant with uniform
inversion weight 12 on each leg.
Under dual conformal inversion, the superamplitudes of 4d N = 4 SYM transform
covariantly with uniform inversion weight 1 on each leg. In Section 5.3 we argued that,
as a result, the dual conformal boosts Kμ annihilate the superamplitudes only after the
non-trivial weights have been compensated by a shift of Kμ , as below (5.42). This shift
is crucial for defining the dual superconformal symmetry and extending it together with
the ordinary superconformal symmetry to the SU (2, 2|4) Yangian of the 4d planar N = 4
SYM superamplitudes.
In 3d, the dual conformal symmetry can be enlarged into the dual superconformal
symmetry group O Sp(6|4) [163]. The symmetry group acts on the dual space that consists
of coordinates (yiab , θiaA , ri AB ), where the extra R-symmetry coordinate ri AB is defined by:
ri AB − ri+1,AB = ηi A ηi B . (11.90)
The group O Sp(6|4) is also the supergroup for the ordinary superconformal symmetry of
the ABJM Lagrangian in (11.53). The combination of the dual and ordinary superconformal
symmetries forms an infinite dimensional O Sp(6|4) Yangian algebra [157], very similar in
nature to the SU (2, 2|4) Yangian symmetry of 4d planar N = 4 SYM.
The super-BCFW construction (11.77) actually gives the tree-level 6-point superampli-
tude in terms of two Yangian invariants,
Atree
6 ≡ A6 ( 1 2 3 4 5 6 ) = Y1 + Y2 .
tree ¯ ¯ ¯ (11.91)
The two Yangian invariants Y1 and Y2 arise precisely from the two BCFW-terms in (11.77).
We will not need their explicit form; they can be found in [162].
It will be relevant for us to also consider the tree superamplitude with shifted sites,
Atree
6,shifted ≡ A6 (1 2 3 4 5 6 ) = A6 ( 2 3 4 5 6 1 ) .
tree ¯ ¯ ¯ tree ¯ ¯ ¯ (11.92)
It has a super-BCFW representation that can be written as
6,shifted = A6 (1 2 3 4 5 6 ) = Y1 − Y2 .
Atree tree ¯ ¯ ¯ (11.93)
Now the important point is that two physical objects, Atree 6 and Atree
6,shifted , are written as
distinct linear combinations of the same two Yangian invariants: this is only possible if each
of the two Yangian invariants is local, i.e. free of spurious poles. We have already noted the
locality for the particular component amplitude (11.81). Now you see why it was needed.
Note that this distinguishes 3d ABJM theory from 4d planar N = 4 SYM in which the dual
superconformal invariant 5-brackets had spurious poles.
11.3.9 Loops and on-shell diagrams in ABJM

The loop-level superamplitudes can be explored using unitarity methods (Chapter 6). Using
the dual inversion property of the tree-level superamplitudes, it can be shown that the planar
loop superamplitudes of ABJM, prior to integration, are dual conformal covariant, i.e. they
satisfy (11.89). Thus perturbatively, planar ABJM has a structure very similar to planar
N = 4 SYM, they are almost baby brothers/sisters. This is rather surprising given that the
two theories have very distinct Lagrangians and live in different spacetime dimensions.
Moreover, in quantum field theory textbooks, one learns that D < 4 theories generically
have more severe IR divergences compared to D = 4. Thus one might expect that although
planar ABJM is very similar to N = 4 SYM at the pre-integrated level, the similarity would
be completely scrambled by the potentially more severe IR-divergence in D = 3.
To see if this is the case, let us take a look at the planar loop amplitudes in detail. The 1-
loop amplitudes in ABJM are purely rational functions [164–166]. This can be understood
as a consequence of dual conformal symmetry, since the only dual conformal covariant
scalar integral is the massive triangle integral, and it integrates to
K2
iπ 1
I3 (K 1 , K 2 , K 3 ) = − 3 3 3 , (11.94)
2 −K 2 −K 2 −K 2
1 2 3
K1 K3
where K 1 , K 2 , K 3 are the sums of the external momenta going out of each of the three
corners and K i2 = 0. There are no triangle diagrams with massless corners K i2 = 0; this
follows from generalized unitarity methods using the fact that 3-point loop amplitudes
vanish. The integrated result (11.94) has transcendentality 1 thanks to the factor of π .
Since 6-point is the lowest multiplicity at which the triangle integral (11.94) contributes
after integration, we conclude that the 4-point 1-loop amplitude vanishes up to O() in
dimensional regularization D = 3 − 2.9
The 6-point 1-loop superamplitude is [164, 165, 168]10

1-loop N
A6 = −i Atree
6,shifted
k

× 123456 I3 (P12 , P34 , P56 ) + 234561 I3 (P23 , P45 , P61 ) , (11.95)
where N comes from the gauge group U (N ) × U (N ), and k is the Chern–Simons level in
(11.53). The tree superamplitude Atree
6,shifted was defined in (11.92). Using the integrated result
(11.94) for the scalar triangle integrals I3 , we find that the 1-loop 6-point superamplitude
is

1-loop π N
A6 =− Atree
6,shifted
2 k

× sgn(12) sgn(34) sgn(56) + sgn(23) sgn(45) sgn(61) . (11.96)
Here we have introduced

i j i j
sgn(i j) ≡ 3 = , (11.97)
−K i2j |i j|
which equals ±1 depending on the kinematics. Thus, remarkably, the 1-loop 6-point su-
peramplitude can be either zero or non-vanishing depending on the kinematics! This
peculiar behavior has to do with an interesting topological feature of light-like mo-
menta in 3-dimensions. In 3d Minkowski space, a light-like vector can be written as
μ
pi = E i (1, cos θi , sin θi ). This means that light-like vectors can be projected to points on
9 The 4-point 1-loop integrand is non-trivial. It is given by a loop-momentum dependent integrand that integrates
to zero up to O() [167].
10 This result is only valid up to O(). There are additional integrands, whose coefficients are proportional to the
tree amplitude, that integrate to zero up to O().
a circle S 1 . From

θi − θ j
i j = −2 pi · p j = i E i E j sin (11.98)
2
we see that the sign of i j changes whenever the two points that represent pi and p j cross
each other on the S 1 . Thus the 1-loop amplitude encounters a sudden jump, from zero to
non-vanishing or vice versa, whenever two points on the S 1 cross each other:
i i+1 i+1 i
i+2 i+2
i−1 i−1
. (11.99)
n n
1 1
The two configurations (11.99) are topologically inequivalent.11 Thus this sudden jolt
is the amplitude’s way of telling us that we are changing the topology of momentum
space!
Exercise 11.12
From 4d, we are familiar with loop amplitudes being proportional to the tree amplitudes,
so it may seem odd that here in 3d the 1-loop 6-point amplitude (11.96) is proportional
to the shifted tree amplitude. Verify that the LHS and RHS of (11.96) have the same
little group scaling thanks to the sign-functions.
The first non-trivial loop contribution to the 4-point amplitude enters at 2-loop order and
it is given by [167, 169]
2 $
−

2 −
2-loop N −μ−2 y13
2
+ −μ−2 y24
A4 = Atree
4 −
k (2)2
'
2
1 y13
+ ln2 2
+ 4ζ2 − 3 ln 2 + O() . (11.100)
2
2 y24
Notice that the IR-divergent part is equivalent to that of the 1-loop 4-point superamplitude
(6.42) of planar N = 4 SYM, with → 2 because this is 2-loops.12 Not only is the

2 2 of this amplitude identical to that of planar N = 4 SYM, but so is the finite
IR-structure
ln2 y13 /y24 piece!
11 This can be made more precise. By judiciously adding 2π to the angles θi , one can arrange the angles such
that a given kinematics configuration has all angles strictly increasing according to their color ordering, i.e.
0 < θi+1 − θi < 2π . This gives a well defined “winding number” w = (θn − θ1 )/(2π ). Now as two points
cross each other, the winding number changes by one, indicating a distinct topological sector.
12 The overall scaling is not fixed because we have not identified the 4d N = 4 SYM gauge coupling g with a
power of N /k in ABJM.
Moving on to the 6-point 2-loop amplitude, one finds [170]
2 ! & % & %"

2-loop N Atree Atree
6,shifted u2
A6 = 6
BDS6 + R6 + ln ln χ1 + cyclic × 2 .
k 2 4i u3
(11.101)
Here BDS6 is the 1-loop MHV amplitude (6.57) of N = 4 SYM, again with proper rescaling
of the regulator → 2 to account for being at 2-loops. As the remaining pieces are finite,
the BDS Ansatz captures the IR-divergent as well as the resulting non-dual-conformal part
of the amplitude. So once again, we observe that the IR structure of planar ABJM theory is
identical to that of planar N = 4 SYM! The “remainder” function R6 in (11.101) is
3 & %
1 √
R6 = −2π 2 + Li2 (1 − u i ) + ln u i ln u i+1 + (arccos u i )2 , (11.102)
i=1
2
where the u i s are the dual conformal cross-ratios defined in (6.53). The shifted tree Atree
6,shifted
was encountered in (11.95). Finally, the function χ1 in (11.101) is
1245 + i3|P123 |6
χ1 = , (11.103)
1245 − i3|P123 |6
and “cyclic×2” means we sum over all cyclic rotations by two sites, i → i + 2.
Exercise 11.13
Seeing both Atree
6 and A6,shifted in the same amplitude, you should check that the other
tree
factors in (11.101) indeed compensate for the little group weight difference.
Now that we have seen explicit examples of planar loop amplitudes in ABJM theory, let us
turn to the subject of Leading Singularities and on-shell diagrams. We studied these for
planar 4d N = 4 SYM in Chapter 8. Because of the dual superconformal symmetry of the
loop-integrands, multi-loop amplitudes of ABJM theory can be calculated with Leading
Singularity methods. In 3d, a maximal cut takes three propagators on-shell for each loop-
momentum. At 1-loop order, the only dual conformal scalar integral is the massive triangle
(11.94), so this plays the equivalent role of the box-diagram in 4d. In 4d, we build the
Leading Singularity on-shell diagrams from vertices that are the fundamental 3-point
MHV and anti-MHV superamplitudes. These vanish in 3d, so here we use the 4-point
superamplitudes instead. In ABJM theory, the first non-trivial 1-loop Leading Singularity
is the 6-point diagram
. (11.104)
As noted above, each vertex represents an on-shell 4-point superamplitude of ABJM theory.
There is no distinction of “black” and “white” vertices because there is only one type of
4-point superamplitude in ABJM.
In 4d N = 4 SYM, we found that the 4-point Leading Singularity box diagram represents
the 4-point tree amplitudes (see Section 8.3). Similarly, in ABJM, it turns out that the
6-point Leading Singularity triangle diagram (11.104) reproduces the tree-level 6-point
superamplitude. To see this, we isolate the 3rd vertex in (11.104) and label the legs as
4 5
c
3 6
a b . (11.105)
2 1
The internal momenta a and b are on-shell, so they each have two degrees of freedom. Of
the combined 4 = 2 + 2 degrees of freedom in a and b, three are fixed in terms of momenta
1 and 2 by the momentum conservation delta function at the bottom vertex in (11.105).
Thus the spinor variables of a and b can be parameterized in terms of |1 and |2 using a
single free variable. With a little thought – or, even better, a little calculation – one finds
that the following parameterization solves the momentum conservation constraints
|a = cos θ |1 − sin θ |2 , |b = sin θ |1 + cos θ |2 . (11.106)
This is exactly the BCFW deformation (11.65) of legs 1 and 2. Indeed, the final on-shell
condition pc2 = 0 becomes the factorization condition that the parameter θ (i.e. z) must
satisfy.
Exercise 11.14
Verify that the supermomentum delta function on the bottom vertex enforces the
following identification ηa = cos θ η1 − sin θ η2 and ηb = sin θ η1 + cos θ η2 .
Now it is very tempting to conclude that the on-shell diagram (11.105) can be also under-
stood as a BCFW diagram for the 6-point tree superamplitude in ABJM. This is true, but
we have to make sure that we produce the BCFW recursion formula (11.73), including the
weight-factor H (a, b) defined in (11.74) and the propagator of the factorization channel.
Taking into account the Jacobian factors associated with the triple cut and the bottom
vertex, it has been shown [168] that H (a, b) and the factorization propagator are indeed
produced. Thus we have found the 3d analogue of the “BCFW bridge,” namely
4 5
c

3 6
H (z 1∗ , z 2∗ )
∗ ∗
a b = d η A4
3 ¯ ¯ A4 + (z 1 ↔ z 2 ) = Atree
¯ ¯ 6 .
(P234 )2
2 1
(11.107)
Recall that in 4d, the Leading Singularity is closely related to the integral coefficients
in expressions like (6.26). Previously we have seen that the 1-loop 6-point amplitude is
proportional to Atree
6,shifted , so it is perhaps puzzling that the 6-point Leading Singularity
(11.107) is just Atree
6 . This has to do with a subtlety of the Jacobian factors. Recall that
the integral coefficients can be determined by unitarity cuts. When we apply unitarity cuts,
we are substituting the propagators with delta functions, as discussed in Section 6.2. As
we solve the delta function constraints, we generate a Jacobian factor with an absolute
value. On the other hand, when we are computing the Leading Singularity, we treat the
delta functions as contour integrals, thus while localizing on a pole, the Jacobian factor
does not come with an absolute value. In the 1-loop cases that we encountered in 4d, the
Jacobian factors for the two loop-momentum solutions are identical, so the presence of an
absolute value did not make a difference. However, in 3d, the Jacobian factors for the two
loop-momentum solutions differ by a sign, so whether or not there is an absolute value on
the Jacobian makes a difference [170]. The result of this is that the 1-loop 6-point amplitude
6,shifted while the 6-point Leading Singularity is just A6 .
is proportional to Atree tree
Exercise 11.15
The above discussion indicates that if we had a relative plus sign for the two BCFW
terms on the RHS of (11.81), the result would be Atree 6,shifted instead of A6 . Verify that
tree
with a relative plus sign, (11.81) has the correct little group property of Atree6,shifted . [Hint:
you need to take into account that the coefficient for the η-polynomial corresponds to
a different component amplitude in the shifted amplitude.]
Instead of (11.105), we could have computed the on-shell diagram
5 6
c
4 a 1
b (11.108)
3 2
and the result would have been exactly the same, namely Atree
6 . This gives us the ABJM
equivalent of the “square move” (8.48) in 4d N = 4 SYM. The ABJM “triangle move” is
4 5 5 6
3 6=4 1. (11.109)
2 1 3 2
This is dubbed the Yang–Baxter move, because it is precisely the graphical representa-
tion of the Yang–Baxter equation that plays an important role in integrable theories. It is
usually represented as
= . (11.110)
1 2 3 1 2 3
Just as in N = 4 SYM, the on-shell diagrams at higher multiplicity have an interesting

structure. A description of this can be found in [120].
11.3.10 The orthogonal Grassmannian

Let us see if we can reason our way to a proper Grassmannian formula for the 3d superam-
plitudes of ABJM theory. We begin with
#
k #k
δ 2|3 Cai i ≡ δ2 Cai |i δ (3) Cai ηi , (11.111)
a=1 i a=1 i i
where i = (|i, ηi ). As in the 4d case, treated in Chapter 9, the function (11.111) is

invariant under generators in (11.39) which are linear in derivatives. If we consider the
generators that are quadratic in derivatives, for example the conformal boost generator, we
find

n #
k
n #
∂ ∂
δ2 Ca j | j = Ca i Cb i f ab δ2 Ca j |i .
∂|ia ∂|ib
i=1 a=1 j i=1 a=a ,b j
(11.112)
Here f ab is a function that includes either a single derivative or double derivatives of the
delta functions, depending on whether a = b or not. The important piece in (11.112) is
the prefactor: it tells us that to ensure invariance under conformal boosts, we need to dress
(11.111) with an extra term that enforces CC T = 0, i.e.
k(k+1)
#
k
δ 2 CC T δ 2|3 Cai i . (11.113)
a=1 i
The product CC T is a symmetric k × k matrix, so setting it to zero takes k(k + 1)/2

constraints, as indicated in the delta function.
Exercise 11.16
Show that the combination in (11.113) is also invariant under the multiplicative gener-
ators such as P ab .
Now we can readily write down an Ansatz for an N = 6 superconformal invariant integral
over a Grassmannian Gr(k, n) subject to the orthogonal constraint CC T = 0:
#k
k(k+1)
d n×k C f (M) δ 2 CC T δ 2|3 Cai i , (11.114)
a=1 i
where f (M) is a function that only depends on the minors of the Grassmannian, so that it
preserves S L(k) invariance. In order to interpret (11.114) as an integral over a Grassmannian
manifold, it has to be G L(k) invariant. All terms in (11.114) are S L(k) invariant, and the
G L(1) weight count of the delta functions plus the measure gives nk − k(k + 1) − 2k +
3k = k(n − k). This tells us that the function f (M) needs to have G L(1) weight −k(n − k).
We need more input to fix f (M) and the extra information comes from little group
analysis. Under the Z2 little group, we have |i → −|i and ηi → −ηi , so invariance of the
delta functions in (11.114) requires Cai → −Cai . For an amplitude with a -supermultiplet
¯
on the odd-sites, the superamplitude should pick up a minus sign whenever we perform a Z2
transformation on the odd-numbered legs, while it should be inert for even legs with their
-supermultiplet. Take n = 2k, and k = even: then the product of k consecutive minors,
#k
1
f (M) = , (11.115)
i=1
M i
indeed satisfies the little group criteria. (Exercise 11.17 helps you see this.) Furthermore,
since n = 2k, the function (11.115) has G L(1) weight −k 2 , precisely as needed for overall
G L(1) invariance.
Exercise 11.17
For k = 3 (and hence n = 6), verify that (11.115) indeed picks up a minus sign under
little group scaling for odd legs and is invariant for even legs. Show that for k = odd,
k+1 1
the function f (M) = i=2 Mi
does the right job.
We conclude that the 3d Grassmannian formula for ABJM theory is given by an orthogonal
Grassmannian integral [171] which for k = even is
2k 2 # k #k
d C 1 k(k+1)
L2k,k =
O
δ 2 CC T δ 2|3 Cai i . (11.116)
G L(k) i=1 Mi a=1 i
The superscript “O” indicates it is an orthogonal Grassmannian. When k = odd, the product
k+1 1
of minors is replaced by i=2 Mi
, as shown in Exercise 11.17. Some comments are in
order:
• Momentum conservation is enforced in (11.116) in a slightly differently manner than in
the 4d version (9.17) of the Grassmannian integral because we only have the |i-spinors
in 3d. Here is how it goes. The orthogonality condition forces the Grassmannian to
be a collection of null vectors in an n-dimensional space. The bosonic delta function
δ 2 (C · |i) requires the two n-dimensional vectors {|i} to be orthogonal to C. This means
that {|i} lies in the complement of C, which is nothing but C T , and thus {|i} must also

be null: i |ia |ib = 0.
• Two-site cyclicity? The integral LO

2k,k does not appear to have the correct cyclic invariance
by two sites as discussed (11.58). However, thanks to the orthogonality condition it can
be shown that
Mi Mi+1 = (−1)k Mi+k Mi+1+k . (11.117)
Therefore the formula (11.116) is indeed invariant under cyclic rotation by two sites up
to a factor of (−1)k−1 , as required.
Exercise 11.18
Show that (11.117) holds true at 4-points: using G L(2) invariance and the orthogo-
nality condition, we can choose to fix the 2 × 4 matrices C to take the form

1 0 i sin θ −i cos θ
C= . (11.118)
0 1 i cos θ i sin θ
Verify that CC T = 0 and that (11.117) holds.
• The dimension of the integral (11.116) is found by counting how many free variables are
left after localization by the delta functions. To start with, there is a total of 2k 2 integration
variables. The bosonic delta functions fix k(k + 1)/2 + 2k − 3 constraints, with the
−3 coming from the removal of the constraints that enforce momentum conservation.
Subtracting the k 2 redundancy of G L(k), the dimension of the integral is then (k−2)(k−3) 2
.
Thus for 4- and 6-point amplitudes (k = 2, 3), the delta functions completely localize
the Grassmannian integral.
Let us now take a closer look at (11.116) for n = 4. Gauge fix the G L(2) by taking

c21 1 c23 0
C= . (11.119)
c41 0 c43 1

This leaves four parameters that can be fixed by the four delta functions in δ C · |i . Denote
the solutions to C · |i = 0 by cr̄∗s , with barred labels indicating even legs and un-barred
odd legs. Then the delta functions can be rewritten as
∗

1 # 4 ∗ c21 c23∗ 1 23 12
δ 4 C · |i = δ (c r̄ s − c ) , ∗ ∗ = − .
132 r̄ ,s
r̄ s
c41 c43 13 43 14
(11.120)
Exercise 11.19
Which property of the external momenta does it take for the above solution cr̄∗s to solve
the orthogonality constraint? (Show it!)
Localizing the Grassmannian integral on to cr̄∗s , we then find

136 3

243 (6)
1 132
δ 3 CC T = δ P , δ6 C · η = δ Q 2 , = .
243 136 M1 M2 1434
(11.121)
Combining (11.120) and (11.121), we recover the superamplitude (11.52) of ABJM theory.
For n = 6, the integral (11.116) is again completely localized by the bosonic delta
functions onto two solutions, each corresponding to one of the BCFW terms in (11.77).
Recall that these two terms are individually local (see (11.81)). We can now understand
why. At 6-point these are the only possible invariants produced by the Grassmannian
integral, so this means that if the orthogonal Grassmannian integral produces all possible
dual conformal invariants13 the Leading Singularity of the 6-point superamplitude must
be some linear combination of them. However, we already know that there are two distinct
1-loop
n and An
local rational functions for n = 6, namely Atree ∝ Atree
n,shifted . As we noted at the
end of Section 11.3.8, since they are distinct, this can only mean one thing, namely that the
two terms in (11.77) are individually local and free of spurious poles.
We conclude this section with a comparison of the 3d and 4d Grassmannians. In the
4d Grassmannian, the choice of contour that gives the tree superamplitude forces the
Grassmannian Gr(k, n) to be localized to a Gr(2, n) submanifold. For n = 6 we saw how
this is intimately related to locality, since the tree contour selected three residues whose
sum was free of spurious poles. In 3d, on the other hand, the orthogonal Grassmannian
integral localizes completely for n = 6 and gives us two local objects without any need for
us to pick any contour. Does this mean that the localization to the Gr(2, n) submanifold is
not really related to locality? The answer turns out to be “no” in an interesting way. It was
found in [173] that for n = 6, the orthogonality constraint indeed forces the Grassmanian
to localize to a Gr(2, n) submanifold. Thus, the Grassmannian for N = 4 SYM achieves
locality for the 6-point NMHV superamplitude by choosing a particular “tree-contour,”
while for 6-point ABJM superamplitudes, the Grassmannian achieves locality by subjecting
itself to the orthogonal constraint. The invariant between the two cases is the Gr(2, n)
submanifold, which was previously [125, 126] linked to Witten’s twistor string formulation
[54]. So is there perhaps a 3d twistor string theory? A twistor-like string theory with target
space SU (2, 3|5) was constructed in [174] and it reproduces the Gr(2, n) formula of the
ABJM amplitudes. It is quite fascinating how the study of scattering amplitudes reveals the
existence of a new twistor string theory!
There is also a momentum twistor formulation of the Grassmannian integral for
ABJM amplitudes (see Elvang et al. [60]). It could be the first step towards an ABJM
amplituhedron.
13 Using on-shell diagrams, one can show that the Leading Singularities obtained from the result of loop-level
recursion can always be identified with residues of the orthogonal Grassmannian integral [172].
Supergravity amplitudes 12
We have seen that on-shell methods are particularly powerful for theories with gauge
redundancy. Gravity has in a sense even more redundancy than Yang–Mills theory because
of the diffeomorphism invariance. So perhaps there are hidden structures waiting to be
discovered in gravity amplitudes? In this chapter, we discuss what is currently known about
the scattering amplitudes in perturbative supergravity theories, including their UV behavior,
and we review the interesting connections between gauge theory amplitudes and gravity
amplitudes, relations that are often phrased loosely as “gravity = (gauge theory)2 .” These
relations are explored further in Chapter 13.
12.1 Perturbative gravity
In a typical course on General Relativity you learn about Einstein’s equation and its so-
lutions, for example the Schwarzchild black hole and Friedmann–Robertson–Walker cos-
mology. (And if you have a hot course, you will also learn about black rings.) These are
solutions to the classical equations of motion of gravity, just as the point-particle Coulomb
potential, electromagnetic waves, or Dirac monopoles are solutions to the Maxwell equa-
tions in electromagnetism. Here we are interested in the scattering of perturbative states at
weak coupling. From the point of view of perturbation theory, monopoles and black holes
are considered non-perturbative states that are typically suppressed by powers e−1/g in the
2
weak-coupling g 1 scattering processes.

As you know well from your QFT courses, scattering amplitudes are obtained after
quantization of the field theory: start with the Lagrangian, extract the Feynman rules, and
off we go to calculate scattering perturbatively. Of course, in the previous 200-something
pages, we have tried to convince you that recursion relations and other on-shell methods
offer much more insight and efficiency than the good old Feynman rules, but to understand
what we mean by perturbative gravity, let us start with the Lagrangian approach and
Feynman rules. This will also give us a greater appreciation for the powers of the modern
on-shell methods.
The Einstein equation, G μν = 8π Tμν , is the classical equation of motion that follows
from the variational principle applied to the Einstein–Hilbert action

1 √
SEH = dDx −g R + Smatter , (12.1)
2κ 2
250 Supergravity amplitudes
where R is the Ricci scalar and 2κ 2 = 16π G N . We have written the action in D spacetime
dimensions with a D-dimensional Newton’s constant G N . The metric gμν (x) is a field
√
in the field theory described by (12.1). The variation δgμν of −g R gives (after partial
integration and a little work [175–177]) the Einstein tensor part, G μν = Rμν − 12 gμν R, of
Einstein’s equation, while the metric variation of the “matter” action Smatter in (12.1) gives
the stress-tensor part, Tμν = √2−g δSδgmatter
μν . In the following, we use the term pure gravity to
describe the field theory (12.1) without matter fields, Smatter = 0.

Quantum field theory in curved spacetime is a highly non-trivial and interesting subject
which has important consequences such as Hawking radiation of black holes. But this is not
what we are going to discuss here. Our focus is the application of standard quantum field
theory in flat spacetime to scattering of the particles associated with the quantization of the
gravitational field gμν . More precisely, we expand the gravitational field around flat space
gμν = ημν + κh μν and regard the fluctuating field h μν as the graviton field. To start with,
let us just consider pure gravity without matter and expand the Einstein–Hilbert action
in powers of κh μν . Since the Ricci-scalar R involves two derivatives, every term in the
expansion has two derivatives. Suppressing the increasingly intricate index-structure, we
write these terms schematically as h n−1 ∂ 2 h for n = 2, 3, 4, . . . , so that the action becomes

1 √
SEH = 2 d D x −g R = d D x h∂ 2 h + κ h 2 ∂ 2 h + κ 2 h 3 ∂ 2 h + κ 3 h 4 ∂ 2 h + · · · .
2κ
(12.2)
There are infinitely many terms. There are two reasons for this: (a) in R, the series expansion
of the inverse metric generates an infinite series and (b) the expansion of the determinant
√
g = det gμν is finite, but the square root in −g generates an infinite series.1 There are no
mass terms in (12.2), so the particles associated with quantization of the gravitational field
h μν are massless: they have spin-2 and are called gravitons.
In order to extract Feynman rules from (12.2) we first have to gauge fix the action. A
typical choice is the de Donder gauge, ∂ μ h μν = 12 ∂ν h μ μ , which brings the quadratic terms
in the action to the form
1 1
h∂ 2 h → − h μν h μν + h μ μ h ν ν . (12.3)
2 4
The propagator resulting from these quadratic terms is
i 2 1
Pμ1 ν1 ,μ2 ν2 = − ημ1 μ2 ην1 ν2 + ημ1 ν2 ην1 μ2 − η μ 1 ν1 η μ 2 ν2 . (12.4)
2 D−2 k2
Each graviton leg is labeled by two Lorentz-indices. The external line rule is to dot in
graviton polarization vectors. In 4d, the polarizations encode the two helicity h = ±2
physical graviton states. They can be constructed as products of the spin-1 polarization
vectors (2.50):
μν μ μν μ
e− ( pi ) = − ( pi )−ν ( pi ) , e+ ( pi ) = + ( pi )+ν ( pi ) . (12.5)
1 By a field redefinition, we can use gμν = e−h μν instead; this brings the metric and its inverse on an equal footing
and therefore offers a simpler expansion [178].
Note that this ensures the correct little group scaling t −2h i of the on-shell graviton scattering
amplitude.
The infinite set of 2-derivative interaction terms h n−1 ∂ 2 h yields Feynman rules for n-
graviton vertices for any n = 3, 4, 5, . . . For example, the de Donder gauge 3-vertex takes
the form
μ
V3 ( p1 , p2 , p3 ) = p1 3 p2ν3 ημ1 ν2 ημ2 ν1 + (many other terms with various index-structures) .
(12.6)
You can look up the full expression for the 3-vertex in [179].
The 3-term de Donder propagator (12.4) and the infinite set of complicated interaction
terms should make it clear that calculation of even tree-level graviton scattering amplitudes
from Feynman diagrams is not a business for babies. The 4-point graviton tree amplitude was
calculated brute force with Feynman diagrams in [22] where each of the four contributing
Feynman diagrams fill about a page or so of elaborate index-delight. Nonetheless, the final
result can be brought to a very simple form: in 4d, it can be written in spinor helicity
formalism as
127 [12] 124 [34]4

M4tree (1− 2− 3+ 4+ ) = = . (12.7)
13142324342 stu
We have already encountered this expression in Exercise 2.33. Note that we will be using Mn
to denote (super)gravity amplitudes to distinguish them from (super) Yang–Mills amplitudes
An .
Of course, you already know where we are headed: on-shell methods and recursion
relations make the calculation of tree-level graviton scattering amplitudes much more fun
and efficient – and they have the power to reveal structures in the amplitudes that were
not visible at the level of the Lagrangian. The short version of the story is that little group
scaling fixes the possible 3-graviton amplitudes and recursion then allows you to compute
all other tree-level graviton processes. Loop-level amplitudes can be constructed with
unitarity techniques (Chapter 6). Thus the infinite set of interaction terms in the Lagrangian
is not needed from the point of view of the on-shell scattering amplitudes: the terms’ role
in life is to ensure off-shell diffeomorphism invariance of the gravitational action. It is an
interesting aspect of on-shell recursion relations that they eliminate the need for infinitely
many interaction terms.
Let us specialize to D = 4 and be more explicit about the graviton scattering amplitudes.
Dimensional analysis and little group scaling fix the 3-point graviton amplitudes to be
126
M3 (1− 2− 3+ ) = = A3 [1− 2− 3+ ]2 ,
232 312
(12.8)
+ + − [12]6
M3 (1 2 3 ) = = A3 [1+ 2+ 3− ]2 .
[23]2 [31]2
At tree-level, the graviton amplitudes with all-plus or all-minus helicity arrangements

vanish in pure gravity as do those with just one ±-helicity:
Mntree (1+ 2+ . . . n + ) = Mntree (1− 2+ . . . n + ) = Mntree (1+ 2− . . . n − )

= Mntree (1− 2− . . . n − ) = 0 . (12.9)
This is most easily proven using the supersymmetry Ward identities, just as we did in
(4.26), (4.27) for gluon amplitudes. The tree gravity amplitudes have to obey these same
Ward identities as in a supergravity theory because the supersymmetric partners couple
quadratically; hence it is only at loop-level the pure graviton amplitudes can distinguish
themselves from the supergravity amplitudes. In particular, (12.9) has to hold at tree-level.
Exercise 12.1
For simplicity, we dropped the explicit powers of the gravitational coupling κ in (12.7)
and (12.8), and we continue to do so henceforth. What is the mass-dimension of κ
in 4d? Show that the 4-graviton amplitude (12.7) has the correct mass-dimension
(cf. (2.102)).
We categorize graviton amplitudes the same way as gluon amplitudes with designation
N K MHV. An important difference is that the graviton scattering amplitudes are not color-
ordered. Using BCFW recursion relations,2 relatively compact graviton amplitudes can be
found for the MHV sector. One of the earliest formulas is BGK (Berends, Giele, and Kuijf)
[180] written here in the form presented in [84] valid for n > 4:
n−1
128 l=3 n|2
+ 3 + · · · + (l − 1)|l]
Mntree (1− 2− 3+ +
...n ) =
n−2
n−1 .
i=1 i, i + 1 1, n − 11n 2n l=3 ln
2 2
P(3,4,...,n−1)
(12.10)
The sum is over all permutations of the labels (3, 4, . . . , n − 1).

Another form of the same MHV graviton tree amplitude expresses Mntree in terms of
color-ordered gluon amplitudes Atree
n of Yang–Mills theory:
n−1
# − − + + 2
Mntree (1− 2− 3+ +
...n ) = s1in βk Atree
n 1 2 i 3 i 4 . . . i n+ , (12.11)
P(i 3 ,i 4 ,...,i n ) k=4
where n ≥ 4 and
i k i k+1
βk = − 2|i 3 + i 4 + · · · + i k−1 |i k ] . (12.12)
2i k+1
The result (12.11) can be derived [181, 182] using a [−, − BCFW shift.
2 BCFW for tree graviton amplitudes is valid when based on shifts [−, −, [−, +, and [+, +.
There are also other graviton MHV formulas available in the literature, for example
the “soft-factor” formula [183]. You may find that these MHV expressions are terribly
complicated compared with Parke–Taylor; however, they are remarkably simple when
compared with the impenetrable mess a Feynman diagram calculation would produce.
Beyond the MHV level, one can readily use BCFW to calculate explicit results. You
might be curious if there is also a CSW-like expansion for gravity amplitudes. The MHV
vertex expansion [35] based on the Risager-shift (see discussion below Exercise 3.10) works
for NMHV graviton amplitudes with n < 12 particles. It fails [36] for n ≥ 12 because the
large-z falloff of the n-point amplitude under Risager-shift is 1/z 12−n and the Cauchy
contour deformation argument needed to derive the recursion relations therefore picks up
a term at infinity for n > 11. The all-line shift discussed in Section 3.4 also fails (for
interesting reasons [26]). For further discussion of CSW for gravity, see [17, 20, 26, 36].
The relation between gravity and gauge theory amplitudes is clearly visible in the 4d
MHV expressions (12.8) and (12.11), but there are more general relations available. The
first such example is the KLT relations, derived in string theory by Kawai, Lewellen and
Tye [21]: the KLT relations give the n-point tree-level closed string scattering amplitude as
a sum over products of pairs of n-point open string partial amplitudes, with coefficients that
depend on the kinematic variables as well as the string tension 1/(2π α ). This is natural,
albeit non-trivial, since the closed string vertex operators are products of open string vertex
operators. The non-triviality of the KLT relations is that the factorization into open string
amplitudes survives the integrals over the insertion points of the vertex operators. In the
limit of infinite tension, α → 0, the closed string amplitudes with massless spin-2 string
external states become the regular graviton scattering amplitudes Mn we have discussed
above. And in this limit, the open-string partial amplitudes with external massless spin-1
states become the color-ordered gluon amplitudes An . Thus, in the limit α → 0, KLT offers
a relationship between tree-level Mn and An for each n. For n = 4, 5, and 6 the field theory
KLT relations are
M4tree (1234) = −s12 Atree

4 [1234] A4 [1243] ,
tree
M5tree (12345) = s23 s45 Atree

5 [12345] A5 [13254] + (3 ↔ 4) ,
tree

M6tree (123456) = −s12 s45 Atree
6 [123456] s 35 A tree
6 [153462] + (s 34 + s 35 )A tree
6 [154362]
+ P(2, 3, 4) . (12.13)
In the 6-point case, P(2, 3, 4) stands for the sum of all permutations of legs 2, 3, 4. At
7-point and higher, the KLT relations are more complicated; they can be found in Appendix
A of [184]. The relation between gravity and gluon scattering is not at all visible in the
Lagrangian (12.2), although field redefinitions and clever gauge choices can bring the first
few terms in the gravitational action into a more KLT-like form; see [185–187] and the
review [188].
Note that there is no specification of helicities of the external states in (12.13): this is
because the KLT relations are valid in D-dimensions. In 4d, the KLT relations work for any
helicity assignments of the gravitons on the LHS; if the ith graviton has helicity h i = +2,
then the gluons labeled i in the amplitudes on the RHS have helicity h i = +1; similarly for
negative helicity. This ensures that the little group scaling works out on both sides of the
KLT relations. We may then say that KLT in 4d uses
graviton±2 ( pi ) = gluon±1 ( pi ) ⊗ gluon±1 ( pi ) . (12.14)
This is also encoded in the graviton polarizations (12.5).

Inspecting the relationship (12.14) between gravitons and gluons, we could also ask
what happens if we combine gluons of opposite helicity in the KLT relations. The result is
something that has the little group scaling of a scalar on the gravity side. In fact, what you
get is the dilaton and axion:
"
dilaton
= gluon±1 ( pi ) ⊗ gluon∓1 ( pi ) . (12.15)
axion
This is completely natural from the string theory point of view where the graviton state
comes together with an antisymmetric tensor Bμν and a scalar “trace” mode. The latter is
the dilaton and the former has a 3-form field strength H = d B which means that in 4d it
is dual to a scalar with shift-symmetry, i.e. an axion. Therefore we can write the relation
between the spectra
4d axion-dilaton gravity = (YM theory) ⊗ (YM theory) . (12.16)
Given the attention we have poured into the study of amplitudes in (planar) N = 4 SYM
in 4d, you may also be curious about what we would get if we tensored the 24 states
of N = 4 SYM à la (12.16). The answer is a very good one: we get the 28 states of
N = 8 supergravity, which is the 4d supergravity theory with maximal supersymmetry.
Supergravity amplitudes, especially those in N = 8 supergravity, are the main focus in the
following. We return to the study of “gravity =(gauge theory)2 ” in Chapter 13, though you
will see more of it in the following sections too.
12.2 Supergravity
Supergravity is the beautiful union of gravity and supersymmetry. It is the result of making
the supersymmetry transformations local in the sense that the SUSY parameter is space-
time dependent. If you have not previously studied supergravity, you should immediately
read the essay [189] and then textbooks such as [37, 39, 190].
The supersymmetry partner of the graviton is called the gravitino. It has spin- 23 and
(when supersymmetry is unbroken) it is massless. In 4d we characterize a gravitino by
its two helicity states h = ± 32 ; its Feynman rule for the external line simply combines a
±-helicity spin-1 polarization vector with the ±-helicity spin- 12 fermion wavefunction.
12.2 Supergravity 255
In a 4d supergravity theory with N supercharges, Q A and Q 2 A , the graviton has N

gravitino-partners. We can construct the spectrum by starting with the negative helicity
graviton h − as the highest-weight state and apply the supercharges Q 2 A . At each step, Q
2A
1
raises the helicity by 2 , so when Q 2 A is applied once to h it produces a negative helicity
−
gravitino ψ A− . If N = 1, then the process terminates because of the Grassmann nature of the
supercharge. So the N = 1 pure supergravity multiplet consists of the two CPT conjugate
pairs of graviton and gravitino:
N = 1 supergravity: (h − , ψ − ) and (ψ + , h + ) . (12.17)
When we say pure supergravity we mean that there are no other matter-supermultiplets
included; we only have the states that are related to the graviton via supersymmetry.
Pure N = 2 supergravity has 2 × 22 states
N = 2 supergravity: (h − , ψ A− , v − ) and (v + , ψ A+ , h + ) , (12.18)
where the two gravitinos ψ A− and ψ A+ are labeled by A = 1, 2 and v ± denotes the two
helicity states of the spin-1 gravi-photon.
Fast-forward to pure N = 4 supergravity. Its 2 × 24 states can be characterized as

N = 4 supergravity = N = 4 SYM ⊗ N = 0 SYM . (12.19)
By N = 0 SYM we just mean pure Yang–Mills theory. The spectrum (12.19) should be
read as follows: the two graviton states are given in terms of the gluon states as in (12.14).
Using the spectrum (4.30) of N = 4 SYM, we find:
±
gravitons: h = g± ⊗ g± ,
gravitinos: ψ A+ = λ A+ ⊗ g + and ψ A− = λ̄−A ⊗ g − ,
gravi-photons: v± ±
AB = S AB ⊗ g , (12.20)
gravi-photinos: ψ A− = λ A+ ⊗ g − and ψ A+ = λ̄−A ⊗ g+ ,
± ∓
scalars (dilaton–axion): g ⊗ g ,
where g ± are gluons, λ̄−A and λ A+ are gluinos, and S AB are the six scalars of N = 4 SYM.
Totaling the states, we get 2 × 1 + 2 × 4 + 2 × 6 + 2 × 4 + 2 = 32 = 2 × 24 .
Exercise 12.2

Identify
the supermultiplets in the theory whose spectrum is N = 2 SYM ⊗ N =
0 YM .
What is the difference between the two N = 4 supergravity theories whose spectra
are (N = 4 SYM) ⊗ (N = 0 SYM) and (N = 2 SYM) ⊗ (N = 2 SYM)?
Applying the N supersymmetry generators Q 2 A to the graviton top state h − , we cannot

avoid states with spin greater than 2 if N > 8. There are no consistent interactions in flat
space for particles with spin greater than 2, so that tells us that maximal supersymmetry
in 4d is N = 8. The N = 8 supergravity theory is unique: the ungauged theory, which

is our focus here, was first written down in [191, 192].3 Its spectrum of 28 states forms
a CPT-self-conjugate supermultiplet (just as in N = 4 SYM). As noted at the end of
Section 12.1, the spectrum can be characterized as

N = 8 supergravity = N = 4 SYM ⊗ N = 4 SYM . (12.21)
In any supergravity theory, there are supersymmetry Ward identities that restrict the ampli-
tudes, just as discussed for gauge theories in Sections 4.2 and 4.4. In particular, the graviton
amplitudes in supergravity satisfy
Mn (1+ 2+ . . . n + ) = Mn (1− 2+ . . . n + ) = Mn (1+ 2− . . . n − ) = Mn (1− 2− . . . n − ) = 0

(12.22)
at all orders in perturbation theory. There are also simple Ward identities among graviton
and gravitino MHV amplitudes that give
13
N ≥ 1 supergravity: Mn (1− ψ − ψ + 4+ . . . n + ) = Mn (1− 2− 3+ 4+ . . . n + ) ,
12
(12.23)
just as for gluons and gluinos. In extended (N > 1) supergravity there are further relations,
as you will see shortly from the superamplitudes in N = 8 supergravity. The example
i j8
N = 8 supergravity: Mn (1+ . . . i − . . . j − . . . n + ) = Mn (1− 2− 3+ . . . n + )
128
(12.24)
is the N = 8 supergravity analogue of the supersymmetry Ward identity (4.35) in N = 4

SYM.
12.3 Superamplitudes in N = 8 supergravity
The spectrum (12.21) of N = 8 supergravity consists of 128 bosons and 128 fermions.
Organized by helicity h = 2, 32 , 1, 12 , 0, − 12 , −1, − 32 , −2, we can write it out as
1 graviton h + , 8 gravitinos ψ A , 28 gravi-photons v AB ,

56 gravi-photinos χ ABC , 70 scalars S ABC D , 56 gravi-photinos χ ABCC D E ,
28 gravi-photons v ABC D E F , 8 gravitinos ψ ABC D E F G , 1 graviton h − = h 12345678 .
(12.25)
3 The gauged N = 8 supergravity theory was presented in [193].

12.3 Superamplitudes in N = 8 supergravity 257
Here A, B, . . . = 1, 2, . . . , 8 are SU (8) R-symmetry indices and each state above is fully
antisymmetric in these labels; this simply reflects that the helicity-h state transforms in the
rank r = 4 − 2h fully antisymmetric irreducible representation (irrep) of SU (8) and the
multiplicity given in (12.25) is the dimension of the irrep. The 70 scalars are self-dual and
2 A act on the
satisfy S ABC D = 4!1 ABC D E F G H S E F G H . Supersymmetry generators Q A and Q
states in an obvious generalization of (4.31).
Just as in N = 4 SYM, it is highly convenient to combine the states into a superfield, or
super-wavefunction, with the help of an on-shell superspace with Grassmann variables ηi A
whose index i = 1, . . . , n is a particle label and A = 1, 2, . . . , 8 is a fundamental SU (8)
R-symmetry index. The N = 8 superfield is then
1
i = h + + ηi A ψ A − ηi A ηi B v AB + · · · + ηi1 ηi2 ηi3 ηi4 ηi5 ηi6 ηi7 ηi8 h − . (12.26)
2
The SU (8) R-symmetry requires that the superamplitudes are degree 8k polynomials in
the Grassmann variables. This directly gives us the N = 8 supergravity version of the
N K MHV classification: the K th sector contains the superamplitudes of degree 8(K + 2)
polynomials in the ηi A s. It should be clear from (12.26) that the MHV sector (K = 0)
includes the graviton component amplitude Mn (1− 2− 3+ . . . n + ).
The super-Poincaré generators – momentum P ȧb , rotations/boosts, and the supercharges
A
Q and Q 2 A are given in (5.1) and (4.46), with the only difference that now A = 1, 2, . . . , 8.
Momentum- and supermomentum-conservation require that the general superamplitudes
in N = 8 supergravity are of the form

MN
n
K
MHV 2 P8K ,
= δ 4 P δ (16) Q (12.27)
where P8K is annihilated by Q A which acts by differentiation: Q A P8K = 0.

At MHV level, we are already home safe. The Grassmann delta function (12.27) eats up
all 16 fermionic variables, so P0 is η-independent. It can be fixed by requiring that Mn
projects out the correct pure graviton MHV amplitude Mn (1− 2− 3+ . . . n + ). This is easily
accomplished:

Mn (1− 2− 3+ . . . n + )
MMHV 2
= δ 4 P δ (16) Q . (12.28)
n
128
In particular, the supersymmetry Ward identity (12.24) for graviton amplitudes in N = 8

supergravity follows from (12.28).
Beyond the MHV level, one can solve the supersymmetric Ward identities Q A Mn =
2 A Mn = 0 (just as in the N = 4 SYM case) to find a basis of input-amplitudes that
Q
completely determine the full superamplitude. The basis can be labeled by the K × 8
rectangular Young tableaux of SU (n − 4) irreps [42]. Another approach is to use the
super-BCFW recursion relations; they are valid for super-shift of any two lines [19, 47].4
4 A super-shift version of CSW was discussed in [44] and while it works for all tree superamplitudes in N = 4
SYM, it has limited validity in N = 8 supergravity.
The set-up for the MHV superamplitude (12.28) is perhaps a bit “cheap” because the
component amplitude Mn (1− 2− 3+ . . . n + ), as we have seen in Section 12.1, does not take a
particularly compact form and it does not clearly reflect symmetries such as full permutation
symmetry of identical external states. So there has been quite a lot of effort towards building
an MHV superamplitude that more clearly encodes the symmetries. One representation
[194] of the superamplitude builds on a super-BCFW shift in the N = 7 formulation of N =
8 supergravity.5 Other MHV formulas use the Grassmannian representations [195, 196] or
the twistor string [197, 198]. More recently, new compact formulas for both Yang–Mills and
gravity amplitudes have been proposed to be valid in any spacetime dimensions [199, 200].
N = 8 supergravity has, as we have noted above, a global SU (8) R-symmetry. This
symmetry is realized linearly, as you can see on the spectrum and on the amplitudes which
vanish unless the external states form an SU (8) singlet. However, the theory also has a
“hidden” symmetry: the equations of motion of N = 8 supergravity have a continuous
global E 7(7) (R) symmetry. The group E 7(7) is a non-compact version of the exceptional
group E 7 ; its maximal compact subgroup is SU (8). E 7(7) has rank 7 and is 133 dimensional.
It is not a symmetry of the action of N = 8 supergravity. The best way to think of this is that
the E 7(7) is spontaneously broken to SU (8). There are 133 − 63 = 70 broken generators,
giving 70 Goldstone bosons. Those are exactly the 70 scalars S ABC D in the spectrum (12.25)
and they “live” in the coset space E 7(7) /SU (8).
As a spontaneously broken symmetry, E 7(7) is not linearly realized on the on-shell
scattering amplitudes, but instead it manifests itself via low-energy theorems.6 If the
momentum of an external scalar S ABC D is taken soft, then the amplitude must vanish
because the Goldstone scalars are derivatively-coupled. Basically this says that the moduli
space E 7(7) /SU (8) is homogeneous: it does not matter what the vevs of the scalars are,
all points on moduli space are equivalent. The soft scalar limit probes the neighborhood
of a point in moduli space and since the moduli space is homogeneous, the soft scalar
limit vanishes. There are also double-soft limits that involve the commutator of two coset
generators and these therefore directly reveal, from the on-shell point of view, the coset
structure E 7(7) /SU (8).
Exercise 12.3
Project the amplitude M4 (S 1234 S 5678 h − h + ) out from the MHV superamplitude (12.28).
Show that

lim M4 S 1234 S 5678 h − h + = 0 . (12.29)
p1 →0
In contrast, the scalars in N = 4 SYM are not Goldstone bosons, so the soft-scalar
limits do not have to vanish. For example, show

lim A4 S 12 g − S 34 g + = 0 . (12.30)
p1 →0
5 Just as N = 3 SYM is identical to N = 4 SYM, so is N = 7 supergravity identical to N = 8 supergravity. The

validity of the super-BCFW shifts in N = 7 supergravity is proven in [87].
6 Low-energy theorems were originally developed in pion-physics [201]. For a review, see [202].
12.4 Loop amplitudes in supergravity 259
The soft limit explores the points of moduli space in the neighborhood of the origin:
away from the origin, the N = 4 SYM theory is on the Coulomb branch, part of the
gauge group is broken, and some of the N = 4 supermultiplets become massive. The
non-vanishing limit (12.30) has a nice interpretation. Set p1 = q for some light-like
q = −|q[q| and take the soft limit as → 0. The limit (12.30) then depends on
|q. The soft limit p1 → 0 leaves an object with momentum conservation on three
particles: the result can be interpreted as the small-mass limit of the Coulomb branch
amplitude A3 [W − S 34 W + ], where W ± are the longitudinal modes of the massive spin-
1 W -bosons of a massive N = 4 supermultiplet and S 34 is a massless scalar. From this
point of view, q is a reference vector that allows us to project the massive momenta of
W ± such that the corresponding angle spinors are well-defined. This is actually also
needed to define the helicity basis because helicity is not a Lorentz-invariant concept
for massive particles; but q breaks Lorentz-invariance and allows us to define a suitable
q-helicity basis [26, 140].
The 3-point amplitude A3 [W − S 34 W + ] violates the SU (4) R-symmetry of N = 4
SYM at the origin of moduli space. This is fine, because the Coulomb branch breaks
the R-symmetry. Minimally, one has SU (4) ∼ S O(6) → S O(5) ∼ Sp(4).
The moral of the story is that single-scalar soft limits for N = 8 supergravity
amplitudes vanish because the 70 scalars are Goldstone bosons of E 7(7) → SU (8).
And single-scalar soft limits for N = 4 SYM are non-vanishing and reproduce the
small-mass limit of the Coulomb branch amplitudes [140]. One can in fact re-sum the
entire small-mass expansion from multiple-soft-scalar limits and recover the general-
mass Coulomb branch amplitudes [140, 203].
The single-soft-scalar limits of tree amplitudes in N = 8 supergravity were first studied

in [36]. Single and double soft limits were discussed extensively and clarified in [47]. The
soft-scalar limits play a key role for us in Section 12.5.
12.4 Loop amplitudes in supergravity
It is taught in all good kindergartens that a point-particle theory of gravity is badly UV

divergent and non-renormalizable. This means that it is not a good quantum theory. So what
is perturbative gravity all about?
Naive power-counting gives a clear indication that gravity with its 2-derivative inter-
actions generically has worse UV behavior than for example Yang–Mills theory with its
1- and 0-derivative interactions. Consider for example a generic 1-loop m-gon diagram.
In gravity, the numerator of the loop-integrand can have up to 2m powers of momenta,
while in Yang–Mills theory it is at most m. Both have m propagators, so in gravity this
gives

( 2 )m
gravity 1-loop m-gon diagram ∼ d 4 ∼ 4 . (12.31)
( 2 )m
This is power-divergent as the UV cutoff is taken to ∞ for all m. On the other hand, for
Yang–Mills theory the m-gon integral has at most m in the numerator, so it is manifestly
UV finite for m > 4.
Now, the power-counting is too naive. There can be cancellations within each diagram.
Moreover, we have learned that we should not take individual Feynman diagrams seriously
if they are not gauge invariant. So cancellations of UV divergences can take place in the sum
of diagrams, rendering the on-shell amplitude better behaved than naive power-counting
indicates.
In fact, pure gravity in 4d is finite at 1-loop order [204]: all the 1-loop UV divergences
cancel! This is difficult to see by direct Feynman diagram calculations, but it follows rather
trivially from the absence of any valid counterterms. We review this approach in detail in
Section 12.5.
At 2-loop order, it has been demonstrated by Feynman diagram calculations that pure
gravity indeed has a UV divergence [205, 206]. In Yang–Mills theory we are not too
scared of UV divergences because we know how to treat them with the procedure of
renormalization. However, in gravity, it would take an infinite set of local counterterms
to absorb the divergences and hence the result is unpredictable: pure gravity is a non-
renormalizable theory.
So what is the theory described by the Einstein–Hilbert action? Because it is non-
renormalizable, it is not by itself a well-defined quantum theory of gravity. Instead, we
should regard the field theory defined by the Einstein–Hilbert action as an effective field
theory, valid at scales much smaller than the Planck scale MPlanck ∼ 1019 GeV. To see
this, recall that the 4d gravitational coupling κ has mass-dimension −1. So when we do
perturbation theory, we should really use the dimensionless coupling Eκ where E is the
characteristic energy of the process. At high enough energies, this dimensionless coupling
is no longer small and we cannot trust perturbation theory. So we should not extrapolate
−1/2
to such high energies. In energy units, we have κ −1 ∼ G N = MPlanck , so this tells us to
use gravity, as described by the Einstein–Hilbert action, for energies E MPlanck . As a
classical effective field theory, General Relativity is enormously successful and captures
classical gravitational phenomena stunningly as shown by experimental tests.
Regarding gravity as an effective theory, we can study the low-energy perturbative
amplitudes: the tree amplitudes capture the classical physics and there are no divergences
to worry about. At 1-loop level, we have mentioned that pure gravity is finite. Could we
imagine adding matter to gravity in such a way that its higher-loop amplitudes were also
finite? Gravity with generic matter is 1-loop divergent [204, 207], but we know from gauge
theories that supersymmetry improves the UV behavior of loop amplitudes, even to such an
extreme extent that the maximally supersymmetric Yang–Mills theory, N = 4 SYM, is UV
finite: the UV divergences cancel completely at each order in the loop expansion. Could
something like that also happen in supergravity? If it did, it would eliminate the need for
renormalization and the problems of non-renormalizability would be obsolete. There would
still be important questions unresolved about non-perturbative aspects of supergravity;
perturbative finiteness would not mean that the theory is UV complete. The question
of perturbative UV finiteness of (maximal) supergravity in 4d has received increased
attention in the past few years and the on-shell amplitude techniques have facilitated
multiple explicit calculations of supergravity loop amplitudes. It should be emphasized that

whether or not the perturbative calculations eventually encounter a divergence, one should
appreciate that the study of loop amplitudes in supergravity has resulted in a number of
new insights, of independent value, about gravity scattering amplitudes. An example is the
connection between gravity and Yang–Mills amplitudes via the so-called BCJ dualities (see
Chapter 13).
Pure supergravity in 4d is better behaved in the UV than pure gravity: all pure su-
pergravity theories in 4d are finite at 1-loop [208] and 2-loop order [209–211], i.e. the
first possible UV divergence can appear only at 3-loop order, improving on the 2-loop
UV divergence of pure gravity [205, 206]. In the spirit of “the more supersymmetry, the
better,” it is natural to focus on maximal supersymmetry, i.e. N = 8 supergravity in 4d. An
explicit calculation, using the generalized unitarity method, demonstrated that the 3-loop
4-graviton amplitude is UV finite in N = 8 supergravity in 4d [212–214]. This and related
observations of unexpected cancellations motivated Bern, Dixon, and Roiban [212] to ask
if N = 8 supergravity in 4d is UV finite? They further proposed that the critical dimension
Dc for the first UV divergence of maximal supergravity in D-dimensions follows the same
pattern as for maximal super Yang–Mills theory [184, 215], namely
6
Dc (L) = +4 for L > 1. (12.32)
L
It has been shown [216, 217] that the 4-loop 4-graviton amplitude is UV finite in N = 8
supergravity in 4d and that it follows the pattern (12.32). How about 5-loops? A pure-spinor
based argument [218] leads to (12.32) for L = 2, 3, 4, but implies that the critical dimension
for L = 5 is D = 24/5 and not 26/5 as (12.32) predicts. This question can be settled by
direct computation: at the time of writing, the 5-loop calculation is still in progress, so you
will have to watch the news feeds for a resolution.
For D = 4, the symmetries of N = 8 supergravity can be used to establish that all
amplitudes of the theory are UV finite for L ≤ 6: this explains the finiteness of the 3- and
4-loop 4-graviton amplitudes and predicts that no UV divergence appears in any higher
point amplitudes for L ≤ 6 in 4d. For L ≥ 7, the known symmetries do not suffice to rule
out UV divergences in 4d. The following section reviews how these results are obtained
using an on-shell amplitude-based approach [43, 219–221] to counterterms in N = 8
supergravity. We then provide in Section 12.6 an overview of the current status of the UV
behavior of supergravity as a function of dimensions D, supersymmetries N , and loop
order L.
12.5 N = 8 supergravity: loops and counterterms
Suppose that a supergravity has its first UV divergence in an n-point amplitude at L-loop
order. Then the effective action for the theory must have a local diffeomorphism in-
variant counterterm constructed from n fields (corresponding to the n external states)
and (2L + 2) derivatives. The latter statement follows from dimensional analysis be-
cause the gravitational coupling κ has mass-dimension −1: for given n, the ratio of the
L-loop supergravity amplitude to the tree amplitude has an overall factor of κ 2L , so the
corresponding local counterterm has to make up the mass-dimension by having 2L more
derivatives than the 2-derivative tree-level theory.7
Exercise 12.4
Show that the n-graviton 1-loop amplitude has an overall factor of κ 2 compared with
the n-graviton tree amplitude.
If we consider just pure gravity, the possible local diff-invariant counterterms must be
Lorentz scalars formed from contractions of Riemann tensors and possibly covariant
derivatives. Each Riemann tensor contributes 2-derivatives, so at 1-loop (4-derivatives),
√ √ √
the possible candidates are −g R 2 , −g Rμν R μν , and −g Rμνρσ R μνρσ . If we suppress
the index contractions, we can write schematically

1 √

Seff = 2 d 4 x −g R + κ R + κ
2 2
R + κ
4 3
R + κ 6 D2 R4 + R5
6 4
2κ
L=0 L=1 L=2 L=3 L=4

+ κ D R +D R +R
8 4 4 2 5 6
+ ··· , (12.33)

L=5
where R denote Riemann tensors and D are covariant derivatives. This should be viewed
as a list of possible candidate counterterms; the operators in (12.33) are not necessarily
generated in perturbation theory.
Exercise 12.5
Trivia question: why are we not including operators D 2k R in (12.33)?
How about operators D 2k R 2 ?
[Hint: use the Bianchi identity for the Riemann tensor.]
Show that operators of the form D 2k R 3 have vanishing 3-point matrix elements for
k ≥ 1.
Since we consider on-shell amplitudes, we can enforce the equations of motion on the
candidate counterterms. In pure gravity, the Einstein equation gives Rμν = 0, so this leaves
√
−g Rμνρσ R μνρσ as the only possibility at 1-loop. However, we are free to add zero to
√ √
convert −g Rμνρσ R μνρσ to the Gauss–Bonnet term −g Rμνρσ R μνρσ − 4Rμν R μν + R 2

which equals a total derivative. Therefore there is no local counterterm for pure gravity
at 1-loop order! And since there is no counterterm, pure gravity is not UV divergent
at 1-loop. A cleaner way to say this is that one can do a field redefinition that changes
7 To be a little more precise, the above statement is true for amplitudes with purely bosonic fields. Since external
fermions dress the amplitude with dimensionful wavefunctions, each pair of fermions counts one derivative for
the purpose of dimensional analysis.
√
−g Rμνρσ R μνρσ to the Gauss–Bonnet term, and since a field redefinition does not change
the amplitude there cannot be a 1-loop divergence.
At 2-loop order, the candidate counterterm has to be composed of some index contractions
√
of three Riemann tensors – let us denote it by R 3 , and here and henceforth leave the −g-
factor implicit in candidate counterterm operators. The R 3 counterterm is present for pure
gravity, which (as noted in Section 12.4) is 2-loop divergent.
Counterterms have to respect any non-anomalous symmetries of the theory. Supersym-
metry is preserved at loop-level, so in supergravity theories counterterm candidates must
be supersymmetrizable. We showed in Exercise 2.35 that a matrix element produced by R 3
is fixed by little group scaling to be
M3 (1− 2− 3− ) R 3 = constant × 122 232 132 . (12.34)
But we also know from (12.22) that all-minus amplitudes violate the supersymmetry Ward
identities. So this means that any operator that produces a non-vanishing matrix element
M3 (1− 2− 3− ) R 3 violates supersymmetry. Since R 3 produces a supersymmetry-violating
amplitude, we conclude that R 3 cannot be supersymmetrized [209, 222]. Therefore R 3 is
not a viable counterterm and hence any pure supergravity must be 2-loop finite!
In the above argument, you may object that we may not have to care about the 3-point
amplitude (12.34) since it vanishes in real kinematics. But it is easy to show (using for
example an all-line shift [26]) that a non-vanishing all-minus 3-point amplitude implies
that there is a non-vanishing 4-graviton amplitude M4 (1− 2− 3− 4− ) and clearly this violates
supersymmetry. Another objection could be: “what if the constant in (12.34) is zero, then
there is no contradiction with supersymmetry?” That is true, but if the matrix element
vanishes that means that the 3-field part of R 3 is a total derivative, and then we don’t care
about it anyway because there are no available gravity diff-invariant 4-field operators at
6-derivative order. So, either way, we conclude that pure supergravity is finite at 2-loops.
Let us now specialize to N = 8 supergravity in 4d. The candidate counterterms have to
respect N = 8 supersymmetry and also be SU (8) invariant, since the global R-symmetry
is non-anomalous [89, 90]. Moreover, they should be compatible with the “hidden” E 7(7)
symmetry [223]; we will come back to E 7(7) later – for now, we explore the constraints of
supersymmetry and R-symmetry on the candidate counterterms in N = 8 supergravity.
It is in general difficult to analyze the candidate counterterm operators directly: a full
field theory N = 8 supersymmetrization of the independent contractions of Rs and Ds is
complicated in component form; for R 4 it has been done explicitly at the linearized level only
[224]. A better approach is to use a superfield formalism. There is no off-shell superfield
formalism for N = 8 supergravity, but harmonic superspace techniques have been used
to constrain the possible counterterms in supergravity theories in various dimensions. We
are going to highlight some of the results of the superspace approach in Section 12.6, but
otherwise we do not discuss these methods here: this is a book about amplitudes and that
is the path we take.
The supersymmetry and R-symmetry constraints on the candidate counterterm operators
translate into Ward-identity constraints on the matrix elements produced by counterterms.
Let us list the translation of constraints between an operator, whose lowest interaction-term
is an n-vertex, and the corresponding n-point matrix element:
n-field operator n-point matrix element

local with 2L + 2 derivatives ↔ polynomial in i j and [kl] of degree 2L + 2
N = 8 SUSY ↔ N = 8 SUSY Ward identities
SU (8) R-symmetry ↔ SU (8) Ward identities
(12.35)
(E 7(7) constraints will be treated separately, starting on page 267.) In addition, the matrix
elements have to respect Bose/Fermi symmetry under exchange of identical external states.
The condition that the matrix element is polynomial follows from the locality of the
operator and the insistence that it corresponds to the leading (i.e. first) UV divergence in
the theory; with other operators present, there could be pole terms. The matrix element
we consider here is strictly the amplitude calculated from the n-point vertex of the given
n-field operator, and therefore it cannot have any poles, i.e. it must be a polynomial in the
kinematic variables i j and [kl]. The degree of the polynomial follows from dimensional
analysis.
Let us be clear about what our approach is: for a given L, we ask if the first UV divergence
could appear in an n-point amplitude. If this is so, then there must be a corresponding n-field
(2L + 2)-derivative counterterm. For example, for L = 3 the lowest-n candidate counter-
term would be an SU (8)-invariant N = 8 supersymmetrization of R 4 . To analyze if such
an operator exists, we write down all possible matrix elements satisfying the constraints
(12.35). If there are no such matrix elements, we conclude there is no corresponding N = 8
SUSY and SU (8)-invariant operator and therefore the first divergence in the theory cannot
be in the n-point L-loop amplitude. On the other hand, if one or more such matrix elements
exist, then the corresponding operator respects linearized N = 8 supersymmetry and SU (8)
and we may consider it as a candidate counterterm. That does not mean that perturbation
theory actually produces the corresponding UV divergence; that would have to be settled by
other means, such as an explicit L-loop computation. Thus, the approach here is to use the
matrix elements to exclude counterterm operators as well as characterize candidate coun-
terterms as operators that respect N = 8 SUSY and SU (8)-symmetry at the linearized level.
To illustrate the idea, consider R 4 .8 Its 4-point matrix element M4 (1− 2− 3+ 4+ ) R 4 has
to be a degree 8 polynomial in angle and square brackets. Taking into account the little
group scaling, there is only one option: the matrix element has to be M4 (1− 2− 3+ 4+ ) R 4 =
c R 4 124 [34]4 , where c R 4 is some undetermined constant. By the same arguments as
above, dimensional analysis and little group scaling, we know that M4 (1− 2+ 3− 4+ ) R 4 =
c R 4 134 [24]4 . We can then check the MHV-level N = 8 SUSY Ward identity (12.24)
which at n-point reads
i j8
N = 8 supergravity: Mn (1+ . . . i − . . . j − . . . n + )O = Mn (1− 2− 3+ . . . n + )O .
128
(12.36)
8 The relevant contraction of four Riemann tensors is the square of the Bel–Robinson tensor [225].
For our 4-point R 4 matrix elements, we have

138 138
M4 (1− 2+ 3− 4+ ) R 4 = M 4 (1− − + +
2 3 4 ) R 4 = c R 4 124 [34]4 = c R 4 134 [24]4 ,
128 128
(12.37)
thanks to momentum conservation 13[34] = −12[24]. It is not hard to see that the
4-point super-matrix-element

[34]4
2
M4 (1234) R 4 = δ 4 P δ (16) Q (12.38)
124
fulfills all criteria (12.35).9 This means that linear N = 8 supersymmetry and SU (8) do
not rule out R 4 . But it does not mean that it will occur in perturbation theory: in fact, we
know from the explicit 3-loop calculation [212] that the 4-graviton amplitude is finite, so
R 4 does not occur. Why not? Well, read on, we will get to that later in this section.
Let us now see an example of how the on-shell matrix elements can be used to rule
out a counterterm. At 4-loop order, we can write down two pure-gravity 10-derivative
operators, D 2 R 4 and R 5 . The first one, D 2 R 4 , stands for the possible scalar contractions
of two covariant derivatives acting (in some way) on four Riemann tensors. No matter
how the indices are contracted, its matrix element turns out to be proportional to (s + t +
u)M4 (1234) R 4 and therefore it vanishes. This means that the 4-point interaction in the
operator D 2 R 4 is a total derivative when evaluated on the equations of motion. So we can
rule out the 4-loop 4-graviton amplitude as the first instance of a UV divergence in N = 8
supergravity. The second 10-derivative operator R 5 is a little more interesting.
The 5-point MHV matrix element of R 5 is fixed uniquely by locality, dimensional
analysis, and little group scaling up to an overall constant:
M5 (1− 2− 3+ 4+ 5+ ) R 5 = a R 5 124 [34]2 [45]2 [53]2 . (12.39)
Next, we check the supersymmetry Ward identity (12.36) in much the same way as for R 4 .
But now we find
? 138
M5 (1− 2+ 3− 4+ 5+ ) R 5 = M5 (1− 2− 3+ 4+ 5+ ) R 5
128
! 138 [34]2 [45]2 [53]2
=⇒ a R 5 134 [24]2 [45]2 [52]2 = a R 5 . (12.40)
124
This time momentum conservation does not save us. The LHS and RHS of (12.40) are not
equal for generic momenta, in particular the LHS is local (i.e. does not have any poles)
while the RHS has a pole 1/124 . This is a contradiction that can only be resolved when
a R 5 = 0. So that means that the operator R 5 does not have an N = 8 supersymmetrization.
And that in turn rules out the 5-point 4-loop amplitude as the first UV divergence in N = 8
supergravity. One can further argue [219, 226] that there are no other possible 10-derivative
operators compatible with N = 8 supersymmetry and SU (8), so this means that N = 8
supergravity cannot have its first UV divergence at 4-loop order.
9 For details of how the R-symmetry acts on the superamplitudes, see Section 2.2 of [42].
Equation (12.40) illustrates a conflict between supersymmetry and locality, a conflict

that can be exploited to rule out potential counterterms. We will describe the method for
operators of the form D 2k R 5 , then outline the general results. The strategy is to construct
the most general matrix element M5 (1− 2− 3+ 4+ 5+ ) D2k R 5 that respects the little group scal-
ing, has mass-dimension 2k + 10, and is Bose symmetric in exchange of same-helicity
gravitons. Then ask if there exists a linear combination that respects the SUSY Ward iden-
tities. Practically, this was done in [219] using Mathematica. (For more advanced cases,
Gröbner basis techniques are very useful [221], and the results found can be reproduced
and extended by an analysis based on the superconformal group SU (2, 2|8) [221].)
Example. As an example of the procedure in [219], consider D2 R5. First we find that
there are 40 angle-square bracket monomials of degree 12 that have the correct little
group scaling of M5 (1− 2− 3+ 4+ 5+ ) D2 R 5 . This does not take into account redundancy
under Schouten or momentum conservation. Now take linear combinations of the 40
monomials to enforce Bose symmetry: this leaves six polynomials as candidates for
M5 (1− 2− 3+ 4+ 5+ ) D2 R 5 . Then impose Schouten and momentum conservation and one
finds that only one polynomial survives: this means that the MHV matrix element of
D 2 R 5 is unique; it takes the form
M5 (1− 2− 3+ 4+ 5+ ) D2 R 5 = a D2 R 5 s12 124 [34]2 [45]2 [53]2 . (12.41)
This is actually just s12 times M5 (1− 2− 3+ 4+ 5+ ) R 5 in (12.39). Now repeat the SUSY
test (12.40) for M5,D2 R 5 to find that the RHS has a pole 1/123 , contradicting the
locality of the LHS. So the operator D 2 R 5 is excluded as a counterterm for N = 8
supergravity, which means that the first divergence cannot be in the 6-loop 5-graviton
amplitude.
Let us summarize the result of the process outlined above for a few more operators in
the same class – after each step given in the first column, we list how many polynomials
remain:
R5 D2 R5 D4 R5 D6 R5 D8 R5
Little grp 1 40 595 4983 29397
Bose symmetry 1 6 63 454 2562 (12.42)
Schouten, mom-cons 1 1 6 9 24
Weakest pole in SUSY Ward id 12−4 12−3 12−2 12−1 no pole
It follows from the last line that the D 2k R 5 operators are excluded as counterterm
candidates for k = 0, 1, 2, 3, but not for k = 4 where there is one unique matrix element
that solves the supersymmetry Ward identities. Thus there is a unique operator D 8 R 5
that passes the tests of linearized SUSY; if present, this would correspond to a first UV
divergence at 8-loop order in the 5-graviton amplitude.
Exercise 12.6
Show that for the operator R n with n ≥ 3 there are no n-point MHV matrix elements
compatible with the N = 8 supersymmetry Ward identities.
For operators with n > 5 fields, one has to distinguish between the different N K MHV sec-
tors. Beyond MHV level, this can be done using the solutions to the SUSY Ward identities
[42] in N = 8 supergravity. As an example of a non-MHV result, the lowest order permis-
sible NMHV-level operator is D 4 R 6 at 7-loops. Actually, there are two independent NMHV
matrix elements, so this means that there are two independent linearly-supersymmetrizable
operators D 4 R 6 .
A detailed analysis of possible counterterm operators was carried out in [219] (see also
[226]) and it was found that
below 7-loop order, the only operators compatible with linearized N = 8 supersymmetry
and SU (8) R-symmetry are
R4 ,
DR ,
4 4
DR .
6 4
(12.43)
3-loop 5-loop 6-loop
At 7-loop order, an infinite tower of linearized N = 8 supersymmetry and SU (8) R-

symmetry permissible operators was found
7-loops: D8 R4 , D4 R6 , R8 , φ2 R8 , φ4 R8 , ... (12.44)
The operators φ 2k R 8 should be viewed as representatives for the linearized N = 8 super-

symmetrization of some contraction of eight Riemann tensors multiplied by an SU (8)-
singlet combination of 2k scalars S ABC D . These do not have purely gravitational (8+2k)-
point matrix elements, so they cannot be in the MHV or anti-MHV sector. For example, it
was shown [221] that φ 2 R 8 only gives N3 MHV matrix elements (four distinct ones).
At 8-loop order and beyond, there are infinite towers of operators that respect linearized
N = 8 supersymmetry and SU (8).
You can find a detailed characterization of the counterterms in Table 1 of [221].

There is one symmetry we did not use to restrict the candidate counterterm operators
in the above discussion, and that is the E 7(7) “hidden” symmetry.10 As discussed at the
end of Section 12.3, it manifests itself in the amplitudes of N = 8 supergravity through
low-energy theorems. These also have to apply to the matrix elements of any acceptable
candidate counterterm operator O. In particular, the single-soft-scalar limit must vanish,
for example

lim M6 S 1234 S 5678 h − h − h + h + O = 0 . (12.45)
p1 →0
If the matrix element of an operator does not pass the single-soft-scalar test, then it is not
compatible with E 7(7) symmetry. If it does pass the test, then we conclude nothing: it could
be E 7(7) at play or just a coincidence.
10 See also [227, 228] for related aspects of E 7(7) symmetry.

It turns out that the single-soft-scalar test is only non-trivial starting at n = 6. So pre-
cisely (12.45) can be used to test the operators that survived the N = 8 supersymmetry
and SU (8) constraints, for example the L = 3, 5, 6 operators in (12.43). But it is not
easy to use Feynman rules to calculate the 6-point matrix elements of, say, R 4 . However,
M6 (S 1234 S 5678 h − h − h + h + ) R 4 can be extracted from the α -expansion of the closed super-
string theory tree amplitude. This may bother you, because there are no continuous global
symmetries in string theory, and here we are interested in testing global continuous E 7(7) .
However, the 4d tree-level superstring amplitudes have an accidental global SU (4) × SU (4)
symmetry. This is a consequence of T-duality when 10-dimensional superstring theory is
reduced to 4d by compactifying it on a 6-torus. The easiest way to see the SU (4) × SU (4)
symmetry is through the KLT relations: the two open string tree amplitudes on one side of
KLT have SU (4) symmetry, so the closed string tree amplitude on the other side of KLT
inherits SU (4) × SU (4); the SU (4) × SU (4) is enhanced to SU (8) only in the α → 0 limit
[36].
Now, we are going to use the α contributions from the superstring tree amplitudes,
but their SU (4) × SU (4) is not good enough, we need SU (8). So we average the string
amplitude over all 35 independent embeddings of SU (4) × SU (4) into SU (8) to get an
SU (8)-invariant answer. For example, the leading α -correction to the closed superstring
amplitude is order α 3 . Dimensional analysis implies that this comes from an 8-derivative
effective operator with N = 8 supersymmetry and (after averaging) SU (8)-invariance. But
we know from the previous analysis that there is only one such operator, namely R 4 . So
after making it SU (8)-invariant, the α 3 contribution from the SU (8)-averaged superstring
amplitude must be identical (up to an overall constant) to the matrix element of the operator
R 4 in N = 8 supergravity [220]! 11 The open superstring tree amplitudes are known in
the literature [229], so pulling them through KLT, the M6 (S 1234 S 5678 h − h − h + h + ) matrix
element can be extracted, SU (8)-averaged, and then subjected to the single-soft-scalar test
(12.45).12 And R 4 fails this test: so R 4 is not compatible with E 7(7) . Hence E 7(7) excludes R 4
as a candidate counterterm and this explains why the 3-loop 4-graviton amplitude is finite.
The on-shell matrix element technique made it possible to show that the 3-loop divergence
of N = 8 supergravity is excluded without doing any loop-amplitude calculations [220].
Note that this is not a string theory argument – string theory amplitudes were only used as
a tool to calculate the relevant field theory matrix elements – so the argument in [220] is
field theoretical.
The analysis outlined above can be repeated [221] for D 4 R 4 and D 6 R 8 and both are
shown to be excluded by E 7(7) . (See also [231] for a string-based argument for the absence
of R 4 , D 4 R 4 and D 6 R 8 .) This means that the symmetries of N = 8 supergravity ex-
clude the divergences in any amplitudes below 7-loop order [221]. Explicit calculations of
the L = 5, 6 amplitudes are not yet available, but as explained here they are expected to
yield finite results in 4d.
11 In the Einstein frame, the effective operator in superstring theory at order α 3 is actually e−6φ R 4 , where φ
is the dilaton. The presence of the dilaton (which can be identified as a certain linear combination of the 70
scalar scalars in the N = 8 supergravity spectrum) operator breaks SU (8) to SU (4) × SU (4). The average
over embeddings SU (4) × SU (4) → SU (8) eliminates this breaking and restores SU (8).
12 An earlier test [230] did not involve the SU (8)-average.
L
E7(7)
3 R4
4 D2R4 R5
E7(7)
5 D4R4 D2R5 R6
E7(7)
6 D6R4 D4R5 D2R6 R7
7 D8R4 D6R5 D4 6
E7(7) R D2R7 R8
E7(7)
ϕ2D2R7 ϕ2R8 ϕ4D2R7 ϕ4R8 ϕ6D2R7
E7(7) E7(7)
8 D10R4 D8R5 D6R6 D4R7 D2R8 R9 ϕ2D2R8 ϕ2R9 ϕ4D2R8 ϕ4R9

No
4
9 D12R4 D10R5 D8R6 D6R7 D4R8 D2R9 R10 ϕ2D2R9 ϕ2R10 ϕ4D2R9 N MHV
10 D14R4 D12R5 D10R6 D8R7 D6R8 D4R9 D2R10 R11 ϕ2D2R10 ϕ2R11
No MHV No NMHV No N2MHV No
N3MHV
Overview of exclusion of counterterm operators in N = 8 supergravity in 4d. Fig. 12.1
An overview of possible counterterms in 4d N = 8 supergravity is presented in Figure

12.1. Note that at 7-loop order, all but the D 8 R 4 operator are excluded. This means
that calculation of the 4-graviton 7-loop amplitude can completely settle the question of
finiteness at 7-loop order. But this is not known to be the case at 8-loops or higher.
In the following section, we give a brief survey of the current status of the UV behavior
of loop amplitudes in supergravity theories.
12.6 Supergravity divergences for various N , L, and D
Current approaches to examining the possible UV divergences of perturbative (super)-

gravity can be categorized as follows:
• Direct computation. The explicit computations of loop amplitudes are made possible
by increasingly sophisticated applications of the generalized unitarity method; this in
itself advances the technical tools for attacking higher-loop computations in general
field theories. For the purpose of exploring UV divergences in supergravity theories,
most efforts focus on the 4-graviton amplitude; this is because (as we have seen in the
previous section) the lowest counterterm for pure supergravity is of the form D 2k R 4 .
See [232] for a discussion on the various details of obtaining multi-loop integrands and
extracting the UV divergences.
• Symmetry analysis. The leading UV divergence of a theory has a corresponding local
gauge-invariant operator that must respect all non-anomalous global symmetries of the
theory. Analyzing the symmetry properties of operators, one can rule out UV divergences
and identify candidate counterterms. In Section 12.5, we took an approach based on the
on-shell matrix elements of the candidate operators, and used it to rule out divergences in
4d N = 8 supergravity for L ≤ 6. Alternatively, one can also use extended-superspace
to construct possible invariant operators; for early constructions, see [233, 234]. If the
invariant operator can be expressed as a superspace integral over a subset of superspace
coordinates, then it is considered a “BPS operator” and it is subject to non-renormalization
theorems. If it is given as a full superspace integral, then it is non-BPS and expected to

receive quantum corrections, thus serving as a candidate counterterm. The distinction
between BPS and non-BPS invariants relies on subtle assumptions about the number of
supersymmetries that can be linearly realized off-shell. This lies outside the scope of this
presentation, so we refer you to [235] and references therein.
• Pure spinor formalism. The “pure spinor formalism” [236] is a first-quantized approach
(in contrast, QFT is a second-quantized approach) to scattering processes in 10d maximal
supersymmetric theories. Using the 10d loop-integrand obtained in the pure spinor
formalism, one can infer properties of the loop amplitudes in D ≤ 10 and this can be
helpful for assessing potential UV divergences. For more details of this approach to
multi-loop amplitudes, see [218].
• Role of non-perturbative states? The 28 massless states of ungauged N = 8 supergravity
in 4d match exactly the spectrum of massless states of closed Type IIB superstring theory
compactified on a six-torus T 6 . At the classical level, N = 8 supergravity in 4d can be
viewed as the low-energy (α → 0) limit of Type IIB superstring theory on T 6 . However,
it was pointed out in [237] that one cannot obtain perturbative N = 8 supergravity in
4d as a consistent truncation of the string spectrum: in the limit α → 0, keeping the 4d
coupling small forces infinite towers of additional states to become light, e.g. Kaluza–
Klein states, winding modes, Kaluza–Klein monopoles, and/or wrapped branes. Thus
one obtains from string theory not just the spectrum of N = 8 supergravity, but a slew
of additional light states. This argument is independent of whether N = 8 supergravity
is finite in 4d or not. However, it does mean that even if pure N = 8 supergravity in 4d
were to be a well-defined theory of quantum gravity, its UV completion would not be
Type IIB string theory.
A related objection [238]13 to the program of studying finiteness of perturbative su-
pergravity is that even the 4d N = 8 supergravity theory itself contains non-perturbative
states, namely BPS black holes, that in certain regions of moduli-space become light
enough that they may enter the perturbative expansion [238]. Such contributions would
never enter the unitarity method approach to explicit calculation of amplitudes.
As we have noted, there are certainly examples of divergences in various perturbative (non-
super)gravity theories in 4d: 1-loop in gravity with generic matter [204, 207], 2-loop in
pure gravity [205, 206], and at 1-loop [240] in dilaton–axion gravity whose spectrum was
given in (12.16). The first example of a UV divergence in pure supergravity was found at
4-loop order in the 4-graviton amplitude of N = 4 supergravity [241].
We end this chapter by summarizing what explicit computations of supergravity loop
amplitudes have revealed so far about the critical dimension Dc of supergravity with various
numbers of supersymmetry:
Maximal supergravity (32 supercharges)
Loop-order 1 2 3 4
. (12.46)
11
Dc 8 [82] 7 [242] 6 [213] 2
[216]
13 See also [239].

Half-maximal supergravity (16 supercharges)
Loop-order 1 2 3 4
. (12.47)
Dc 8 [243] 6 [244] > 4 [245] ≤ 4 [241]
Half-maximal supergravity with matter (both with 16 supercharges)
Loop-order 1 2 3
. (12.48)
Dc 4 [246] ≤ 4 [247] ≤ 4 [247]
In the above, “≤ 4” indicates an upper bound for the critical dimension and we have given
a literature reference for each result.
The absence of a UV divergence for half-maximal supergravity at 3-loops in 4d [245]
as well as 2-loops in 5d [244], was not anticipated by superspace-based analyses. This
unexpected result prompted a conjecture of the existence of an off-shell formalism that
preserves the full 16 supersymmetries [248]. This would imply finiteness for half-maximal
supergravity with matter at 2-loop in 5d, but explicit calculations [247] have shown that a
UV divergence is actually present, thus contradicting the conjecture.
The study of the UV structure of perturbative supergravity theories in diverse dimensions
has resulted in some interesting insights into the relation between gravity and gauge theory
scattering amplitudes: this includes the color-kinematics duality that is the subject of the
next chapter. It is relevant to note that no matter what one thinks of the program to study
the UV-behavior of supergravity, these new insights would have been difficult to come by
without the effort put into explicit calculations and the lessons learned in the process.
13 A colorful duality
A recurring theme in our discussion of perturbative supergravity in Chapter 12 is captured by

the abstract expression “gravity = (gauge theory)2 .” It enters in the context of the spectrum
of states, for example as

N = 8 supergravity = N = 4 SYM ⊗ N = 4 SYM , (13.1)
and also carries over to the scattering amplitudes. For instance, the gravity 3-point amplitude
equals the square of Yang–Mills 3-gluon amplitude (12.8). That is a special case of the KLT
formula (12.13) which expresses the tree-level n-graviton amplitude as sums of products of
color-ordered n-gluon Yang–Mills amplitudes. The KLT formula extends to all tree-level
amplitudes in N = 8 supergravity following the prescription (13.1) for “squaring” the
spectrum.
While the KLT formula follows the “gravity = (gauge theory)2 ” storyline, it is unsat-
isfactory in some respects. First, the formula becomes highly tangled at higher points, as
it involves nested permutation sums and rather complicated kinematic invariants. Second,
since it involves products of different color-ordered amplitudes, it is not really a squaring
relation (except at 3-points). Finally, it is only valid at tree-level. You may think that it
is asking too much to have gravity amplitudes, arising from the complicated Einstein–
Hilbert Lagrangian, closely related to amplitudes of the much simpler Yang–Mills theory.
But one lesson we have learned so far is not to let the Lagrangians obscure our view!
In this chapter, we explore a form of “gravity = (gauge theory)2 ” that makes the ampli-
tude squaring relation more direct and has been proposed to be valid at both tree- and
loop-level.
We begin by giving a qualitative answer to the simple question: why is the KLT formula
so complicated? In our study of color-ordered amplitudes, we often exploit the fact that the
2
allowed physical poles are those that involve adjacent momenta, e.g. 1/Pi,i+1,..., j−1, j . This is
a special feature linked to the color-ordered Feynman rules. But for gravity, there is no color-
structure and hence no canonical sense of ordering of the external states. Thus the poles
that appear in a gravity amplitude can involve any combination of external momenta. This
tells us that in order to faithfully reproduce the pole structure of a gravity amplitude from
“gravity = (gauge theory)2 ,” we need color-ordered Yang–Mills amplitudes with different
orderings. For higher points, the proliferation of physical poles in the gravity amplitude
forces us to include more and more Yang–Mills amplitudes with distinct orderings. This
is why the KLT formula involves a sum over a growing number of different color-ordered
Yang–Mills amplitudes. The complicated kinematic factors in KLT are needed to cancel
double poles.
Exercise 13.1
Justify the kinematic factors and distinct orderings of the Yang–Mills amplitude in the
KLT relations (12.13).
The above discussion suggests that for the comparison with gravity, it may be more useful to
consider the fully color-dressed Yang–Mills amplitudes instead of the color-ordered partial
amplitudes. The former has the same physical poles as the gravity amplitude. Indeed, this is
a productive path, so to get started, we review some useful properties of the color-structure
of Yang–Mills amplitudes.
13.1 The color-structure of Yang–Mills theory
The full color-dressed n-point tree amplitude of Yang–Mills theory can be conveniently
organized in terms of diagrams with only cubic vertices, such as
. (13.2)
The tree amplitude is then written as a sum over all distinct trivalent diagrams,
labeled by i,
ci n i
n =
Atree 2 . (13.3)
i∈trivalent αi pαi
The denominator is given by the product of all propagators (labeled by αi ) of a given

diagram. The numerators factorize into a group-theoretic color-part ci , which is a polyno-
mial of structure constants f abc , and a purely kinematic part n i , which is a polynomial of
Lorentz-invariant contractions of polarization vectors i and momenta pi . As an example,
the 4-point amplitude is
2 3
2 3 2 3
cs n s cu n u ct n t
Atree
4 = + + = + + ,
s u t
1 4 1 4
1 4
(13.4)
sj
si sk
ci + cj + ck = 0
Fig. 13.1 Three trivalent diagrams whose color factors ci , c j , and ck are related by the Jacobi identity. Note that the diagrams
share the same propagators except one, indicated by a thicker gray line. We denote the unshared inverse propagators
as si , s j , and sk .
where the color factors
cs ≡ fã1 a2 b f˜b a3 a4 , ct ≡ fã1 a3 b f˜b a4 a2 , cu ≡ fã1 a4 b f˜b a2 a3 , (13.5)
as already introduced in (2.68). The normalization of the structure constants fãbc was
discussed in footnote 5 on page 30 of Chapter 2.
The numerators n i can be constructed straightforwardly using Feynman rules. Feynman
diagrams with only cubic vertices directly contribute terms of the form ci npi 2 . The Yang–
αi αi
Mills 4-point contact terms can be “blown up” into s-, t- or u-channel 3-vertex pole diagrams
by trivial multiplication by 1 = t/t = s/s = u/u. Note that since cs + ct + cu = 0, this
does not give a unique prescription for how to assign a given contact term into the cubic
diagrams, so the numerators in (13.3) are not uniquely defined.
We can deform the numerators n i in several ways without changing the result of the
amplitude. For example, one can trivially shift the polarization vectors as i ( pi ) → i ( pi ) +
αi pi ; this changes the kinematic numerator factors n i , but not the overall amplitude because
it is gauge invariant. A more non-trivial deformation uses the color factor Jacobi identity
cs + ct + cu = 0: taking n s → n s + s, n t → n t + t, and n u → n u + u, where is an
arbitrary function, leaves the amplitude invariant since the net deformation is proportional
to cs + ct + cu .
In general, for any set of three trivalent diagrams whose color factors are related through
a Jacobi identity,
ci + c j + ck = 0 , (13.6)
the following numerator-deformation leaves the amplitude invariant:
n i → n i + si , n j → n j + s j , n k → n k + sk . (13.7)
Here 1/si , 1/s j , and 1/sk are the unique propagators that are not shared among the three
diagrams, as shown in Figure 13.1. Since can be an arbitrary function, it is similar to a
gauge parameter, except that now it is not a transformation of the gauge field, but rather a
transformation of the numerator factors n i . Because of this similarity, the freedom (13.6),
(13.7) is often called generalized gauge transformation [9]. It plays an important role in
linking Yang–Mills and gravity.
The fact that the numerators n i are not unique nor gauge invariant should not raise
any alarm. After all, the individual Feynman diagrams are not physical observables. For
practical purposes, it is useful to focus on gauge-invariant quantities. Note that if the color
factors are organized in a basis that is independent under Jacobi identities, the coeffi-
cient in front of each basis element is necessarily gauge invariant. These coefficients then
serve as “partial-amplitudes” that constitute part of the full amplitude, but are fully gauge
invariant.
A straightforward way to obtain such partial-amplitudes is to start with the full color-
dressed amplitude in (13.3) and use the color Jacobi identity to systematically disentangle
the color factors. This can be achieved in a graphical fashion introduced in [8]: start with
the color factor of an arbitrary Feynman diagram (with all contact vertices blown up into
two cubic vertices as discussed earlier) and convert it by repeated use of the Jacobi identity
into a sum of color factors in multi-peripheral form
σ1 σ2 σn–2
....... → fã1 aσ1 b1 f˜b1 aσ2 b2 · · · f˜bn−3 aσn−2 an ,
1 n
(13.8)
where the positions of legs 1 and n are fixed and σ represents a permutation of the remaining
n−2 legs. As an example, consider a color diagram that has a Y-fork extending from the
baseline. Applying the Jacobi identity on the propagator in the Y-fork, the diagram is
converted to a linear combination of two diagrams in multi-peripheral form:
2 3
4 2 34 2 3 4
=− − . (13.9)
1 5 1 5 1 5
Any trivalent diagram can be cast into a linear combination of diagrams of multi-peripheral
form (13.8). The important point is then that the color factors in multi-peripheral form are
not related by any Jacobi identities, so there is a total of (n − 2)! independent color factors
at a given n. The full color-dressed tree amplitude can be expressed in terms of this color
basis and then the coefficient of each color factor is a gauge-invariant quantity, denoted
for now by Ãn . We write the full color-dressed amplitude in the multi-peripheral basis
as

n =
Atree fã1 aσ1 b1 f˜b1 aσ2 b2 · · · f˜bn−3 aσn−2 an Ãn (1, σ1 , σ2 , . . . , σn−2 , n) , (13.10)
σ ∈Sn−2
where the sum is over all permutations of lines 2, 3, . . . , n − 1.

Recall from Section 2.5 that we introduced an alternative, manifestly crossing, symmetric
representation that uses trace factors of generators as the basis for the color factor. In this
trace-basis the color-dressed amplitude is

n =
Atree Tr(T aσ1 T aσ2 · · · T aσn−1 T an )An σ1 , σ2 , . . . , σn−1 , n , (13.11)
σ ∈Sn−1
where one sums over all permutations of lines 1, 2, . . . , n − 1 and An [. . .] is our familiar
color-ordered amplitude. Note that there are (n − 1)! distinct traces in (13.11), but since
there are only (n − 2)! independent color factors, the trace “basis” is over-complete and
the color-ordered partial amplitudes satisfy special linear relations. These linear relations
are the Kleiss–Kuijf relations [7], the simplest of which is the U (1) decoupling identity
shown in (2.85). These relations were discussed in Section 2.5; they reduce the number of
independent color-ordered amplitudes from (n−1)! to (n−2)!.
The (n−2)! partial amplitudes Ãn are exactly the color-ordered partial amplitudes that
are independent under the Kleiss–Kuijf relations. Ãn is not unique, since we could have
chosen any other two legs to replace 1 and n as reference legs in the multi-peripheral
color-basis (13.10). This reflects the fact that there is no unique choice of independent
color-ordered partial amplitudes under the Kleiss–Kuijf relations.
In summary, we have reviewed that for a given Yang–Mills n-point tree amplitude, there is
a total of (n−2)! color factors that are independent under the Jacobi identities. A convenient
choice of independent color factors is those that appear in a multi-peripheral representation,
and they can be chosen to be a suitable subset of the (n − 1)! color-ordered amplitudes
An [. . .].
13.2 Color-kinematics duality: BCJ, the tree-level story
The discussion in Section 13.1 may appear to be a deviation from our path to gravity, but it
is a useful detour, as we will see shortly. We begin with the only amplitude we know where
gravity is given as a direct square of Yang–Mills, namely the 3-point amplitude. We would
like to understand what is so special about the 3-point amplitude that is not shared by its
higher-point counterparts. The 3-point superamplitude of N = 4 SYM is

δ (8) Q̃
A3 = . (13.12)
122331
This amplitude is cyclic invariant, as required for a color-ordered superamplitude. But note
that it is also totally antisymmetric, exactly as the 3-point color factor f abc . Hence, the
3-point superamplitude has kinematics that reflect the structure of the color factor of a
3-vertex. Taking a leap of faith, we might wish to generalize this to higher-points, such
that the kinematics of each individual trivalent diagram satisfies the same properties as its
color factor, including Jacobi identities as in Figure 13.1. But what do we mean by the
kinematics of each diagram? As we have seen, this is not a gauge-invariant statement. It

certainly cannot include propagators, as you can see from Figure 13.1. Thus we jump to
the conclusion that perhaps the numerator factors n i in (13.3) can be arranged to have the
same properties as the corresponding color factors ci ?
The color-kinematics duality was first proposed for Yang–Mills theories by Bern,
Carrasco, and Johansson (BCJ) [9]. The duality states that scattering amplitudes of
Yang–Mills theory, and its supersymmetric extensions, can be given in a representation
where the numerators n i have the same algebraic properties as the corresponding color
factors ci . More precisely, using the representation (13.3), the BCJ proposal is that one can
always find a representation such that the following parallel relations hold for the color and
kinematic factors:
ci = −c j ⇔ n i = −n j
(13.13)
ci + c j + ck = 0 ⇔ n i + n j + n k = 0 .
The duality does not state that the numerator factors (13.13) have to be local; they are
allowed to have poles.
To illustrate the identity in the first line of (13.13), consider the two diagrams
. (13.14)
ci cj
They are related by simply switching two lines on a 3-point vertex, highlighted in gray solid
and dashed lines in (13.14). The color factors of the diagrams are related by a minus sign:
ci = −c j , so the color-kinematics duality (13.13) states that there exists a representation
where the numerator factors of the two diagrams respect the same antisymmetry property:
n i = −n j .
The second line in (13.13) signifies that the numerator factors must satisfy exactly
the same linear relations as their associated color factors. For example, since the color
factors of the three diagrams in Figure 13.1 satisfy the Jacobi relation ci + c j + ck = 0, the
color-kinematics duality states that there is a representation of the numerators of the three
diagrams such that n i + n j + n k = 0.
At first sight this duality may seem implausible. While the underlying reason for the
Jacobi identity to hold for the ci s is the non-abelian gauge algebra defined by two-brackets,
[T a , T b ] = f ab c T c , there appears to be no reason for the kinematic numerators to satisfy
the same relations. As we now show, the color-kinematics duality is not as impossible as it
seems.
Recall that there are (n − 2)! independent color factors under the Jacobi identities. If we
require that the numerator factors n i satisfy the same Jacobi identities, then there will only
be (n − 2)! independent numerators as well. Since there are also only (n − 2)! independent
color-ordered partial amplitudes, we can express the set of linearly independent partial
amplitudes in terms of the (n − 2)! numerators:

(n−2)!
(i) =
Atree i j n̂ j . (13.15)
j=1
Here i, j = 1, . . . , (n − 2)!, and Atree

(i) and n̂ j are the independent color-ordered amplitudes
and numerators, respectively. The (n − 2)! × (n − 2)! matrix i j consists solely of massless
scalar propagators. (This matrix was first introduced in [249] as the “propagator matrix.”) As
an example, for n = 4 we choose Atree tree
4 [1, 2, 3, 4] and A4 [1, 3, 2, 4] as the two independent
color-ordered amplitudes. Expanding the color factors in (13.4) in terms of traces (as in
Exercise 2.31), we find
ns nu nu nt
4 [1, 2, 3, 4] = −
Atree + , 4 [1, 3, 2, 4] = −
Atree + . (13.16)
s u u t
Enforcing the color-kinematics duality (13.13) on the numerators gives n t = −n s − n u .
Choosing n s , n u as (n̂ 1 , n̂ 2 ) in (13.15), we can now identify the 2 × 2 matrix i j from:
1 1
Atree
4 [1, 2, 3, 4] =
−s 1
n̂ 1 −s 1
tree
u
−→ ij = u
. (13.17)
A4 [1, 3, 2, 4] − 1t −u − t
1 1 n̂ 2 − 1t − u1 − 1
t
As advertised, the matrix i j consists of propagators. The construction generalizes to

higher-points. The explicit form of the 6 × 6 matrix i j for n = 5 was given in [249]. For
related work, see [250].
Inverting the matrix i j would give us numerators n̂ i expressed in terms of the color-
independent amplitudes, and from the Jacobis one can generate the rest of the numerator
factors, thus trivially obtaining a representation that satisfies the color-kinematics duality.
If this were true, the color-kinematic duality would be trivial and this doesn’t quite smell
right. And it isn’t: in 4d Yang–Mills theory, the matrix i j does not have full rank, so it
cannot be inverted, and we do not have unique numerators n̂ i . Indeed looking back at our
4-point example in (13.17), one can easily verify that the 2 × 2 matrix i j only has rank 1.
There is also another way to see that the numerator factors cannot be uniquely determined.
Suppose we have obtained a set of numerators that satisfy color-kinematic identity. Let us
assume that we have achieved this at 5-points and then add the following term [251] to the
Yang–Mills action:
D5 = f a1 a2 b f ba3 c f ca4 a5

a1 a2 a3 μ a2 a1 μ a3 a1 a2 μ a3 1 a4 ν a5 ρ
× ∂(μ Aν) Aρ A + ∂(μ Aν) A Aρ + Aρ A ∂(μ Aν) (A A ) .

(13.18)
The presence of the trivial 1 =

facilitates the identification of n i s in the cubic diagram
expansion. The term (13.18) is identically zero thanks to the Jacobi identity, so it does not
change the theory. It does, however, modify the Feynman rules so it changes the numerator
factors, but in such a way that they still satisfy the color-kinematic Jacobi identity. We
conclude that the duality-satisfying numerators n̂ i cannot be unique; and this is why the
matrix i j is not invertible in 4d Yang–Mills theory.
Given that the matrix i j has lower rank, there must be linear relations among the color-
independent partial amplitudes. To expose such relations for n = 4, use the first row of
(13.17) to express n̂ 1 in terms of the partial amplitude Atree
4 [1, 2, 3, 4] and n̂ 2 ,
s
n̂ 1 = −s Atree
4 [1, 2, 3, 4] + n̂ 2 . (13.19)
u
Substituting this solution into the second row of (13.17), we find that

n̂ 1 n̂ 2 n̂ 2 s tree s 1 1
A4 [1, 3, 2, 4] = − −
tree
− = A4 [1, 2, 3, 4] − + + n̂ 2 . (13.20)
t t u t ut t u
The coefficient of n̂ 2 is proportional to s + t + u = 0. Thus, imposing color-kinematics

duality gives the following relation among color-ordered amplitudes:
4 [1, 3, 2, 4] = s A4 [1, 2, 3, 4] .
t Atree tree
(13.21)
This is an example of the BCJ relations that we discussed earlier in Section 2.5. In fact,
what we did above is equivalent to your calculation in Exercise 2.31. Since the matrix i j
is defined with respect to color-ordered amplitudes that are independent under the Kleiss–
Kuijf relations, the BCJ amplitude relations are new relations beyond the consequences
of the color-structure. These novel relations reflect that color-kinematics duality exists
in Yang–Mills theory. It is known [9] that for general n, the (n − 2)!-by-(n − 2)! matrix
i j has rank (n − 3)!, thus implying (n − 2)! − (n − 3)! = (n − 3)(n − 3)! BCJ relations
among n-point color-ordered amplitudes. The simplest type of such relations (sometimes
called fundamental BCJ relations) can be nicely condensed to the form [9]
⎛ ⎞

n i

⎝ s2 j ⎠ Atree
n 1, 3, . . . , i, 2, i + 1, . . . , n = 0 . (13.22)
i=3 j=3
In our 4-point example, n̂ 2 dropped out of the final equation (13.20). This means that
no consistency conditions can be put on n̂ 2 , so we can take n̂ 2 to be anything we want
without affecting the physical amplitude. Such free numerators are sometimes referred to
as “pure-gauge.” In practice, it is often convenient to set them to zero.
Now you may say that this is all very interesting, but have we lost sight of our original
motivation, to get gravity amplitudes from Yang–Mills theory!? No worries, we are already
there. A remarkable proposed consequence of the color-kinematics duality is that once
duality-satisfying numerators n i are obtained, the formula
n i2
Mntree = 2
(13.23)
i∈cubic αi pαi
calculates the n-point tree amplitude in the (super)gravity whose spectrum is given by
squaring the (super) Yang–Mills spectrum. That is, we simply take the Yang–Mills amplitude
formula in (13.3) and replace each color factor ci with the corresponding duality-satisfying
numerator n i . And, boom, that is gravity! This relation is called the BCJ double-copy
relation. The formula (13.23) manifestly reproduces all possible poles that should appear
in the gravity amplitude. Furthermore, the mass-dimension matches on both sides of the
equation.
Example. Let us check (13.23) at 4-points. From (13.23), we find that

n 2s n2 n2 n 2s n2 (n s + n u )2
M4tree = + u + t = + u + , (13.24)
s u t s u t
where in the second equality, we use color-kinematics duality to set n t = −n s − n u .
Now use (13.19) and (13.21), remembering that (n̂ 1 , n̂ 2 ) is identified with (n s , n u ) and
that we can freely set n̂ 2 = 0: we then find
su tree
M4tree = − A [1, 2, 3, 4]2 = −u Atree
4 [1, 2, 3, 4] A4 [1, 3, 2, 4] .
tree
(13.25)
t 4
This is just a different form of the KLT formula we encountered previously in (12.13)!
A more involved 5-point example was worked out in detail in [9]. Thus, by reproducing
the correct KLT relations, the validity of (13.23) is verified at 4- and 5-points.
Exercise 13.2
What if we choose a gauge where n̂ 2 is not zero? Show that when substituting (13.19)
and (13.21) into (13.23), n̂ 2 drops out in the final result. This shows that the gravity
formula (13.23) is gauge invariant in the sense of (13.7).
The BCJ squaring relations (13.23) can be exploited to determine an explicit representation
of color-kinematics duality-satisfying numerators of the tree amplitude [252, 253]. Recall
that by using the color Jacobi relations, we can convert the fully dressed amplitude in
(13.3) into the multi-peripheral form (13.8). Assuming that we have duality-satisfying
numerators, the double copy representation of the gravity amplitude in (13.23) can now
go through exactly the same steps as those that converted (13.3) into (13.8), and obtain a
multi-peripheral form of gravity amplitude

Mntree = n (1, σ1 , σ2 , . . . , σn−2 , n) ,
n 1|σ1 ,σ2 ,...,σn−2 |n Atree (13.26)
σ ∈Sn−2
where n 1|σ1 ,σ2 ,...,σn−2 |n are the duality-satisfying numerators for the cubic diagrams in multi-
peripheral form with legs 1 and n held fixed. Thus we have an expression for the gravity
amplitude in terms of a sum of Yang–Mills color-ordered amplitudes times kinematic
factors. Note that it looks a lot like the KLT formula. Indeed, as realized first in [252],
by lining up a copy of the KLT formula with the color-ordered amplitudes in (13.26), one
can readily read-off duality-satisfying n i s. Take, for example, the n = 4, 5 KLT formulas
(12.13):
M4tree (1234) = −s12 Atree

4 [1234] A4 [1243] ,
tree
(13.27)
M5tree (12345) = s23 s45 Atree
5 [12345] A5 [13254] + (3 ↔ 4) .
tree
Choosing legs 1 and n to be fixed in our multi-peripheral form, we can readily read off:
n = 4: n 1|2,3|4 = −s12 Atree
4 [1243], n 1|3,2|4 = 0 ;
n = 5: n 1|2,3,4|5 = s23 s45 Atree
5 [13254] , n 1|2,4,3|5 = s24 s35 A5 [14253] ,
tree (13.28)
n 1|3,4,2|5 = n 1|4,2,3|5 = n 1|4,3,2|5 = n 1|3,2,4|5 = 0 .
From these (n − 2)! independent numerators, all remaining numerators can be obtained by
applying the Jacobi identities.
If you think that the BCJ double-copy relation (13.23) is too good to be true, have no
fear: things are about to get even better! It turns out that the squaring relation (13.23) can
be generalized to
n i ñ i
Mntree = 2 , (13.29)
i∈cubic αi pαi
where the gravity numerators are given as the product of two possibly distinct Yang–Mills
numerators n i and ñ i . Only one set of numerators, say n i , has to satisfy the color-kinematics
duality (13.13), while the other copy, ñ i , can be an arbitrary representation of the Yang–
Mills amplitude. To understand why this is so, let us assume that the n i s respect the duality
(13.13) while the ñ i s do not. Define the difference of the two distinct numerators to be
i ≡ n i − ñ i . (13.30)
Since n i and ñ i are both valid representations of the same Yang–Mills amplitude, it follows
from (13.3) that
ci i
2 = 0. (13.31)
i∈cubic αi pαi
In the discussion so far, we have not specified the gauge group, just that it is non-abelian
with structure constants that satisfy the Jacobi identities. Thus the only property of the
color factors ci that can make (13.31) hold is the Jacobi relation. Since the n i s satisfy
color-kinematics duality, they satisfy the exact same algebraic properties as the ci s, so we
conclude that
n i i
2 = 0. (13.32)
i∈cubic αi pαi
This establishes the equivalence of (13.23) and (13.29).

Why bother with the existence of (13.29) vs. (13.23)? – we need one set of duality-
satisfying numerators for (13.29) anyway, so why not simply square them and just use
(13.23)? Well, recall from Section 12.2 that the spectrum of many supergravity theories can
be obtained from tensoring two different Yang–Mills theories. For example, the spectrum of
pure N = 4 supergravity is the product of N = 4 SYM and pure Yang–Mills theory (12.19).
The point is then that the BCJ double-copy relation (13.29) can be used to construct the
supergravity scattering amplitude by using the numerators of two distinct (S)YM theories.
And importantly, only one copy of the numerators needs to satisfy the duality, not both. Thus,
if we have a set of duality-satisfying numerators for N = 4 SYM, by simply combining them
in (13.29) with the numerators of ordinary Yang–Mills, say obtained from explicit Feynman
diagram computation, we directly get the scattering amplitudes of N = 4 supergravity.
This convenient result has powerful consequences as we move on to loop amplitudes in
Section 13.3.
We end this section with some concluding remarks regarding the tree-level BCJ color-
kinematics duality (13.13) and the double-copy relations (13.23) and (13.29).
First, the existence of numerators that satisfy (13.13) was exemplified for any n in
[252, 253] (see also [254]).
Second, the BCJ relations in (13.22) have been successfully derived from string theory
using monodromy relations [255, 256] and in field theory using the improved large-z fall-off
of non-adjacent BCFW shifts [257]. Another, quite elegant, derivation was given in [258]. In
our discussion, the BCJ relations were a consequence of imposing color-kinematic duality
on the numerator. Given that the BCJ relations can be proven via string- and field-theory
arguments, one can reverse the argument and show that the existence of BCJ relations and
Kleiss–Kuijf identities gives rise to algebraic relations on the kinematics [259].
Third, assuming that there exists a duality-satisfying set of local numerators for the Yang–
Mills tree amplitude, one can rigorously prove that the doubling-relation (13.23) produces
the correct gravity tree amplitude for any n [251]. The proof is established inductively
by showing that the difference between (13.23) and the gravity amplitude obtained from
BCFW recursion vanishes if one assumes (13.23) to hold for all lower-point amplitudes.
Finally, you may wonder if duality-satisfying numerators can be obtained directly from
the Feynman rules of some Lagrangian. In 4d, this can be done for MHV amplitudes [260],
but difficulties arise beyond MHV. In general dimensions, a straightforward construction of
the cubic diagrams using Feynman rules does not give duality-satisfying numerators, even
if the freedom of how to assign contact terms is taken into account. However, modification
of the action by non-local terms can give duality-satisfying numerators straight from the
Feynman rules of the deformed action [251],
LY M = L + L5 + L6 + · · · (13.33)
Here L is the conventional Yang–Mills Lagrangian and Ln , n > 4 are terms that involve
n fields and vanish by the Jacobi identity so that the theory is actually not changed. As an
example, the quintic terms are
1
L5 ∼ Tr [Aν , Aρ ] [∂μ Aν , Aρ ], Aμ + [Aρ , Aμ ], ∂μ Aν + [Aμ , ∂μ Aν ], Aρ .

(13.34)
Even though the deformation is non-local, it is completely harmless: L5 is simply zero
because the terms in the parentheses vanish by the Jacobi identity. Thus by adding a
particular zero to the action, one can expose the intricate relation between gravity and
Yang–Mills theory. (One may say that this points to a curious deficiency of the action,
namely that it treats all zeros in the same way.) A systematic approach to generating
explicit higher-order deformations Ln is given in [261].
We have seen how the tree-level squaring relation between gravity and Yang–Mills is
more straightforward when phrased in terms of the non-gauge-invariant numerators n i than
in terms of the gauge-invariant color-ordered Yang–Mills amplitudes, as in KLT. The true

power of this is revealed when it is applied to loop-integrands: the KLT relations hold only
at tree-level but, as we will see next, the BCJ squaring relations survive at loop-level.
13.3 Color-kinematics duality: BCJ, the loop-level story
We begin with color-kinematic duality for loop amplitudes of Yang–Mills theory. Any
diagram involving the 4-point contact term can be blown up into cubic vertices, as discussed
in Section 13.1, so we consider only trivalent loop-diagrams. The full L-loop color-dressed
Yang–Mills amplitude can then be written as
L
# d D l 1 n j cj
AnL-loop = D
2
, (13.35)
j∈cubic l=1
(2π ) S j α j pα j
where the notation follows that defined for (13.3) and S j is the symmetry factor of the dia-
gram. It was proposed in [262] that there exist representations (13.35) where the kinematic
numerators n i satisfy the same algebraic relation as that of color factors, i.e. (13.13). And
once such numerators are found, the gravity amplitude is given by the double-copy formula
[262]
L
# d D l 1 n j ñ j
MnL-loop = D
, (13.36)
j∈cubic l=1
(2π ) S j α j pα2 j
in which only one of the two copies n i , ñ i is required to satisfy color-kinematics duality.
The validity of (13.36) can be justified through unitarity cuts [67]: assuming that gauge-
theory numerators n i satisfy the duality, the gravity integrand built by taking double copies
of numerators has the correct cuts in all channels. To see this, consider a set of generalized
unitarity cuts that break the loop amplitude down to products of tree amplitudes. On the cut,
the gauge-theory integrand factorizes into products of tree amplitudes whose numerator
factors satisfy all color-kinematics dualities relevant for each tree amplitude, because they
are merely a subset of the relations required by the loop-level duality. As an example
consider the following unitarity cut of the 3-loop cubic diagram:

2 4 n 2 4 na nb nc
1 = 2 2 2 2 2 = 1 = .
3 5 p1 p2 p3 p4 p5 cut 3 5 p p2 p3 p42 p52
2
states 1
2 2
n na nb nc
(13.37)
We have labeled the uncut propagators 1/ pi2 by i = 1, . . . , 5 and the numerator of the
diagram is denoted by n. On the cut, the LHS must be equivalent to the RHS, which
is the product of kinematic factors of the factorized tree-diagrams. If the numerators n

satisfy all Jacobi identities associated with the parent diagram, the numerators n a , n b , and
n c must satisfy the Jacobi identities of the individual tree-diagrams. Now squaring the
duality-satisfying numerators in the Yang–Mills tree amplitude, one obtains the gravity tree
amplitude. Thus, squaring the Yang–Mills loop-numerators, one is guaranteed to obtain the
correct cut for the gravity loop amplitude. In other words, (13.36) is guaranteed to satisfy
all unitarity cuts and therefore give the correct answer [262].
If any readers have come all the way with us to this page, then they may question if the
above argument only justifies (13.36) as the correct answer for the cut-constructible part of
the (super)gravity loop amplitude: what about rational terms that are not cut constructible?
Recall that rational terms can be obtained by considering the unitarity cuts in higher-
dimensions, where the extra-dimensional momenta can be interpreted as the regulator (see
(6.34)). But in the discussion so far, there was no specification of the spacetime dimension:
the color-kinematics duality and the double-copy relation are valid in arbitrary dimensions!
In other words, (13.36) produces the correct cut in any spacetime dimension, and thus it
also faithfully reproduces the rational terms.
Now the validity of (13.36) relies on the existence of duality-satisfying numerators n i .
Do we know that there always exists such a representation – and, if so, how to systemat-
ically construct it? Indeed, this is the million dollar question.1 Unlike at tree-level, there
is currently not a formal proof of the existence of duality-satisfying numerators. How-
ever, we have explicit examples of such numerators in multiple cases, as summarized in
Section 13.3.4. For now, let us see some non-trivial examples.
13.3.1 1-loop 4-point N = 4 SYM

In Section 6.3, we used the unitarity method to compute the color-ordered 1-loop 4-point
superamplitude of N = 4 SYM and found the result to be
1-loop
4 [1234] I4 ( p1 , p2 , p3 , p4 ) , (13.38)
where I4 ( p1 , p2 , p3 , p4 ) is the scalar box integral. The fully color-dressed 1-loop amplitude
can be written in terms of color-ordered amplitudes as [40, 103]
1-loop, full (1) 1-loop (1) 1-loop (1) 1-loop

A4 = c1234 A4 [1234] + c1342 A4 [1342] + c1423 A4 [1423] , (13.39)
(1)
with ci jkl the 1-loop color factor of a box diagram with consecutive external legs (i, j, k, l),
e.g.
(1)
c1234 = f˜ea1 b f˜ba2 c f˜ca3 d f˜da4 e . (13.40)
Let us now show that (13.39) does satisfy color-kinematic duality. Take one of the four
propagators in the box diagram and apply the Jacobi identity to, say, the propagator between
1 Let’s be clear: the authors (H.E. and Y.-t.H.) will not pay anyone a million dollars – or any other amount.
legs 1 and 2:
2 3 1 3 3
2
= − . (13.41)
1
1 4 2 4 4
The relevant Jacobi identity is
(1) (1) (1)
c1234 − c2134 + cTri:34 = 0 , (13.42)
(1)
where cTri:34 ≡ fã1 ba2 f˜ebc f˜ca3 d f˜da4 e is the color factor for a triangle diagram. The duality
then states that the numerators of the integrals are related as
(1) (1) (1)
n 1234 − n 2134 + n Tri:34 = 0 . (13.43)
Looking back at the 1-loop color-dressed result (13.39), we immediately identify
(1) (1) (1)
n 1234 = su Atree
4 [1234], n 2134 = st Atree
4 [2134], n Tri:34 = 0 . (13.44)
4 [1234] is fully permutation invariant, it equals st A4 [2134], and therefore

Since su Atree tree
the numerator Jacobi identity (13.43) is trivially satisfied! Applying the Jacobi identity to
any other propagator of the box diagram, one arrives at the same result.
In principle we should also check the Jacobi identity
2 1
2 3 3 3
= − . (13.45)
1 4 4 4
1 2
However, for N = 4 SYM, this is trivially satisfied since the numerator associated with
each of the above diagrams is zero.

At 2-loop order, the color-dressed Yang–Mills amplitude is [103, 242]

2-loop, full
A4 = c1234
P
AP4 [1234] + c3421
P
AP4 [3421] + c1234
NP
ANP
4 [1234] + c3421 A4 [3421]
NP NP
+ cyclic(2, 3, 4) , (13.46)
where “cyclic(2, 3, 4)” indicates a sum over the remaining two cyclic permutations of legs 2,
P NP
3 and 4. The color factors c1234 and c1234 are obtained by dressing the planar and non-planar
double-box diagrams with structure constants fãbc :
2 3 2 3
. (13.47)
1 4 1 4
The explicit amplitudes AP4 [1234] and ANP

4 [1234] are given by the 2-loop scalar-box inte-
grals corresponding to the diagrams (13.47), multiplied by the numerator factors
n P1234 = s 2 u Atree
4 [1234] , 1234 = s t A4 [1234] .
n NP 2 tree
(13.48)
Color-kinematic duality now imposes the following linear relations among the numerators
of the scalar integrals:
2 3 1 3 2 3
= − ,
1
1 4 2 4 4
2 3 1 3 2 3
= + .
1 4 2 4 1 4
(13.49)
The above two identities are satisfied by the numerators in (13.48) by virtue of the absence
of integrals with triangle sub-loops as well as the permutation invariance of su Atree
4 [1234].
Exercise 13.3
What is the identity associated with the Jacobi relation applied to the thick gray
propagator in the following diagram:
2 3
? (13.50)
1 4
Do the numerators in (13.48) satisfy this identity?

Thus far in our examples, the duality-satisfying numerators n i have been independent of the
loop-momenta. At 3-loop order, the representation given in [262] for the 4-point amplitude
uses duality-satisfying numerators that do depend on the loop-momenta. The scalar integrals
that participate in the 3-loop amplitude of N = 4 SYM answer are shown in Figure 13.2.
The full amplitude is given as a sum of these integrals (along with permutations of the
external legs), their associated color factors and kinematic numerators, suitably normalized
2 3 2 3 2 3
1 (a) 4 1 (b) 4 1 (c) 4
2 3 2 3
2 3
1 5 5
(d) 4
1 (e) 4 1 (f ) 4
2 3 2 3 2 3
7 6 6
5 5 5
1 (g) 4 1 (h) 4 1 (i) 4
2 3 2 3 3
2
1 (j) 4 1 (k) 1 (l)

4 4
The scalar integrals that are involved in the BCJ calculation of the 3-loop amplitude of N = 4 SYM. (Reprinted figure Fig. 13.2
with permission from Z. Bern, J. J. M. Carrasco, and H. Johansson, Phys. Rev. Lett. 105, 061602 (2010) [262]. Copyright
2010 by the American Physical Society. http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.105.061602.)
by the symmetry factor of the diagram. The kinematic numerators are [262]
Integral N = 4 SYM numerator
(a)–(d) s2

(e)–(g) s (−τ35 + τ45 + u) − u (τ25 + τ45 ) + t (τ25 + τ35 ) − s 2 /3

(h) s (2τ15 − τ16 + 2τ26 − τ27 + 2τ35 + τ36 + τ37 − t) (13.51)

+u (τ16 + τ26 − τ37 + 2τ36 − 2τ15 − 2τ27 − 2τ35 − 3τ17 ) + s 2 /3

(i) s (−τ25 − τ26 − τ35 + τ36 + τ45 + 2u)

+u (τ26 + τ35 + 2τ36 + 2τ45 + 3τ46 ) + t τ25 + s 2 /3
(j)–(l) s(u − t)/3
An overall factor of suAtree

4 has been removed, and τi j = 2ki · l j , where ki and l j are the
momenta labeled in each diagram in Figure 13.2. The numerators in the table satisfy all
Jacobi identities of the corresponding color factors. For example,
= + (13.52)
is trivially satisfied because diagram (e) and (f) have the same numerator factor and the
third diagram in (13.52) vanishes.
Exercise 13.4
Verify that the numerators (13.51) satisfy the following identity:
2 1
2
= − . (13.53)
1
1 2
Note that in this example all three numerators are non-zero.
13.3.4 Summary
There is no formal proof that duality-satisfying numerators can always be found for loop
amplitudes in (super) Yang–Mills, however, there is considerable evidence in favor of this
property. We present here a list of non-trivial examples for which the BCJ duality-satisfying
numerators have been constructed:
• Up to 4-loops for 4-point in N = 4 SYM [232, 262].

• Up to 2-loops for 5-point in N = 4 SYM [263].
• At 1-loop up to 7-points in N = 4 SYM [264].
• Up to 2-loops for 4-point for the all-plus pure Yang–Mills amplitude [262].
• 1-loop 4-point for pure Yang–Mills theory in arbitrary dimensions [240].
• 1-loop n-point all-plus or single-minus helicity amplitudes in pure Yang–Mills theory
[265].
• 1-loop 4-point amplitudes in theories with less than maximal supersymmetry [266].
• 1-loop 4-point for an abelian orbifold of N = 4 SYM [267].
• 1-loop 4-point Yang–Mills theory with matter [268].
Although most progress has been made for N = 4 SYM, the examples are not restricted to
the maximally supersymmetric theory or to 4d.
13.4 Implications for UV behavior of supergravity
With duality-satisfying numerators for (N = 4) super Yang–Mills amplitudes, we do not

need to do much work to compute supergravity amplitudes! We now give several examples
of this application of BCJ.
N = 8 supergravity
By squaring the duality-satisfying numerators of the 1-, 2- and 3-loop 4-point ampli-
tudes in Sections 13.3.1–13.3.3, we immediately obtain the integrands of the 1-, 2- and
3-loop 4-point amplitudes in N = 8 supergravity. At 1- and 2-loop order, the N = 4
SYM numerators are independent of the loop-momentum, so since the scalar box and
the scalar double box integrals are UV finite in 4d, we immediately see that the 4-
point amplitudes in N = 8 supergravity are finite in 4d at 1- and 2-loops. The same
conclusion extends to 3-loop order by simple power-counting even though the 3-loop
numerators in (13.51) depend on the loop-momenta. Thus, as promised, without further
calculations, we have just reproduced the result that the 4-point amplitudes of N = 8
supergravity are finite in 4d up to and including 3-loop order. Of course, this agrees
with previous explicit calculations [208–214] and the counterterm analysis discussed in
Section 12.5.
Critical dimension for maximal pure supergravity

As mentioned in Section 12.4, the critical dimension for UV divergences of maximal
supergravity is proposed [212] to match that of maximal SYM,
6
Dc (L) = +4 for L > 1. (13.54)
L
At 1- and 2-loop orders, the duality-satisfying numerators of N = 4 SYM are loop-
independent, so the critical dimension is simply determined by the scalar integrals; it
is therefore universal between N = 4 SYM and N = 8 supergravity. At 3-loops, one can
use power-counting to see that the integrals with the worst UV behavior are the three dia-
grams ( j), (k), (l) in Figure 13.2. These diagrams dictate the critical dimension at 3-loops
for N = 4 SYM. But it follows from (13.51) that the numerator factors of these three
diagrams are loop-independent, so squaring them to get the N = 8 supergravity amplitude
does not change the critical dimension. One can check that none of the other diagrams has
worse behavior than ( j), (k), (l) after squaring. Hence the relation (13.54) for the critical
dimension also holds for N = 8 supergravity at 3-loop order. A similar BCJ argument
extends this result to 4-loop order [232].
Exercise 13.5
Use the explicit integrands given in Sections 13.3.2 and 13.3.3 to verify that the critical
dimension at 2- and 3-loops is (13.54) for both N = 4 SYM and N = 8 supergravity.
What is the critical dimension at 1-loop order?
N ≥ 4 supergravity
We obtained N ≤ 8 supergravity amplitudes by tensoring two sets of numerators from loop
amplitudes in N ≤ 4 supersymmetric Yang–Mills theories. Only one of the two copies
of numerators needs to satisfy the color-kinematic duality, as discussed around (13.31).
Since we already have duality-satisfying numerators for the 1-, 2- and 3-loops 4-point
N = 4 SYM amplitudes, we can just tensor them with any Yang–Mills or SYM numerators
we like, and obtain N ≥ 4 supergravity amplitudes. Since only a small number of cubic
diagrams have non-vanishing numerators in the N = 4 SYM copy, we only need a few of
the numerators of the other copy.
We begin at 1-loop. Suppose we have an explicit representation of the 1-loop integrand
of N ≤ 4 SYM computed from Feynman rules. Such a representation usually involves
triangles and bubbles as well as diagrams that are not 1-particle-irreducible, but it can be
converted into a representation that only involves the box integrals. The price one pays is
that the numerators will in general be non-local, but that is not a problem for our application.
As an example, the following triangle- and bubble-diagrams can be converted to boxes by
introducing inverse propagators:
pa
n1 n2 n3
p pa pb
pb
pb2
n 1 = n 1 p 2 , n 2 = n 1 pa2 pb2 , n 3 = n 1 . (13.55)
pa2
To obtain the N ≤ 4 supergravity amplitude, we “tensor” the new numerators n i with the
duality-satisfying N = 4 SYM numerators n i in (13.44). Specializing to 4-point, recall that
the numerators in (13.44) are not only loop-momentum independent, but also permutation
invariant. Independence of the loop-momentum means that the n i s move outside the box-
integral. Permutation invariance for the n i s then tells us that the n i -factor is uniform for each
box-integral. In other words, it is just an overall factor, su Atree4,Q=16 [1, 2, 3, 4], multiplying
the entire 1-loop N ≤ 4 SYM amplitude! Thus we have found the following very simple
formula for the 4-point 1-loop amplitude in N ≥ 4 supergravity [269],
& %
(1) (1) (1) (1)
M4,Q+16 = su Atree
4,Q=16 [1, 2, 3, 4] A4,Q [1, 2, 3, 4] + A4,Q [1, 3, 4, 2] + A4,Q [1, 4, 2, 3] .
(13.56)
The subscript Q indicates the number of supercharges, with Q = 16 corresponding to

N = 4 supersymmetry. It is remarkable that the 1-loop amplitude of a non-renormalizable
gravity theory can be given by a sum of 1-loop amplitudes of a renormalizable one. Note
that this relation was exposed only after imposing color-kinematics duality.
Consider now the UV structure in 4d. The 1-loop amplitudes of N = 0, 1, 2 SYM have
UV divergences. In 4d SYM, the UV divergence must be proportional to the tree amplitude,
otherwise it would imply that a new operator is needed to renormalize the 1-loop divergence.
Thus we conclude that

M4,Q+16
(1)
∼ su Atree
4,Q=16 [1, 2, 3, 4]
D=4 div.
& %
× Atree
4,Q [1, 2, 3, 4] + Atree
4,Q [1, 3, 4, 2] + Atree
4,Q [1, 4, 2, 3] = 0.
(13.57)
This result vanishes due to the U (1) decoupling identity (2.85) for the Yang–Mills color-
ordered tree amplitudes. Therefore, with the help of BCJ color-kinematics duality, we have
shown that pure N ≥ 4 supergravity is finite at 1-loop in 4d.
1-loop UV divergence in N = 4 supergravity-matter theory

The simple relation (13.56) can also be used to demonstrate UV divergences. Consider
N = 4 supergravity coupled to N = 4 Maxwell theory (i.e. N = 4 SYM with U (1) gauge
group). The spectrum of this N = 4 supergravity-matter theory is given by tensoring N = 4
SYM with Yang–Mills minimally coupled to an adjoint scalar. The 1-loop amplitude is
(1)
exactly the same as (13.56), except that A4,Q is now the 1-loop amplitude of the Yang–Mills-
scalar theory. The 4-point 1-loop amplitude in Yang–Mills-scalar theory is UV divergent
and is renormalized by a 4-scalar counterterm
(1)
L = cabcd φ a φ b φ c φ d , (13.58)
(1)
where cabcd is the color factor for the box-diagram. Putting this divergence into (13.56) the
sum of the three terms is now non-vanishing because the color-structure of (13.58) is not
that of a tree amplitude. This then shows that there is a 1-loop UV divergence in the N = 4
supergravity-matter model. This divergence was first computed by standard methods in
[246].
Exercise 13.6
Use (13.56) to show that the 1-loop UV divergence of the N = 4 gravity-matter system
corresponds to an F 4 = (Fμν F μν )2 operator.
Color-kinematics constraints on candidate counterterms

In Section 12.5, we approached UV divergences of supergravity from the viewpoint of
local counterterms: we ruled out candidate counterterms based on the known symmetries
in N = 8 supergravity. It is reasonable to say that “everything not forbidden is compulsory”2
and therefore expect that if the known symmetries do not rule out a certain counterterm,
then it is likely to appear in the perturbation theory. In this section, we have studied
a new structure, BCJ color-kinematics duality, that is very different in nature from the
other symmetries imposed on the local counterterms. In examples, we have seen how the
BCJ doubling relation reveals the “true” power-counting (at least “truer” than Feynman
diagrams) for UV divergences in supergravity amplitudes. So one may now wonder if it is
possible that there exist counterterms that respect all known “ordinary” symmetries of the
theory and yet are ruled out by color-kinematics duality? We now present such a case.
2 Also known as Gell-Mann’s Totalitarian Principle (from T. H. White’s The Once and Future King).
The 2-loop duality-satisfying numerators (13.48) of N = 4 SYM are momentum inde-

pendent. Following the same arguments that gave us the 1-loop result (13.56), we find that
the 2-loop 4-point amplitude of N ≥ 4 supergravity is given as a sum of 2-loop (S)YM
amplitudes:
&
(2)
M4,Q+16 [1, 2, 3, 4] = su A4,Q=16 [1, 2, 3, 4] s AP4,Q [1, 2, 3, 4] + ANP
tree
4,Q [1, 2, 3, 4]
%
+ A4,Q [3, 4, 2, 1] + A4,Q [3, 4, 2, 1] + cyclic(2, 3, 4) .
P NP
(13.59)
Now let us see what (13.59) says about the UV divergence of supergravity. Consider N = 4
supergravity, hence Q = 0 for AP4,Q and ANP 4,Q . In 4d Yang–Mills theory, no counterterm
operators are needed to regularize the UV divergence of the 4-point 2-loop amplitude, so
the coefficient of the divergence must be generated by the F 2 operator of the classical
action. In particular, the 2-loop UV divergence must have tree-level color factors. In 5d,
dimension-counting shows that F 3 is the only allowed counterterm; it again only has tree-
level color factors.3 This is because the divergence in 5d is renormalized by a tree diagram
with one F 3 counterterm insertion,
3
F . (13.60)
With the intention of using this information, we expand the color-structure cP and cNP in
(13.59) into a basis of color factors that are independent under the Jacobi relation. Such
a basis is given in Appendix B of [217], and it consists of two tree, one 1-loop, and two
2-loop color factors. So casting (13.59) into the said basis, we know that the coefficients
of the 1- and 2-loop color factors have to vanish for the UV-divergent part, since in 4d and
5d only the tree color-structure is generated. The requirement of vanishing 2-loop color
factors is [244]
&
0 = t APQ [1, 3, 4, 2] + APQ [1, 4, 2, 3] + APQ [3, 1, 4, 2] + APQ [3, 2, 1, 4]

+ ANP Q [1, 3, 4, 2] + A Q [1, 4, 2, 3] + A Q [3, 1, 4, 2] + A Q [3, 2, 1, 4]
NP NP NP
%
+ s A Q [1, 3, 4, 2] + A Q [3, 1, 4, 2] + A Q [1, 3, 4, 2] + A Q [3, 1, 4, 2]
P P NP NP
,
D=4,5 div.
&
0 = s APQ [1, 2, 3, 4] + APQ [1, 3, 4, 2] + APQ [3, 1, 4, 2] + APQ [3, 4, 2, 1] (13.61)

+ ANP
Q [1, 2, 3, 4] + A Q [1, 3, 4, 2] + A Q [3, 1, 4, 2] + A Q [3, 4, 2, 1]
NP NP NP
%
+ t A Q [1, 3, 4, 2] + A Q [3, 1, 4, 2] + A Q [1, 3, 4, 2] + A Q [3, 1, 4, 2]
P P NP NP
.
D=4,5 div.
3 “Tree-level color factor” denotes a contraction of fãbc s such that no closed loops are formed. For example
fãbc f˜cde is a tree color factor for a 4-point amplitude. On the other hand, fãbc f˜cde f˜e f g f˜gha is a 4-point
“1-loop color factor.”
13.5 Extensions 293
Substituting the above into (13.59), one finds that

(2)
M4,16 [1, 2, 3, 4] = 0. (13.62)
D=4,5 div.
So the 2-loop 4-point amplitude in pure N ≥ 4 supergravity is finite in both 4d and 5d.
In 4d, the 2-loop R 3 operator can be ruled out by supersymmetry, as we have seen in
Section 12.5 from the spinor helicity violating amplitude it generates. This is of course
perfectly compatible with the UV finiteness of the 2-loop amplitude (13.62).
However, in 5d the relevant operator at 2-loop order is R 4 , and it is compatible with
supersymmetry. It has been argued to be duality invariant [248, 270]. So this is an explicit
5d example where a counterterm appears to respect all known symmetries of the theory,
yet is not generated because the corresponding 2-loop 4-point amplitude is UV finite.
In 4d, it has been argued [235] that the 3-loop candidate counterterm operator R 4 is
allowed by all symmetries of N = 4 supergravity. (It was ruled out in N = 8 supergravity
by E 7(7) .) Yet, by explicit computation, utilizing color-kinematic duality, it has been shown
that N = 4 supergravity is actually finite at 3-loop order [245]. So what is going on? The
absence of a divergence could be coincidence. Or it could indicate that there is a hidden
symmetry that is violated by the would-be-counterterm [271]. The issue is not settled at the
time of finishing this book.
13.5 Extensions
We end this section with brief mention of other applications of the BCJ color-kinematics du-
ality. Color-kinematic duality has been extended to scattering amplitudes involving higher-
dimension operators [272] and also to form factors [273].
In Section 11.3.5, we encountered the Lie 3-algebra for 3d Chern–Simons matter theory
(BLG): it involved 4-index structure constants that, in place of the usual Jacobi identity,
satisfy a 4-term fundamental identity (11.49). Surprisingly, color-kinematic duality can
also be established for such 3-algebra theories [274] with the basic diagrams built from
4-point vertices only. Which supergravity amplitude is calculated by the BCJ double-copy
of duality-satisfying numerators from the 3-algebra Chern–Simons matter theory? At first
sight it seems that the answer has to be different from the supergravity amplitude obtained
from “squaring” 3d Yang–Mills amplitudes, because diagrams built from quartic vertices
must have an even number of external legs n, while trivalent diagrams can have even or
odd n. But one can use the double-copy based on 3d maximal (N = 8) super Yang–Mills
theory (or the KLT formula) to show that the odd-n supergravity amplitudes vanish in 3d,
even though the odd-n Yang–Mills amplitudes are non-vanishing. It has in fact been shown
that applying the double-copy trick to 3d N = 8 super Yang–Mills and 3d BLG theory
remarkably results in the same supergravity tree amplitudes up to 12 points [274, 275].
By dimensional analysis, 3d gravity is non-renormalizable. It is curious that 3d super-
gravity amplitudes can be constructed from two distinct color-kinematic dualities. Perhaps
this puts constraints on the UV behavior of supergravity in 3d.
14 Further reading
In this chapter, we list references to other reviews and we highlight a few subjects that were
not covered in the main text.
Reviews on on-shell methods for scattering amplitudes

• Introduction to on-shell methods
The QFT textbooks by Srednicki [2], Zee [276], and Schwartz [277] provide brief
introductions to the spinor helicity formalism and on-shell recursion relations. In
addition, the book [278] offers a comprehensive introduction to on-shell recursion and
to loop-integrals and also covers some aspects of scattering amplitudes in N = 4 SYM.
The following reviews cover various aspects of on-shell methods (the most recent
reviews are listed first):
“A brief introduction to modern amplitude methods” [11].
“Scattering amplitudes: the most perfect microscopic structures in the universe” [4].
“Hidden simplicity of gauge theory amplitudes” [52].
“A first course on twistors, integrability and gluon scattering amplitudes” [279].
“On-shell methods in perturbative QCD” [280].
“Calculating scattering amplitudes efficiently” [3].
“Multiparton amplitudes in gauge theories” [1].
• Numeric methods and applications in phenomenology
“Susy theories and QCD: Numerical approaches” [70].
“One-loop calculations in quantum field theory: from Feynman diagrams to unitarity
cuts” [77].
“Simplifying multi-jet QCD computation” [281].
“Loop amplitudes in gauge theories: Modern analytic approaches” [71].
• Gravity
“Introduction to the effective field theory description of gravity” [282].
“Perturbative quantum gravity and its relation to gauge theory” [188].
“Ultraviolet behavior of N = 8 supergravity” [283].
Less supersymmetry: 1 ≤ N < 4

An obvious arena for on-shell methods is theories with non-maximal supersymmetry.
Scattering amplitudes in N = 1 supersymmetric theories have been of particular interest
in particle phenomenology and on-shell methods were used in such studies [48, 67, 284–
286]. The superamplitude and on-shell superspace formalism generalizes to 1 ≤ N < 4
SYM [87] and has also been used for SYM coupled with matter [85, 287]. It is expected
14 Further reading 295
that the on-shell diagram approach to planar amplitudes of N = 4 SYM has a natural
extension to N < 4 theories [120].
Amplitudes with massive particles

We have focused on amplitudes with massless particles, but there are also efficient on-shell
methods available for amplitudes involving massive particles. The spinor helicity formalism
for massless particles, introduced in Section 2, can be generalized to massive particles. There
are two approaches to this. In the first, one studies the eigenvectors of the momentum matrix
μ
pαβ̇ = pμ σαβ̇ to directly get solutions to the massive Dirac equation. In the second approach,
the time-like momentum pi is decomposed along two light-like directions by introducing
μ μ m2
a null reference vector qi for each state: pi = pi⊥ − 2qi ·i pi q μ . The familiar spinor helicity
formalism can then be used for qi = −|qi [qi | and pi⊥ = −|i ⊥ [i ⊥ |. Helicity is only a
Lorentz-invariant quantity for massless particles, but qi breaks Lorentz-invariance and can
therefore be used to define a helicity basis in which we can calculate helicity amplitudes.
For an introduction to both approaches, see [288]. The papers [135, 289, 290] studied
applications of BCFW recursion relations with massive particles. A recent discussion of
the latter approach, as well as applications to CSW-like recursion relations, was given
in [26]. Finally, let us mention that there are simple amplitudes with a pair of massive
particles. Examples of such “towers” of amplitudes – in a sense massive versions of Parke–
Taylor – were presented for scattering processes on the Coulomb branch of N = 4 SYM
in [140, 203] (see also [291–293]).
Extensions of recursion relations

Attempts have been made to generalize various forms of recursion relations beyond gauge
theory and gravity, for example to string theory amplitudes [294, 295] and to non-linear
sigma models [6, 296]. Recursion relations can also be utilized to obtain rational functions
that appear at 1-loop [297]. A review of various recursion relations is given in [298].
Triality: Wilson loop, correlation function, amplitude

In Section 5.3 we discussed the emergence of dual superconformal symmetry in planar
N = 4 SYM. It states that the scattering amplitude in the dual coordinates is supercon-
formal covariant. Could we define N = 4 SYM directly in these new coordinates? In dual
coordinates, the kinematic setup for the on-shell momenta is a polygon with null edges.
A similar physical quantity is a Wilson-loop specialized to a null-polygon contour. So
could the amplitude in momentum space be dual to a null-polygon Wilson-loop in the
dual space? Indeed it is. This duality has been established at strong coupling by Alday and
Maldacena [107] as well as at weak coupling by Drummond, Korchemsky, and Sokatchev
[299].
At strong coupling, the duality can be understood as a consequence of T-duality in
string theory [300, 301]. At weak coupling, evidence for such duality was first reported by
[299, 302]. It was later proven by defining the action of N = 4 SYM directly in supertwistor
space [303]. Remarkably, using the duality, the first computation of the 6-point 2-loop MHV
amplitude was done by computing the 6-edged Wilson loop [110]. The amplitude/Wilson-
loop duality was first established between the bosonic Wilson-loop and the ratio of the
296 Further reading
MHV scattering amplitude, divided by the MHV tree amplitude. It can be generalized to
N K MHV amplitudes by supersymmetrizing the bosonic Wilson-loop [303–305].
Another extension is the realization that the super Wilson-loop in N = 4 SYM is also
dual to the correlation function of operators with light-like separation [306]. This can also be
proven in supertwistor space [307]. For a review on the (MHV)amplitude/(bosonic)Wilson-
loop duality see [308–310], for the general amplitude/super-Wilson loop duality in the
framework of supertwistor space, see [311].
Based on an operator-product-expansion approach first developed in the perturba-
tive computation of null Wilson-loops [312], a non-perturbative formulation of the S-
matrix/Wilson-loop for planar N = 4 SYM has been proposed in [313–315]. Its perturba-
tive expansion gives important predictions for the explicit loop amplitude, and as a result
the integrated functional form of the 6-point 3- and 4-loop MHV amplitude for planar
N = 4 SYM [316, 317] was found.
Twistors
Standard reviews of twistor space include [318–320]. For amplitude-friendly reviews, we
suggest [321] as well as [54, 60].
What’s next?
We hope you have found this book useful. We look forward to seeing what the future will
bring for studies of scattering amplitudes.
Appendix Conventions for 4d spinor
helicity formalism
The conventions of these notes follow those in Srednicki’s QFT textbook [2].
We use a “mostly-plus” metric, ημν = diag(−1, +1, +1, +1) and define
(σ μ )a ḃ = (1, σ i )a ḃ , (σ̄ μ )ȧb = (1, −σ i )ȧb , (A.1)
with Pauli matrices

0 1 0 −i 1 0
σ1 = , σ2 = , σ3 = . (A.2)
1 0 i 0 0 −1
Two-index spinor indices are raised/lowered using

0 1
εab = εȧ ḃ = = −εab = −εȧ ḃ , (A.3)
−1 0
which obey εab εbc = δa c .

We list the following properties
(σ̄ μ )ȧa = εab εȧ ḃ (σ μ )bḃ , (A.4)

(σ μ )a ȧ (σμ )bḃ = −2εab εȧ ḃ , (A.5)

μ ν
σ σ̄ + σ ν σ̄ μ a b = −2ημν δa b , (A.6)
Tr (σ μ σ̄ ν ) = Tr (σ̄ μ σ ν ) = −2ημν . (A.7)
Define γ -matrices:

0 (σ μ )a ḃ
γμ = μ ȧb , {γ μ , γ ν } = −2ημν , (A.8)
(σ̄ ) 0
and

−1 0 1 1
γ5 ≡ iγ 0 γ 1 γ 2 γ 3 = , PL = (1 − γ5 ) , PR = (1 + γ5 ) . (A.9)
0 1 2 2
For a momentum 4-vector p μ = ( p 0 , pi ) = (E, pi ) with p μ pμ = −m 2 , we define momen-
tum bi-spinors
pa ḃ ≡ pμ (σ μ )a ḃ , pȧb ≡ pμ (σ̄ μ )ȧb . (A.10)
For example,
0
− p + p3 p1 − i p2
pa ḃ = . (A.11)
p1 + i p2 − p0 − p3
298 Conventions for 4d spinor helicity formalism
Taking the determinant of this 2×2 matrix gives

det p = − p μ pμ = m 2 . (A.12)
The Dirac conjugate is defined as

0 δ ȧ ḃ
= †β , β= . (A.13)
δa b 0
The 4 × 4 matrix β is the same as γ 0 but has a different index structure.
For convenience, we collect here some useful spinor helicity identities
[ p|a = ab | p]b , | p]a = ab [ p|b ,
| pȧ = ȧ ḃ p|ḃ p|ȧ = ȧ ḃ | pḃ ,
pa ḃ = −| p]a p|ḃ , pȧb = −| pȧ [ p|b ,
p q = p|ȧ |qȧ , [ p q] = [ p|a |q]a ,
(A.14)
p q [ p q] = 2 p · q = ( p + q)2 ,
[k|γ μ | p = p|γ μ |k] , [k|γ μ | p∗ = [ p|γ μ |k for real momenta ,
p|P|k] = p|ȧ P ȧb |k]b , p|y1 .y2 |k = p|ȧ (y1 )ȧb (y2 )bċ |kċ ,
p|q|k] = − pq[qk] , 1|γ μ |2]3|γμ |4] = 213[24] .
We also use the analytic continuation
| − p = −| p , | − p] = +| p] . (A.15)
The angle and square spinor are related by complex conjugation:

∗
∗
[ p|a = | pȧ , p|ȧ = | p]a , (A.16)
μ
if the momentum p is real.
These identities are used in multiple places in the text and exercises.
References
[1] M. L. Mangano and S. J. Parke, “Multiparton amplitudes in gauge theories,” Phys.

Rept. 200, 301 (1991) [hep-th/0509223].
[2] M. Srednicki, Quantum Field Theory, Cambridge, UK: Cambridge University Press
(2007).
[3] L. J. Dixon, “Calculating scattering amplitudes efficiently,” In Boulder 1995, QCD
and beyond, 539–582 [hep-ph/9601359].
[4] L. J. Dixon, “Scattering amplitudes: the most perfect microscopic structures in the
universe,” J. Phys. A 44, 454001 (2011) [arXiv:1105.0771 [hep-th]].
[5] S. J. Parke and T. R. Taylor, “An amplitude for n gluon scattering,” Phys. Rev. Lett.
56, 2459 (1986).
[6] K. Kampf, J. Novotny, and J. Trnka, “Tree-level amplitudes in the nonlinear sigma
model,” JHEP 1305, 032 (2013) [arXiv:1304.3048 [hep-th]].
[7] R. Kleiss and H. Kuijf, “Multi-gluon cross-sections and five jet production at hadron
colliders,” Nucl. Phys. B 312, 616 (1989).
[8] V. Del Duca, L. J. Dixon, and F. Maltoni, “New color decompositions for gauge
amplitudes at tree and loop level,” Nucl. Phys. B 571, 51 (2000) [hep-ph/9910563].
[9] Z. Bern, J. J. M. Carrasco, and H. Johansson, “New relations for gauge-theory
amplitudes,” Phys. Rev. D 78, 085011 (2008) [arXiv:0805.3993 [hep-ph]].
[10] C. F. Berger, V. Del Duca, and L. J. Dixon, “Recursive construction of Higgs-plus-
multiparton loop amplitudes: The last of the phi-nite loop amplitudes,” Phys. Rev. D
74, 094021 (2006) [Erratum-ibid. D 76, 099901 (2007)] [hep-ph/0608180];
S. D. Badger, E. W. N. Glover, and K. Risager, “One-loop phi-MHV amplitudes
using the unitarity bootstrap,” JHEP 0707, 066 (2007) [arXiv:0704.3914 [hep-ph]];
L. J. Dixon and Y. Sofianatos, “Analytic one-loop amplitudes for a Higgs boson plus
four partons,” JHEP 0908, 058 (2009) [arXiv:0906.0008 [hep-ph]];
S. Badger, E. W. Nigel Glover, P. Mastrolia, and C. Williams, “One-loop Higgs
plus four gluon amplitudes: Full analytic results,” JHEP 1001, 036 (2010)
[arXiv:0909.4475 [hep-ph]].
[11] L. J. Dixon, “A brief introduction to modern amplitude methods,” [arXiv:1310.5353
[hep-ph]].
[12] F. A. Berends and W. T. Giele, “Recursive calculations for processes with n gluons,”
Nucl. Phys. B 306, 759 (1988).
F. A. Berends, W. T. Giele, and H. Kuijf, “Exact and approximate expressions for
multi-gluon scattering,” Nucl. Phys. B 333, 120 (1990).
[13] R. Britto, F. Cachazo, and B. Feng, “New recursion relations for tree amplitudes of
gluons,” Nucl. Phys. B 715, 499 (2005) [hep-th/0412308].
300 References
[14] R. Britto, F. Cachazo, B. Feng, and E. Witten, “Direct proof of tree-level recursion
relation in Yang–Mills theory,” Phys. Rev. Lett. 94, 181602 (2005) [hep-th/0501052].
[15] F. Cachazo, P. Svrcek, and E. Witten, “MHV vertices and tree amplitudes in gauge
theory,” JHEP 0409, 006 (2004) [hep-th/0403047].
[16] B. Feng, J. Wang, Y. Wang, and Z. Zhang, “BCFW recursion relation with nonzero
boundary contribution,” JHEP 1001, 019 (2010) [arXiv:0911.0301 [hep-th]].
[17] E. Conde and S. Rajabi, “The twelve-graviton next-to-MHV amplitude from Ris-
ager’s construction,” JHEP 1209, 120 (2012) [arXiv:1205.3500 [hep-th]].
[18] N. Arkani-Hamed and J. Kaplan, “On tree amplitudes in gauge theory and gravity,”
JHEP 0804, 076 (2008) [arXiv:0801.2385 [hep-th]].
[19] C. Cheung, “On-shell recursion relations for generic theories,” JHEP 1003, 098
(2010) [arXiv:0808.0504 [hep-th]].
[20] P. Benincasa, C. Boucher-Veronneau, and F. Cachazo, “Taming tree amplitudes in
general relativity,” JHEP 0711, 057 (2007) [hep-th/0702032 [HEP-TH]].
[21] H. Kawai, D. C. Lewellen, and S. H. H. Tye, “A relation between tree amplitudes of
closed and open strings,” Nucl. Phys. B 269, 1 (1986).
[22] S. Sannan, “Gravity as the limit of the type II superstring theory,” Phys. Rev. D 34,
1749 (1986).
[23] N. Arkani-Hamed, F. Cachazo, C. Cheung, and J. Kaplan, “A duality for the S matrix,”
JHEP 1003, 020 (2010) [arXiv:0907.5418 [hep-th]].
[24] A. Hodges, “Eliminating spurious poles from gauge-theoretic amplitudes,” JHEP
1305, 135 (2013) [arXiv:0905.1473 [hep-th]].
[25] M. Spradlin, A. Volovich, and C. Wen, “Three applications of a bonus relation for
gravity amplitudes,” Phys. Lett. B 674, 69 (2009) [arXiv:0812.4767 [hep-th]].
[26] T. Cohen, H. Elvang, and M. Kiermaier, “On-shell constructibility of tree amplitudes
in general field theories,” JHEP 1104, 053 (2011) [arXiv:1010.0257 [hep-th]].
[27] H. Elvang, D. Z. Freedman, and M. Kiermaier, “Recursion relations, generating
functions, and unitarity sums in N = 4 SYM theory,” JHEP 0904, 009 (2009)
[arXiv:0808.1720 [hep-th]].
[28] K. Risager, “A direct proof of the CSW rules,” JHEP 0512, 003 (2005) [hep-th/
0508206].
[29] H. Elvang, D. Z. Freedman, and M. Kiermaier, “Proof of the MHV vertex ex-
pansion for all tree amplitudes in N = 4 SYM theory,” JHEP 0906, 068 (2009)
[arXiv:0811.3624 [hep-th]].
[30] L. J. Dixon, E. W. N. Glover, and V. V. Khoze, “MHV rules for Higgs plus multi-gluon
amplitudes,” JHEP 0412, 015 (2004) [hep-th/0411092].
[31] S. D. Badger, E. W. N. Glover, and V. V. Khoze, “MHV rules for Higgs plus multi-
parton amplitudes,” JHEP 0503, 023 (2005) [hep-th/0412275].
[32] A. Brandhuber, B. Spence, and G. Travaglini, “Tree-level formalism,” J. Phys. A 44,
454002 (2011) [arXiv:1103.3477 [hep-th]].
[33] A. Gorsky and A. Rosly, “From Yang–Mills Lagrangian to MHV diagrams,” JHEP
0601, 101 (2006) [hep-th/0510111];
P. Mansfield, “The Lagrangian origin of MHV rules,” JHEP 0603, 037 (2006) [hep-th/
0511264];
References 301
J. H. Ettle and T. R. Morris, “Structure of the MHV-rules Lagrangian,” JHEP 0608,

003 (2006) [hep-th/0605121];
J. H. Ettle, C.-H. Fu, J. P. Fudger, P. R. W. Mansfield, and T. R. Morris, “S-matrix
equivalence theorem evasion and dimensional regularisation with the canonical MHV
Lagrangian,” JHEP 0705, 011 (2007) [hep-th/0703286];
H. Feng and Y.-t. Huang, “MHV Lagrangian for N = 4 super Yang–Mills,” JHEP
0904, 047 (2009) [hep-th/0611164].
[34] R. Boels, L. J. Mason, and D. Skinner, “From twistor actions to MHV diagrams,”
Phys. Lett. B 648, 90 (2007) [hep-th/0702035].
[35] N. E. J. Bjerrum-Bohr, D. C. Dunbar, H. Ita, W. B. Perkins, and K. Risager, “MHV-
vertices for gravity amplitudes,” JHEP 0601, 009 (2006) [hep-th/0509016].
[36] M. Bianchi, H. Elvang, and D. Z. Freedman, “Generating tree amplitudes in N = 4
SYM and N = 8 SG,” JHEP 0809, 063 (2008) [arXiv:0805.0757 [hep-th]].
[37] J. Wess and J. Bagger, Supersymmetry and Supergravity, Princeton, USA: University
Press (1992).
[38] M. T. Grisaru and H. N. Pendleton, “Some properties of scattering amplitudes in
supersymmetric theories,” Nucl. Phys. B 124, 81 (1977);
M. T. Grisaru, H. N. Pendleton, and P. van Nieuwenhuizen, “Supergravity and the S
matrix,” Phys. Rev. D 15, 996 (1977).
[39] D. Z. Freedman and A. Van Proeyen, Supergravity, Cambridge, UK: Cambridge
University Press (2012).
[40] L. Brink, J. H. Schwarz, and J. Scherk, “Supersymmetric Yang–Mills theories,” Nucl.
Phys. B 121, 77 (1977).
[41] A. Ferber, “Supertwistors and conformal supersymmetry,” Nucl. Phys. B 132, 55
(1978).
[42] H. Elvang, D. Z. Freedman, and M. Kiermaier, “Solution to the Ward identities for
superamplitudes,” JHEP 1010, 103 (2010) [arXiv:0911.3169 [hep-th]].
[43] H. Elvang, D. Z. Freedman, and M. Kiermaier, “SUSY Ward identities, super-
amplitudes, and counterterms,” J. Phys. A 44, 454009 (2011) [arXiv:1012.3401
[hep-th]].
[44] M. Kiermaier and S. G. Naculich, “A super MHV vertex expansion for N = 4 SYM
theory,” JHEP 0905, 072 (2009) [arXiv:0903.0377 [hep-th]].
[45] N. Arkani-Hamed, “What is the simplest QFT?,” talk given at the Paris Workshop
Wonders of Gauge Theory and Supergravity, 24 June 2008.
[46] A. Brandhuber, P. Heslop, and G. Travaglini, “A note on dual superconformal sym-
metry of the N = 4 super Yang–Mills S-matrix,” Phys. Rev. D 78, 125005 (2008)
[arXiv:0807.4097 [hep-th]].
[47] N. Arkani-Hamed, F. Cachazo, and J. Kaplan, “What is the simplest quantum field
theory?,” JHEP 1009, 016 (2010) [arXiv:0808.1446 [hep-th]].
[48] Z. Bern, J. J. M. Carrasco, H. Ita, H. Johansson, and R. Roiban, “On the structure of
supersymmetric sums in multi-loop unitarity cuts,” Phys. Rev. D 80, 065029 (2009)
[arXiv:0903.5348 [hep-th]].
[49] J. M. Drummond and J. M. Henn, “All tree-level amplitudes in N = 4 SYM,” JHEP
0904, 018 (2009) [arXiv:0808.2475 [hep-th]].
302 References
[50] J. M. Drummond, J. Henn, G. P. Korchemsky, and E. Sokatchev, “Dual superconfor-

mal symmetry of scattering amplitudes in N = 4 super-Yang–Mills theory,” Nucl.
Phys. B 828, 317 (2010) [arXiv:0807.1095 [hep-th]].
[51] J. M. Drummond, M. Spradlin, A. Volovich, and C. Wen, “Tree-level amplitudes in
N = 8 supergravity,” Phys. Rev. D 79, 105018 (2009) [arXiv:0901.2363 [hep-th]].
[52] J. M. Drummond, “Hidden simplicity of gauge theory amplitudes,” Class. Quant.
Grav. 27, 214001 (2010) [arXiv:1010.2418 [hep-th]].
[53] R. Penrose, “Twistor algebra,” J. Math. Phys. 8, 345 (1967).
[54] E. Witten, “Perturbative gauge theory as a string theory in twistor space,” Commun.
Math. Phys. 252, 189 (2004) [hep-th/0312171].
[55] P. A. M. Dirac, “Wave equations in conformal space,” Annals Math. 37, 429 (1936).
[56] W. Siegel, “Embedding versus 6D twistors,” [arXiv:1204.5679 [hep-th]].
[57] W. Siegel, “Fields,” [hep-th/9912205].
[58] J. M. Drummond, J. Henn, V. A. Smirnov, and E. Sokatchev, “Magic identities for
conformal four-point integrals,” JHEP 0701, 064 (2007) [hep-th/0607160].
[59] J. M. Drummond, J. M. Henn, and J. Plefka, “Yangian symmetry of scattering ampli-
tudes in N = 4 super Yang–Mills theory,” JHEP 0905, 046 (2009) [arXiv:0902.2987
[hep-th]].
[60] L. J. Mason and D. Skinner, “Dual superconformal invariance, momentum twistors
and grassmannians,” JHEP 0911, 045 (2009) [arXiv:0909.0250 [hep-th]].
N. Arkani-Hamed, F. Cachazo, and C. Cheung, “The Grassmannian origin of dual
superconformal invariance,” JHEP 10033, 036 (2010) [arXiv: 0909.0483 [hep-th]].
H. Elvang, Y.-t. Huang, C. Keeler, et al., “Grassmannians for scattering amplitudes
in 4d N = 4 SYM and 3d ABJM,” 2014 [arXiv: 1410.0621 [hep.th]].
[61] R. Roiban, “Review of AdS/CFT integrability, Chapter V.1: Scattering amplitudes –
a brief introduction,” Lett. Math. Phys. 99, 455 (2012) [arXiv:1012.4001 [hep-th]].
[62] Lorenzo Magnea, Lecture notes on Perturbative QCD at the National School of
Theoretical Physics of the University of Parma (2008). http://personalpages.to.infn.
it/∼magnea/QCD.pdf
[63] L. V. Bork, D. I. Kazakov, G. S. Vartanov, and A. V. Zhiboedov, “Construction of
infrared finite observables in N = 4 super Yang–Mills theory,” Phys. Rev. D 81,
105028 (2010) [arXiv:0911.1617 [hep-th]].
[64] Z. Bern, G. Chalmers, L. J. Dixon, and D. A. Kosower, “One loop N gluon amplitudes
with maximal helicity violation via collinear limits,” Phys. Rev. Lett. 72, 2134 (1994)
[hep-ph/9312333].
[65] G. Mahlon, “Multi-gluon helicity amplitudes involving a quark loop,” Phys. Rev. D
49, 4438 (1994) [hep-ph/9312276].
[66] Z. Bern, L. J. Dixon, and D. A. Kosower, “Dimensionally regulated pentagon inte-
grals,” Nucl. Phys. B 412, 751 (1994) [hep-ph/9306240].
[67] Z. Bern, L. J. Dixon, D. C. Dunbar, and D. A. Kosower, “One loop n point gauge
theory amplitudes, unitarity and collinear limits,” Nucl. Phys. B 425, 217 (1994)
[hep-ph/9403226];
Z. Bern, L. J. Dixon, D. C. Dunbar, and D. A. Kosower, “Fusing gauge theory tree
amplitudes into loop amplitudes,” Nucl. Phys. B 435, 59 (1995) [hep-ph/9409265].
References 303
[68] J. J. M. Carrasco and H. Johansson, “Generic multiloop methods and application to

N = 4 super-Yang–Mills,” J. Phys. A 44, 454004 (2011) [arXiv:1103.3298 [hep-th]].
[69] Z. Bern and Y.-t. Huang, “Basics of generalized unitarity,” J. Phys. A 44, 454003
(2011) [arXiv:1103.1869 [hep-th]].
[70] H. Ita, “Susy theories and QCD: Numerical approaches,” J. Phys. A 44, 454005
(2011) [arXiv:1109.6527 [hep-th]].
[71] R. Britto, “Loop amplitudes in gauge theories: Modern analytic approaches,” J. Phys.
A 44, 454006 (2011) [arXiv:1012.4493 [hep-th]].
[72] W. L. van Neerven and J. A. M. Vermaseren, “Large loop integrals,” Phys. Lett. B
137, 241 (1984).
[73] Z. Bern, L. J. Dixon, and D. A. Kosower, “Dimensionally regulated one loop inte-
grals,” Phys. Lett. B 302, 299 (1993) [Erratum-ibid. B 318, 649 (1993)] [hep-ph/
9212308].
[74] Z. Bern, L. J. Dixon, and D. A. Kosower, “Dimensionally regulated pentagon inte-
grals,” Nucl. Phys. B 412, 751 (1994) [hep-ph/9306240].
[75] L. M. Brown and R. P. Feynman, “Radiative corrections to Compton scattering,”
Phys. Rev. 85, 231 (1952);
G. ’t Hooft and M. J. G. Veltman, “Scalar one loop integrals,” Nucl. Phys. B 153, 365
(1979).
[76] G. Passarino and M. J. G. Veltman, “One loop corrections for e+ e− annihilation
into mu+ mu− in the Weinberg model,” Nucl. Phys. B 160, 151 (1979).
[77] R. K. Ellis, Z. Kunszt, K. Melnikov, and G. Zanderighi, “One-loop calculations in
quantum field theory: from Feynman diagrams to unitarity cuts,” Phys. Rept. 518,
141 (2012) [arXiv:1105.4319 [hep-ph]].
[78] H. Johansson, D. A. Kosower, and K. J. Larsen, “An overview of maximal unitarity
at two loops,” PoS LL 2012, 066 (2012) [arXiv:1212.2132 [hep-th]].
[79] C. Anastasiou, R. Britto, B. Feng, Z. Kunszt, and P. Mastrolia, “D-dimensional
unitarity cut method,” Phys. Lett. B 645, 213 (2007) [hep-ph/0609191];
R. Britto and B. Feng, “Integral coefficients for one-loop amplitudes,” JHEP 0802,
095 (2008) [0711.4284 [hep-ph]];
R. Britto and B. Feng, “Unitarity cuts with massive propagators and alge-
braic expressions for coefficients,” Phys. Rev. D 75, 105006 (2007) [hep-ph/
0612089];
G. Ossola, C. G. Papadopoulos, and R. Pittau, “Reducing full one-loop amplitudes
to scalar integrals at the integrand level,” Nucl. Phys. B 763, 147 (2007) [hep-ph/
0609007];
R. Britto, B. Feng, and P. Mastrolia, “Closed-form decomposition of one-loop mas-
sive amplitudes,” Phys. Rev. D 78, 025031 (2008) [0803.1989 [hep-ph]];
D. Forde, “Direct extraction of one-loop integral coefficients,” Phys. Rev. D 75,
125019 (2007) [0704.1835 [hep-ph]];
[80] Z. Bern and A. G. Morgan, “Massive loop amplitudes from unitarity,” Nucl. Phys. B
467, 479 (1996) [hep-ph/9511336];
Z. Bern, L. J. Dixon, and D. A. Kosower, “Progress in one-loop QCD computations,”
Ann. Rev. Nucl. Part. Sci. 46, 109 (1996) [hep-ph/9602280].
304 References
[81] S. D. Badger, “Direct extraction of one loop rational terms,” JHEP 0901, 049 (2009)
[0806.4600 [hep-ph]].
[82] M. B. Green, J. H. Schwarz, and L. Brink, “N = 4 Yang–Mills and N = 8 supergravity
as limits of string theories,” Nucl. Phys. B 198, 474 (1982).
[83] Z. Bern, N. E. J. Bjerrum-Bohr, and D. C. Dunbar, “Inherited twistor-space structure
of gravity loop amplitudes,” JHEP 0505, 056 (2005) [hep-th/0501137];
N. E. J. Bjerrum-Bohr, D. C. Dunbar, H. Ita, W. B. Perkins, and K. Risager, “The
no-triangle hypothesis for N = 8 supergravity,” JHEP 0612, 072 (2006) [hep-
th/0610043];
N. E. J. Bjerrum-Bohr and P. Vanhove, “Absence of triangles in maximal supergravity
amplitudes,” JHEP 0810, 006 (2008) [arXiv:0805.3682 [hep-th]].
[84] Z. Bern, J. J. Carrasco, D. Forde, H. Ita, and H. Johansson, “Unexpected cancellations
in gravity theories,” Phys. Rev. D 77, 025010 (2008) [arXiv:0707.1035 [hep-th]].
[85] S. Lal and S. Raju, “The next-to-simplest quantum field theories,” Phys. Rev. D 81,
105002 (2010) [arXiv:0910.0930 [hep-th]].
[86] D. C. Dunbar, J. H. Ettle, and W. B. Perkins, “Perturbative expansion of N < 8
supergravity,” Phys. Rev. D 83, 065015 (2011) [arXiv:1011.5378 [hep-th]].
[87] H. Elvang, Y.-t. Huang, and C. Peng, “On-shell superamplitudes in N < 4 SYM,”
JHEP 1109, 031 (2011) [arXiv:1102.4843 [hep-th]].
[88] Y.-t. Huang, D. A. McGady, and C. Peng, “One-loop renormalization and the S-
matrix,” [arXiv:1205.5606 [hep-th]].
[89] N. Marcus, “Composite anomalies in supergravity,” Phys. Lett. B 157, 383 (1985).
[90] P. di Vecchia, S. Ferrara, and L. Girardello, “Anomalies of hidden local chiral
symmetries in sigma models and extended supergravities,” Phys. Lett. B 151, 199
(1985).
[91] J. M. Drummond, J. Henn, G. P. Korchemsky, and E. Sokatchev, “Generalized uni-
tarity for N = 4 super-amplitudes,” Nucl. Phys. B 869, 452 (2013) [arXiv:0808.0491
[hep-th]].
[92] A. Brandhuber, P. Heslop, and G. Travaglini, “One-loop amplitudes in N = 4 super
Yang–Mills and anomalous dual conformal symmetry,” JHEP 0908, 095 (2009)
[arXiv:0905.4377 [hep-th]].
[93] H. Elvang, D. Z. Freedman, and M. Kiermaier, “Dual conformal symmetry of 1-loop
NMHV amplitudes in N = 4 SYM theory,” JHEP 1003, 075 (2010) [arXiv:0905.4379
[hep-th]].
[94] G. P. Korchemsky and E. Sokatchev, “Symmetries and analytic properties of scatter-
ing amplitudes in N = 4 SYM theory,” Nucl. Phys. B 832, 1 (2010) [arXiv:0906.1737
[hep-th]].
[95] J. Gluza, K. Kajda, and D. A. Kosower, “Towards a basis for planar two-loop inte-
grals,” Phys. Rev. D 83, 045012 (2011) [arXiv:1009.0472 [hep-th]].
[96] D. A. Kosower and K. J. Larsen, “Maximal unitarity at two loops,” Phys. Rev. D 85,
045017 (2012) [arXiv:1108.1180 [hep-th]];
H. Johansson, D. A. Kosower, and K. J. Larsen, “Two-loop maximal unitarity with
external masses,” Phys. Rev. D 87, 025030 (2013) [arXiv:1208.1754 [hep-th]].
[97] S. Badger, H. Frellesvig, and Y. Zhang, “Hepta-cuts of two-loop scattering ampli-
tudes,” JHEP 1204, 055 (2012) [arXiv:1202.2019 [hep-ph]].
References 305
[98] Y. Zhang, “Integrand-level reduction of loop amplitudes by computational algebraic

geometry methods,” JHEP 1209, 042 (2012) [arXiv:1205.5707 [hep-ph]].
[99] M. Sgaard, “Global residues and two-loop hepta-cuts,” JHEP 1309, 116 (2013)
[arXiv:1306.1496 [hep-th]].
[100] A. V. Smirnov and A. V. Petukhov, “The number of master integrals is finite,” Lett.
Math. Phys. 97, 37 (2011) [arXiv:1004.4199 [hep-th]].
[101] C. Anastasiou, Z. Bern, L. J. Dixon, and D. A. Kosower, “Planar amplitudes in
maximally supersymmetric Yang–Mills theory,” Phys. Rev. Lett. 91, 251602 (2003)
[hep-th/0309040].
[102] Z. Bern, L. J. Dixon, and V. A. Smirnov, “Iteration of planar amplitudes in maximally
supersymmetric Yang–Mills theory at three loops and beyond,” Phys. Rev. D 72,
085001 (2005) [hep-th/0505205].
[103] Z. Bern, J. S. Rozowsky, and B. Yan, “Two loop four gluon amplitudes in N = 4
superYang–Mills,” Phys. Lett. B 401, 273 (1997) [hep-ph/9702424].
[104] Z. Bern, M. Czakon, D. A. Kosower, R. Roiban, and V. A. Smirnov, “Two-loop
iteration of five-point N = 4 super-Yang–Mills amplitudes,” Phys. Rev. Lett. 97,
181601 (2006) [hep-th/0604074].
[105] F. Cachazo, M. Spradlin, and A. Volovich, “Iterative structure within the five-particle
two-loop amplitude,” Phys. Rev. D 74, 045020 (2006) [hep-th/0602228].
[106] L. F. Alday and J. Maldacena, “Comments on gluon scattering amplitudes via
AdS/CFT,” JHEP 0711, 068 (2007) [arXiv:0710.1060 [hep-th]].
[107] L. F. Alday and J. M. Maldacena, “Gluon scattering amplitudes at strong coupling,”
JHEP 0706, 064 (2007) [arXiv:0705.0303 [hep-th]].
[108] Z. Bern, L. J. Dixon, D. A. Kosower, et al., “The two-loop six-gluon MHV amplitude
in maximally supersymmetric Yang–Mills theory,” Phys. Rev. D 78 (2008) 045007
[arXiv:0803.1465 [hep-th]].
[109] F. Cachazo, M. Spradlin, and A. Volovich, “Leading singularities of the two-loop
six-particle MHV amplitude,” Phys. Rev. D 78, 105022 (2008) [arXiv:0805.4832
[hep-th]].
[110] V. Del Duca, C. Duhr, and V. A. Smirnov, “An analytic result for the two-loop hexagon
Wilson loop in N = 4 SYM,” JHEP 1003, 099 (2010) [arXiv:0911.5332 [hep-ph]].
[111] V. Del Duca, C. Duhr, and V. A. Smirnov, “The two-loop hexagon Wilson loop in
N = 4 SYM,” JHEP 1005, 084 (2010) [arXiv:1003.1702 [hep-th]].
[112] A. B. Goncharov, M. Spradlin, C. Vergu, and A. Volovich, “Classical polylog-
arithms for amplitudes and Wilson loops,” Phys. Rev. Lett. 105, 151605 (2010)
[arXiv:1006.5703 [hep-th]].
[113] E. W. Nigel Glover and C. Williams, “One-loop gluonic amplitudes from single
unitarity cuts,” JHEP 0812, 067 (2008) [0810.2964 [hep-th]];
I. Bierenbaum, S. Catani, P. Draggiotis, and G. Rodrigo, “A tree-loop duality relation
at two loops and beyond,” JHEP 1010, 073 (2010) [arXiv:1007.0194 [hep-ph]];
H. Elvang, D. Z. Freedman, and M. Kiermaier, “Integrands for QCD rational terms
and N = 4 SYM from massive CSW rules,” JHEP 1206, 015 (2012) [arXiv:1111.0635
[hep-th]].
[114] S. Caron-Huot, “Loops and trees,” JHEP 1105, 080 (2011) [arXiv:1007.3224 [hep-
ph]].
306 References
[115] N. Arkani-Hamed, J. L. Bourjaily, F. Cachazo, S. Caron-Huot, and J. Trnka, “The

all-loop integrand for scattering amplitudes in planar N = 4 SYM,” JHEP 1101, 041
(2011) [arXiv:1008.2958 [hep-th]].
[116] R. H. Boels, “On BCFW shifts of integrands and integrals,” JHEP 1011, 113 (2010)
[arXiv:1008.3101 [hep-th]].
[117] Z. Bern, J. J. M. Carrasco, H. Johansson, and D. A. Kosower, “Maximally supersym-
metric planar Yang–Mills amplitudes at five loops,” Phys. Rev. D 76, 125020 (2007)
[0705.1864 [hep-th]].
[118] R. Britto, F. Cachazo, and B. Feng, “Generalized unitarity and one-loop amplitudes
in N = 4 super-Yang–Mills,” Nucl. Phys. B 725, 275 (2005) [hep-th/0412103];
E. I. Buchbinder and F. Cachazo, “Two-loop amplitudes of gluons and octa-cuts in
N = 4 super Yang–Mills,” JHEP 0511, 036 (2005) [hep-th/0506126].
[119] N. Arkani-Hamed, J. L. Bourjaily, F. Cachazo, and J. Trnka, “Local integrals
for planar scattering amplitudes,” JHEP 1206, 125 (2012) [arXiv:1012.6032
[hep-th]].
For extensions, see J. L. Bourjaily, S. Caron-Huot, and J. Trnka, “Dual-conformal
regularization of infrared loop divergences and the chiral box expansion,”
[arXiv:1303.4734 [hep-th]].
[120] N. Arkani-Hamed, J. L. Bourjaily, F. Cachazo, et al., “Scattering amplitudes and the
positive Grassmannian,” [arXiv:1212.5605 [hep-th]].
[121] F. Cachazo, “Sharpening the leading singularity,” [arXiv:0803.1988 [hep-th]].
[122] J. M. Drummond and L. Ferro, “Yangians, Grassmannians and T-duality,” JHEP
1007, 027 (2010) [arXiv:1001.3348 [hep-th]].
[123] J. M. Drummond and L. Ferro, “The Yangian origin of the Grassmannian integral,”
JHEP 1012, 010 (2010) [arXiv:1002.4622 [hep-th]];
G. P. Korchemsky and E. Sokatchev, “Superconformal invariants for scattering am-
plitudes in N = 4 SYM theory,” Nucl. Phys. B 839, 377 (2010) [arXiv:1002.4625
[hep-th]].
[124] R. Roiban, M. Spradlin, and A. Volovich, “On the tree level S matrix of Yang–Mills
theory,” Phys. Rev. D 70, 026009 (2004) [hep-th/0403190].
[125] M. Spradlin and A. Volovich, “From twistor string theory to recursion relations,”
Phys. Rev. D 80, 085022 (2009) [arXiv:0909.0229 [hep-th]].
[126] N. Arkani-Hamed, J. Bourjaily, F. Cachazo, and J. Trnka, “Unification of residues
and Grassmannian dualities,” JHEP 1101, 049 (2011) [arXiv:0912.4912 [hep-th]].
[127] J. L. Bourjaily, J. Trnka, A. Volovich, and C. Wen, “The Grassmannian and the
twistor string: Connecting all trees in N = 4 SYM,” JHEP 1101, 038 (2011)
[arXiv:1006.1899 [hep-th]].
[128] M. Bullimore, L. J. Mason, and D. Skinner, “Twistor-strings, Grassmannians and
leading singularities,” JHEP 1003, 070 (2010) [arXiv:0912.0539 [hep-th]].
[129] L. Dolan and P. Goddard, “Complete equivalence between gluon tree amplitudes in
twistor string theory and in gauge theory,” JHEP 1206, 030 (2012) [arXiv:1111.0950
[hep-th]].
[130] N. Arkani-Hamed, J. L. Bourjaily, F. Cachazo, A. Hodges, and J. Trnka, “A note
on polytopes for scattering amplitudes,” JHEP 1204, 081 (2012) [arXiv:1012.6030
[hep-th]].
References 307
[131] M. Bullimore, L. J. Mason, and D. Skinner, “MHV diagrams in momentum twistor

space,” JHEP 1012, 032 (2010) [arXiv:1009.1854 [hep-th]].
[132] N. Arkani-Hamed and J. Trnka, “The amplituhedron,” [arXiv:1312.2007 [hep-th]];
N. Arkani-Hamed and J. Trnka, “Into the amplituhedron,” [arXiv:1312.7878
[hep-th]].
[133] L. Mason and D. Skinner, “Amplitudes at weak coupling as polytopes in AdS5 ,”
J. Phys. A 44, 135401 (2011) [arXiv:1004.3498 [hep-th]].
[134] H. Nastase and H. J. Schnitzer, “Twistor and polytope interpretations for sublead-
ing color one-loop amplitudes,” Nucl. Phys. B 855, 901 (2012) [arXiv:1104.2752
[hep-th]].
[135] R. Boels, “Covariant representation theory of the Poincaré algebra and some of its
extensions,” JHEP 1001, 010 (2010) [arXiv:0908.0738 [hep-th]].
[136] S. Caron-Huot and D. O’Connell, “Spinor helicity and dual conformal symmetry in
ten dimensions,” JHEP 1108, 014 (2011) [arXiv:1010.5487 [hep-th]].
[137] R. H. Boels and D. O’Connell, “Simple superamplitudes in higher dimensions,”
JHEP 1206, 163 (2012) [arXiv:1201.2653 [hep-th]].
[138] S. Davies, “One-loop QCD and Higgs to partons processes using six-dimensional
helicity and generalized unitarity,” Phys. Rev. D 84, 094016 (2011) [arXiv:1108.0398
[hep-ph]].
[139] Z. Bern, J. J. Carrasco, T. Dennen, Y.-t. Huang, and H. Ita, “Generalized unitarity
and six-dimensional helicity,” Phys. Rev. D 83, 085022 (2011) [arXiv:1010.0494
[hep-th]].
[140] N. Craig, H. Elvang, M. Kiermaier, and T. Slatyer, “Massive amplitudes on the
Coulomb branch of N = 4 SYM,” JHEP 1112, 097 (2011) [arXiv:1104.2050
[hep-th]].
[141] C. Cheung and D. O’Connell, “Amplitudes and spinor-helicity in six dimensions,”
JHEP 0907, 075 (2009) [arXiv:0902.0981 [hep-th]].
[142] T. Dennen, Y.-t. Huang, and W. Siegel, “Supertwistor space for 6D maximal super
Yang–Mills,” JHEP 1004, 127 (2010) [arXiv:0910.2688 [hep-th]].
[143] A. Brandhuber, D. Korres, D. Koschade, and G. Travaglini, “One-loop amplitudes in
six-dimensional (1,1) theories from generalised unitarity,” JHEP 1102, 077 (2011)
[arXiv:1010.1515 [hep-th]];
C. Saemann, R. Wimmer, and M. Wolf, “A twistor description of six-dimensional
N = (1,1) super Yang–Mills theory,” JHEP 1205, 020 (2012) [arXiv:1201.6285
[hep-th]].
[144] T. Dennen and Y.-t. Huang, “Dual conformal properties of six-dimensional maximal
super Yang–Mills amplitudes,” JHEP 1101, 140 (2011) [arXiv:1010.5874 [hep-th]].
[145] T. Chern, “Superconformal field theory in six dimensions and supertwistor,”
[arXiv:0906.0657 [hep-th]];
M. Chiodaroli, M. Gunaydin, and R. Roiban, “Superconformal symmetry, and maxi-
mal supergravity in various dimensions,” JHEP 1203, 093 (2012) [arXiv:1108.3085
[hep-th]];
L. J. Mason, R. A. Reid-Edwards, and A. Taghavi-Chabert, “Conformal field theories
in six-dimensional twistor space,” J. Geom. Phys. 62, 2353 (2012) [arXiv:1111.2585
[hep-th]];
308 References
C. Saemann and M. Wolf, “On twistors and conformal field theories from six dimen-
sions,” J. Math. Phys. 54, 013507 (2013) [arXiv:1111.2539 [hep-th]].
[146] B. Czech, Y.-t. Huang, and M. Rozali, “Amplitudes for multiple M5 branes,” JHEP
1210, 143 (2012) [arXiv:1110.2791 [hep-th]].
[147] C. Saemann and M. Wolf, “Non-Abelian tensor multiplet equations from twistor
space,” [arXiv:1205.3108 [hep-th]].
[148] Y.-t. Huang and A. E. Lipstein, “Amplitudes of 3D and 6D maximal superconformal
theories in supertwistor space,” JHEP 1010, 007 (2010) [arXiv:1004.4735 [hep-th]].
[149] L. F. Alday, J. M. Henn, J. Plefka, and T. Schuster, “Scattering into the fifth dimension
of N = 4 super Yang–Mills,” JHEP 1001, 077 (2010) [arXiv:0908.0684 [hep-th]].
[150] A. Agarwal, N. Beisert, and T. McLoughlin, “Scattering in mass-deformed N ≥ 4
Chern–Simons models,” JHEP 0906, 045 (2009) [arXiv:0812.3367 [hep-th]].
[151] A. Gustavsson, “Algebraic structures on parallel M2-branes,” Nucl. Phys. B 811, 66
(2009) [arXiv:0709.1260 [hep-th]].
[152] J. Bagger and N. Lambert, “Gauge symmetry and supersymmetry of multiple M2-
branes,” Phys. Rev. D 77, 065008 (2008) [arXiv:0711.0955 [hep-th]].
[153] M. A. Bandres, A. E. Lipstein, and J. H. Schwarz, “N = 8 superconformal Chern–
Simons theories,” JHEP 0805, 025 (2008) [arXiv:0803.3242 [hep-th]].
[154] J. Gomis, G. Milanesi, and J. G. Russo, “Bagger-Lambert theory for general Lie
algebras,” JHEP 0806, 075 (2008) [arXiv:0805.1012 [hep-th]].
[155] S. Benvenuti, D. Rodriguez-Gomez, E. Tonni, and H. Verlinde, “N = 8 supercon-
formal gauge theories and M2 branes,” JHEP 0901, 078 (2009) [arXiv:0805.1087
[hep-th]].
[156] P. -M. Ho, Y. Imamura, and Y. Matsuo, “M2 to D2 revisited,” JHEP 0807, 003 (2008)
[arXiv:0805.1202 [hep-th]].
[157] T. Bargheer, F. Loebbert, and C. Meneghelli, “Symmetries of tree-level scattering
amplitudes in N = 6 superconformal Chern–Simons theory,” Phys. Rev. D 82, 045016
(2010) [arXiv:1003.6120 [hep-th]].
[158] O. Aharony, O. Bergman, D. L. Jafferis, and J. Maldacena, “N = 6 superconformal
Chern–Simons-matter theories, M2-branes and their gravity duals,” JHEP 0810, 091
(2008) [arXiv:0806.1218 [hep-th]].
[159] M. Benna, I. Klebanov, T. Klose, and M. Smedback, “Superconformal Chern–
Simons theories and AdS(4)/CFT(3) correspondence,” JHEP 0809, 072 (2008)
[arXiv:0806.1519 [hep-th]].
[160] M. A. Bandres, A. E. Lipstein, and J. H. Schwarz, “Studies of the ABJM the-
ory in a formulation with manifest SU(4) R-symmetry,” JHEP 0809, 027 (2008)
[arXiv:0807.0880 [hep-th]].
[161] A. Gustavsson, “Selfdual strings and loop space Nahm equations,” JHEP 0804, 083
(2008) [arXiv:0802.3456 [hep-th]];
J. Bagger and N. Lambert, “Three-algebras and N = 6 Chern–Simons gauge theories,”
[162] D. Gang, Y.-t. Huang, E. Koh, S. Lee, and A. E. Lipstein, “Tree-level recursion
relation and dual superconformal symmetry of the ABJM theory,” JHEP 1103, 116
(2011) [arXiv:1012.5032 [hep-th]].
References 309
[163] Y.-t. Huang and A. E. Lipstein, “Dual superconformal symmetry of N = 6 Chern–

Simons theory,” JHEP 1011, 076 (2010) [arXiv:1008.0041 [hep-th]].
[164] T. Bargheer, N. Beisert, F. Loebbert, and T. McLoughlin, “Conformal anomaly for
amplitudes in N = 6 superconformal Chern–Simons theory,” J. Phys. A 45, 475402
(2012) [arXiv:1204.4406 [hep-th]].
[165] M. S. Bianchi, M. Leoni, A. Mauri, S. Penati, and A. Santambrogio, “One loop
amplitudes In ABJM,” JHEP 1207, 029 (2012) [arXiv:1204.4407 [hep-th]].
[166] A. Brandhuber, G. Travaglini, and C. Wen, “All one-loop amplitudes in N = 6
superconformal Chern–Simons theory,” JHEP 1210, 145 (2012) [arXiv:1207.6908
[hep-th]].
[167] W. -M. Chen and Y.-t. Huang, “Dualities for loop amplitudes of N = 6 Chern–Simons
matter theory,” JHEP 1111, 057 (2011) [arXiv:1107.2710 [hep-th]];
[168] A. Brandhuber, G. Travaglini, and C. Wen, “A note on amplitudes in N = 6 supercon-
formal Chern–Simons theory,” JHEP 1207, 160 (2012) [arXiv:1205.6705 [hep-th]].
[169] M. S. Bianchi, M. Leoni, A. Mauri, S. Penati, and A. Santambrogio, “Scatter-
ing amplitudes/Wilson loop duality in ABJM theory,” JHEP 1201, 056 (2012)
[arXiv:1107.3139 [hep-th]].
[170] S. Caron-Huot and Y.-t. Huang, “The two-loop six-point amplitude in ABJM theory,”
JHEP 1303, 075 (2013) [arXiv:1210.4226 [hep-th]].
[171] S. Lee, “Yangian invariant scattering amplitudes in supersymmetric Chern–Simons
theory,” Phys. Rev. Lett. 105, 151603 (2010) [arXiv:1007.4772 [hep-th]].
[172] Y. -t. Huang, C. Wen, and D. Xie, “The positive orthogonal Grassmannian and loop
amplitudes of ABJM,” [arXiv:1402.1479 [hep-th]].
[173] Y.-t. Huang and S. Lee, “A new integral formula for supersymmetric scattering am-
plitudes in three dimensions,” Phys. Rev. Lett. 109, 191601 (2012) [arXiv:1207.4851
[hep-th]].
[174] O. T. Engelund and R. Roiban, “A twistor string for the ABJ(M) theory,”
[arXiv:1401.6242 [hep-th]].
[175] S. Weinberg, Gravitation and Cosmology: Principles and Applications of the General
Theory of Relativity, USA: John Wiley & Sons (1972).
[176] R. M. Wald, General Relativity, Chicago, USA: University Press (1984).
[177] S. M. Carroll, Spacetime and Geometry: An Introduction to General Relativity, San
Francisco, USA: Addison-Wesley (2004).
[178] H. Elvang and D. Z. Freedman, unpublished notes (2007).
[179] B. S. DeWitt, “Quantum theory of gravity. 2. The manifestly covariant theory,” Phys.
Rev. 162, 1195 (1967);
B. S. DeWitt, “Quantum theory of gravity. 3. Applications of the covariant theory,”
Phys. Rev. 162, 1239 (1967);
M. J. G. Veltman, “Quantum theory of gravitation,” Conf. Proc. C 7507281, 265
(1975).
[180] F. A. Berends, W. T. Giele, and H. Kuijf, “On relations between multi-gluon and
multigraviton scattering,” Phys. Lett. B 211, 91 (1988).
[181] J. Bedford, A. Brandhuber, B. J. Spence, and G. Travaglini, “A recursion relation for
gravity amplitudes,” Nucl. Phys. B 721, 98 (2005) [hep-th/0502146].
310 References
[182] H. Elvang and D. Z. Freedman, “Note on graviton MHV amplitudes,” JHEP 0805,
096 (2008) [arXiv:0710.1270 [hep-th]].
[183] D. Nguyen, M. Spradlin, A. Volovich, and C. Wen, “The tree formula for MHV
graviton amplitudes,” JHEP 1007, 045 (2010) [arXiv:0907.2276 [hep-th]].
[184] Z. Bern, L. J. Dixon, M. Perelstein, and J. S. Rozowsky, “Multileg one loop gravity
amplitudes from gauge theory,” Nucl. Phys. B 546, 423 (1999) [hep-th/9811140].
[185] Z. Bern and A. K. Grant, “Perturbative gravity from QCD amplitudes,” Phys. Lett.
B 457, 23 (1999) [hep-th/9904026].
[186] Z. Bern, L. J. Dixon, D. C. Dunbar, et al., “On perturbative gravity and gauge theory,”
Nucl. Phys. Proc. Suppl. 88, 194 (2000) [hep-th/0002078].
[187] W. Siegel, “Two vierbein formalism for string inspired axionic gravity,” Phys. Rev.
D 47, 5453 (1993) [hep-th/9302036].
[188] Z. Bern, “Perturbative quantum gravity and its relation to gauge theory,” Living Rev.
Rel. 5, 5 (2002) [gr-qc/0206071].
[189] D. Z. Freedman, “Some beautiful equations of mathematical physics,” In ICTP
(ed.): The Dirac Medals of the ICTP 1993 25–53, and CERN Geneva - TH.-7367
(94/07,rec.Sep.) [hep-th/9408175].
[190] S. J. Gates, M. T. Grisaru, M. Rocek, and W. Siegel, “Superspace or one thousand
and one lessons in supersymmetry,” Front. Phys. 58, 1 (1983) [hep-th/0108200].
[191] B. de Wit and D. Z. Freedman, “On SO(8) extended supergravity,” Nucl. Phys. B
130, 105 (1977).
[192] E. Cremmer and B. Julia, “The N = 8 supergravity theory. 1. The Lagrangian,” Phys.
Lett. B 80, 48 (1978);
E. Cremmer and B. Julia, “The SO(8) supergravity,” Nucl. Phys. B 159, 141
(1979).
[193] B. de Wit and H. Nicolai, “N = 8 supergravity,” Nucl. Phys. B 208, 323 (1982).
[194] A. Hodges, “A simple formula for gravitational MHV amplitudes,” [arXiv:1204.1930
[hep-th]].
[195] F. Cachazo, L. Mason, and D. Skinner, “Gravity in twistor space and its Grassmannian
formulation,” [arXiv:1207.4712 [hep-th]].
[196] S. He, “A link representation for gravity amplitudes,” [arXiv:1207.4064 [hep-th]].
[197] F. Cachazo and Y. Geyer, “A ‘twistor string’ inspired formula for tree-level scattering
amplitudes in N = 8 SUGRA,” [arXiv:1206.6511 [hep-th]].
[198] D. Skinner, “Twistor strings for N = 8 supergravity,” [arXiv:1301.0868 [hep-th]].
[199] F. Cachazo, S. He, and E. Y. Yuan, “Scattering equations and KLT orthogonality,”
[arXiv:1306.6575 [hep-th]].
[200] F. Cachazo, S. He, and E. Y. Yuan, “Scattering of massless particles in arbitrary
dimension,” [arXiv:1307.2199 [hep-th]].
[201] S. L. Adler, “Consistency conditions on the strong interactions implied by a partially
conserved axial vector current,” Phys. Rev. 137, B1022 (1965).
[202] S. R. Coleman, “Secret symmetry: An introduction to spontaneous symmetry break-
down and gauge fields,” Subnucl. Ser. 11, 139 (1975).
[203] M. Kiermaier, “The Coulomb-branch S-matrix from massless amplitudes,”
[arXiv:1105.5385 [hep-th]].
References 311
[204] G. ’t Hooft and M. J. G. Veltman, “One loop divergencies in the theory of gravitation,”
Annales Poincaré Phys. Theor. A 20, 69 (1974).
[205] M. H. Goroff and A. Sagnotti, “Quantum gravity at two loops,” Phys. Lett. B 160,
81 (1985).
[206] A. E. M. van de Ven, “Two loop quantum gravity,” Nucl. Phys. B 378, 309 (1992).
[207] S. Deser and P. van Nieuwenhuizen, “One loop divergences of quantized Einstein–
Maxwell fields,” Phys. Rev. D 10, 401 (1974).
[208] M. T. Grisaru, P. van Nieuwenhuizen, and J. A. M. Vermaseren, “One loop renor-
malizability of pure supergravity and of Maxwell–Einstein theory in extended su-
pergravity,” Phys. Rev. Lett. 37, 1662 (1976).
[209] M. T. Grisaru, “Two loop renormalizability of supergravity,” Phys. Lett. B 66, 75
(1977).
[210] E. Tomboulis, “On the two loop divergences of supersymmetric gravitation,” Phys.
Lett. B 67, 417 (1977).
[211] S. Deser, J. H. Kay, and K. S. Stelle, “Renormalizability properties of supergravity,”
Phys. Rev. Lett. 38, 527 (1977).
[212] Z. Bern, L. J. Dixon, and R. Roiban, “Is N = 8 supergravity ultraviolet finite?,” Phys.
Lett. B 644, 265 (2007) [hep-th/0611086].
[213] Z. Bern, J. J. Carrasco, L. J. Dixon, et al., “Three-loop superfiniteness of N = 8
supergravity,” Phys. Rev. Lett. 98, 161303 (2007) [hep-th/0702112].
[214] Z. Bern, J. J. M. Carrasco, L. J. Dixon, H. Johansson, and R. Roiban, “Manifest
ultraviolet behavior for the three-loop four-point amplitude of N = 8 supergravity,”
[215] P. S. Howe and K. S. Stelle, “Supersymmetry counterterms revisited,” Phys. Lett. B
554, 190 (2003) [hep-th/0211279].
[216] Z. Bern, J. J. Carrasco, L. J. Dixon, H. Johansson, and R. Roiban, “The ultraviolet
behavior of N = 8 supergravity at four loops,” Phys. Rev. Lett. 103, 081301 (2009)
[arXiv:0905.2326 [hep-th]].
[217] Z. Bern, J. J. M. Carrasco, L. J. Dixon, H. Johansson, and R. Roiban, “The complete
four-loop four-point amplitude in N = 4 super-Yang–Mills theory,” Phys. Rev. D 82,
125040 (2010) [arXiv:1008.3327 [hep-th]].
[218] J. Bjornsson and M. B. Green, “5 loops in 24/5 dimensions,” JHEP 1008, 132 (2010)
[arXiv:1004.2692 [hep-th]].
[219] H. Elvang, D. Z. Freedman, and M. Kiermaier, “A simple approach to counterterms
in N = 8 supergravity,” JHEP 1011, 016 (2010) [arXiv:1003.5018 [hep-th]].
[220] H. Elvang and M. Kiermaier, “Stringy KLT relations, global symmetries, and E 7(7)
violation,” JHEP 1010, 108 (2010) [arXiv:1007.4813 [hep-th]].
[221] N. Beisert, H. Elvang, D. Z. Freedman, et al., “E7(7) constraints on counterterms in
N = 8 supergravity,” Phys. Lett. B 694, 265 (2010) [arXiv:1009.1643 [hep-th]].
[222] P. van Nieuwenhuizen and C. C. Wu, “On integral relations for invariants constructed
from three riemann tensors and their applications in quantum gravity,” J. Math. Phys.
18, 182 (1977).
[223] G. Bossard, C. Hillmann, and H. Nicolai, “E7(7) symmetry in perturbatively quan-
tised N = 8 supergravity,” JHEP 1012, 052 (2010) [arXiv:1007.5472 [hep-th]].
312 References
[224] D. Z. Freedman and E. Tonni, “The D 2k R 4 invariants of N = 8 supergravity,” JHEP

1104, 006 (2011) [arXiv:1101.1672 [hep-th]].
[225] S. Deser and J. H. Kay, “Three loop counterterms for extended supergravity,” Phys.
Lett. B 76, 400 (1978).
[226] J. M. Drummond, P. J. Heslop, and P. S. Howe, “A note on N = 8 counterterms,”
[arXiv:1008.4939 [hep-th]].
[227] G. Bossard and H. Nicolai, “Counterterms vs. dualities,” JHEP 1108, 074 (2011)
[arXiv:1105.1273 [hep-th]].
[228] R. Kallosh and T. Kugo, “The footprint of E(7(7)) amplitudes of N = 8 supergravity,”
JHEP 0901, 072 (2009) [arXiv:0811.3414 [hep-th]];
R. Kallosh, “E 7(7) symmetry and finiteness of N = 8 supergravity,” JHEP 1203, 083
(2012) [arXiv:1103.4115 [hep-th]];
R. Kallosh, “N = 8 counterterms and E 7(7) current conservation,” JHEP 1106, 073
(2011) [arXiv:1104.5480 [hep-th]];
R. Kallosh and T. Ortin, “New E77 invariants and amplitudes,” JHEP 1209, 137
(2012) [arXiv:1205.4437 [hep-th]];
M. Gunaydin and R. Kallosh, “Obstruction to E 7(7) deformation in N = 8 super-
gravity,” [arXiv:1303.3540 [hep-th]];
J. J. M. Carrasco and R. Kallosh, “Hidden supersymmetry may imply duality invari-
ance,” [arXiv:1303.5663 [hep-th]].
[229] S. Stieberger and T. R. Taylor, “Complete six-gluon disk amplitude in superstring
theory,” Nucl. Phys. B 801, 128 (2008) [arXiv:0711.4354 [hep-th]].
[230] J. Broedel and L. J. Dixon, “R∗∗ 4 counterterm and E(7)(7) symmetry in maximal
supergravity,” JHEP 1005, 003 (2010) [arXiv:0911.5704 [hep-th]].
[231] N. Berkovits, “New higher-derivative R∗∗ 4 theorems,” Phys. Rev. Lett. 98, 211601
(2007) [arXiv:hep-th/0609006];
M. B. Green, J. G. Russo, and P. Vanhove, “Non-renormalisation conditions in type
II string theory and maximal supergravity,” JHEP 0702, 099 (2007) [arXiv:hep-
th/0610299];
M. B. Green, J. G. Russo, and P. Vanhove, “Ultraviolet properties of maximal super-
gravity,” Phys. Rev. Lett. 98, 131602 (2007) [arXiv:hep-th/0611273];
M. B. Green, J. G. Russo, and P. Vanhove, “Modular properties of two-loop max-
imal supergravity and connections with string theory,” JHEP 0807, 126 (2008)
[arXiv:0807.0389 [hep-th]];
N. Berkovits, M. B. Green, J. G. Russo, and P. Vanhove, “Non-renormalization
conditions for four-gluon scattering in supersymmetric string and field theory,” JHEP
0911, 063 (2009) [arXiv:0908.1923 [hep-th]];
P. Vanhove, “The critical ultraviolet behaviour of N = 8 supergravity amplitudes,”
[arXiv:1004.1392 [hep-th]].
[232] Z. Bern, J. J. M. Carrasco, L. J. Dixon, H. Johansson, and R. Roiban, “Simplify-
ing multiloop integrands and ultraviolet divergences of gauge theory and gravity
amplitudes,” Phys. Rev. D 85, 105014 (2012) [arXiv:1201.5366 [hep-th]].
[233] P. S. Howe and U. Lindstrom, “Higher order invariants in extended supergravity,”
Nucl. Phys. B 181, 487 (1981).
References 313
[234] R. E. Kallosh, “Counterterms in extended supergravities,” Phys. Lett. B 99, 122

(1981).
[235] G. Bossard, P. S. Howe, K. S. Stelle, and P. Vanhove, “The vanishing volume of
D = 4 superspace,” Class. Quant. Grav. 28, 215005 (2011) [arXiv:1105.6087
[hep-th]].
[236] N. Berkovits, “Super Poincaré covariant quantization of the superstring,” JHEP 0004,
018 (2000) [hep-th/0001035].
[237] M. B. Green, H. Ooguri, and J. H. Schwarz, “Nondecoupling of maximal super-
gravity from the superstring,” Phys. Rev. Lett. 99, 041601 (2007) [arXiv:0704.0777
[hep-th]].
[238] T. Banks, “Arguments against a finite N = 8 supergravity,” [arXiv:1205.5768 [hep-
th]].
[239] M. Bianchi, S. Ferrara, and R. Kallosh, “Perturbative and non-perturbative N = 8
supergravity,” Phys. Lett. B 690, 328 (2010) [arXiv:0910.3674 [hep-th]].
[240] Z. Bern, S. Davies, T. Dennen, Y.-t. Huang, and J. Nohle, “Color-kinematics du-
ality for pure Yang–Mills and gravity at one and two loops,” [arXiv:1303.6605
[hep-th]].
[241] Z. Bern, S. Davies, T. Dennen, A. V. Smirnov, and V. A. Smirnov, “The ultraviolet
properties of N = 4 supergravity at four loops,” Phys. Rev. Lett. 111, 231302 (2013)
[arXiv:1309.2498 [hep-th]].
[242] Z. Bern, L. J. Dixon, D. C. Dunbar, M. Perelstein, and J. S. Rozowsky, “On the
relationship between Yang–Mills theory and gravity and its implication for ultraviolet
divergences,” Nucl. Phys. B 530, 401 (1998) [hep-th/9802162].
[243] D. C. Dunbar, B. Julia, D. Seminara, and M. Trigiante, “Counterterms in type I
supergravities,” JHEP 0001, 046 (2000) [hep-th/9911158].
[244] Z. Bern, S. Davies, T. Dennen, and Y.-t. Huang, “Ultraviolet cancellations in half-
maximal supergravity as a consequence of the double-copy structure,” Phys. Rev. D
86, 105014 (2012) [arXiv:1209.2472 [hep-th]].
[245] Z. Bern, S. Davies, T. Dennen, and Y.-t. Huang, “Absence of three-loop four-
point divergences in N = 4 supergravity,” Phys. Rev. Lett. 108, 201301 (2012)
[arXiv:1202.3423 [hep-th]];
[246] M. Fischler, “Finiteness calculations for O(4) through O(8) extended supergrav-
ity and O(4) supergravity coupled to selfdual O(4) matter,” Phys. Rev. D 20, 396
(1979).
[247] Z. Bern, S. Davies, and T. Dennen, “The ultraviolet structure of half-maximal super-
gravity with matter multiplets at two and three loops,” [arXiv:1305.4876 [hep-th]].
[248] G. Bossard, P. S. Howe, and K. S. Stelle, “Invariants and divergences in half-maximal
supergravity theories,” [arXiv:1304.7753 [hep-th]].
[249] D. Vaman and Y.-P. Yao, “Constraints and generalized gauge transformations on
tree-level gluon and graviton amplitudes,” JHEP 1011, 028 (2010) [arXiv:1007.3475
[hep-th]].
[250] R. H. Boels and R. S. Isermann, “On powercounting in perturbative quan-
tum gravity theories through color-kinematic duality,” JHEP 1306, 017 (2013)
[arXiv:1212.3473].
314 References
[251] Z. Bern, T. Dennen, Y.-t. Huang, and M. Kiermaier, “Gravity as the square of gauge
theory,” Phys. Rev. D 82, 065003 (2010) [arXiv:1004.0693 [hep-th]].
[252] M. Kiermaier, Talk at Amplitudes 2010, May 2010 at QMUL, London, UK.
http://www.strings.ph.qmul.ac.uk/∼theory/Amplitudes2010/
[253] N. E. J. Bjerrum-Bohr, P. H. Damgaard, T. Sondergaard, and P. Vanhove, “The momen-
tum kernel of gauge and gravity theories,” JHEP 1101, 001 (2011) [arXiv:1010.3933
[hep-th]].
[254] C. R. Mafra, O. Schlotterer, and S. Stieberger, “Explicit BCJ numerators from pure
spinors,” JHEP 1107, 092 (2011) [arXiv:1104.5224 [hep-th]];
C.-H. Fu, Y.-J. Du, and B. Feng, “An algebraic approach to BCJ numerators,” JHEP
1303, 050 (2013) [arXiv:1212.6168 [hep-th]].
[255] N. E. J. Bjerrum-Bohr, P. H. Damgaard, and P. Vanhove, “Minimal basis for gauge
theory amplitudes,” Phys. Rev. Lett. 103, 161602 (2009) [0907.1425 [hep-th]];
S. Stieberger, “Open & closed vs. pure open string disk amplitudes,” [arXiv:
0907.2211 [hep-th]];
C. R. Mafra and O. Schlotterer, “The structure of n-point one-loop open superstring
amplitudes,” [arXiv:1203.6215 [hep-th]];
O. Schlotterer and S. Stieberger, “Motivic multiple zeta values and superstring am-
plitudes,” [arXiv:1205.1516 [hep-th]];
J. Broedel, O. Schlotterer, and S. Stieberger, “Polylogarithms, multiple zeta values
and superstring amplitudes,” [arXiv:1304.7267 [hep-th]].
[256] S. H. Henry Tye and Y. Zhang, “Dual identities inside the gluon and the graviton
scattering amplitudes,” JHEP 1006, 071 (2010) [Erratum-ibid. 1104, 114 (2011)]
[arXiv:1003.1732 [hep-th]].
[257] B. Feng, R. Huang, and Y. Jia, “Gauge amplitude identities by on-shell recursion
relation in s-matrix program,” Phys. Lett. B 695, 350 (2011) [arXiv:1004.3417 [hep-
th]].
[258] F. Cachazo, “Fundamental BCJ relation in N = 4 SYM from the connected formu-
lation,” [arXiv:1206.5970 [hep-th]].
[259] N. E. J. Bjerrum-Bohr, P. H. Damgaard, R. Monteiro, and D. O’Connell, “Algebras
for amplitudes,” JHEP 1206, 061 (2012) [arXiv:1203.0944 [hep-th]].
[260] R. Monteiro and D. O’Connell, “The kinematic algebra from the self-dual sector,”
JHEP 1107, 007 (2011) [arXiv:1105.2565 [hep-th]].
[261] M. Tolotti and S. Weinzierl, “Construction of an effective Yang–Mills Lagrangian
with manifest BCJ duality,” [arXiv:1306.2975 [hep-th]].
[262] Z. Bern, J. J. M. Carrasco, and H. Johansson, “Perturbative quantum gravity as a
double copy of gauge theory,” Phys. Rev. Lett. 105, 061602 (2010) [arXiv:1004.0476
[hep-th]].
[263] J. J. Carrasco and H. Johansson, “Five-point amplitudes in N = 4 super-Yang–Mills
theory and N = 8 supergravity,” Phys. Rev. D 85, 025006 (2012) [arXiv:1106.4711
[hep-th]].
[264] N. E. J. Bjerrum-Bohr, T. Dennen, R. Monteiro, and D. O’Connell, “Integrand oxida-
tion and one-loop colour-dual numerators in N = 4 gauge theory,” [arXiv:1303.2913
[hep-th]].
References 315
[265] R. H. Boels, R. S. Isermann, R. Monteiro, and D. O’Connell, “Colour-kinematics

duality for one-loop rational amplitudes,” JHEP 1304, 107 (2013) [arXiv:1301.4165
[hep-th]].
[266] J. J. M. Carrasco, M. Chiodaroli, M. Günaydin, and R. Roiban, “One-loop four-point
amplitudes in pure and matter-coupled N ≤ 4 supergravity,” JHEP 1303, 056 (2013)
[arXiv:1212.1146 [hep-th]].
[267] M. Chiodaroli, Q. Jin, and R. Roiban, “Color/kinematics duality for general
abelian orbifolds of N = 4 super Yang–Mills theory,” JHEP 1401, 152 (2014)
[arXiv:1311.3600 [hep-th]].
[268] J. Nohle, “Color-kinematics duality in one-loop four-gluon amplitudes with matter,”
[arXiv:1309.7416 [hep-th]].
[269] Z. Bern, C. Boucher-Veronneau, and H. Johansson, “N ≥ 4 supergravity amplitudes
from gauge theory at one loop,” Phys. Rev. D 84, 105035 (2011) [arXiv:1107.1935
[hep-th]];
C. Boucher-Veronneau and L. J. Dixon, “N ≥ 4 supergravity amplitudes from gauge
theory at two loops,” JHEP 1112, 046 (2011) [arXiv:1110.1132 [hep-th]].
[270] M. T. Grisaru and W. Siegel, “Supergraphity. 2. Manifestly covariant rules and higher
loop finiteness,” Nucl. Phys. B 201, 292 (1982) [Erratum-ibid. B 206, 496 (1982)].
[271] S. Ferrara, R. Kallosh, and A. Van Proeyen, “Conjecture on hidden superconformal
symmetry of N = 4 supergravity,” Phys. Rev. D 87, 025004 (2013) [arXiv:1209.0418
[hep-th]].
[272] J. Broedel and L. J. Dixon, “Color-kinematics duality and double-copy construc-
tion for amplitudes from higher-dimension operators,” JHEP 1210, 091 (2012)
[arXiv:1208.0876 [hep-th]].
[273] R. H. Boels, B. A. Kniehl, O. V. Tarasov, and G. Yang, “Color-kinematic duality for
form factors,” JHEP 1302, 063 (2013) [arXiv:1211.7028 [hep-th]];
[274] T. Bargheer, S. He, and T. McLoughlin, “New relations for three-dimensional
supersymmetric scattering amplitudes,” Phys. Rev. Lett. 108, 231601 (2012)
[arXiv:1203.0562 [hep-th]].
[275] Y.-t. Huang and H. Johansson, “Equivalent D = 3 supergravity amplitudes from
double copies of three-algebra and two-algebra gauge theories,” [arXiv:1210.2255
[hep-th]].
[276] A. Zee, “Quantum field theory in a nutshell,” Princeton, USA: Princeton University
Press (2010).
[277] M. D. Schwartz, Quantum Field Theory and the Standard Model, Cambridge, UK:
Cambridge University Press (2013).
[278] J. M. Henn and J. C. Plefka, “Scattering amplitudes in gauge theories,” Lecture Notes
in Physics 883 Heidelbera: Springer (2014).
[279] M. Wolf, “A first course on twistors, integrability and gluon scattering amplitudes,”
J. Phys. A 43, 393001 (2010) [arXiv:1001.3871 [hep-th]].
[280] Z. Bern, L. J. Dixon, and D. A. Kosower, “On-shell methods in perturbative QCD,”
Annals Phys. 322, 1587 (2007) [arXiv:0704.2798 [hep-ph]].
[281] M. E. Peskin, “Simplifying multi-jet QCD computation,” [arXiv:1101.2414
[hep-ph]].
316 References
[282] J. F. Donoghue, “Introduction to the effective field theory description of gravity,”

gr-qc/9512024.
[283] L. J. Dixon, “Ultraviolet behavior of N = 8 supergravity,” [arXiv:1005.2703
[hep-th]].
[284] Z. Bern, P. Gondolo, and M. Perelstein, “Neutralino annihilation into two photons,”
Phys. Lett. B 411, 86 (1997) [hep-ph/9706538];
Z. Bern, A. De Freitas, and L. J. Dixon, “Two loop helicity amplitudes for gluon–
gluon scattering in QCD and supersymmetric Yang–Mills theory,” JHEP 0203, 018
(2002) [hep-ph/0201161];
Z. Bern, A. De Freitas, and L. J. Dixon, “Two loop helicity amplitudes for quark
gluon scattering in QCD and gluino gluon scattering in supersymmetric Yang–Mills
theory,” JHEP 0306, 028 (2003) [hep-ph/0304168].
[285] S. J. Bidder, N. E. J. Bjerrum-Bohr, D. C. Dunbar, and W. B. Perkins, “One-loop
gluon scattering amplitudes in theories with N < 4 supersymmetries,” Phys. Lett. B
612, 75 (2005) [hep-th/0502028].
[286] R. Britto, E. Buchbinder, F. Cachazo, and B. Feng, “One-loop amplitudes of gluons
in SQCD,” Phys. Rev. D 72, 065012 (2005) [hep-ph/0503132].
[287] S. Lal and S. Raju, “Rational terms in theories with matter,” JHEP 1008, 022 (2010)
[arXiv:1003.5264 [hep-th]].
[288] S. Dittmaier, “Weyl-van der Waerden formalism for helicity amplitudes of massive
particles,” Phys. Rev. D 59, 016007 (1998) [hep-ph/9805445].
[289] R. Boels and C. Schwinn, “CSW rules for massive matter legs and glue loops,” Nucl.
Phys. Proc. Suppl. 183, 137 (2008) [arXiv:0805.4577 [hep-th]].
[290] R. H. Boels, “No triangles on the moduli space of maximally supersymmetric gauge
[291] P. Ferrario, G. Rodrigo, and P. Talavera, “Compact multigluonic scattering am-
plitudes with heavy scalars and fermions,” Phys. Rev. Lett. 96, 182001 (2006)
[hep-th/0602043].
[292] D. Forde and D. A. Kosower, “All-multiplicity amplitudes with massive scalars,”
Phys. Rev. D 73, 065007 (2006) [hep-th/0507292].
[293] G. Rodrigo, “Multigluonic scattering amplitudes of heavy quarks,” JHEP 0509, 079
(2005) [hep-ph/0508138].
[294] C. Cheung, D. O’Connell, and B. Wecht, “BCFW recursion relations and string
[295] R. H. Boels, D. Marmiroli, and N. A. Obers, “On-shell recursion in string theory,”
JHEP 1010, 034 (2010) [arXiv:1002.5029 [hep-th]].
[296] K. Kampf, J. Novotny, and J. Trnka, “Recursion relations for tree-level ampli-
tudes in the SU(N) non-linear sigma model,” Phys. Rev. D 87, 081701 (2013)
[arXiv:1212.5224 [hep-th]].
[297] Z. Bern, L. J. Dixon, and D. A. Kosower, “The last of the finite loop amplitudes in
QCD,” Phys. Rev. D 72, 125003 (2005) [hep-ph/0505055].
[298] B. Feng and M. Luo, “An introduction to on-shell recursion relations,”
[arXiv:1111.5759 [hep-th]].
References 317
[299] J. M. Drummond, G. P. Korchemsky, and E. Sokatchev, “Conformal properties of

four-gluon planar amplitudes and Wilson loops,” Nucl. Phys. B 795, 385 (2008)
[arXiv:0707.0243 [hep-th]].
[300] N. Berkovits and J. Maldacena, “Fermionic T-duality, dual superconformal sym-
metry, and the amplitude/Wilson loop connection,” JHEP 0809, 062 (2008)
[arXiv:0807.3196 [hep-th]].
[301] N. Beisert, R. Ricci, A. A. Tseytlin, and M. Wolf, “Dual superconformal symmetry
from AdS(5) x S∗∗ 5 superstring integrability,” Phys. Rev. D 78, 126004 (2008)
[arXiv:0807.3228 [hep-th]].
[302] A. Brandhuber, P. Heslop, and G. Travaglini, “MHV amplitudes in N = 4 super
Yang–Mills and Wilson loops,” Nucl. Phys. B 794, 231 (2008) [arXiv:0707.1153
[hep-th]].
[303] L. J. Mason and D. Skinner, “The complete planar S-matrix of N = 4 SYM as a
Wilson loop in twistor space,” JHEP 1012, 018 (2010) [arXiv:1009.2225 [hep-th]].
[304] S. Caron-Huot, “Notes on the scattering amplitude / Wilson loop duality,” JHEP
1107, 058 (2011) [arXiv:1010.1167 [hep-th]].
[305] B. Eden, P. Heslop, G. P. Korchemsky, and E. Sokatchev, “The super-correlator/super-
amplitude duality: Part I,” Nucl. Phys. B 869, 329 (2013) [arXiv:1103.3714 [hep-th]];
Part II, Nucl. Phys. B 869, 378 (2013) [arXiv:1103.4353 [hep-th]].
[306] L. F. Alday, B. Eden, G. P. Korchemsky, J. Maldacena, and E. Sokatchev, “From
correlation functions to Wilson loops,” JHEP 1109, 123 (2011) [arXiv:1007.3243
[hep-th]].
[307] T. Adamo, M. Bullimore, L. Mason, and D. Skinner, “A proof of the supersym-
metric correlation function / Wilson loop correspondence,” JHEP 1108, 076 (2011)
[arXiv:1103.4119 [hep-th]].
[308] L. F. Alday and R. Roiban, “Scattering amplitudes, Wilson loops and the string/gauge
theory correspondence,” Phys. Rept. 468, 153 (2008) [arXiv:0807.1889 [hep-th]].
[309] R. M. Schabinger, “One-loop N = 4 super Yang–Mills scattering amplitudes in d
dimensions, relation to open strings and polygonal Wilson loops,” J. Phys. A 44,
454007 (2011) [arXiv:1104.3873 [hep-th]].
[310] J. M. Henn, “Duality between Wilson loops and gluon amplitudes,” Fortsch. Phys.
57, 729 (2009) [arXiv:0903.0522 [hep-th]].
[311] T. Adamo, M. Bullimore, L. Mason, and D. Skinner, “Scattering amplitudes and
Wilson loops in twistor space,” J. Phys. A 44, 454008 (2011) [arXiv:1104.2890
[hep-th]].
[312] L. F. Alday, D. Gaiotto, J. Maldacena, A. Sever, and P. Vieira, “An operator product ex-
pansion for polygonal null Wilson loops,” JHEP 1104, 088 (2011) [arXiv:1006.2788
[hep-th]].
[313] B. Basso, A. Sever, and P. Vieira, “Space-time S-matrix and flux-tube S-matrix at
finite coupling,” Phys. Rev. Lett. 111, 091602 (2013) [arXiv:1303.1396 [hep-th]].
[314] B. Basso, A. Sever, and P. Vieira, “Space-time S-matrix and flux tube S-matrix
II. Extracting and matching data,” JHEP 1401, 008 (2014) [arXiv:1306.2058
[hep-th]].
318 References
[315] B. Basso, A. Sever, and P. Vieira, “Space-time S-matrix and flux-tube S-matrix III.
The two-particle contributions,” [arXiv:1402.3307 [hep-th]].
[316] L. J. Dixon, J. M. Drummond, M. von Hippel, and J. Pennington, “Hexagon functions
and the three-loop remainder function,” JHEP 1312, 049 (2013) [arXiv:1308.2276
[hep-th]].
[317] L. J. Dixon, J. M. Drummond, C. Duhr, and J. Pennington, “The four-loop remainder
function and multi-Regge behavior at NNLLA in planar N = 4 super-Yang–Mills
theory,” [arXiv:1402.3300 [hep-th]].
[318] R. Penrose and W. Rindler, Spinors and Space-Time, vol. 2, Cambridge: Cambridge
[319] R. Ward and R. Wells, Twistor Geometry and Field Theory, Cambridge: Cambridge
[320] S. Huggett and P. Tod, An Introduction to Twistor Theory, Student Texts 4, London:
London Mathematical Society (1985).
[321] F. Cachazo and P. Svrcek, “Lectures on twistor strings and perturbative Yang–Mills
theory,” PoS RTN 2005, 004 (2005) [hep-th/0504194].
Index
3-algebra, 230 NMHV 5-point, 57, 58

3-gluon amplitude shift, 53
all-minus from F 3 , 40 spurious poles, 61
anti-MHV, 33 supershift, 86
MHV, 33, 38 BCJ
3-graviton amplitude, 40, 251 (fundamental) amplitude relations, 35, 279
all-minus from R 3 , 40 color-kinematics duality, 277
3-particle special kinematics, 27–28 double-copy relation, 280, 281
3-point anti-MHV superamplitude, 85 generalized gauge transformation, 275
4-bracket, 109 BDS/ABDK Ansatz, 136–137
5-bracket, 111 BGK graviton amplitude, 252
Bhabha scattering, 2, 29
ABDK/BDS Ansatz, 136–137 BLG theory, 230
ABJM theory, 231 bonus relation, 62
all-line shift, 64, 65–68 Bose symmetry, 5, 22, 39, 40
all-minus gluon amplitude, 40, 43, 76, 122 boundary operation, 210
all-minus graviton amplitude, 40, 252 box integral, 123, 124, 126, 141
all-plus gluon amplitude, 43, 76 BPS operator, 269
all-plus graviton amplitude, 252 bubble integral, 123, 124
amplitugicians, 64
amplituhedron, 214 Cauchy’s theorem, 52
analytic continuation, 55 central charge, 73
analytic properties Chern–Simons
loop amplitudes, 118–126 Lagrangian, 224
tree amplitudes, 44–49, 51, 61 level, 224
angle bracket, 20 matter Lagrangian, 224
angle spinor, 17 propagator, 226
angle-square bracket, 21 χi A , 110
anti-MHV collinear limits
3-graviton, 251 loop-level, 121
3-point gluon, 33 tree-level, 47–49, 61
3-point superamplitude, 85 color factors, 31, 274
definition, 44 color-kinematics duality, 36, 277
Parke–Taylor formula, 44, 58 color-ordered amplitude, 31
auxiliary field, 72 BCJ, 276
axion, 254 gluon MHV, 33, 44, 96
gluon NMHV, 57–61
basis of 1-loop scalar integrals, 126 N = 4 super Yang–Mills, 64
BCFW, 50, 53–62 properties, 34
bridge, 174, 198 color-ordered Feynman rules, 32, 118
bridge in 3d, 243 color-structure, 31
in 3d, 233–238 completeness relation, 18
in momentum twistor space, 142–159 complex conjugation, 17
n-point tree gluon MHV, 54–57 complex momentum, 17, 37
n-point tree graviton MHV, 252 component amplitude, 82
n-point tree NMHV superamplitude, 111 Compton scattering, 2, 28
320 Index
conformal boosts, 97 Gervais–Neveu gauge, 30

conformal symmetry, 78 gluino, 39
convex hull, 202 gluino–gluon interaction, 39
Coulomb branch, 78, 259 gluon amplitudes
counterterm, 261 MHV classification, 43–44
CPT, 72, 73, 231 N = 1 supersymmetry Ward identity, 76
critical dimension Dc , 261 N = 4 supersymmetry Ward identity, 80
cross-section, 1, 6 gluon tree amplitudes
IR finite, 122 3-point, 33
spin-sum, 23 CSW, 65–67
CSW, 50, 65–68, 253 helicity violating, 76
CSW prescription, 66, 67 MHV, 33, 44, 96
cut conditions, 124 NMHV 5-point from BCFW, 57, 58
cyclic identity, 34 NMHV 6-point from BCFW, 59–61
Gr(k, n), 185
D 2k R n , 262, 266 Grassmann delta function, 84, 85
de Donder gauge, 250 Grassmann variables η A , 80
dilatation, 97 Grassmannian, 185
dilaton, 254 orthogonal, 246
dilaton–axion gravity, 254, 270 residue theorem, 60, 193
dilogarithm, 118 Grassmannian integral
dimensional analysis, 39 momentum space form, 185, 186
dimensional regularization, 119 supertwistor form, 184
double-copy relation, 280, 281 gravi-photon, 255
dual inversion, 104 gravitino, 38, 254
dual space yi , 103, 104 graviton tree amplitudes
dual superconformal symmetry, 95, 102–113 3-point, 40, 251
in 3d ABJM theory, 238 4-point, 40, 59, 251
all-minus from R 3 , 40
e− e+ annihilation, 2, 6 BCFW, 65
E 7(7) symmetry, 258, 267, 269 CSW, 68
effective field theory, 260 MHV classification, 44
Einstein–Hilbert action, 40, 65, 249 n-point MHV, 252
embedding formalism, 102 gravitons, 38, 40, 250
extended supersymmetry, 72, 73 gravity=(gauge theory)2 , 36, 41, 59,
external data, 2, 3 253–256
F 3 , 40 helicity, 16, 19
factorization, 49, 52 helicity amplitude, 20
Fermi symmetry, 5 helicity operator, 98
Feynman rules helicity violating, 42–44, 76–77, 252
color-ordered, 32, 118 Higgs-gluon fusion, 41
fermions, 18 higher derivative operators
gluons, 32, 118 D 2k R n , 262, 266
gravitinos, 38 F 3 , 40
gravitons, 38, 250 R 3 , 40, 263
vectors, 25 R 4 , 263, 265, 267–268
Yang–Mills theory, 32, 118 R 5 , 265
Fierz identity, 22
fine structure constant, 3 incidence relations, 101, 107
forward limit, 139, 151–155 infinity twistor, 202
inversion, 99
gauge invariance inversion (dual space), 104
polarization vectors, 25 IR divergence, 120–124
generalized gauge transformation (BJC), 275
generalized unitarity, 125, 162 Jacobian, 165
Index 321
κ (gravitational coupling), 40 graviton tree amplitude, 252–253

κCS (Chern–Simons level), 224 N = 8 supergravity superamplitude, 257
Klein–Gordon equation, 41 superamplitude tree-level, 84, 96
Kleiss–Kuijf relations, 35, 276 superamplitude 1-loop 4-point, 130,
KLT relations, 41, 59, 253 132
Kostka number, 85 superamplitude 2-loop 4-point, 135
superamplitude 3-loop 4-point, 136
Lagrangian MHV vertex expansion, 50, 65–68
gravity, 40, 65, 249, 250 moduli space, 78
N = 4 SYM, 77 Møller scattering, 2
φ 4 , 64 momentum
QCD, 30 complex, 17, 37
QED, 26 real, 17, 37
scalar-QED, 29, 62 momentum conservation, 3, 23
supersymmetric chiral model, 69, 72 momentum supertwistors, 110, 142
Yang–Mills theory, 30, 62 momentum twistors, 107–113, 138–159
Yukawa theory, 19 multi-peripheral form, 275
Yukawa with Majorana-fermion, 24
λi λ̃i , 18 n-gluon tree amplitude, 33, 44, 96
large z, 52 n-graviton tree amplitude, 252–253
gluon amplitudes, 54, 61 n-simplex, 203
graviton amplitudes, 252 volume, 205
Leading Singularity, 162, 164 N = 1 super Yang–Mills, 73
composite, 173 N = 1 supergravity, 255
in 3d, 242 N = 2 super Yang–Mills, 73
link variables, 183 N = 2 supergravity, 255
little group N = 4 super Yang–Mills, 64, 73, 77–86
3d, 218, 223 spectrum, 79
4d, 37, 218 Ward identities, 79–80
6d, 218, 219 N = 4 supergravity, 255, 289, 293
D dimensions, 216 N = 8 supergravity, 256–258, 261–269
little group scaling, 37–41 UV structure, 261
determines 3-point amplitudes, 38–39 natural units, 2
locality, 39, 44, 51, 181, 193, 194, 206, 214 N K MHV
loop amplitude, 118 definition, 44, 80, 252
loop-integral, 118 sector, 80
loop-integrand, 118, 138–141 superamplitude, 83
4-point MHV, 162–167 tree superamplitude, 146–149
5-point MHV, 170 NMHV
n-point MHV, 171 definition, 44
low-energy theorems, 258, 267, 269 gluon amplitudes, 57–61
superamplitude (tree), 93, 111, 112
Mandelstam invariants, 23 superamplitude 1-loop 6-point, 134
massive particles, 259, 295 no-triangle, 131
matrix element, 41 non-linear sigma models, 295
maximal cut, 127, 161
MHV off-shell superspace, 81
3-gluon, 33 on-shell amplitude, 3
3-graviton, 251 on-shell condition, 2, 15
ABDK/BDS Ansatz, 136–137 on-shell diagram, 175
BCFW proof (gluons), 54–57 in 3d, 242
classification gluons, 41–44 on-shell superspace, 80–82
classification gravitons, 44, 252 orthogonal Grassmannian, 246
classification with SU (4), 80
definition, 43 Parke–Taylor formula, 33, 44, 96
gluon tree amplitude, 33, 44, 96 inductive proof of, 54–57
322 Index
partial amplitude, 31 scalar theory λφ 4 , 64

Passarino–Veltman reduction, 127 scalar-QED, 29, 58, 62
pentagon integral, 169 Schouten identity, 24, 25
perturbative gravity, 249 selectrons-spositrons scattering, 29
perturbative N = 8 supergravity, 261 shifted amplitude, 51
φ 4 , 64 shifted momentum, 50, 51
photon decoupling identity, 35 BCFW, 53
planar, 118 valid (good) shift, 52
Poincaré algebra, 69 sigma model, 41
Poincaré generators, 95 simplex, 203
polarizations volume, 205
spin-1, 25, 38 soft factor, 46
spin-2, 38, 250 soft function, 46
pole at infinity, 52, 144 soft limits
polylogarithms, 118 loop-level, 121
polytopes, 202 tree-level, 46, 47, 61
propagator matrix, 278 special kinematics, 27
pure gravity, 250 3d, 223, 225
pure spinor formalism, 270 6d, 220
pure supergravity, 255 BCFW, 55
spin sum, 23
QCD, 8, 30 spin-sum completeness, 18
QED processes, 2, 25–30 spinor helicity formalism
3d, 223
R-invariant Ri jk , 92, 104, 109–111 4d, 17
R-symmetry, 73, 78, 258 6d, 219
R 3 , 40, 263 massive particles, 295
R 4 , 263, 265, 267–268 split-helicity, 59, 61
R 5 , 265 splitting function, 48
ratio function, 133 spurious pole, 45, 61
rational terms, 129 square bracket, 20
real momentum, 17, 37 square move, 167, 175
reality condition, 17 square spinor, 17
recursion relations string theory amplitudes, 295
BCFW, 50, 53–62 subamplitude, 49, 52
boundary term (absence of), 52 super Yang-Mills, 63, 64
CSW, 50 N = 1, 73, 75, 77
in momentum twistor space, 142–159 N = 2, 73
off-shell vs. on-shell, 50 N = 4, 73, 77–86
shifted momentum, 51 N < 4, 294
super-BCFW, 86–94 super-BCFW, 86–94
reference spinor, 25 in 3d, 233–238
choosing, 27 in momentum twistor space,
in CSW, 65–67 142–159
independence of, 26, 29, 45 shift, 86
reflection identity, 34 super-Poincaré, 70
region variables, 103 super-QED, 73
relation angle square spinors, 17 super-sum, 86, 87, 130
remainder function, 136 superamplitude, 82–86
renormalization, 120 ABDK/BDS Ansatz, 136–137
Risager shift, 65, 68 MHV classification, 82
RSV connected prescription, 194 MHV tree, 84, 96
MHV 1-loop 4-point, 130, 132
s, t, u, 23 MHV 2-loop 4-point, 135
S-matrix, 7, 122, 125 MHV 3-loop 4-point, 136
scalar integrals, 122 N K MHV tree, 94, 146–149
Index 323
NMHV tree, 93, 111, 112 solution, 85

MHV 1-loop 6-point, 134 supergravity, 252, 256
supercharges, 69–74, 76, 79, 81–84 supertwistor, 100
superconformal symmetry, 78 superwavefunction, 80–82
3d, 226
dual, 95, 102–113 the Symbol, 137
ordinary, 95–99 θi j
superfield, 81 definition, 93
supergravity dual space θi , 104
N = 1, 255 transcendentality, 119, 134, 137, 240
N = 2, 255 triangle integral, 123, 124
N = 4, 255, 289, 293 twistor, 100, 102, 296
N = 8, 256–258, 261–269 twistor string, 101, 193, 248
N = 8 superamplitude, 257
N = 8 UV structure, 261 U (1) decoupling identity, 34, 276
UV structure for various N , L, and D, unitarity, 7, 125, 214
269–271 unitarity cut, 125
supermomentum conservation, 84 UV divergence, 119, 120
supermultiplet, 72–73
supershift, 86 vevs, 78, 250
superspace
off-shell, 72 Ward identity (gauge), 25
on-shell, 80–82 Ward identity (supersymmetry), 73–77, 79–80
supersymmetry, 69 Weyl equation, 17
algebra, 70 Wilson loops, 295,
central charge, 73
extended, 72, 73 i jk (definition), 92
R-symmetry, 73
transformation, 69 Yang–Baxter move, 244
supersymmetry Ward identities, 73–77, Yang–Mills theory, 30–37, 62
79–80 Yangian, 106, 181, 238
gluon MHV amplitudes, 80 yi j
gluon–gluino MHV, 77 definition, 90, 93
helicity violation, 76, 252 dual space yi , 103
N = 1 chiral scalar model, 73–75 Young diagram, 85
N = 1 SYM, 75–77 Yukawa theory, 19
N = 4 SYM, 79–80
NMHV, 77 zone variables, 103

Henriette Elvang, Yu-Tin Huang - Scattering Amplitudes in Gauge Theory and Gravity-Cambridge University Press (2015) - 1

Uploaded by

Copyright:

Available Formats

Henriette Elvang, Yu-Tin Huang - Scattering Amplitudes in Gauge Theory and Gravity-Cambridge University Press (2015) - 1

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Henriette Elvang, Yu-Tin Huang - Scattering Amplitudes in Gauge Theory and Gravity-Cambridge University Press (2015) - 1

Uploaded by

Copyright:

Available Formats

Scattering Amplitudes in Gauge Theory and Gravity

Providing a comprehensive, pedagogical introduction to scattering amplitudes in gauge

Henriette Elvang is Associate Professor in the Department of Physics, University of Michigan.

Cambridge University Press is a part of the University of Cambridge.

Thank you, Mormor.

2 Spinor helicity formalism 15

3 On-shell recursion relations at tree-level 50

6 Loop amplitudes and generalized unitarity 117

7 BCFW recursion for loops 138

8 Leading Singularities and on-shell diagrams 161

Part III Topics

11 Amplitudes in dimensions D = 4 216

11.3.2 3d SYM and Chern–Simons theory 224

12 Supergravity amplitudes 249

13 A colorful duality 272

14 Further reading 294

Appendix Conventions for 4d spinor helicity formalism 297

Yu-tin Huang and Henriette Elvang

What is a scattering amplitude?

tree-level 1-loop 2-loop

From this, one can build diagrams such as

One converts a Feynman diagram to a mathematical expression by tracing through the

Here ( p1 + p2 )2 = −( p10 + p20 )2 + | p1 + p2 |2 etc.

We have used momentum conservation p1 + p2 + p3 + p4 = 0 to simplify the fourth

Beyond Feynman diagrams

Some of the keywords for the topics we explore are:

1. spinor helicity formalism;

7. Leading Singularities and on-shell blob-diagrams ;

8. the Grassmannian, polytopes, amplituhedrons, and mathematicians;

and much more.

Chapter 4 introduces supersymmetry and superamplitudes. In Chapter 5 we explore

2.1 Dirac spinors

The Lagrangian for a free massive 4-component Dirac field is

with ∂/ = ∂μ γ μ . Multiplying the Dirac equation by (i ∂/ + m) gives the Klein-Gordon equa-

(x) ∼ u( p) ei p.x + v( p) e−i p.x , (2.3)

( p/ + m)u( p) = 0 and p + m)v( p) = 0 ,

= d 3 p is the 3d Lorentz-invariant momentum measure. For one finds a

2.2 Spinor helicity notation

We focus on v± ( p) and u ± ( p) as the wavefunctions associated with outgoing anti-fermions

We write the two independent solutions to the Dirac equation (2.9) as

pȧb | p]b = 0 , pa ḃ | pḃ = 0 , [ p|b pbȧ = 0 ,  p|ḃ p ḃa = 0 . (2.12)

[ p|a =  ab | p]b , | pȧ =  ȧ ḃ  p|ḃ . (2.13)

p μ real : [ p|a = (| pȧ )∗ and  p|ȧ = (| p]a )∗ . (2.14)

Spinor completeness relation. The spin-sum completeness relation with m = 0 reads

− /p = | p[ p| + | p] p| . (2.15)

pa ḃ = −| p]a  p|ḃ , pȧb = −| pȧ [ p|b . (2.16)

p μ = (E, E sin θ cos φ, E sin θ sin φ, E cos θ ) . (2.17)

Express pa ḃ and pȧb in terms of E, sin θ2 , cos θ2 and e±iφ .

2.3 Examples from Yukawa theory

 p q =  p|ȧ |qȧ , [ p q] = [ p|a |q]a . (2.22)

 p q = −q p , [ p q] = −[q p] . (2.23)

In amplitudes with light-like momenta p1 , p2 , . . . we use the short-hand notation |1 =

theory. In our notation, f denotes an outgoing fermion and f¯ an outgoing anti-fermion.

Translating the uv-products to spinor brackets we find

12[12] = 2 p1 . p2 = ( p1 + p2 )2 = ( p3 + p4 )2 = 2 p3 . p4 = 34[34] . (2.28)

Angle-square brackets appear often, so it is useful to record the following properties:

[k|γ μ | p =  p|γ μ |k] , (2.33)

[k|γ μ | p∗ = [ p|γ μ |k (for real momenta). (2.34)

 p|P|k] =  p|ȧ P ȧb |k]b =  p|ȧ (−|Pȧ [P|b ) |k]b = − p P[Pk] , (P 2 = 0) .

Here ( p1 + p2 )2 = −( p10 + p20 )2 + | p1 + p2 |2 etc.

pȧb | p]b = 0 , pa ḃ | pḃ = 0 , [ p|b pbȧ = 0 , p|ḃ p ḃa = 0 . (2.12)

[ p|a = ab | p]b , | pȧ = ȧ ḃ p|ḃ . (2.13)

p μ real : [ p|a = (| pȧ )∗ and p|ȧ = (| p]a )∗ . (2.14)

− /p = | p[ p| + | p] p| . (2.15)

pa ḃ = −| p]a p|ḃ , pȧb = −| pȧ [ p|b . (2.16)

p q = p|ȧ |qȧ , [ p q] = [ p|a |q]a . (2.22)

p q = −q p , [ p q] = −[q p] . (2.23)

In amplitudes with light-like momenta p1 , p2 , . . . we use the short-hand notation |1 =

12[12] = 2 p1 . p2 = ( p1 + p2 )2 = ( p3 + p4 )2 = 2 p3 . p4 = 34[34] . (2.28)

[k|γ μ | p = p|γ μ |k] , (2.33)

[k|γ μ | p∗ = [ p|γ μ |k (for real momenta). (2.34)

p|P|k] = p|ȧ P ȧb |k]b = p|ȧ (−|Pȧ [P|b ) |k]b = − p P[Pk] , (P 2 = 0) .

1|γ μ |2]3|γμ |4] = 213[24] (2.36)

|k = a|i + b| j for some a and b. (2.47)

|i jk + | jki + |ki j = 0 . (2.48)

ri jk + r jki + r ki j = 0 . (2.49)

μ p|γ μ |q] μ q|γ μ | p]

(a) Show that ˜± ( p)2 = 0 and ˜± ( p) · p = 0.

|1 ∝ |2 ∝ |3 . (2.59)

We have introduced | p and | p] as solutions to the massless Weyl equation, p| p = 0 and

A tree-level amplitude is a rational function of Lorentz invariants, such as pi μ p μj and pi μ μj ,

|î] = |i] + z | j] , | ĵ] = | j] , |î = |i , | ĵ = | j − z|i . (3.9)

[i, j [−, − [−, + [+, + [+, −

| P̂1n [ P̂1n n] = − P̂1n |n] = −( p̂1 + pn )|n] = |1[1̂n] = 0 , (3.19)

1̂ P̂23 4 1 [3 P̂23 ]3

1̂ P̂3 1 2̂34

• 1̂ P̂[ P̂2] = −1̂|1̂+5|2] = 1̂5[52] = −15[25] (since |1̂ = |1).