1

Quantum Field Theory

A Diagrammatic Approach

Ronald Kleiss
2

Simon Partlic (T`ŕe`

śt́,1590- ?, 1649)
astronomer, mathematician and physician

”...The Ancients were wont to draw Diagrams &

thus divine Predictions for future Happenings, by
Arts magickal or conjectural... likewise the Sa-
vants of the Future will learn to employ Diagrams ;
yet not by Arts magickal, rather by Arts arith-
metickal, algebraickal & by Geometrie and the
Quadrature will they study to foretell the Events
of Nature...”

(attributed to Simon Partlić)

Contents

0 Introductory remarks 19
0.1 Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
0.2 Basic tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
0.2.1 Units and fundamental units . . . . . . . . . . . . . . . 20
0.2.2 Planck units . . . . . . . . . . . . . . . . . . . . . . . 21
0.2.3 Charges . . . . . . . . . . . . . . . . . . . . . . . . . . 22
0.2.4 Conventions . . . . . . . . . . . . . . . . . . . . . . . . 23
0.3 The P 4 Hall of Fame . . . . . . . . . . . . . . . . . . . . . . . 26
0.4 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

1 QFT in zero dimensions 29

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.2 Probabilistic considerations . . . . . . . . . . . . . . . . . . . 30
1.2.1 Quantum field and action . . . . . . . . . . . . . . . . 30
1.2.2 Green’s functions, sources and the path integral . . . . 30
1.2.3 Connected Green’s functions . . . . . . . . . . . . . . . 31
1.2.4 The free theory . . . . . . . . . . . . . . . . . . . . . . 32
1.2.5 The ϕ4 model and perturbation theory . . . . . . . . . 33
1.2.6 The Schwinger-Dyson equation . . . . . . . . . . . . . 35
1.2.7 The Schwinger-Dyson equation for the field function . . 36
1.3 Diagrammatics . . . . . . . . . . . . . . . . . . . . . . . . . . 37
1.3.1 Feynman diagrams . . . . . . . . . . . . . . . . . . . . 37
1.3.2 Feynman rules . . . . . . . . . . . . . . . . . . . . . . . 38
1.3.3 Symmetries and multiplicities . . . . . . . . . . . . . . 39
1.3.4 Vacuum bubbles . . . . . . . . . . . . . . . . . . . . . 42
1.3.5 An equation for connected graphs . . . . . . . . . . . . 42
1.3.6 Semi-connected graphs and the SDe . . . . . . . . . . . 44
1.3.7 The path integral as a set of diagrams . . . . . . . . . 45

3
4 CONTENTS

1.3.8 Dyson summation . . . . . . . . . . . . . . . . . . . . . 46

1.4 Planck’s constant . . . . . . . . . . . . . . . . . . . . . . . . . 48
1.4.1 The loop expansion, and reverse engineering . . . . . . 48
1.4.2 The classical limit . . . . . . . . . . . . . . . . . . . . . 50
1.4.3 On second quantisation . . . . . . . . . . . . . . . . . . 52
1.4.4 Instanton contributions . . . . . . . . . . . . . . . . . . 52
1.5 The effective action . . . . . . . . . . . . . . . . . . . . . . . . 53
1.5.1 The effective action as a Legendre transform . . . . . . 53
1.5.2 Diagrams for the effective action . . . . . . . . . . . . . 55
1.5.3 Computing the effective action . . . . . . . . . . . . . . 56
1.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

2 On renormalization 65
2.1 Doing physics : mentality against reality . . . . . . . . . . . . 65
2.1.1 Physics vs. Mathematics . . . . . . . . . . . . . . . . . 65
2.1.2 The renormalisation program : an example . . . . . . . 66
2.2 A handle on loop divergences . . . . . . . . . . . . . . . . . . 68
2.2.1 A toy : the dot model . . . . . . . . . . . . . . . . . . 68
2.2.2 Nonrenormalizable theories . . . . . . . . . . . . . . . . 71
2.3 Scale dependence . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.3.1 Scale-inpendent scale dependence . . . . . . . . . . . . 72
2.3.2 Low-order approximation to the renormalised coupling 75
2.3.3 Scheme dependence . . . . . . . . . . . . . . . . . . . . 77
2.3.4 Theories with more parameters . . . . . . . . . . . . . 78
2.3.5 Failure of the dot model . . . . . . . . . . . . . . . . . 79
2.4 Asymptotics of renormalisation in ϕ4 theory . . . . . . . . . . 79
2.5 The method of counterterms . . . . . . . . . . . . . . . . . . . 81
2.5.1 Counterterms in the action . . . . . . . . . . . . . . . . 81
2.5.2 Return to the dot model, and a preview . . . . . . . . 82
2.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

3 More fields in zero dimensions 85

3.1 The action and the path integral . . . . . . . . . . . . . . . . 85
3.2 Connected Green’s functions and field functions . . . . . . . . 86
3.3 The Schwinger-Dyson equation . . . . . . . . . . . . . . . . . 87
3.4 The sum rules revisited . . . . . . . . . . . . . . . . . . . . . . 88
3.5 A zero-dimensional toy for QED . . . . . . . . . . . . . . . . 89
3.5.1 Fields and sources . . . . . . . . . . . . . . . . . . . . 89
CONTENTS 5

3.5.2 Bald, furry and quenched toys . . . . . . . . . . . . . . 91

3.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

4 QFT in Euclidean spaces 95

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.2 One-dimensional discrete theory . . . . . . . . . . . . . . . . . 95
4.2.1 An infinite number of fields . . . . . . . . . . . . . . . 95
4.2.2 Introducing the propagator . . . . . . . . . . . . . . . . 97
4.2.3 Computing the propagator . . . . . . . . . . . . . . . . 98
4.2.4 Figments of the imagination : a sermon . . . . . . . . . 100
4.3 One-dimensional continuum theory . . . . . . . . . . . . . . . 102
4.3.1 The continuum limit for the propagator . . . . . . . . . 102
4.3.2 The continuum limit for the action . . . . . . . . . . . 103
4.3.3 The continuum limit of the classical equation . . . . . 105
4.3.4 The continuum Feynman rules and SDe . . . . . . . . . 105
4.3.5 Field configurations in one dimension . . . . . . . . . . 106
4.4 The momentum representation . . . . . . . . . . . . . . . . . . 109
4.4.1 Fourier transforming the SDe . . . . . . . . . . . . . . 109
4.5 Doing it in momentum space . . . . . . . . . . . . . . . . . . . 110
4.5.1 The Feynman rules . . . . . . . . . . . . . . . . . . . . 110
4.5.2 Some example diagrams . . . . . . . . . . . . . . . . . 111
4.6 More-dimensional theories . . . . . . . . . . . . . . . . . . . . 113
4.6.1 The more-dimensional continuum . . . . . . . . . . . . 113
4.6.2 The propagator, explicitly . . . . . . . . . . . . . . . . 116
4.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

5 QFT in Minkowski space 121

5.1 Moving to Minkowski space : making time . . . . . . . . . . . 121
5.1.1 Distance in Minkowski space . . . . . . . . . . . . . . . 121
5.1.2 Farewell probability, hello SDe . . . . . . . . . . . . . . 122
5.1.3 A close look at almost nothing: i and − . . . . . . . . 124
5.1.4 The need for quantum transition amplitudes . . . . . . 125
5.1.5 Feynman rules for Minkowskian theories . . . . . . . . 126
5.1.6 The propagator, explicitly . . . . . . . . . . . . . . . . 126
5.2 Moving in Minkowski space : particles . . . . . . . . . . . . . 131
5.2.1 The Klein-Gordon equation . . . . . . . . . . . . . . . 131
5.2.2 Enter the particle ! . . . . . . . . . . . . . . . . . . . . 132
5.2.3 Unstable particles, i and the flow of time . . . . . . . 133
6 CONTENTS

5.2.4 The Yukawa potential . . . . . . . . . . . . . . . . . . 135

5.2.5 Kinematics and Newton’s First Law . . . . . . . . . . . 136
5.2.6 Antimatter . . . . . . . . . . . . . . . . . . . . . . . . 139
5.2.7 Counting states : the phase-space integration element . 143
5.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144

6 Scattering processes 147

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.2 Incursion into the scattering process . . . . . . . . . . . . . . . 147
6.2.1 Diagrammatic picture of scattering . . . . . . . . . . . 147
6.2.2 The argument for connectedness . . . . . . . . . . . . . 149
6.3 Building predictions . . . . . . . . . . . . . . . . . . . . . . . 151
6.3.1 General formulæ for decay widhts and cross sections . . 151
6.3.2 The truncation bootstrap . . . . . . . . . . . . . . . . 152
6.3.3 A check on dimensionalities . . . . . . . . . . . . . . . 157
6.3.4 Crossing symmetry . . . . . . . . . . . . . . . . . . . . 158
6.4 Unitarity issues . . . . . . . . . . . . . . . . . . . . . . . . . . 159
6.4.1 Unitarity of the S matrix . . . . . . . . . . . . . . . . 159
6.4.2 The cutting rules . . . . . . . . . . . . . . . . . . . . . 163
6.4.3 Infrared cancellations in QED . . . . . . . . . . . . . . 165
6.5 Some example calculations . . . . . . . . . . . . . . . . . . . . 167
6.5.1 The FEE model . . . . . . . . . . . . . . . . . . . . . . 167
6.5.2 Two-body phase space . . . . . . . . . . . . . . . . . . 168
6.5.3 A decay process . . . . . . . . . . . . . . . . . . . . . . 169
6.5.4 A scattering process . . . . . . . . . . . . . . . . . . . 170
6.6 The one-loop cookbook . . . . . . . . . . . . . . . . . . . . . . 172
6.6.1 The one-loop calculation . . . . . . . . . . . . . . . . . 172
6.6.2 Dispersion relations . . . . . . . . . . . . . . . . . . . . 179
6.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

7 Dirac particles 185

7.1 Pimp my propagator . . . . . . . . . . . . . . . . . . . . . . . 185
7.1.1 Down with dyads ! . . . . . . . . . . . . . . . . . . . . 186
7.1.2 The spin interpretation . . . . . . . . . . . . . . . . . . 187
7.2 The Dirac algebra . . . . . . . . . . . . . . . . . . . . . . . . . 188
7.2.1 The Dirac matrices . . . . . . . . . . . . . . . . . . . 188
7.2.2 The Clifford algebra . . . . . . . . . . . . . . . . . . . 191
7.2.3 Trace identities . . . . . . . . . . . . . . . . . . . . . . 192
CONTENTS 7

7.2.4 Pauli and Chisholm, their identities . . . . . . . . . . . 194

7.2.5 Hermite no, Dirac yes . . . . . . . . . . . . . . . . . . 195
7.2.6 Flipping in and flipping out . . . . . . . . . . . . . . . 197
7.2.7 A Fierz identity . . . . . . . . . . . . . . . . . . . . . . 198
7.3 Dirac spinors . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
7.3.1 Chirality spinors . . . . . . . . . . . . . . . . . . . . . 199
7.3.2 Chirality spinors are massless . . . . . . . . . . . . . . 200
7.3.3 Phase conventions, spinor products and Weyl . . . . . 201
7.3.4 General spinors and their dyads . . . . . . . . . . . . . 202
7.3.5 General spinors and Dirac spinors . . . . . . . . . . . . 204
7.3.6 Particular general spinors . . . . . . . . . . . . . . . . 205
7.4 Dirac particles . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
7.4.1 Casimir does tricks . . . . . . . . . . . . . . . . . . . . 207
7.4.2 The Dirac propagator, and a convention . . . . . . . . 209
7.4.3 Truncating Dirac particles : external Dirac lines . . . . 211
7.4.4 Lorentz transformations in Dirac space . . . . . . . . . 213
7.4.5 The spin of Dirac particles . . . . . . . . . . . . . . . . 214
7.4.6 Massless Dirac particles ; helicity states . . . . . . . . 215
7.4.7 The parity transform . . . . . . . . . . . . . . . . . . . 217
7.5 The Feynman rules for Dirac particles . . . . . . . . . . . . . . 218
7.5.1 Dirac loops. . . . . . . . . . . . . . . . . . . . . . . . . . 218
7.5.2 . . . and Dirac loops only . . . . . . . . . . . . . . . . . 219
7.5.3 Interchange signs . . . . . . . . . . . . . . . . . . . . . 220
7.5.4 The Pauli principle . . . . . . . . . . . . . . . . . . . . 222
7.6 The Dirac equation . . . . . . . . . . . . . . . . . . . . . . . . 223
7.6.1 The classical limit . . . . . . . . . . . . . . . . . . . . . 223
7.6.2 The free Dirac action . . . . . . . . . . . . . . . . . . . 224
7.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225

8 Helicity techniques for Dirac particles 229

8.1 The standard form for spinors . . . . . . . . . . . . . . . . . . 229
8.1.1 Opting for helicities, opting for antisymmetry . . . . . 229
8.1.2 The standard form for helicity spinors . . . . . . . . . 230
8.1.3 Some useful identities . . . . . . . . . . . . . . . . . . . 231
8.1.4 How to compute spinor products . . . . . . . . . . . . 232
8.1.5 The standard form for massive particles . . . . . . . . 233
8.1.6 The standard form for complex momenta . . . . . . . . 234
8.2 Summary of tools for spinor techniques . . . . . . . . . . . . 235
8 CONTENTS

8.3 Fermionic decays : the Fermi model . . . . . . . . . . . . . . . 236

8.3.1 The amplitude for muon decay . . . . . . . . . . . . . 236
8.3.2 Three-body phase space . . . . . . . . . . . . . . . . . 238
8.3.3 The muon decay width . . . . . . . . . . . . . . . . . . 239
8.3.4 Observable distributions in muon decay . . . . . . . . . 240
8.3.5 Charged pion decay: helicity suppression . . . . . . . . 242
8.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243

9 Vector particles 247

9.1 Massive vector particles . . . . . . . . . . . . . . . . . . . . . 247
9.1.1 The propagator . . . . . . . . . . . . . . . . . . . . . . 247
9.1.2 The Feynman rules for external vector particles . . . . 248
9.1.3 The spin of vector particles . . . . . . . . . . . . . . . 249
9.1.4 Polarisation vectors for helicity states . . . . . . . . . . 251
9.1.5 The Proca equation . . . . . . . . . . . . . . . . . . . . 252
9.2 The spin-statistics theorem . . . . . . . . . . . . . . . . . . . . 253
9.2.1 Spinorial form of vector polarisations . . . . . . . . . . 253
9.2.2 Proof of the spin-statistics theorem . . . . . . . . . . . 255
9.3 Massless vector particles . . . . . . . . . . . . . . . . . . . . . 256
9.3.1 Polarisations of massless vector particles . . . . . . . . 256
9.3.2 Current conservation from the polarisation . . . . . . . 257
9.3.3 Handlebar condition for massive vector particles . . . . 259
9.3.4 Helicity states for massless vectors . . . . . . . . . . . 259
9.3.5 The massless propagator : the axial gauge . . . . . . . 261
9.3.6 Gauge vector shift . . . . . . . . . . . . . . . . . . . . 262
9.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263

10 Quantum Electrodynamics 267

10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
10.2 Constructing QED . . . . . . . . . . . . . . . . . . . . . . . . 267
10.2.1 The QED vertex . . . . . . . . . . . . . . . . . . . . . 267
10.2.2 Handlebars : a first look . . . . . . . . . . . . . . . . . 268
10.2.3 Handlebar diagrammatics . . . . . . . . . . . . . . . . 270
10.2.4 The Ward-Takahashi identity . . . . . . . . . . . . . . 272
10.2.5 The charged Dirac equation . . . . . . . . . . . . . . . 275
10.2.6 The Gordon decomposition . . . . . . . . . . . . . . . 276
10.2.7 Furry’s theorem . . . . . . . . . . . . . . . . . . . . . . 277
10.3 Some QED processes . . . . . . . . . . . . . . . . . . . . . . . 279
CONTENTS 9

10.3.1 A classic calculation : muon pair production . . . . . . 279

10.3.2 Compton and Thomson scattering . . . . . . . . . . . 283
10.3.3 Electron-positron annihilation . . . . . . . . . . . . . . 286
10.3.4 Bhabha scattering . . . . . . . . . . . . . . . . . . . . 288
10.3.5 Bremsstrahlung in Mœller scattering . . . . . . . . . . 290
10.4 Scalar electrodynamics . . . . . . . . . . . . . . . . . . . . . . 298
10.4.1 The vertices . . . . . . . . . . . . . . . . . . . . . . . . 298
10.4.2 Proof of current conservation in sQED . . . . . . . . . 300
10.5 The Coulomb potential . . . . . . . . . . . . . . . . . . . . . . 302
10.6 Electrons in external fields : g = 2 . . . . . . . . . . . . . . . . 304
10.6.1 The charged Klein-Gordon equation . . . . . . . . . . . 304
10.6.2 The relativistic Pauli equation . . . . . . . . . . . . . . 305
10.6.3 A constant magnetic field . . . . . . . . . . . . . . . . 306
10.7 Selected topics in QED . . . . . . . . . . . . . . . . . . . . . . 307
10.7.1 Three-photon production . . . . . . . . . . . . . . . . . 307
10.7.2 The Thomson limit : scalar vs spinor . . . . . . . . . . 310
10.7.3 The Landau-Yang theorem . . . . . . . . . . . . . . . 314
10.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316

11 Loop effects in QED 321

11.1 One-loop effects in QED . . . . . . . . . . . . . . . . . . . . . 321
11.2 The photon self-energy . . . . . . . . . . . . . . . . . . . . . . 321
11.2.1 Current conservation . . . . . . . . . . . . . . . . . . . 321
11.2.2 Using the optical theorem . . . . . . . . . . . . . . . . 322
11.2.3 Getting the divergence . . . . . . . . . . . . . . . . . . 323
11.2.4 The vacuum polarization . . . . . . . . . . . . . . . . . 324
11.2.5 Hadronic vacuum polarization . . . . . . . . . . . . . . 325
11.3 The fermion self-energy . . . . . . . . . . . . . . . . . . . . . . 327
11.3.1 A look at gauge invariance . . . . . . . . . . . . . . . . 327
11.3.2 Summing the self-energies . . . . . . . . . . . . . . . . 328
11.3.3 The loop calculation . . . . . . . . . . . . . . . . . . . 329
11.3.4 The curious incident of the divergences in the nighttime 331
11.4 Infrared singularities in Bremsstrahlung . . . . . . . . . . . . . 332
11.5 The vertex correction . . . . . . . . . . . . . . . . . . . . . . . 335
11.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
10 CONTENTS

12 Quantum Chromodynamics 341

12.1 Introduction: coloured quarks and gluons . . . . . . . . . . . . 341
12.2 Quarks and gluons : first Feynman rules . . . . . . . . . . . . 342
12.2.1 The propagators . . . . . . . . . . . . . . . . . . . . . 342
12.2.2 The quark-gluon vertex . . . . . . . . . . . . . . . . . . 343
12.2.3 A closer look at the T matrices . . . . . . . . . . . . . 343
12.2.4 The Fierz identity for T matrices . . . . . . . . . . . . 346
12.3 The three-gluon interaction . . . . . . . . . . . . . . . . . . . 348
12.3.1 The need for three-gluon vertices . . . . . . . . . . . . 348
12.3.2 Furry’s failure . . . . . . . . . . . . . . . . . . . . . . . 350
12.3.3 The ggg vertex and its handlebar . . . . . . . . . . . . 351
12.3.4 On coupling quantisation . . . . . . . . . . . . . . . . . 355
12.4 The four-gluon interaction . . . . . . . . . . . . . . . . . . . . 355
12.4.1 Colourful manipulations . . . . . . . . . . . . . . . . . 355
12.4.2 A purely gluonic process . . . . . . . . . . . . . . . . . 356
12.5 Current conservation in QCD . . . . . . . . . . . . . . . . . . 358
12.5.1 More vertices ? . . . . . . . . . . . . . . . . . . . . . . 358
12.5.2 The Antkaz . . . . . . . . . . . . . . . . . . . . . . . . 359
12.5.3 Proof of current conservation . . . . . . . . . . . . . . 360
12.6 Selected topics in QCD . . . . . . . . . . . . . . . . . . . . . . 361
12.6.1 White and coloured states . . . . . . . . . . . . . . . . 361
12.6.2 The QCD Coulomb interaction . . . . . . . . . . . . . 362
12.6.3 The process q q̄ → gg . . . . . . . . . . . . . . . . . . . 363
12.6.4 The Landau-Yang theorem revisited . . . . . . . . . . . 366
12.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369

13 Electroweak theory 373

13.1 Muon decay . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
13.1.1 The Fermi coupling constant . . . . . . . . . . . . . . . 373
13.1.2 Failure of the Fermi model in µ− ν µ → e− ν e . . . . . 374
13.2 The W particle . . . . . . . . . . . . . . . . . . . . . . . . . . 375
13.2.1 The IVB strategy . . . . . . . . . . . . . . . . . . . . . 375
13.2.2 The cross section for µ− ν µ → e− ν e revisited . . . . . . 378
13.2.3 The W W γ vertex . . . . . . . . . . . . . . . . . . . . . 379
13.3 The Z particle . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
13.3.1 W pair production . . . . . . . . . . . . . . . . . . . . 384
13.3.2 The weak mixing angle for couplings . . . . . . . . . . 388
13.3.3 W, Z and γ four-point interactions . . . . . . . . . . . 389
CONTENTS 11

13.4 The Higgs sector . . . . . . . . . . . . . . . . . . . . . . . . . 394

13.4.1 The Higgs hypothesis . . . . . . . . . . . . . . . . . . . 394
13.4.2 Predictions from the Higgs hypothesis . . . . . . . . . 400
13.4.3 W, Z and H four-point interactions . . . . . . . . . . . 401
13.4.4 Higgs-fermion couplings . . . . . . . . . . . . . . . . . 403
13.4.5 Higgs self-interactions . . . . . . . . . . . . . . . . . . 405
13.5 About anomalies . . . . . . . . . . . . . . . . . . . . . . . . . 409
13.6 Conclusions and remarks . . . . . . . . . . . . . . . . . . . . . 411
13.7 A look at non-minimal models . . . . . . . . . . . . . . . . . . 412
13.7.1 Non-minimal Higgs sector . . . . . . . . . . . . . . . . 412
13.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 418

14 Example computations 421

14.1 Neutrino production in e+ e− scattering . . . . . . . . . . . . . 421
14.1.1 The cross section . . . . . . . . . . . . . . . . . . . . . 421
14.1.2 Unitarity considerations . . . . . . . . . . . . . . . . . 424
14.2 W pair production in e+ e− scattering . . . . . . . . . . . . . . 425
14.2.1 Setting up the amplitude . . . . . . . . . . . . . . . . . 425
14.2.2 Momenta and polarisations . . . . . . . . . . . . . . . 426
14.2.3 Working out the amplitudes . . . . . . . . . . . . . . . 427
14.2.4 W pair production at very high energy . . . . . . . . . 430
14.3 Higgs coupling to massless vectors . . . . . . . . . . . . . . . . 431
14.3.1 The γγH vertex . . . . . . . . . . . . . . . . . . . . . . 431
14.3.2 The ggH amplitude and the Next Generation . . . . . 435

15 Appendices 437
15.1 Perturbative (non)convergence issues . . . . . . . . . . . . . . 437
15.1.1 Punishment at the singular point . . . . . . . . . . . . 437
15.1.2 Borel summation . . . . . . . . . . . . . . . . . . . . . 438
15.2 More on symmetry factors . . . . . . . . . . . . . . . . . . . . 441
15.2.1 The origin of symmetry factors . . . . . . . . . . . . . 441
15.2.2 Explicit computation of symmetry factors . . . . . . . 442
15.3 Derivation of the diagrammatic sum rules . . . . . . . . . . . . 445
15.4 Alternative solutions to the Schwinger-Dyson equation . . . . 448
15.4.1 Alternative contours for general theories . . . . . . . . 448
15.4.2 Alternative contours for ϕ3 theory . . . . . . . . . . . . 449
15.4.3 Alternative contours for ϕ4 theory . . . . . . . . . . . . 450
15.5 Diagram counting . . . . . . . . . . . . . . . . . . . . . . . . . 450
12 CONTENTS

15.5.1 Tree graphs and asymptotics . . . . . . . . . . . . . . . 450

15.5.2 Counting one-loop diagrams . . . . . . . . . . . . . . . 455
15.6 Concavity of the effective action . . . . . . . . . . . . . . . . . 461
15.7 Functional derivatives . . . . . . . . . . . . . . . . . . . . . . . 463
15.8 Frustrated and unusual actions . . . . . . . . . . . . . . . . . 464
15.8.1 Frustrating your neighbours . . . . . . . . . . . . . . . 464
15.8.2 Increasing frustration . . . . . . . . . . . . . . . . . . . 467
15.9 Newton’s First Law revisited . . . . . . . . . . . . . . . . . . 469
15.9.1 Introduction : the matter of sources . . . . . . . . . . . 469
15.9.2 Slow, fast and abrupt . . . . . . . . . . . . . . . . . . . 471
15.9.3 Conclusion : general effect of the sources . . . . . . . . 473
15.10 Unitarity bounds . . . . . . . . . . . . . . . . . . . . . . . . . 473
15.10.1 Resonances . . . . . . . . . . . . . . . . . . . . . . . . 473
15.10.2 Preliminaries : decay widths . . . . . . . . . . . . . . 474
15.10.3 The rôle of angular momentum conservation . . . . . . 475
15.10.4 The unitarity bound . . . . . . . . . . . . . . . . . . . 476
15.11 The fundamental theorem for Dirac matrices . . . . . . . . . 477
15.11.1 Proof of the fundamental theorem . . . . . . . . . . . 477
15.11.2 The charge conjugation matrix . . . . . . . . . . . . . 480
15.12 States of higher spin . . . . . . . . . . . . . . . . . . . . . . . 480
15.12.1 The spin algebra for integer spins . . . . . . . . . . . . 480
15.12.2 Rank one for spin one . . . . . . . . . . . . . . . . . . 482
15.12.3 Rank-2 tensors . . . . . . . . . . . . . . . . . . . . . . 483
15.12.4 Rank-3 tensors . . . . . . . . . . . . . . . . . . . . . . 485
15.12.5 Massless particles : surviving states . . . . . . . . . . 489
15.12.6 The Bermuda triangle . . . . . . . . . . . . . . . . . . 490
15.12.7 Massless propagators . . . . . . . . . . . . . . . . . . . 491
15.12.8 Spin of the Kalb-Ramond state . . . . . . . . . . . . . 493
15.12.9 Spin 1 from Dirac particles . . . . . . . . . . . . . . . 494
15.12.10 Spin 3/2 particles . . . . . . . . . . . . . . . . . . . . 496
15.12.11 The spin algebra for Dirac particles . . . . . . . . . . 498
15.13 Generating three-particle kinematics . . . . . . . . . . . . . . 499
15.14 The CPT theorem . . . . . . . . . . . . . . . . . . . . . . . . 500
15.14.1 Transforming spinors . . . . . . . . . . . . . . . . . . 501
15.14.2 CPT transformation on sandwiches . . . . . . . . . . . 502
15.14.3 CPT transformation on diagrams . . . . . . . . . . . . 503
15.14.4 How to kill CPT, and what it costs . . . . . . . . . . . 504
15.15 Mathematical Miscellanies . . . . . . . . . . . . . . . . . . . . 506
CONTENTS 13

15.15.1 The Gaussian doubling trick . . . . . . . . . . . . . . 506

15.15.2 Stirling’s approximation for n! . . . . . . . . . . . . . 507
15.15.3 Manipulating asymptotic series . . . . . . . . . . . . . 508
15.15.4 Gamma, Digamma and Bernoulli . . . . . . . . . . . . 511
15.15.5 The Dirac delta distribution . . . . . . . . . . . . . . 513
15.15.6 The principal-value distribution . . . . . . . . . . . . 514
15.15.7 Generating the Bell numbers . . . . . . . . . . . . . . 514
15.15.8 The exponential integral E1 and the Bessel K functions 515
15.15.9 The Kramers-Kronig relation . . . . . . . . . . . . . . 518
15.15.10 The dilogarithm function . . . . . . . . . . . . . . . . 519
15.15.11 Some values of the ζ function . . . . . . . . . . . . . 521
15.15.12 The Lagrange expansion . . . . . . . . . . . . . . . . 522
15.15.13 Determinants from traces . . . . . . . . . . . . . . . 524
14 CONTENTS
List of Figures

1.1 The developing of instanton effects . . . . . . . . . . . . . . . 53

2.1 Improvement factor for ϕ4 theory in zero dimensions . . . . . 81

4.1 Zooming out from a fractal path . . . . . . . . . . . . . . . . . 108

5.1 Closing the contour for positive times . . . . . . . . . . . . . . 127

5.2 Closing the contour for spacelike separation . . . . . . . . . . 129
5.3 Bessel functions K1 for real and imaginary argument . . . . . 130
5.4 Closing the contour for Newton’s first . . . . . . . . . . . . . . 137
5.5 Moving forward ; moving backward . . . . . . . . . . . . . . . 140
5.6 An annihilating spacetime path . . . . . . . . . . . . . . . . . 142

6.1 The function R(m2 , s) of Eq.(6.89) . . . . . . . . . . . . . . . 178

7.1 A Mental Flip : one vase or two faces ? . . . . . . . . . . . . . 197

8.1 Elecrton-mass dependence of the muon decay width . . . . . . 239

10.1 Threshold behaviour for fermions and scalars . . . . . . . . . . 283

11.1 The fermion-induced vacuum polarization . . . . . . . . . . . 325

13.1 Monte Carlo investigation of cancellations in ZZ → ZZH . . 407

13.2 Monte Carlo investigation of cancellations in ZZ → HHH . . 409

14.1 The absolute value of F(M 2 /mH 2 ) defined in Eq.(14.58) . . . . 435

15.1 Example of Borel summation . . . . . . . . . . . . . . . . . . 440

15.2 Propagators for frustrated actions . . . . . . . . . . . . . . . . 469
15.3 Contour shift for a Gaussian source . . . . . . . . . . . . . . . 472

15
16 LIST OF FIGURES

15.4 The functions K0 , K1 , K2 and E1 . . . . . . . . . . . . . . . . 517

15.5 Integration contour for the ramers-Kronig relation . . . . . . . 518
15.6 The dilogarithm function . . . . . . . . . . . . . . . . . . . . . 520
List of Tables

13.1 Contributions to the triangle anomalies . . . . . . . . . . . . . 410

15.1 Exact and asymptotic number of tree diagrams in ϕ3/4 theory 454
15.2 Computing the average symmetry factors . . . . . . . . . . . . 460
15.3 Multiplication table for the Clifford algebra . . . . . . . . . . 478

17
18 August 31, 2019
Chapter 0

Introductory remarks

0.1 Preface

In what follows, whatever is correct I owe to many other people ; that which
is wrong I managed on my own. I am perpetually in need of, and grateful
to, those pointing out typing or thinking errors in these notes1 .

About the appendices : I have collected material that, when included

in the main text, would interfere with the narrative flow. The appendices
relevant to each chapter are indicated below the chapter heading.
About the exercises : these are intended to encourage you towards com-
petence in actually computing things. Theoretical knowledge without dirty
hands is not very good2 . The exercises apposite to a topic are indicated by a
box like Eℵ in the margin, where ‘ℵ’ is the number of the exercise, situated
at the end of the chapter. Finally, a word of caution : I am a physicist, not E0
a mathematician : I will worry about mathematical rigour but not so much
as to become paralysed3 .

1
I cordially invite all and sundry to do so. The P 4 Hall of Fame collects the names of
friends who have helped me in learning about, formulating, contemplating, or execrating
one or several issues.
2
Labori inutilis manus salubris (attr. to S. Partlic)
3
Rigor mathesis protractus mortis rigorem inducet (attr. to S. Partlic).

19
20 August 31, 2019

0.2 Basic tools

0.2.1 Units and fundamental units
The fundamental constants of relativistic quantum field theory are the speed
of light in vacuo 4 :
m
c = 299792458 ,
sec
and Planck’s (or rather Dirac’s) constant5 :
~ = 1.05457182 × 10−34 Joule sec .
Compared to the scales of our everyday experiences, ~ is miniscule and c
is huge ; in the world of elementary particles, they are just about right.
We can see this as follows. It is useful to replace our human-scale meters6 ,
kilograms7 , and seconds8 , by what may be called fundamental HEP units of
mass, length and time9 :
Mf = 1.7827 10−27 kg ,
Lf = 1.9733 10−16 m ,
Tf = 6.5821 10−25 sec .
In terms of these units, we have precisely
Mf Lf 2 Lf
~= , c= ,
Tf Tf
so that both ~ and c have the numerical value one ; and the unit of energy
turns out to be
Mf Lf 2
2 = 1.6022 10−10 Joule = 1 GeV .
Tf
The mass and size of the proton are of the same order as Mf and Lf , re-
spectively, and Tf is roughly the time scale of strong interactions. The use
4
This is exact since it establishes the definition of the metre as of October 21, 1983.
5
This is exact since it establishes the definition of the kilogram as of May 20, 2019.
6
Typically, how far you can reach.
7
Typically, what you would not mind lifting.
8
Typically, how fast you can count things if you are not a teller.
9
I give these numbers with only a few digits, for simplicity. What matters is of course
not the precise value, but rather that such units exist.
August 31, 2019 21

of fundamental units is attractive since you won’t have to write factors of c

and ~, and can express both length and time in inverse GeV, and mass in
GeV. This is customarily one of the first statements in advanced courses :
“we shall adopt units such that c = ~ = 1”. This usage denies the various
objects their dimensionality10 , and I have decided to try to retain the ~’s
and c’s where they belong ; after all, it is much easier to erase them from
formulæ than to put them back in11 .

0.2.2 Planck units

Along with ~ and c there exists a third fundamental constant of nature,
namely Newton’s (or rather Cavendish’s) gravitational constant12 :

m3
G = 6.67408(31) 10−11 .
kg sec2

The truly, ultimately fundamental units of mass, length and time that can
be recovered from c, ~ and GN are then the Planck units :
r
~
MP = = 2.1764 10−8 kg ,
r Gc
~G
LP = 3
= 1.6162 10−35 m ,
r c
~G
TP = 5
= 5.3912 10−44 sec . (1)
c
These values are outrageously far removed from the typical scales of particle
phenomenology. We may interpret this as an indication that in what follows
the gravitational interaction will not play any part. In fact, in any case we do
10
This has the unfortunate consequence of making it impossible to check the (at least)
grammatical correctness of an expression by dimensional analysis, and even leads to erro-
neous statements about ‘classical limits’ and the like : see our discussion of particle masses
in chapter 6.
11
You would be surprised how difficult it is for even well-trained physicists to put the ~
and c back, when asked to do so in a hurry !
12
Taken from P.J. Mohr, D.B. Newell, B.N. Taylor, CODATA recommended values of
the fundamental physical constants: 2014, Rev. Mod. Phys. 88, July-September 2016.
The uncertainty in the last two digits is bracketed.
22 August 31, 2019

not (yet) have a satisfactory quantum theory of gravitation leading to specific

and falsifiable predictions for particle phenomenology. To bring the Planck
units close to the fundamental units we need to increase the value of G, thus
the strength of gravity, by a factor of about 1038 . On a more philosophical
note, consider that we have, basically, only the three measures, of length, of
time, and of mass, to do physics with ; and there are precisely three constants
of nature that we know of. Suppose a fourth one were discovered : it would
be only natural that we would try very hard to relate it to the other three
in some way ; this is how fundamental physics must work.

0.2.3 Charges
The electrostatic charge is adopted to the Gaussian system, so as to have
no truck with the ‘permittivity of the vacuum’ and suchlike : that is, two
charges e1 and e2 separated by a distance r feel a mutual Coulomb force F~
characterized by
1 | e1 e2 |
| F~ | = .
4π r2
√
This implies that the charge has the dimensionality of ~c. It follows that, if
we choose the proton charge as the unit charge e, the fine structure constant

e2
αe =
4π ~ c

is a dimensionless number13 . Experimentally14

1
αe = ,
137.035999139(31)

which yields the result

√ kg1/2 m3/2
e = 0.30282 ~c = 5.3844 10−14 .
sec
13
Meaning that it has the same value in all possible systems of units ! An alien civiliza-
tion in outer space will find the same value.
14
Taken from P.J. Mohr, D.B. Newell, B.N. Taylor, CODATA recommended values of
the fundamental physical constants: 2014, Rev. Mod. Phys. 88, July-September 2016.
The uncertainty in the last two digits is bracketed.
August 31, 2019 23

0.2.4 Conventions
Forgetting phases
If two quantities A and B are equal up to a disregarded overall complex
phase we shall write
AlB .

The step function

The Whittaker (or step) function is a function of a real number :

1 if x ≥ 1
θ(x) = . (2)
0 if x < 0
We extend this to a logical step function of a predicate P :

1 if P is true
θ(P) = . (3)
0 if P is false

The metric
By convention, the Minkowski metric15 has the form

g µν = gµν = diag(1, −1, −1, −1) .

In many textbooks the metric tensor is introduced as a diagonal matrix.

This is of course misleading since the covariant metric tensor has only lower
indices, whereas a matrix has one upper and one lower index. Unfortunately,
the ‘correct’ matrix form of the metric, which would be g µ ν , equals the unit
matrix whatever the metric !

Kronecker’s symbol
The Kronecker symbol is defined by

α 1 if α = µ
δ µ= .
6 µ
0 if α =
15
In the usual Cartesian coordinate systems. Since we are not considering curved space-
time in these notes, we shall adhere to this simplest of coordinate system throughout,
except when discussing phase space integration where polar coordinates often come in
handy. But there we shall not refer to the metric.
24 August 31, 2019

Note : in Minkowski space, Kronecker symbols tend to carry one upper, and
one lower index. Kronecker symbols with two upper or two lower indices are
slightly suspect and to be treated with care unless they refer to some other
label such as color.

The antisymmetrizer
Using Kronecker symbols we can build the following antisymmetric objects :
   
µ1 µ1 µ1 µ2
= δ ν1 , = δ µ1 ν1 δ µ2 ν2 − δ µ1 ν2 δ µ2 ν1 ,
ν1 ν1 ν2
 
µ1 µ2 µ3
= δ µ1 ν1 δ µ2 ν2 δ µ3 ν3 + δ µ1 ν2 δ µ2 ν3 δ µ3 ν1 + δ µ1 ν3 δ µ2 ν1 δ µ3 ν2
ν1 ν2 ν3
− δ µ1 ν1 δ µ2 ν3 δ µ3 ν2 − δ µ1 ν2 δ µ2 ν1 δ µ3 ν3 − δ µ1 ν3 δ µ2 ν2 δ µ3 ν1

and so on. Here we encounter all signed permutations16 of the lower indices.
These are compuationally handy as we see below.

The Levi-Civita symbol

The totally antisymmetric Levi-Civita symbol is defined by

0123 = + 1 hence 0123 = − 1 .

This implies the following identities :

 
ν1 ν2 ν3 ν4 ν1 ν2 ν3 ν4
µ1 µ2 µ3 µ4  = − ,
µ1 µ2 µ3 µ4
 
ν1 ν2 ν3 µ4 ν1 ν2 ν3
µ1 µ2 µ3 µ4  = − ,
µ1 µ2 µ3
 
ν1 ν2 µ3 µ4 ν1 ν2
µ1 µ2 µ3 µ4  = −2 ,
µ1 µ2
 
ν1 µ2 µ3 µ4 ν1
µ1 µ2 µ3 µ4  = −6 = −6 δ ν1 µ1 ,
µ1
µ1 µ2 µ3 µ4 µ1 µ2 µ3 µ4 = −24 .
16
Even permutations occur with a +, and odd permutations with a – sign.
August 31, 2019 25

In order not to lumber ourselves with too many explicit Lorentz indices, we
shall use the following shorthand :
µ (a, b, c) means µνρσ aν bρ cσ and so on.
Finally, although it falls outside the scope of these notes, it is useful to note
that the Kronecker and antisymmetrizer symbols are fully-fledged tensors
in a much wider class of spaces than Minkowski space, but the Levi-Civita
symbols themselves are not. For instance, if we move from Cartesian to
polar coordinates, say, the Kronecker symbol remains unaffected but the
Levi-Civita is changed.

Minus in the momentum

There is a subtlety : the contravariant partial derivative contains a possibly
surprising minus sign :
 
µ ∂ 1∂ ~
∂ = = , −∇ . (4)
∂xµ c ∂t
This explains why in nonrelativistic quantum mechanics the momentum op-
~ whereas in the relativistic theory we use pµ = i~ ∂ µ .
erator is p~ = −i~ ∇

The polyparenthetophobia rule

This deals with the notation for compound vector products. If pµ1,2,3,4 are
four-vectors, then the expression
(p1 + p2 · p3 + p4 )
must be understood to mean
(p1 · p3 ) + (p2 · p3 ) + (p1 · p4 ) + (p2 · p4 )
The more rigorously correct form


(p1 + p2 ) · (p3 + p4 )
is in my opinion less easily readable, unless by true parenthetophiliacs. In
the same spirit, a slight tendency to periodophobia will lead us to write,
e.g., (pq) as shorthand for (p · q) where no risk of confusion is likely. Also, k 2
occurs when it is clear that (k ·k) is intended rather than the second spacelike
component of k µ .
26 August 31, 2019

In favor of loose terminology

Among particle physicists there exist a tendency to be sloppy with some
terms. In particular this holds for the usage of the words ‘mass’ and ‘mo-
mentum’. Strictly speaking, in the Feynman rules to be discussed the ‘mass’
m and the ‘momentum’ k µ have dimensions of inverse length and therefore
cannot be the same as the notions of the classical mass M , expressed in kg,
and the classical momentum pµ , expressed in (kg m/sec) . As discussed in
section 5.2.5, these various notions are related by
m = M c/~ , k µ = pµ /~ .
Not wishing to succumb to pedantry, I shall use ‘mass’ and ‘momentum’
insouciantly. Experience shows that one easily gets used to it.

0.3 The P 4 Hall of Fame

Ernestos Argyres Wolfgang Hollik Harald Niederreiter
Dima Bardin Gerard ’t Hooft Gijs van der Oord
Wim Beenakker Staszek Jadach Kostas Papadopoulos
Frits Berends Frederick James Simon Partlic
Alain Blondel Tim Janssen Giampiero Passarino
Oscar Boher Luna Sijbrand de Jong Roberto Pittau
Michael Borinsky Martijn Jongen Marcel Raas
Stefan Brinck de Heren Jos Frank Redig
Kristof de Bruyn Marcel van Kessel Robbert Rietkerk
Dirk van Buul Jochem Kip Chris Ripken
Sascha Caron Stefan Krieg Tom Rijken
Chris Dams Hans Kühn Bert Schellekens
Pertros Draggiotis Hans Kuijf James Stirling
Helmut Eberl Zoltan Kunszt John Swain
Steve Ellis Achilleas Lazopoulos Oleg Teryaev
Raymond Gastmans Joep Leenaarts Theodor Todorov
Edward Gibbon Yannis Malamos Tini Veltman
Walter Giele Mich. Mangano Rob Verheyen
Jeroen de Groot John March-Russell Jos Vermaseren
André van Hameren Melvin Meijer Tai Tsun Wu
Lisa Hartgring Mark Netjes Sjoerd Ypma
Thank you for allowing me to think together with you !
August 31, 2019 27

0.4 Exercise
Excercise 0 Would-be exercise
This might have been an exercise.
28 August 31, 2019
Chapter 1

QFT in zero dimensions

1.1 Introduction

For the description of elementary particles, a theory including both special

relativity and quantum mechanics is necessary ; we shall introduce relativ-
ity further on, and concentrate in this chapter on the quantum-mechanical
nature of nature. The fundamental object used for describing the particles
is a quantum field. In many treatments quantum fields are considered to be
operator-valued entities ; we shall rather adhere to Feynman’s approach and
use what are called c-number fields. Such a field assigns one or more num-
bers to every point in spacetime, and is hence a pretty complicated subject
the behaviour of which is not to be characterized trivially, especially when
it also undergoes quantum fluctuations. It is therefore useful to first build
up expertise in the various necessary techniques in a more controllable sit-
uation. To this end we shall simplify the whole four-dimensional spacetime
arena of particle physics to a lower-dimensional system ; in fact, we shall
reduce spacetime to a single point, hence a zero-dimensional arena. The
quantum fields are then assignments of a single number ; the simplest quan-
tum field is, in this case, a single stochastic, or random, number. Many of the
techniques of quantum field theory do apply to this case : in particular the
notion of path integrals, Green’s functions, the Schwinger-Dyson equation,
and Feynman diagrams come up naturally.

29
30 August 31, 2019

1.2 Probabilistic considerations

1.2.1 Quantum field and action
Let us imagine a quantum field ϕ that takes its values on the whole real axis
from −∞ to +∞. Since it is a random variable, the most we can specify
about it is its probability density P (ϕ), which we write, for now, as


P (ϕ) = N exp − S(ϕ) . (1.1)

The function S(ϕ) is called the action of the particular quantum field theory :
in a sense, it defines the theory. For the probability density to be acceptable,
S(ϕ) must go to infinity sufficiently fast as |ϕ| → ∞. The normalization
factor N is defined by1
Z
−1


N = exp − S(ϕ) dϕ . (1.2)

1.2.2 Green’s functions, sources and the path integral

Since the quantum field is a random variable, the most that can be known
about it2 is the collection of its moments, in the jargon called Green’s func-
tions3 :
Z
n
exp − S(ϕ) ϕn dϕ , n = 0, 1, 2, 3, . . . .


Gn ≡ hϕ i ≡ N (1.3)

1
If not explicitly indicated otherwise, integrals run from −∞ to +∞.
2
You are here approaching a career decision ! You may decide simply to measure
the value of ϕ : in that case you have decided to become an experimentalist rather than
a theorist.
3
A clarifying remark is necessary here. In this text, the Green’s functions are simply
defined to be expectation values. This may appear to contrast with the use of Green’s
functions in the solution of inhomogeneous linear differential equations such as are encoun-
tered in classical electrodynamics where one uses them to compute the electromagnetic
field configurations for given sources. The difference is only apparent since, as we shall
recognize, the latter type of Green’s functions are in our treatment simply the two-point
Green’s functions ; and for theories such as electrodynamics, where the electromagnetic
fields do not undergo self-interaction, the two-point functions are in fact the only nonzero
connected Green’s functions. Be not, therefore, misled into thinking that there are some-
how two sorts of Green’s functions. The Green’s function formulation of electrodynamics
will in fact appear as the classical limit of the Schwinger-Dyson equation discussed below.
August 31, 2019 31

We shall assume that Gn exists for all n. By construction, we must always

have
G0 = ϕ0 = h1i = 1 .


(1.4)
The most fruitful way4 of discussing the set of all Green’s functions is in
terms of a generating function :
X 1
Z(J) = J n Gn . (1.5)
n≥0
n!

This is called the path integral, for reasons that will become clear later. It
can be written as
Z


Z(J) = N exp − S(ϕ) + Jϕ dϕ . (1.6)

The number J, which here serves purely as a device to distinguish the various
Green’s functions, is called a source, again for reasons that will become ap-
parent later. Once Z(J) is known, an individual Green’s function is extracted
by differentiation :  n 
∂
Gn = Z(J) . (1.7)
(∂J)n J=0

1.2.3 Connected Green’s functions

The path integral Z(J) contains all the information about the Green’s func-
tions, and hence about the probability density P (ϕ). The same information
is, therefore, also contained in its logarithm. We write
X 1
W (J) = log Z(J) ≡ J n Cn . (1.8)
n≥1
n!

The quantities Cn (with, obviously C0 = 0 since G0 = 1) are called the

connected Green’s functions of the theory, and will play an important rôle in
what follows. The connected Green’s functions can be recognized to be the
cumulants of the probability density : E1
4
Kids ! Do this at home. Whenever an infinite collection of objects with some kind of
structure between them occurs, generating functions are always a good idea.
32 August 31, 2019

C0 =0
C1 = hϕi the mean
C2 = h(ϕ − hϕi)2 i the variance (1.9)
C3 = h(ϕ − hϕi)3 i the skewness
C4 = h(ϕ − hϕi)4 i − 3C2 2 the kurtosis
and so on. Since W (0) = C0 = 0, the same information about the probability
density is also contained in the field function :
∂ X 1
φ(J) ≡ W (J) = J n Cn+1 . (1.10)
∂J n≥0
n!

Since from its definition, we have

Z  Z −1



φ(J) = exp − S(ϕ) + Jϕ ϕ dϕ exp − S(ϕ) + Jϕ dϕ ,
(1.11)
we can say that φ(J) is the expectation value of the quantum field ϕ in the
presence of sources: to denote this, we might write
φ(J) = hϕiJ , (1.12)
which explains the similar typographies for the quantum field and the field
function5 . We should not, however, forget the difference in status of these
objects : ϕ is the physical entity, an unknowable, fluctuating random field ;
but φ(J) is an perfectly well-defined function that contains all the informa-
tion about the probability density of ϕ, and is6 computable once the action is
given.

1.2.4 The free theory

The simplest probability density is probably7 the Gaussian one, given by the
action
1
S(ϕ) = µ ϕ2 , (1.13)
2
5
Try this out : “phi of J equals phi of J”.
6
In principle, if not in practice easily or completely.
7
A uniform density may be thought even simpler, but then it cannot run from ϕ = −∞
to ϕ = +∞. As a matter of fact, ask any mathematician or physicist to name you a nice
proability density over the whole real line, and she will almost without fail quote the
Gaussian.
August 31, 2019 33

with µ a positive real number. For any action, we shall call the part quadratic
in the fields (or bilinear in the case of several fields) the kinetic part. This
action, called the free action, consists of only a kinetic part. The path integral
is now simply computed by
Z  
1 2
Z(J) = N exp − µϕ + Jϕ dϕ
2
2 !
J2
Z   2
1 J J
= N exp − µ ϕ − + dϕ = exp . (1.14)
2 µ 2µ 2µ

It is not even necessary8 to actually calculate the value of N . By Taylor

expansion of the exponential, we immediately find that
(2n)! 1
G2n = , G2n+1 = 0 , n = 0, 1, 2, . . . , . (1.15)
2n n! µn
The connected Green’s functions follow from
J2 J
W (J) = log Z(J) = , φ(J) = , (1.16)
2µ µ
so that the only nonvanishing connected Green’s function is C2 = 1/µ. The
fact that here only the two-point connected Green’s function is nonvanishing
is the reason for calling this model the free theory. Again, things will become
clearer later on, in a more realistic spacetime — after all, what does ‘freedom’
mean in a zero-dimensional prison ?

1.2.5 The ϕ4 model and perturbation theory

An action S(ϕ) may contain other terms than just the quadratic one. Such
terms are called interaction terms : they may be linear, but more usually
they are of higher power in the field ϕ. The simplest acceptable interacting
theory in our probabilistic setting is therefore given by the action
1 1
S(ϕ) = µϕ2 + λ4 ϕ4 . (1.17)
2 4!
The (nonnegative !) real number λ4 is called a coupling constant : this model
is called the ϕ4 theory9 . E2
8
Because we must always have Z(0) = 1.
 34 August 31, 2019

Computing the path integral is now a much less trivial matter. A possible
approach is to assume that, in some sense, the ϕ4 theory is close to a free
theory, that is, in the same some sense, λ4 is a small number. We can then
expand the probability density in powers of λ4 :
   k
1 2 X 1 λ4
exp(−S(ϕ)) = exp − µϕ − ϕ4k . (1.18)
2 k≥0
k! 24

This procedure is called perturbation theory. Having thus reduced the prob-
lem to the previous case of the free theory, we cavalierly10 interchange the
series expansion in λ4 with the integration over ϕ and arrive at the following
expression for the Green’s functions :
G2n = H2n /H0 ,
 k
1 X (4k + 2n)! λ4
H2n = − . (1.19)
µn k≥0 22k+n (2k + n)!k! 24µ2

For example, we have

1 35 2 385 3
H0 = 1 − u + u − u + ··· ,
8 384 3072
1 29 2 107 3
1/H0 = 1 + u − u + u + ··· , (1.20)
8 384 1024
with u ≡ λ4 /µ2 . Note that, in this theory, also the normalization N has to
E3 be treated perturbatively, which explains the expression for 1/H0 . For the
first few nonvanishing Green’s functions we find
G0 = 1 ,
 
1 1 2 2 11 3
G2 = 1 − u + u − u + ··· ,
µ 2 3 8
 
1 33 2 68 3
G4 = 3 − 4u + u − u + · · · ,
µ2 4 3
 
1 75 445 2 1585 3
G6 = 15 − u + u − u + ··· . (1.21)
µ3 2 4 4
9
An action in which ϕ3 is the highest power does not lead to a convergent integral
over the real axis (see, however, Appendix 15.4). Of course, an action of the form S(ϕ) =
µϕ2 /2 + λ3 ϕ3 /3! + λ4 ϕ4 /4! is perfectly acceptable, and we shall consider this ‘ϕ3/4 model’
later on.
10
And not with impunity ! See Appendix 15.1.
August 31, 2019 35

The corresponding nonzero connected Green’s functions are given by

 
1 1 2 2 11 3
C2 = 1 − u + u − u + ··· ,
µ 2 3 8
 
1 7 2 149 3
C4 = −u + u − u + ··· ,
µ2 2 12
1 2 3


C6 = 10u − 80u + · · · . (1.22)
µ3
Note that, whereas the Green’s functions all have a perturbation expansion
starting with terms containing no λ4 , the connected Green’s functions of in-
creasing order are also of increasingly high order in λ4 : the higher connected
Green’s functions need more interactions than the lower ones.

1.2.6 The Schwinger-Dyson equation

Although the path integral is, generally, a very complicated function of J,
it is nevertheless easy to find an equation describing it completely. This is
the Schwinger-Dyson equation (SDe), which we construct as follows. Let the
action be given by the general expression :
X 1
S(ϕ) = λk ϕ k , (1.23)
k≥1
k!

where λ2 = µ11 . Now, from the observation that

∂p
Z
exp − S(ϕ) + Jϕ ϕp dϕ ,


p Z(J) = N p = 0, 1, 2, 3, . . . (1.24)
(∂J)
we immedately deduce that
" #
X λk+1 ∂ k
−J + Z(J) =
k≥0
k! (∂J)k
Z " #
X λk+1
ϕk dϕ


= N exp − S(ϕ) + Jϕ −J +
k≥0
k!
Z  
0


= N exp − S(ϕ) + Jϕ S (ϕ) − J dϕ = 0 , (1.25)

11
The sum starts at 1 since a constant, ϕ-independent term in the action is always
immediately swallowed up by the normalization factor N .
 36 August 31, 2019

where in the last lemma we have used partial integration, and the fact that
the integrand vanishes at the endpoints. Symbolically, we may write the SDe
as    
∂ 0 ∂
S(ϕ) Z(J) = S Z(J) = JZ(J) . (1.26)
∂ϕ ϕ=∂/∂J ∂J
E4 For our sample model, the ϕ4 theory, the SDe reads12
1
λ4 Z 000 (J) + µZ 0 (J) − JZ(J) = 0 . (1.27)
6
Using the series expansion of the path integral we can express this as a
relation between different Green’s functions :
λ4
Gn+3 + µGn+1 − nGn−1 = 0 , n ≥ 1 . (1.28)
6
This relation may usefully be rewritten as follows :
 
1 λ4
Gn = (n − 1)Gn−2 − Gn+2 , n≥2 . (1.29)
µ 6
If we start by assigning to the Green’s functions the values Gn = δ0,n ,
then repeated applications of Eq.(1.29) will precisely reproduce the Green’s
functions of Eq.(1.21)13 .

1.2.7 The Schwinger-Dyson equation for the field func-

tion
From the definition of φ(J) as the logarithmic derivative of the path integral,
we can infer that
p
∂p

1 ∂
Z(J) = φ(J) + e(J) . (1.30)
Z(J) (∂J)p ∂J
Here, e(J) is the unit function: e(J) ≡ 1. We immediately arrive at the form
E5 of the SDe for the field function:
12
The SD equation is, in general, of higher than the first order. It therefore has several
independent solutions, only one of which corresponds to the usual perturbative expansion.
The nature of the other solutions is discussed in Appendix 2.
13
The correct way to do this is to subsequently evaluate G2 , G4 , G6 , . . .. On the first
iteration, the lowest-order expressions are obtained. Each subsequent iteration gives one
higher order in perturbation theory. Note that if we want to obtain the k th order term in
Gn , the (k + 1)th order term in Gn+2 is needed, and so on. It is therefore necessary to
compute the lower-order terms for more Gn ’s.
August 31, 2019 37
 
0 ∂
S φ(J) + e(J) = J . (1.31)
∂J

For the ϕ3/4 theory, it reads

 
J λ3 2 ∂
φ(J) = − φ(J) + φ(J)
µ 2µ ∂J
∂2
 
λ4 3 ∂
− φ(J) + 3φ(J) φ(J) + φ(J) . (1.32)
6µ ∂J (∂J)2

Although this leads to very nonlinear relations between the various connected
Green’s functions this form of the SD equation is actually even simpler to
apply : with φ(J) = 0 as a starting pont, iterating the assignment (1.32) then E6
results14 in the correct form of φ(J), giving the connected Green’s functions
of Eq.(1.22).

1.3 Diagrammatics
1.3.1 Feynman diagrams
An extremely useful tool for computing Green’s functions and connected
Green’s functions is at hand in the form of Feynman diagrams. In this section
we shall first introduce these diagrams and their concomitant Feynman rules.
Only after that shall we prove that these diagrams do, indeed, correctly
describe Green’s functions. Feynman diagrams are constructs of lines and
vertices. A vertex is a meeting point for one or more lines. Diagrams are
allowed in which one or more lines do not end in a vertex but, in a sense sie
wandern ins Blaue hinein : such lines are called external lines. Lines that are
not external lines, and end up at vertices at both ends, are called internal
lines. Diagrams may be connected, in which case one can move between any
two points in the diagram following lines of that diagram ; or they may be
disconnected, in which case it consists of two or more disjoint pieces that are
themselves connected. Any graph15 consists of a finite number of connected
14
For this approach to work in practice, it turns out to be useful to truncate φ(J) as a
power series in J, the truncation order increasing by one with each iteration. If you don’t
do this, each iteration triples the highest power in J, leading to very unwieldy expressions
with only the first few terms being actually correct.
15
For us, the terms ‘diagram’ and ‘graph’ are interchangeable.
38 August 31, 2019

subgraphs. The ‘empty’ graph E, containing no lines or vertices whatsoever,

also exists ; it does not count as connected16 . Diagrams containing one or
more closed loops are perfectly allowed. Diagrams with no closed loops are
called tree diagrams. Some examples of Feynman diagrams are

a connected graph a disconnected graph a connected tree graph

Note that the precise shape of the lines and the precise position of the ver-
tices are irrelevant. The important thing is the way in which the lines are
connected to the vertices17 .

1.3.2 Feynman rules

The noteworthy thing about Feynman diagrams is that they have an alge-
braic interpretation; that is, they correspond to numbers that may be added
and multiplied. The assignment of a number to a Feynman diagram is gov-
erned by the Feynman rules, which postulate a numerical object for every
ingredient of a Feynman graph. In the simple zero-dimensional theories that
we consider here the Feynman rules are just numbers. We may use, for
instance, the following rules :

16
Sophistry : it has no points between which one might wish to move.
17
As you will discover, I have endeavoured in these notes to avoid drawing straight
lines, or to draw blobs or closed loops as circles. Many texts do employ only straight lines
and circles. This not only leads to awfully unæsthetic-looking pictures, but is also deeply
misleading. There is a (natural) tendency to look at Feynman diagrams with the idea
that the lines represent ‘particles moving freely through space’ so that the lines ‘ought’
to be straight according to Newton’s first law. That this is completely wrong becomes
immediately clear if we realize that, in the zero-dimensional world we are dealing with for
now, there cannot be any notion of movement yet, let alone any Newton to pronounce on
it. In fact, Newton’s first law ought to be derived from our theory, and we shall do so in
due course.
August 31, 2019 39

↔ 1/µ

↔ −λ3

↔ −λ4

↔ +J
Feynman rules, version 1.1 (1.33)

A vertex at which a single line ends (and which carries a Feynman rule
factor +J) is called a source vertex. A disconnected diagram evaluates to
the product of the values of its disjunct connected pieces. Because of this
multiplicative rule, the value of the empty diagram E is taken to be unity.
In addition, we assign to every Feynman diagram a symmetry factor. The
symmetry factor is the single most nontrivial ingredient of the diagrammatic
approach. We shall therefore devote a separate section to this issue.

1.3.3 Symmetries and multiplicities

Feynman diagrams have, in general, an ‘inner’ and an ‘outer’ part. The
‘inner’ part consists of the various vertices and internal lines : the ‘outer’ part
is made up from the external lines (if any). The inner part concomitates with
the symmetry factor of the diagram, and for the outer part we have what
may be called the multiplicity, to be discussed below. Let us first turn to the
symmetry factor. The rules are :

• for every set of k lines that may be permuted without changing the
diagram, there will be a factor 1/k! ;

• for every set of m vertices that may be permuted without changing the
diagram, there will be a factor 1/m! ;

• for every set of p disjunct connected pieces that maybe interchanged

without changing the diagram, there will be a factor 1/p!;
40 August 31, 2019

• a factor 1/k for every k-fold rotational symmetry18 ;

• a factor 1/2 for every mirror symmetry.
External lines cannot be permuted without changing the diagram. Since
external lines cannot be permuted, only vacuum diagrams, that is diagrams
without any external lines, can have a rotational symmetry. It is important
to note that the symmetry factor cannot be read off from the individual
components of the diagram, but depends on the topology of the whole di-
agram19 . As our universe grows from zero to more dimensions, and as the
particles considered acquire more properties, the Feynman rules will grow in
complication ; but the symmetry factors remain the same20 .
A few examples of diagram values are presented here. First, consider the
diagram
λ3 2
= . (1.34)
µ5
In this case, the symmetry factor is 1, since for a tree diagram no internal
lines or vertices can be interchanged with impunity. The similar-looking
diagram
1 λ3 2 3
= J . (1.35)
2 µ5
has a symmetry factor 1/2! since the upper two one-point vertices are inter-
changeable. Then, there is the graph
1 λ4
= −
2 µ3
Here, there is a symmetry factor 1/2 because the ‘leaf’ can be flipped over
without changing the diagram. The diagram
1 λ4 2
=
6 µ5
18
Note: 1/k, not 1/(k!).
19
This is what makes the automated evaluation of diagrams a nontrivial task : compo-
nent factors of diagrams can be easily assigned, but working out the symmetry factor of a
diagram calls for for very complicated computer algorithms indeed.
20
This is only modified if we include lines of different types, or oriented lines. Then again,
the more-dimensional diagrams have the same symmetry factors as their zero-dimensional
siblings.
August 31, 2019 41

carries a symmetry factor of 1/3! because the three internal lines are inter-
changeable. The graph

1 λ4 3
= −
4 µ7

carries a symmetry factor (1/2!)(1/2!) since there are now only two inter-
changeable internal lines, and a single ‘leaf’. Finally, the diagram

1 λ4 2
=
48 µ4
has a symmetry factor (1/4!)(1/2!) since there are 4 equivalent internal lines,
and moreover the diagram can be ‘flipped over’ without changing it. E7

Next, we address the multiplicity. This is the number of different ways

the external lines (that each have their own ‘individuality’) can be attached.
To determine the multiplicity we must imagine that the whole diagram, or a
part of it, can be ‘flipped over’ while retaining the same attachement of the
external lines. To illustrate this, we temporarily denote the external lines
with a letter, and then notice that the two diagrams
a c d a
and
b d c b
are, in fact, identical ; the multiplicity of this graph is therefore 3, since
there are 3 ways to group four letters into two groups of two without regard
to ordering. We see that the diagram of Eq.(1.34) has, also, multiplicity 3,
while that of Eq.(1.35) has multiplicity 1. We see that, if we include the
multiplicity, the replacing of p external lines with p one-point source vertices
induces a factor of 1/p!, which will become important later on.
The determination of symmetry factors may appear somewhat fanciful21 ,
but of course it has a solid and unambiguous basis ; the symmetry factor
21
This is due to the fact that the line in the loop is not oriented: for oriented lines it
will no longer hold. The discussion of symmetry factors of Feynman diagrams goes, in
practice, with a lot of remarks like ‘... so you flip over this leaf, you wriggle this set of
internal lines, you shove these vertices back and forth ... see ?’ Although the symmetry
factor is totally unambiguous, the arguments for a symmetry factor often come with a lot
of prestidigitatorial hand-waving and finger-wriggling in front of a blackboard.
42 August 31, 2019

(and the multiplicity) can always be computed. The procedure is somewhat

involved, and is outlined in section 15.2.1.

1.3.4 Vacuum bubbles

Feynman diagrams exist that contain neither external lines nor source ver-
tices. These are called vacuum bubbles!definition ; the empty graph E is
obviously a vacuum bubble. We may consider the set of all vacuum bubbles,
which we denote by H0 . Let us assume that only four-point vertices occur.
Then, H0 , given by

H0 = E + + + + + ··· (1.36)

(where the ellipsis denotes diagrams with more four-vertices) evaluates to

2
1 λ4 2 1 λ4 2

1 λ4 1 1 λ4
H0 = 1− + + + + ···
8 µ2 2 8 µ2 16 µ4 48 µ4
1 λ4 35 λ4 2
= 1− + + ··· , (1.37)
8 µ2 384 µ4

which, indeed, looks suspiciously like H0 for the ϕ4 theory.

1.3.5 An equation for connected graphs

We shall now construct an equation for a special set of diagrams. We do this
for the set of Feynman rules of section 1.3.2. First, let us denote by Cn the
set of all connected graphs with no source vertices and precisely n external
lines. Clearly this is a enumerably infinite set. Next, we define the object
Ψ(J), denoted by the symbol
11111
00000
00000
11111
00000
11111
00000
11111
Ψ(J) ≡ 00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
(1.38)
00000
11111
00000
11111

to be the set of all connected diagrams with precisely one external line, and
any number of source vertices. The shading indicates that all the diagrams
in the blob must be connected . Clearly, then, we have
X 1
Ψ(J) = J n Cn+1 , (1.39)
n≥0
n!
August 31, 2019 43

where the extra factor 1/n! is the additional symmetry/multiplicity factor

for n source vertices.
Let us now consider what can happen if we enter the blob of Eq.(1.38)
along the single external line. In the first place, we can simply encounter a
source vertex, so that the diagram is
J
= . (1.40)
µ
Alternatively, we may encounter a vertex. If this is a three-point vertex, the
line splits into two. Taking one of these branches, we may be able to come
back to the vertex via the other branch. In that case, the diagram has the
form
11111
00000
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111

On the other hand, it may happen that the two branches end up in disjunct
connected pieces of the diagram, which then looks like
00000
11111
11111
00000
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
111111
000000
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111

Note that these two alternative cases can be unambiguously distinguished

because we have restricted ourselves to using only connected graphs. Another
important insight is that, in the above diagram, the two final blobs (with their
attached lines) are both exactly identical to the original Ψ(J) of Eq.(1.38),
and therefore also to each other : a situation that is of course only possible
because the blobs represent infinite sets of diagrams. In contrast, the closed-
loop blob of the first alternative is not equal to Ψ(J) since it has not one
but two lines sticking out ; but then again these two lines are completely
equivalent.
If we encounter a four-point rather than a three-point vertex, the line
splits into three, with three alternatives : no branches meeting again fur-
ther on, all three meeting again, or only two out of the three. We find the
diagrammatic equation
11111
00000
00000
11111
00000
11111
00000
11111
11111
00000
00000
11111 00000
11111
00000
11111 11111
00000
00000
11111
00000
11111 00000
11111 00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
= + 00000
11111
111111
000000
+ 00000
11111
00000
11111
00000
11111
00000
11111
+
00000
11111 000000
111111
000000
111111 00000
11111
00000
11111
000000
111111 00000
11111
000000
111111 00000
11111
000000
111111
 44 August 31, 2019
00000
11111
11111
00000
00000
11111
00000
11111 111111
000000 000000
111111
00000
11111 000000
111111 111111
000000
000000
111111
00000
11111 000000
111111 000000
111111
00000
11111 000000
111111
000000
111111 000000
111111
00000
11111
0000
1111
0000
1111
00001111
11110000
0000
1111
+ 000000
111111
000000
111111
000000
111111
000000
111111
+ 000000
111111
000000
111111
000000
111111 . (1.41)
0000
1111 000000
111111
0000
1111
00001111
1111
0000
1111
0000
0000
1111 0000
1111
00001111
11110000 0000
1111
0000
1111 0000
1111
0000
1111
0000
1111 0000
1111
0000
1111
0000
1111

Now, realize that

0000000
1111111
0000000
1111111
1111111
0000000
X 1 ∂
J n Cn+2 =
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
= Ψ(J) (1.42)
0000000
1111111

n≥0
n! ∂J

and 1111111
0000000
0000000
1111111
0000000
1111111
0000000
1111111
X 1 ∂2
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111 = J n Cn+3 = Ψ(J) , (1.43)
0000000
1111111
0000000
1111111
0000000
1111111
n≥0
n! (∂J)2
so that we can translate the diagrammatic equation (1.41) into an algebraic
equation for Ψ(J) by carefully implementing the correct Feynman rules, in-
cluding the symmetry factors for equivalent blobs and lines:
 
J λ3 1 2 1 ∂
Ψ(J) = − Ψ(J) + Ψ(J)
µ µ 2 2 ∂J
1 ∂2
 
λ4 1 3 1 ∂
− Ψ(J) + Ψ(J) Ψ(J) + Ψ(J) . (1.44)
µ 6 2 ∂J 6 (∂J)2

Now Eq.(1.44), obtained from the Feynman diagrams via the Feynman rules,
has exactly the same form as Eq.(1.32), valid for the field function φ(J) – note
the importance of the symmetry factors ! Moreover, the iterative solution
for φ(J) starts with φ(J) = J/µ, also identical to the diagrammatic starting
E8 point . We therefore conclude that Ψ(J) = φ(J), in other words Cn = Cn
(n ≥ 1). This proves that connected Green’s functions can be obtained
E9
by the following recipe: to obtain Cn (n ≥ 1), write out all connected
E10 Feynman diagrams with no source vertices and precisely n external
lines. Evaluate the diagrams using the Feynman rules, and sum
them.

1.3.6 Semi-connected graphs and the SDe

A useful notion, which allows us to write SDe’s more compactly, is that of
semi-connected graphs. We shall denote these with a lightly shaded blob,
and they are defined as follows : a semi-connected graph with n ≥ 1 lines
at the left is a general graph with n lines on the left (and any number of
August 31, 2019 45

other external lines), with the constraint that each connected piece of the
semi-connected graph is attached to at least one of the lines indicated on the
left. This may sound more intimidating that is actually is : an example is
1 1 1 1
2 = 2 + 2 + 3
3 3 3 2

1
2
+ 3 + 2 . (1.45)
1
3

A single semi-connected graph with n indicated lines stands for B(n) dia-
grams with explicit connected graphs, where B(n) is the so-called Bell num-
ber : the number of ways to divide n distinct objects into non-empty groups22 .
For ϕ3/4 theory, the SDe then becomes simply

= + + . (1.46)

We shall use semi-connected diagrams to good effect in later chapters. Note

that the sum of the symmetry factors of all connected diagrams arising from
a ϕp vertex must be equal to B(p − 1)/(p − 1)!, which may serve as a check
on your SDe’s.

1.3.7 The path integral as a set of diagrams

By affixing a source vertex to the single external line of Ψ(J), we immediately
have the result that the generating function W (J) is the sum of all
connected Feynman diagrams without external lines and at least
one source vertex. If we explicitly indicate the source vertices, and recall
that n source vertices in a diagram imply a factor 1/n!, we can write
111111
000000 111111
000000
000000
111111
11111
00000
00000
11111
000000
111111
000000
111111 000000
111111 00000
11111 00000
11111
000000
111111 11111
00000 00000
11111
W (J) = 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
+ 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
+ 00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
+ 00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
+ · · · , (1.47)
00000
11111

22
For small n we have B(0) = 1, B(1) = 1, B(2) = 2, B(3) = 5, B(4) = 15, and
B(5) = 52 ; more general values can be obtained from the identity
X xn
B(n) = exp (ex − 1) .
n!
n≥0

which is derived in Appendix 15.15.7.

 46 August 31, 2019

where the ellipsis contains connected contributions with more source vertices.
Vacuum bubbles do not contribute to W (J). By taking careful account of
the symmetry factor assigned to identical connected parts of a disconnected
diagram, we can see that
111111
000000
000000
111111
111111
000000 000000
111111 000000
111111
000000
111111 111111
000000 000000
111111
000000
111111 000000
111111 000000
111111
000000
111111 000000
111111 000000
111111
000000
111111 000000
111111 000000
111111
000000
111111 000000
111111

1 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111

W (J)2
000000
111111

= 000000
111111
111111
000000
000000
111111
000000
111111
+ 111111
000000
000000
111111
000000
111111
+ 11111
00000
00000
11111
00000
11111
00000
11111
2! 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111

000000
111111 00000
11111
111111
000000 11111
00000
00000
11111
000000
111111 000000
111111 00000
11111
111111
000000
000000
111111 000000
111111 00000
11111
000000
111111 000000
111111 00000
11111
000000
111111 000000
111111
000000
111111 00000
11111
000000
111111 000000
111111 00000
11111
00000
11111
111111
000000
000000
111111 000000
111111 00000
11111
000000
111111

+ 000000
111111
111111
000000
000000
111111
000000
111111
+ 00000
11111
11111
00000
00000
11111
00000
11111
00000
11111
+ 00000
11111
00000
11111
11111
00000
000000
111111 00000
11111 00000
11111
000000
111111
000000
111111 00000
11111 00000
11111
000000
111111 00000
11111 00000
11111
00000
11111
000000
111111 00000
11111
00000
11111 00000
11111
00000
11111
00000
11111

+ (lots of other diagrams) . (1.48)

Similar arguments hold for higher powers of W (J). In addition, W (J)0 =

E = 1. From this it easy to see that the path integral Z(J) consists of
all Feynman diagrams without external lines, and without vacuum
bubbles, but including the empty diagram.
We might wonder why the vacuum bubbles are so conspicously absent .
Suppose that we would allow the inclusion of arbitrary numbers of vacuum
bubbles in Z(J). Then the Green’s function G0 = 1 would be represented
not by the single empty graph but by the whole set H0 discussed before:
indeed, H0 is proportional to H0 . In fact, any Green’s function Gn would
acquire exactly the same additional factor H0 . The normalization factor N ,
that must be chosen such as to make G0 equal to unity, therefore extracts
exactly the factor H0 from any Green’s function. In the jargon, the vacuum
bubbles ‘disappear into the normalization of the path integral’. This is not
E11 to say that vacuum diagrams are never important ; but in our approach
to computing Green’s functions and connected Green’s functions they are
indeed irrelevant. Another way of seeing this is very simple : if we take
our diagrammatic prescription of Z(J) and then take J = 0, all diagrams
disappear except the empty one, and we find Z(0) = E = 1, just as we must.

1.3.8 Dyson summation

Why is the Feynman rule for lines, stemming from the quadratic part of
the action, so different from those for the vertices, that come from the non-
quadratic terms ? To see that our treatment is actually a consistent one, let
August 31, 2019 47

us consider an action is given by

1 1 1
S(ϕ) = µϕ2 + λ2 ϕ2 + λ4 ϕ4 . (1.49)
2 2 4!
If we wish, we may treat the λ2 term as an interaction, described by a vertex
with two legs. the SDe is then seen to be
11111
00000 1111111
0000000
00000
11111 0000000
1111111
00000
11111 0000000
1111111
00000
11111 0000000
1111111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
= + 0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
00000
11111 0000000
1111111
0000000
1111111
00000
11111
00000
11111
11111
00000
00000
11111
00000
11111 111111
000000 000000
111111
00000
11111 000000
111111 111111
000000
000000
111111
00000
11111 000000
111111 000000
111111
00000
11111 000000
111111
000000
111111 000000
111111
00000
11111
+ 0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
+ 000000
111111
000000
111111
000000
111111
000000
111111
+ 000000
111111
000000
111111
000000
111111 . (1.50)
0000
1111 000000
111111
0000
1111
00001111
11110000
0000
1111 0000
1111
0000
1111 0000
1111
00001111
11110000 0000
1111
0000
1111
0000
1111 0000
1111
0000
1111
0000
1111
0000
1111

corresponding to
∂2
 
J λ2 λ4 3 ∂
φ(J) = − φ(J) − φ(J) + 3φ(J) φ(J) + φ(J) . (1.51)
µ µ 6µ ∂J (∂J)2
Multiplying the equation by µ and transposing the λ2 term to the left, we
obtain
∂2
 
J λ4 3 ∂
φ(J) = − φ(J) + 3φ(J) φ(J) + φ(J) ,
µ + λ2 6(µ + λ2 ) ∂J (∂J)2
(1.52)
precisely what we woud have obtained by taking the combination (µ + λ2 ) as
the kinetic part from the start. This procedure, by which the effect of two-
point (effective) vertices is subsumed in a redefinition of the kinetic part,
is called Dyson summation. In the present example, the summation is of
course trivial ; but we shall see that two-point interactions can also arise
from more complicated Feynman diagrams corresponding to higher orders
in perturbation theory. The manner in which Dyson summation is usually
treated is by explicitly writing out the propagator, ‘dressed’ with two-point
vertices in all possible ways :
+ + + + ···
1 1 1 1 1 1 1 1 1 1
= − λ2 + λ2 λ2 − λ2 λ2 λ2 + ···
µ µ µ µ µ µ µ µ µ µ
 k
1 X λ2
= −
µ k≥0 µ
1 1 1
= = , (1.53)
µ 1 + λ2 /µ µ + λ2
48 August 31, 2019

where it should come as no surprise that we cheerfully ignore all issues about
convergence, in the spirit of perturbation theory. Every propagator line can
(and must ! ) be dressed in this way once any two-point vertex (elementary
of effective, that is, as the result of a collection of closed loops with two legs
sticking out) occurs.

1.4 Planck’s constant

1.4.1 The loop expansion, and reverse engineering
As we have seen, Green’s functions can be computed in a perturbative ex-
pansion in which the coupling constant λ4 is in some sense a small number.
Now consider doing perturbation theory in the ϕ3/4 theory. We then have
to decide on the relative order of magnitude of the two coupling constants
λ3 and λ4 : are they of the same order, or should we take, say, λ4 to be
of the same order as λ3 2 ? And what if even more coupling constants are
involved ? We shall adopt the approach that the order of magnitude of the
various diagrams should depend not on their coupling-constant content but,
rather, on their complexity , in particular on the number of closed loops. That
is, the more closed loops a diagram contains, the smaller it is considered to
be ; and perturbation theory then prescribes the perturbation expansion to
be truncated at a given number of closed loops.
To quantify these ideas we shall assign to every closed loop a factor ~,
where ~ is a (small) number23 . That is, we define the following ratios :
00000
11111 000000
111111
11111
00000 111111
000000
000000
111111
00000
11111 000000
111111
00000
11111
00000
11111 000000
111111
00000
11111 000000
111111
000000
111111
00000
11111
00000
11111 000000
111111
00000
11111
00000
11111
00000
11111
11111
00000
00000
11111
00000
11111
∼ ~ , 00000
11111
11111
00000
00000
11111
00000
11111
∼ ~2 ,
00000
11111 00000
11111
00000
11111 00000
11111
00000
11111 00000
11111
00000
11111 00000
11111
00000
11111 00000
11111
00000
11111 00000
11111
00000
11111 00000
11111

etcetera. This implies, of course, a modification of the Schwinger-Dyson

equation from the form (1.32) into
 
J λ3 2 ∂
φ(J) = − φ(J) + ~φ(J) φ(J)
µ 2µ ∂J
2
 
λ4 3 ∂ 2 ∂
− φ(J) + 3~φ(J) φ(J) + ~ φ(J) . (1.54)
6µ ∂J (∂J)2
23
As the notation suggests, it will develop into Planck’s (or Dirac’s) constant as our
universe increases in complexity.
August 31, 2019 49

In turn, we shall have to modify everything else as well : we must re-define

∂
φ(J) = ~ log Z(J) , (1.55)
∂J
so that the SDe for the path integral must read
 
0 ∂
S ~ Z(J) = JZ(J) . (1.56)
∂J
The path integral must therefore be re-defined with inclusion of ~ :
Z  
1

Z(J) = N exp − S(ϕ) − Jϕ dϕ , (1.57)
~
and for the Green’s functions we have
n n
   
n ∂ n ∂
Gn = ~ Z(J) , Cn = ~ log Z(J) . (1.58)
(∂J)n J=0 (∂J)n J=0

The Feynman rules must, therefore, take the form

~
↔
µ

λ3
↔ −
~

λ4
↔ −
~

J
↔ +
~
Feynman rules, version 1.2 (1.59)

The introduction of ~ as the perturbation expansion parameter allows us

to determine the relative orders of magnitude of coupling constants. Since
under our decree all tree diagrams are of the same order, the two graphs

and
50 August 31, 2019

tell us that λ4 is of the same order as λ3 2 . Similarly, a k-point coupling

constant λk is of the same order as λ3 k−2 . As a last point, you may note that
the introduction of ~ does not influence the Dyson summation of sec.1.3.8,
since every extra two-point vertex (with 1/~) also gives an extra propagator
(with ~).

Since in the Feynman rules ~ appears all over the place, it is advisable to
check that the ~-behaviour of the Feynman graphs is indeed as desired. For
an arbitrary given diagram let us define the characteristics E : the number
of external lines, I : number of internal lines, Vq : the number of vertices
of q-point type, L : the number of closed loops, and P : the number of
disjunct connected pieces. There are precisely two ‘sum rules’ that involve
linear combinations of these quantities. They are derived in section 15.3 and
read X X
Vq = I + P − L , qVq = 2I + E . (1.60)
q q

We are now able to read off the power of ~ associated with an arbitrary
connected diagram (with P = 1). From the Feynman rules, we infer that
every line contributes a factor ~ and every vertex a factor 1/~. The total
power of ~ is, therefore
X
E+I − Vq = E + L − 1 . (1.61)
q

Independently of its precise form, the power of ~ of any connected diagram

depends only on the number of its external lines and the number of loops,
and indeed each extra loop leads to an additional factor ~, as advertised.

1.4.2 The classical limit

Since in perturbation theory ~ is taken to be an infinitesimally small quantity,
the limit ~ → 0 is of automatic interest. This limit has to be taken with
some care since ~ = 0 strictly would imply that only Green’s functions with
E + L = 1 would survive24 . Instead, the classical limit ~ → 0 is meant to be
the result of leaving out diagrams containing closed loops. The diagrammatic
24
Later on, the discussion about truncation will clarify how this is not inconsistent.
August 31, 2019 51

SDe will, for the ϕ3/4 theory, then take the form
11111
00000
00000
11111 11111
00000
00000
11111 00000
11111
00000
11111 00000
11111
00000
11111
11111
00000
00000
11111 00000
11111
00000
11111 00000
11111
00000
11111
00000
11111 00000
11111 00000
11111
00000
11111 00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
= + 00000
11111
111111
000000
+ 0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
. (1.62)
00000
11111 000000
111111
000000
111111 00001111
1111
00000000
1111
00000
11111 0000
1111
000000
111111
000000
111111
0000
1111
00001111
11110000
000000
111111 0000
1111
0000
1111

The corresponding solution will be denoted by φc (J) (with c for ‘classical’),

and the classical SDe is written as
J λ3 λ4
φc (J) = − φc (J)2 − φc (J)3 . (1.63)
µ 2µ 6µ
The classical field function is exclusively built up from tree diagrams : this
is called the tree approximation. Note that it obeys an algebraic, rather than
a differential, equation, that can be written as
S 0 (φc (J)) = J . (1.64)
This is called the classical field equation. This is not to be confused with
equations from classical, nonquantum physics. In fact, the classical field
equations will turn out to be the Klein-Gordon, Dirac, Proca and Maxwell
equations. Of these, only the Maxwell equations can be considered classical,
since they do not contain a particle mass.. Note that such equations have, in
general, more than a single solution. Here, however, we are interested in that
solution that vanishes as J → 0, which may be written out using Lagrange
expansion25 :
  n 
J X 1 n−1 ∂ n−1 J
φc (J) = + µ n−1 S0 . (1.65)
µ n≥1 n! (∂J) µ

Let us now look at the path-integral picture of the classical limit. When ~
becomes small, the fluctuations in the path integrand
  
1
exp − S(ϕ) − Jϕ
~
become extremely exaggerated. The main contribution to hϕi therefore
comes from that value ϕc of ϕ for whichthe probability distribution attains
its maximum, that is,
hϕiJ ≈ ϕc , where S 0 (ϕc ) = J , S 00 (ϕc ) > 0 . (1.66)
Also in the classical limit, we therefore have φc (J) = ϕc .
25
see Appendix 15.15.12.
52 August 31, 2019

1.4.3 On second quantisation

The ‘classical’ approximations of our quantum field theory are26 quantum
equations. In fact, this is not so very surprising. In ordinary quantum me-
chanics, the classical variables such as position, momentum, etcetera are
identified with the expectation values of their quantum-mechanical counter-
parts, and considered a useful approximation of reality as long as they are
reasonably well-defined27 . So it is here as well : the field generating function
φ(J) is considered as the expectation value of the quantum field ϕ, and it is
identified with the quantum-mechanical wave function of whatever object it
is we are studying. In this sense, to go from ϕ to a classical observable we
have to ‘classicize’ two times. The transition from ordinary quantum me-
chanics to what we are doing here is therefore dubbed ‘second quantization’.
Of course, from the point of view we have taken here, this is simply a matter
of taking limits (expectation value upon expectation value), but if one comes
in from the classical side it may look quite mysterious. This is another re-
minder that one should not try to build a more fundamental theory from a
limiting case. Limiting cases are only hints.

1.4.4 Instanton contributions

As mentioned, for a non-free action S(ϕ), the equation (1.66) has, of couse,
more than a single solution28 . Suppose that we have several such solutions,
(0) (1) (2)
denoted by ϕc , ϕc , ϕc ,. . ., and that the minimal value of S(ϕ) − Jϕ is
(0)
attained for ϕc . Then, the other classical solutions will give contributions
that, relative to the dominant one, are suppressed by exponential factors of
order
  
1 (k) (0) (k) (0)
exp − S(ϕc ) − S(ϕc ) − Jϕc + Jϕc , k = 1, 2, . . . .
~

In figure 1.1 we plot the (normalized) form of exp(−S(φ)/~) for an action

with two minima, for two values of ~ different by a factor 6. It is seen how the
lowest minimum of S(ϕ) starts to dominate the integral as ~ becomes small ;
the contribution from the subleading maximum decreases nonperturbatively
26
Will be found to be ; see the later chapters of these notes.
27
With small uncertainty, that is, the variance of their statistical distribution around
the expectation value.
28
Since the action is at least of order ϕ3 , the classical field equation is at least quadratic.
August 31, 2019 53

Figure 1.1: The developing of instanton effects

fast. Such subdominant solutions to the classical field equations are called
instantons. Their contribution to Green’s functions do not have a series
expansion around ~ = 0. Such effects are therefore not accessible using
Feynman diagrams : they are called nonpertubative. This is not to say that
they are irrelevant. Indeed, we usually have a finite value for ~ ; more
(1)
dramatically, if we let J vary as a parameter, ϕc , say, may for some value
(0)
of J take over from ϕc as the true maximum position of the probability
(0) (1)
density, causing a sudden shift in the value of φc (J) from ϕc to ϕc .

1.5 The effective action

1.5.1 The effective action as a Legendre transform
Since perturbation theory presumes that higher orders in the loop expansion
are small compared to lower orders, the following question suggests itself : is
it possible to find, for a given action S(ϕ), another action, called the effective
action, with the property that its tree approximation reproduces the full field
function of the original action S ? If such an effective action, denoted by
54 August 31, 2019

Γ(φ), exists, we must have

Γ0 (φ(J)) = J , (1.67)
where φ(J) is the full solution to the SDe belonging with S(ϕ). We can use
partial integration to find
Z Z
Γ(φ) = J dφ = J φ − φ dJ = J φ − ~W , (1.68)

where J is now to be interpreted as a function of φ. The transition from

W (J) to Γ(φ) is called the Legrendre transform. In classical mechanics, we
have the same situation : there, ~W would be the Lagrangian with J as the
velocity and φ as the momentum, and then the effective action would turn
out to be the Hamiltonian.

An important fact to be noted about the effective action can be inferred

as follows. Let us consider the derivative of φ(J). If we denote the probability
density (including the sources) of the quantum field ϕ by PJ (ϕ), that is,
 
A(ϕ) 1
PJ (ϕ) = R , A(ϕ) = exp − (S(ϕ) − Jϕ) , (1.69)
dϕA(ϕ) ~
we can write this derivative as
R  R R
2
1 0 1 d PJ (ϕ) ϕ dϕ PJ (ϕ) ϕ2 dϕ PJ (ϕ) ϕ dϕ
φ (J) = R = R −
2
~ dJ PJ (ϕ) dϕ PJ (ϕ) dϕ
R
~ PJ (ϕ) dϕ
R
PJ (ϕ1 )PJ (ϕ2 ) (ϕ1 2 − ϕ1 ϕ2 ) dϕ1 dϕ2
= R
2 . (1.70)
PJ (ϕ) dϕ
By symmetry, we can replace the factor (ϕ1 2 − ϕ1 ϕ2 ) by (ϕ1 − ϕ2 )2 /2, so as
to see that dφ(J)/dJ is positive. This implies that
∂2 dJ
2 Γ(φ) = >0 . (1.71)
(∂φ) dφ
In other words, the effective action is concave everywhere29 . Whereas one
would assume that the effective action Γ would differ only slightly from the
original action S, this can obviously no longer hold in situations where the
action S is not concave.
29
This concavity persists in case there are more than just a single field involved. By
extension, it also holds for Euclidean theories in more dimensions ; see also Appendix 15.6.
August 31, 2019 55

1.5.2 Diagrams for the effective action

A tree approximation consists of tree diagrams only. To see how the loop
effects of the action S end up in Γ, we define a new concept, that of a one-
particle irreducible (1PI) diagram. A connected Feynman graph is 1PI if
it contains no internal line such that cutting that line makes the diagram
disconnected.

1PI diagrams a non-1PI diagram

External lines, of course, do not enter in the 1PI criterion30 . Note that a
diagram consisting in only external lines and a single vertex also counts as
1PI, since it does not have any internal lines to be cut. A typical one-loop
1PI diagram looks like this :

(1.72)
Let us denote the set of all 1PI graphs with precisely n external lines by
−γn /~, where the convention is that the Feynman factors for the external
lines are not included. Consider, now, what happens if we enter the field
function by way of its single external leg, as in the SDe. If we encounter
a vertex, that vertex is part of a 1PI subdiagram (possibly consisting of
only the vertex itself). Indicating the 1PI property with cross-hatches, we
therefore obtain the diagrammatic equation
11111
00000
00000
11111
00000
11111
111111
000000
000000
111111
00000
11111
00000
11111
00000
11111 = + 000000
111111
000000
111111
00000
11111
00000
11111
00000
11111
00000
11111
000000
111111
000000
111111
00000
11111
00000
11111
11111
00000
000000
111111 00000
11111
00000
11111
111111
000000 00000
11111
000000
111111
000000
111111 00000
11111

0000000
1111111
000000
111111 11111111
00000000
00000
11111
00000000
11111111
00000
11111
00000
11111
111111
000000
000000 111111
111111 000000
111111
000000 1111111
0000000
000000
111111
000000
111111
000000
111111
000000
111111 00000000
11111111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111

+ 000000
111111 000000
111111
000000
111111
+ 0000000
1111111
0000000
1111111
000000
111111
+ 00000000
11111111
00000000
11111111
00000
11111
00000
11111
00000
11111
· · · . (1.73)
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111 0000000
1111111
0000000
1111111 00000000
11111111
00000
11111
00000
11111
00000
11111
000000 111111
111111 000000
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111 0000000011111
11111111 00000
11111
00000
11111
00000
11111
00000
1111111
0000000 00000
11111
0000000
1111111
0000000
1111111 00000
11111
0000000
1111111 11111
00000
00000
11111
0000000
1111111 00000
11111
00000
11111
00000
11111
00000
11111
00000
11111

Algebraically, it reads
 
J 1 1 2 1 3
φ(J) = − γ1 + γ2 φ(J) + γ3 φ(J) + γ4 φ(J) + · · · , (1.74)
µ µ 2! 3!
30
Including them would be silly, since any diagram falls apart if we chop through an
external line.
 56 August 31, 2019

in other words
Γ0 (φ) = J , (1.75)
where
1 1 1
Γ(ϕ) = γ1 ϕ + (γ2 + µ)ϕ2 + γ3 ϕ3 + γ4 ϕ4 + · · · . (1.76)
2! 3! 4!
We conclude that the vertices of the effective action are determined
by the 1PI diagrams. It must be noted that, in general, the effective
action contains vertices with an arbitrarily large number of legs, even if the
E12 original action S goes up only to ϕ3 or ϕ4 , say.

1.5.3 Computing the effective action

We shall now describe a computation of the effective action
Γ(φ) = Γ0 (φ) + ~Γ1 (φ) + ~2 Γ2 (φ) + · · · , (1.77)
from its Feynman diagrams, for a theory with arbitrary couplings :
1 X λk
S(ϕ) = µ ϕ2 + ϕk . (1.78)
2 k≥3
k!

We start by considering a general one-loop 1PI diagram such as that of

Eq.(1.72), and cutting through the loop at some arbitrary place. We then
have a propagator ‘dressed’ with zero or more vertices where external lines
are ‘radiated off’. If there are precisely n external lines we can denote this
by
n

Such an object has, of course, its own SDe. Taking careful account of all
possibilities to attach external lines, we can write it as
 
n n n−1
= θ(n = 0) +
1
   
n n−2 n n−3
+ + + · · ·(1.79)
2 3
We define the generating function for such dressed propagators as
X zn n
P (z) = = , (1.80)
n≥0
n!
August 31, 2019 57

and see that the SDe reads

~ λ3 z 2 λ4 z 3 λ5
P (z) = − z P (z) − P (z) − P (z) − · · · , (1.81)
µ µ 2! µ 3! µ
in other words,
~
P (z) = . (1.82)
S 00 (z)
We now close the loop again with an arbitrary vertex, at which vertex at least
one other external line is included. By the same combinatorial arguments as
above we can find the generating function L(z) for such loops :

L(z) = + + + ···

z 2 λ5 S 000 (z)
 
1 λ3 λ4
= − P (z) − z P (z) − P (z) − · · · = − 00 (1.83).
2 ~ ~ 2! ~ 2S (z)

The symmetry factor 1/2 arises from the fact that the propagator is not
oriented and thus we have to avoid double-counting. Considering that a
propagator with n external legs leads to a closed loop with at least n + 1
external legs, we see that the one-loop effective action is given by
Z
~
Γ1 (φ) = −~ dφ L(φ) = log(S 00 (φ)) . (1.84)
2
A few remarks are in order here. In the first place, we see that the effective
action obtained in this way is only well-defined where the action itself is
concave, in agreement with the discussion in 1.5. In the second place, the
trick of closing the loop with an extra vertex, rather than just trying to
‘glue’ the endpoints of P (z) together, is technically useful since it avoids
potential problems with the symmetry factors. In the last place, the above
calculation is possible since all external lines are, so to speak, identical. In
more dimensions, where external lines can carry momentum, this is no longer
true. However, the effective potential, that is the effective action at zero
momentum, does lend itself to such a calculation in higher dimensions31 .
31
For simple scalar theories. Of course external lines may carry more than just momen-
tum information, that is, they can also carry spin/charge/colour· · · information. Then the
calculation is again more difficult.
 58 August 31, 2019

We can extend this treatment to higher loop orders as well. Let us denote
a vertex where at least n + 1 lines come together by

n = n+ n+1 + n+2 + ··· (1.85)

and assign to this dressed vertex the Feynman rule

1
n = − S (n+1) (z) . (1.86)
~
Now, we introduce the notion of a tadpole diagram : this is a connected di-
agram with precisely one external line and no source vertices. The effective
action as given above then follows from writing out the 1PI tadpole dia-
grams, replacing propagators by dressed propagators and vertices by dressed
vertices ; we can then simply read off the result.
111
000 S (3) (z) S (3) (φ)
000
111
→ 000
111
000
111 = − ⇒ Γ01 (φ) = , (1.87)
000
111 2S (2) (z) 2S (2) (φ)
as before. In two loops, the 1PI tadpole is given by the diagrams

+ + +

Dressing these tadpole diagrams gives us

1111
0000
0000
1111
11111
00000 111
000 0000
1111
00000
11111
00000
11111 000
111
000
111 1111
0000 111
000
000
111
000 111
111 000 0000
1111 000
111
000
111
00
11 + 000
111 000
111
000
111 + 000
111 000
111 +
00
11 000
111
000 111
111 000 000
111
00
11 000
111 000
111
000000
111111
000000
111111 111
000
000
111
S (3) (z) S (4) (z) S (3) (z)3 S (3) (z) S (4) (z) S (5) (z)
 
= ~ − + −
6 S (2) (z)3 4 S (2) (z)4 4 S (2) (z)3 8 S (2) (z)2
(1.88)

The two-loop contribution to the effective action is therefore

d S (5) (φ) 5 S (3) (φ) S (4) (φ) S (3) (φ)3

Γ2 (φ) = − + . (1.89)
dφ 8 S (2) (φ)2 12 S (2) (φ)4 4 S (2) (φ)4
E13 The effective action itself, the integral over the above experession, does not
August 31, 2019 59

have so simple a form as in Eq.(1.84), but is of course calculable as soon as

S(φ) is explicitly given ; moreover, we see that it, too, becomes undefined
where S 00 (φ) vanishes. From our diagrammatic approach we see that this will
persist in all loop orders32 .

1.6 Exercises
Excercise 1 Green’s functions and connected Green’s functions
We have X Jn X Jn
Z(J) = Gn , W (J) = Cn
n≥0
n! n≥0
n!
and
W (J) = log(Z(J)) .
Using that G0 = 1, and the expansion
X xn x2 x3 x4 x5
− log(1 − x) = =x+ + + + − ··· ,
n≥1
n 2 3 4 5

express C0,...,5 in terms of the G’s.

Excercise 2 The problem with ϕ4 theory

By examining the case where λ is infinitesimally small, either positive and
negative, argue that the limit λ → 0 establishes, in fact, an extremely singular
theory.

Excercise 3 Actually doing it for ϕ4

This exercise shows how to express the path integral for the ϕ4 theory in
terms of ‘known functions’. We consider
Z  
µ 2 λ 4
H = dϕ exp − ϕ − ϕ
2~ 24~
1. Show that the combination
~λ
g=
µ2
is dimensionless.
32
Because in all 1PI diagrams we have to dress the propagators, which implies lots of
00
S (φ) in the denominators.
60 August 31, 2019

2. Perform the substitution

s
3~ 1/4
ψ − ψ −1/4


ϕ=
µg

to derive
s Z∞   
3~ 3/(4g) −3/4 3 1
H= e dψ ψ exp − ψ+
4µg 8g ψ
0

3. The so-called modified Bessel functions of the second kind , denoted by

Kν (z), are discussed in appendix 15.15.8. Use this to express H in
terms of these Bessel functions.

4. The function Kν (z) has two expansions: one, an unattractive-looking

but convergent series in positive powers of z, and a nonconvergent,
asymptotic series in terms of negative powers of z, given in Eq.(15.305).
Show that H therefore has a regular series expansion in 1/g and an
asymptotic one in g. Show that the leading term in the asymptotic
expansion is independent of g (but not of µ), and compute the next
two terms. Compare your result with Eq.(1.20).

Excercise 4 The SDe for Z in another action

Find the SDe for the path integral Z(J) for the action
6
µ 2 X λk k
S(ϕ) = ϕ + ϕ ,
2! k=3
k!

Excercise 5 Writing the SDe for φ

Prove Eq.(1.30). Do this using the fact that
 Z 
1
Z(J) = exp dJ φ(J) ,
~

and then considering Z 0 , Z 00 and so on.

Excercise 6 The SDe for φ in another action

For the action of Exercise 4, derive the SDe for φ(J).
August 31, 2019 61

Excercise 7 The symmetry factor of life, the universe, and every-

thing
Devise a diagram (or set of diagrams) that has a symmetry factor of 1/42.

Excercise 8 Diagrammatic SDe for ϕ6 theory

Give the diagrammatic SDE for ϕ6 theory, and write it out algebraically.

Excercise 9 Actually doing it

Consider the diagrammatic SDe of Eq.(1.41). Starting with the empty dia-
gram, iterate this diagrammatic equation three or four times, and write down
the resulting set of diagrams.

Excercise 10 Some diagrams to consider

Of the following 12 diagrams, determine the multiplicity factor, the symmetry
factor, and the number of loops :

Excercise 11 Vacuum diagrams

Consider a theory with couplings ϕn , n = 3, 4, 5, 6, . . .. Show that the single
vacuum diagram at one loop is actually zero. At two loops there are 3 vacuum
diagrams, and at 3 loops there are 15 vacuum graphs. Write these down, and
determine the symmetry factors.
62 August 31, 2019

Excercise 12 SDe for the effective action

The effective action is defined such that Γ0 (φ(J)) = J. Show that the SDe
for pure ϕ4 theory can be written as
000
 
0 λ4 3 φ 2 Γ (φ)
µφ = Γ (φ) − φ + 3~ 00 − ~ 00 3
6 Γ (φ) Γ (φ)
Let the effective action have a perturbation expansion :
Γ(φ) = Γ0 (φ) + ~Γ1 (φ) + ~2 Γ2 (φ) + · · ·
Use the SDe to find Γ0 , Γ1 and Γ2 . Compare the results with the correspond-
ing 1PI diagrams with up to 6 external legs.

Excercise 13 A wonderful logarithmic action

Consider the action
µ µ
S(ϕ) = − 2 log(1 − aϕ) − ϕ ,
a a
1. The domain of ϕ is now no longer (−∞, ∞). Determine which domain
is appropriate, and prove that the path integrand indeed vanishes at
the endpoints of that domain.
2. Prove that this theory has ϕn interactions for all n = 3, 4, 5, . . . and
that the vertex for a ϕn interaction reads −(n − 1)!µan−2 /~.
3. Prove that the SDe for this theory reads
µ~Z 0 (J) = JZ(J) − a~Z(J) − a~JZ 0 (J) .

4. Prove that the field function is given by

J − a~
φ(J) = .
µ + aJ

5. The field function has only one-loop corrections ! Verify that the two-
loop correction to the propagator vanishes.
6. Show that the effective action is
µ  µ
Γ(φ) = − 2 + ~ log(1 − aφ) − φ
a a
It is also free from higher-order corrections beyond one loop !
August 31, 2019 63

7. At J → −µ/a, the field function diverges. Show that this corresponds

precisely to the situation where the path integrand no longer vanishes
at one of the endpoints.

Excercise 14 A wonderful exponential action

Consider the action
µ
S(ϕ) = 2 (eaϕ − 1 − aϕ)
a
1. Show that the SDe for the path integral takes the functional form
 µ
Z(J + a~) = J + Z(J)
a

2. Show that the solution for the path integral has the form
aJ/~ 
a2 ~
  
µ + aJ µ 
Z(J) = Γ /Γ
µ a2 ~ a2 ~

3. Show that the field function reads

  2   
1 a~ µ + aJ
φ(J) = log +ψ
a µ a2 ~

where the digamma function ψ is defined in appendix 15.15.4.

4. Show that the perturbation expansion of the field function is given by

a2k−1 Bk
 
1 aJ a~ X
φ(J) = log 1 + − − ~k
a µ 2(µ + aJ) k≥2 k(µ + aJ)k

5. Show that all odd loop orders beyond the first one vanish.

6. At J → −µ/a, the field function diverges. Show that this corresponds

precisely to the situation where the path integrand no longer vanishes
at one of the endpoints.
64 August 31, 2019
Chapter 2

On renormalization

2.1 Doing physics : mentality against reality

2.1.1 Physics vs. Mathematics
In this chapter, we digress a bit into a discussion of what it is that particle
physicists claim to be doing : confronting theory with reality. This leads to
some interesting subtleties. If we were mathematicians, the subject matter
in the previous chapter might be formulated as the following task : given
the parameters µ, λ3 and λ4 of the action, to compute the connected Green’s
functions. This ‘mathematician’s scheme’ may be depicted as follows :

µ , λ3 , λ 4 −→ C1 , C 2 , C 3 , C 4 , . . .

In this set-up, the parameters are supplied from outside the computational
and experimental context. Since, however, as (I hope) we are physicists
the situation is somewhat different : we first have to measure the values
of the parameters from inside the experimental context, using some of the
connected Green’s functions as measurement processes, and then predict some
other connected Green’s functions, which we shall call prediction processes.
This rather different situation may be depicted by the scheme

Ek = Ck , k = 1 . . . 4 −→ µ , λ3 , λ 4 −→ C5 , C 6 , C 7 , . . .

Here, the quantities E1,2,3,... stand for the experimentally observed values of
the connected Green’s functions : barring experimental errors or refinements,
these numerical values do not change under any improvement of the theory.

65
66 August 31, 2019

Now consider the fact that we are doing perturbation theory. That is, both
the measurement and the prediction processes are known only as truncated
series in ~. Let us suppose that by stolidity and perseverance a next higher
order in perturbation theory for the prediction processes has become avail-
able. Is this any good ? Obviously not, unless a similar increased level of
precision has been attained for the measurement processes. Only in that case
a new ‘fit’ of the parameters of the action can be made, and improved values
of the prediction processes can usefully be obtained. This order-by-order
improvement is called renormalisation. Let us denote by a superscript the
order to which the connected Green’s functions have been computed. The
‘physicist’s scheme’ can then be envisaged as follows :

(0) (0) (0) (0) (0)

Ek = Ck , k = 1 . . . 4 −→ µ(0) , λ3 , λ4 −→ C5 , C 6 , . . .
(1) (1) (1) (1) (1)
Ek = Ck , k = 1 . . . 4 −→ µ(1) , λ3 , λ4 −→ C5 , C 6 , . . .
(2) (2) (2) (2) (2)
Ek = Ck , k = 1 . . . 4 −→ µ(2) , λ3 , λ4 −→ C5 , C 6 , . . .
(3) (3) (3) (3) (3)
Ek = Ck , k = 1 . . . 4 −→ µ(3) , λ3 , λ4 −→ C5 , C 6 , . . .
.. .. ..
. . .

Order by order, the parameters keep getting updated, but in the overall pic-
ture they are just bookkeeping devices that allow one to go from measurements
to predictions of the more physically interesting connected Green’s func-
tions. It should not come as a surprise that in the measurement-parameter-
prediction protocol, a higher-order correction in the parameters due to an
improved measurement expression is cancelled again, to some extent, in the
prediction. In fact, for certain classes of theories, which are called renormal-
isable, these cancellations may be quite extreme.

2.1.2 The renormalisation program : an example

As an example of the renormalisation program, we shall investigate ϕ4 the-
ory, with coupling constant λ. In perturbation theory, the first few nonzero
connected Green’s functions are given by
 
~ 1 2 2 11 3
C2 = 1 − u + u − u + ··· ,
µ 2 3 8
August 31, 2019 67

~3 λ
 
7 149 2 197 3
C4 = − 4 1− u+ u − u + ··· ,
µ 2 12 4
~5 λ2
 
1535 2 6405 3
C6 = 10 − 80u + u − u + ··· ,
µ7 3 2
~7 λ3
 
111755 2 3
C8 = − 10 280 − 3815u + u − 330925u + · · · ,
µ 3
~9 λ4
 
12672800 2 3
C10 = 15400 − 310940u + u − 49859600u + · · · ,
µ13 3
(2.1)
where u is the ubiquitous combination ~λ/µ2 . Now suppose that the two
measurements give C2 = ~/m and C4 = −~3 g/m3 . In lowest order of per-
turbation theory, this would imply that the action parameters are µ = m
and λ = g. However, in higher orders we would no longer have C2 = E2 and
C4 = E4 . Instead, we must choose
 
1 7 2 1 3
µ = m 1 − y − y − y + ··· ,
2 12 8
 
3 3 2 11 3
λ = g 1 + y + y + y + ··· , (2.2)
2 4 8
with y = ~g/m2 . The properly renormalised Green’s functions then become
~
C2 = ,
m
~3 g
C4 = − 4 ,
m
5 2

~g 3 375 3
C6 = 10 − 15y + 45y − y + ··· ,
m7 2
~7 g 3
 
2 67725 3
C8 = − 10 280 − 1155y + 5775y − y + ··· ,
m 2
~9 g 4 2 3


C10 = 15400 − 118440y + 882000y − 6963075y + · · · . (2.3)
m13
The difference between the ‘naive’ and the renormalised connected Green’s
functions is quite evident. In particular C2 and C4 are completely free of
higher-order corrections. For the other connected Green’s functions the co-
efficients in the perturbation expansion are smaller in absolute value than in
the ‘naive’ expressions (see also below, section 2.4).
68 August 31, 2019

This discussion is obviously only a drastically simplified example of a

phenomenological situation that is in practice much more complicated. For
instance, one does not, usually, renormalise connected Green’s functions but
rather quantities extracted from scattering matrix elements, that are them-
selves not identical to, but extracted from connected Green’s functions. The
experimental observables E therefore do not take the simple form given here.
The higher-order corrections themselves are typically much more compli-
cated, and not completely free from ambiguities, nor necessarily finite. Nev-
ertheless, the operational scheme outlined above is essentially the same as
those that are employed in real-life physics. In particular, I cannot stress
often enough that the renormalisation procedure is necessary simply because
we do perturbation theory, not because loop corrections may contain infini-
ties1 .

2.2 A handle on loop divergences

2.2.1 A toy : the dot model
Notwithstanding the above remarks on the per se necessity of renormalisa-
tion, the fact that, in nontrivial theories, loop diagrams often contain in-
finities makes the need to do something about them all the more urgent.
Loop divergences arise from summation over internal degrees of freedom of
Feynman diagrams. In zero dimensions there are no such internal degrees
of freedom, and all diagrams are finite. We can, however, introduce the fol-
lowing toy model. Consider, as before, our working-horse ϕ3/4 theory. Let
us assume that we introduce yet another Feynman rule : we shall apply a
factor 1 + c1 to every closed loop that contains precisely one vertex, and a
factor 1 + c2 to every closed loop that contains precisely two vertices. Loops
with more vertices remain unaffected2 . The numbers c1 and c2 may depend
on the parameters of the theory, or on other parameters. In the spirit of
‘loop divergences’ we shall envisage that c1,2 → ∞ at some stage. In terms of
Feynman diagrams, this rule amounts to duplicating each one- or two-vertex

1
This insight is, even at present, not as endemic as it ought to be.
2
This rule accords with ‘naive power counting’ for four-dimensional scalar theories
without derivative couplings, the most direct four-dimensional extension of the zero-
dimensional theories we are discussing in this chapter.
August 31, 2019 69

loop with a ‘dotted’ loop :

= + , ≡ c1 × ,
= + , ≡ c2 × . (2.4)

For example, under this rule the following two-loop diagrams are modified
accordingly :

→ + + +

= (1 + c1 )(1 + c2 ) ,

→ +

= (1 + c2 ) . (2.5)

The Feynman diagrams are governed by the Schwinger-Dyson equation. Our

new rule must therefore be implemented, somehow, into a modified SDe.
Some reflection tells us that the necessary new ingredients are made up out
of those Feynman diagrams that contain only dotted loops. Fortunately, E15
these form a manageable set, where we differentiate between 1PI diagrams
with up to 4 legs3 :

≡ + + + ··· , (2.6)

≡ + + + ···

+ + + + ···

+ , (2.7)

≡ + + + ··· , (2.8)

≡ + + + ··· . (2.9)

3
With 5 or more legs our rule does not allow for diagrams with only dotted loops.
70 August 31, 2019

The only diagram that does not carry a ‘tower’ of loops on its back is the
last diagram in the two-point dotted series. Using these artefacts, we can
now rewrite the appropriate SDe for our ϕ3/4 theory with the added dotting
rule :
00000
11111 000000
111111
11111
00000 111111
000000
00000
11111 000000
111111
00000
11111 000000
111111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
= + + 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
+
00000
11111
00000
11111 000000
111111
000000
111111
000000
111111

11111
00000
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111 11111
00000
00000
11111 00000
11111
00000
11111
00000
11111
000000
111111
+ 00000
11111
00000
11111
00000
11111
00000
11111
+
000000
111111
000000
111111 00000
11111
000000
111111 00000
11111
00000
11111
000000
111111
000000
111111
000000
111111
000000
111111 000000
111111
000000
111111
111111
000000
000000
111111 111111
000000
000000
111111 000000
111111
000000
111111 000000
111111 111111
000000 000000
111111
000000
111111 000000
111111 000000
111111 111111
000000
000000
111111 000000
111111 000000
111111 000000
111111
000000
111111 000000
111111 000000
111111 000000
111111
000000
111111 000000
111111 000000
111111 000000
111111
000000
111111 000000
111111
000000
111111
000000
111111
111111
000000
000000
111111
+ 000000
111111
000000
111111
111111
000000
000000
111111
+ 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
+ 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
+
000000
111111
000000
111111 000000
111111
000000
111111 000000
111111 000000
111111
000000
111111 000000
111111 000000
111111 000000
111111
000000
111111 000000
111111 000000
111111
000000
111111
000000
111111 000000
111111
000000
111111
11111
00000
00000
11111
00000
11111
00000
11111 000000
111111 000000
111111
111111
000000
00000
11111 111111
000000 000000
111111
00000
11111 000000
111111
000000
111111 000000
111111
00000
11111 000000
111111 000000
111111
00000
11111

0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
+ 000000
111111
000000
111111
000000
111111
000000
111111
+ 000000
111111
000000
111111
000000
111111 +
0000
1111 000000
111111
0000
1111
00001111
11110000
0000
1111 0000
1111
0000
1111 0000
1111
00001111
11110000 0000
1111
0000
1111
0000
1111 0000
1111
0000
1111
0000
1111
0000
1111
00000
11111 00000
11111
11111
00000 00000
11111
00000
11111 11111
00000
00000
11111 00000
11111
00000
11111
00000
11111 00000
11111
00000
11111
00000
11111 00000
11111
00000
11111 00000
11111
00000
11111 00000
11111

00000000
11111111 00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
+ 00000000
11111111
11111111
00000000
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
+
000000
111111
111111
000000
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111

00000
11111
00000
11111
11111
00000
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111

11111111
00000000
00000000
11111111
00000000
11111111
00000000
11111111
+ 11111111
00000000
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
. (2.10)
00000000
11111111
00000000
11111111 00000000
11111111
00000000
11111111
00000000
11111111

We can readily translate this SDe into algebraic form. If we take out the
external propagators from the ‘black box’ graphs, we can write

= B1 , = B2 , = B3 , = B4 . (2.11)

We shall leave the actual evaluation of these sets of graphs for later (cf
section 2.3.5): at this point, we shall simply treat them as effective vertices.
The ‘dotted-loop’-modified SDe then reads, when we work out the graphs
one after the other, in the order in which they are displayed above :
J B1 B2 λ3 2 ~λ3 0
φ = − − φ− φ − φ
µ µ µ 2µ 2µ
B3 B3 2 ~B3 0 ~B3 0
− φ2 − φ − φ − φ
2µ µ 2µ µ
λ4 ~λ4 0 ~2 λ4 00
− φ3 − φφ − φ
6µ 2µ 6µ
August 31, 2019 71

B4 3 ~B4 0 ~B4 0 ~2 B4 00
− φ − φφ − φφ − φ . (2.12)
2µ 2µ µ 2µ
We can simply rewrite this SDe as
(µ + B2 )φ = (J − B1 ) − (λ3 + 3B3 )(φ2 + ~φ0 )
−(λ4 + 3B4 )(φ3 + 3~φφ0 + ~2 φ00 ) . (2.13)
But hold it ! This is just the SDe equation belonging to the action
1 1 1
S(ϕ) = B1 ϕ + (µ + B2 )ϕ2 + (λ3 + 3B3 )ϕ3 + (λ4 + 3B4 )ϕ4 . (2.14)
2 6 24
Therefore, the spirit of renormalisation tells us that in every application the
bare parameters µ, λ3 and λ4 will never occur on their own, but always only
in the renormalised combinations µR = µ + B2 , λR R
3 = λ3 + 3B3 , and λ4 =
λ4 + 3B4 ; and that therefore, whatever the values of B2,3,4 , the combination E17
will automatically be finite if the experimental quantities in which they enter
are finite. We can therefore choose the action’s parameters such that all
Green’s functions come out finite ; and the remaining B1 can always be
completely compensated by an extra ϕ1 in the action4 . Indeed, this is the way
in which the notorious ‘loop divergences’ are absorbed into the bare action :
infinite loop corrections are compensated for by infinite bare parameters5 .

2.2.2 Nonrenormalizable theories

The significant point in the discussion above is the fact that all dotted-
loop contributions can be absorbed into a finite number of terms of the bare
action. We may formulate the requirement of a renormalizable theory as that
which states that a finite number of measured quantities6 suffice to
make all other predictions of the theory well-defined. If an infinite
number of measured quantities would be necessary, the theory would be
called non-renormalizable : but, worse, from the operational point of view it
would be worthless7 . As an example of a non-renormalizable situation, let us
4
Here we introduce the notion of counterterms. More about this in section 2.5.
5
Or, in other words, at any loop order there are only finitely many counterterms to be
determined.
6
Think of E2,3,4, .
7
This train of thought might be relaxed. If the necessary additional experimental
values are only relevant at some very high energy scale, the theory would be effectively
renormalizable. It is a matter of taste whether you feel comfortable with this, or not.
72 August 31, 2019

consider a Feynman rule in which a loop with three vertices acquires a dotted
counterpart : that is, we would have a (potentially infinite) contribution of
the form

This can, of course, be repaired by introducing into the bare action a ϕ6

interaction ; but in that case there would arise dotted loops with eight ex-
ternal legs :

which would necessitate a ϕ8 interaction — and so on. A theory would arise in

which an infinite number of measured quantities would be needed before any
consistent8 prediction could be made : a non-renormalizeable situation ! The
same problem occurs in a theory with a bare ϕ6 interaction. It is seen that
the requirement of renormalizability puts constraints on the bare action9 .

2.3 Scale dependence

2.3.1 Scale-inpendent scale dependence
As mentioned above, the parameters of the action have to be determined
by comparison to experimentally measured quantities. Such measurement
experiments do not take place in some abstract realm10 , but rather in a
concrete physical situation. This experimental context partially determines
the measurement result. A very concrete example is the measurement of
the coupling constant using a scattering process : in that case, one of the
determining factors is the energy at which the scattering takes place. Also
choices made in the theoretical computation of the measured quantities play
their rôle : for example, in dimensional regularization11 an energy scale must
be introduced, and this scale can to a large extent be chosen arbitrarily. We
8
i.e. finite in high orders of perturbation theory.
9
It must come as no surprise that the Higgs potential of the Standard Model has no
interaction terms for the Higgs field (which is scalar) more complicated than the four-point
coupling.
10
Pace Platone.
11
To be discussed later on.
August 31, 2019 73

shall lump all these effects together into a quantity s, which we shall call the
scale. It must be stressed that the scale also contains the (regularized) loop
divergences, and may be expected to become infinite at some stage.
Let us consider a theory with only one parameter : an example of such
a theory is massless QCD, that is the theory of massless quarks and gluons
and their interactions. The single parameter is then the coupling constant.
Let the bare parameter, as it occurs in the action, be denoted by v. The
renormalised parameter, extracted from experiment, will be denoted by w.
The renormalised coupling is then given by the bare coupling and the exper-
imental context, embodied by the scale s :

w = F (v, s) . (2.15)

This relation ought to be invertible12 so that we can find v given w :

v = G(w, s) . (2.16)

Obviously we have

w = F (G(w, s), s) , v = G(F (v, s), s) . (2.17)

By differentiation we find the following relations between the derivatives of

F and G :

∂v F (v, s) ∂w G(w, s) = 1 , ∂s F (v, s) + ∂v F (v, s) ∂s G(w, s) = 0 , (2.18)

where ∂x stands for the partial derivative ∂/∂x.

Let us now consider an infinitesimal change in the scale. Since v is inde-
pendent13 of s, the value of w has to adapt itself in a manner prescribed by
F :
d
w = ∂s F (v, s) . (2.19)
ds
This expression contains the (divergent) scale s and the (divergent) value of
the bare coupling v. We can, of course, express everything in terms of w :
d
w = ∂s F (G(w, s), s) . (2.20)
ds
12
Usually, F is given as a (formal) series in v, for instance w = v + α2 v 2 + α3 v 3 + · · ·,
and then the inverse always exists , v = w − α2 w2 + (2α22 − α3 )w3 + · · ·
13
After all, a parameter in the action couldn’t possibly anticipate which experiment is
going to measure it !
74 August 31, 2019

This expression now contains the finite number w and the divergent scale s :
it can only make sense if s actually drops out. Note that this is not a proof,
it is a requirement that we make of the theory ! We therefore demand that
F (v, s) be such that

∂s F (G(w, s), s) = 0 . (2.21)
In other words,

∂s2 F (v, s) + ∂s ∂v F (v, s) ∂s G(w, s) = 0 . (2.22)

Using Eq.(2.18) and dividing by ∂v F we can write this as

∂s2 F
 
∂s F ∂s ∂v F ∂s F
− = ∂v =0 : (2.23)
∂v F (∂v F )2 ∂v F

this means that there is a function h(v) of v only, such that

∂s F (v, s) = h(v) ∂v F (v, s) . (2.24)

By separation of variables, Eq.(2.24) is easily solved, and we find

d 1
w = F (v, s) = f s + ρ(v) , ρ(v) = , (2.25)
dv h(v)
for some function f . There must be some value s0 for s, such that w and v
precisely coincide. This value can, of course, depend on v, so we write it as
s0 (v). We therefore have


f s0 (v) + ρ(v) = v , s0 (v) + ρ(v) ≡ j(v) , (2.26)

so that f and j are each other’s inverse : f (j(v)) = v. Applying j to Eq.(2.25)

we see that
s + ρ(v) = j(w) = s0 (w) + ρ(w) . (2.27)
In this equation, both terms on the left-hand side are divergent, but on the
right-hand side they are finite. We can now determine the scale dependence
of w. This dependence, called the beta function of the theory, is given by
d ∂

β(w) ≡ w = f s + ρ(v)
ds ∂s
1 h(w)
= f 0 j(w) = 0


= . (2.28)
j (w) 1 + h(w)s00 (w)
August 31, 2019 75

All reference to the bare coupling has been removed : we see that the renor-
malised coupling has a definite, predictable dependence on the energy scale
of the measuring experiment14 . Note, also, that whereas we introduced h as a
function of the bare parameter, it enters in the beta function as a function of
the renormalised parameter. Finally, as we shall see in the next section, the
h functions usually starts at second order (O (v 2 )) in perturbation theory, so
that the two first terms in the beta function are independent of the form of
s0 (v) as long as this is non-singular for v = 0.

2.3.2 Low-order approximation to the renormalised cou-

pling
Let us examine the possible shape of the function F (v, s) in some more detail.
In the spirit of perturbation theory, it will be given by a series expansion like

F (v, s) = v + v 2 α1 (s) + v 3 α2 (s) + v 4 α3 (s) + · · · . (2.29)

Let us assume that the functions αj (s) vanish at s = 0, so that s0 (v) = 0

and h(v) = β(v). The h function is given by

F2 (v, s) v 2 α10 (s) + v 3 α20 (s) + v 4 α30 (s) + · · ·

h(v) = = , (2.30)
F1 (v, s) 1 + 2vα1 (s) + 3v 2 α2 (s) + · · ·

so that we see that the beta function must start with v 2 :

β(v) = β0 v 2 + β1 v 3 + β2 v 4 + · · · (2.31)

The requirement that the beta function depend not on s governs the form of
the functions αj (s) : to low order in v we have from Eq.(2.30)

β(v) = v 2 α10 (s) + v 3 α20 (s) − 2α1 (s)α10 (s) + · · · ,

(2.32)
14
A remark is in order here. What, in these notes, is called the scale is usually under-
stood to be the logarithm of the actual energy scale : indeed, whereas the energy scale has
the dimension of energy (obviously), the number s is, strictly speaking, dimensionless. If
we denote the scale by the conventional symbol µ, the derivative dw/ds should then be
rewritten as
d d
w → µ w
ds dµ
.
76 August 31, 2019

so that we can derive

α1 (s) = β0 s , α2 (s) = (β0 s)2 + β1 s , . . . (2.33)

It is easily derived that the leading term in αn (s) is (β0 s)n .

Let us assume that the beta function is dominated by its lowest-order
term, that is, β(v) = β0 v 2 . Then ρ(v) = −1/(β0 v) and we find
1 1
= − β0 s . (2.34)
w(s) v
We can exchange the bare parameter v for the measured value of w at some
fixed scale s0 , and then the running is given by
1 1
= − β0 (s − s0 ) , (2.35)
w(s) w(s0 )
or
w(s0 )
w(s) = . (2.36)
1 − β0 w(s0 )(s − s0 )
At this point we may start to distinguish between different theories. The
renormalised, physical parameter w is a priori unknown, and has to be de-
termined by experiment ; but the number β0 is perfectly computable from
inside the theory15 . The running of the coupling is therefore determined as
soon as the action has been sufficiently specified. Now, it may happen that β0
is positive : in that case, the effective coupling w(s) increases with increas-
ing s, and will eventually become infinite at some high scale. On the other
hand, when β0 is negative, the effective coupling decreases with increasing
energy scale. This is called asymptotic freedom. It is the phenomenon that
has saved the theory of strong interactions : in the 1960’s when the typical
energy scales of experiments where low, the effective coupling was so high (of
order 10) as to cast doubts on the usefulness of perturbation theory, whereas
at the high energies current from around 197516 the effective coupling has
become small enough (of the order of 0.1) to warrant the use of perturbation
techniques.
15
The number β0 is a combinatorial factor with the addition of some powers of π, and
simple numbers depending on the ingredients and quantum numbers of the particles in
the theory.
16
I take the commissioning of the PETRA (Hamburg, BRD) and PEP (Stanford, USA)
colliders as the definitive starting point of the relevance of perturbative QCD.
August 31, 2019 77

2.3.3 Scheme dependence

We must recognize that not only the scale of a given measurement process is
important, but of course also the nature of the measurement process. That is,
we may define the measured coupling constant w in two different ways, on the
basis of two different measurement processes17 : let us denote the two results
by w and w̃. We say that such different values have been obtained using
different renormalisation schemes. In all cases I have encountered, two such
schemes agree at the tree level18 , and the results are therefore perturbatively
related :
w̃ = w + t1 w2 + t2 w3 + t3 w4 + · · · , (2.37)

with t1,2,3,... computable numbers ; and conversely

w = w̃ − t1 w̃2 + 2t1 2 − t2 w̃3 − 5t1 3 − 5t1 t2 + t3 w̃4 + · · ·

(2.38)

Having computed the beta function for w, we can now simply obtain it for
w̃ :

dw̃ dw̃ dw
β(w̃) = =
ds
 dw ds  
2 3 2 3 4
= 1 + 2t1 w + 3t2 w + 4t3 w + · · · β0 w + β1 w + β2 w + · · ·

= β0 w2 + (β1 − 2t1 β0 ) w3 + β2 − 2t1 β1 + 6t1 2 β0 − 3t2 β0 w4 + · · ·

= β0 w̃2 + β1 w̃3 + β0 t1 2 − β0 t2 + t1 β1 + β2 w̃4 + · · ·

(2.39)

The two beta functions can be transformed from one scheme to another ; for
any scheme dependence for which Eq.(2.37) holds, the first two coefficients,
β0 and β1 , are seen to be independent of the actual scheme, as was to be
expected (cf Eq.(2.28)) : the two schemes correspond to different functions
s0 (v) as defined in section 2.3.

17
In practice, this difference can be quite small, as between the so-called MS and MS
schemes. With ‘different measurement processes’, we here mean two different, complete
operational schemes that both lead to a well-defined value for coupling constants.
18
This rules out possible but, for a practicing physicist useless and/or irrelevant, differ-
ences such as for instance obtained by defining w̃ = 2w. Get a life !
78 August 31, 2019

2.3.4 Theories with more parameters

We can extend the investigation of section 2.3.1 to theories with n ≥ 2
parameters, labelled by a superscript. We then have

wi = F i (v, s) and v i = Gi (w, s) , i = 1, 2, . . . , n , (2.40)

with v = (v 1 , v 2 , . . . , v n ) and w = (w1 , w2 , . . . , wn ). The generalisations of

Eq.(2.18) then read :

∂` F i (v, s) ∂j G` (w, s) = δji , ∂s F i (v, s) + ∂j F i (v, s) ∂s Gj (w, s) = 0 ,

∂` Gi (w, s) ∂j F ` (v, s) = δji , ∂s Gi (w, s) + ∂j Gi (w, s) ∂s F j (v, s) = 0 ,
(2.41)

where ∂j stands for ∂/∂v j , and we adhere to the Einstein summation con-
vention. Note that the pullback function

Pji (v, s) ≡ ∂j Gi (F 1 (v, s), . . . , F n (v, s), s) (2.42)

is the inverse of the derivative of F :

Pji (v, s) ∂i F ` (v, s) = δj` = Pj` (v, s) ∂` F i (v, s) . (2.43)

The scale dependence of wk is given by

d k
w = ∂s F k (w, s) = ∂s F k (G(w, s), s) , (2.44)
ds
and as before we require the last form to be actually independent of s :

0 = ∂s ∂s F k (G(w, s), s) = ∂s2 F k (v, s) + ∂s ∂` F k (v, s) ∂s G` (w, s)

= ∂s2 F k (v, s) − ∂s ∂` F k (v, s) Pj` (v, s) ∂s F j (v, s) . (2.45)

Multiplying this by Pkj and using Eq.(2.43) twice we can derive

0 = Pkj (v, s) ∂s2 F k (v, s) − Pkj (v, s) ∂s ∂` F k (v, s) Pj` (v, s) ∂s F j (v, s)

= Pkj (v, s) ∂s2 F k (v, s) + ∂s Pkj (v, s) ∂` F k (v, s) Pj` (v, s) ∂s F j (v, s)

= Pkj (v, s) ∂s2 F k (v, s) + ∂s Pkj (v, s) ∂s F k (v, s) ,

(2.46)

as the generalisation of Eq.(2.23). Thus there must be functions hi (v) (i =

1, . . . , n) such that
Pkj (v, s) ∂s F k (v, s) = hj (v) . (2.47)
August 31, 2019 79

Upon multiplication by ∂j F m (v, s) we then find the analogue of Eq.(2.24) :

∂ m ∂
F (v, s) = hj (v) j F m (v, s) , m = 1, 2, . . . , n . (2.48)
∂s ∂v
Notice the occurrence of only a single differential operator hj (v)∂j !

2.3.5 Failure of the dot model

Let us apply the above to the ϕ4 theory in the dot model of section 2.2.1.
We can write Eq.(2.48) as follows:
∂ ∂
∂s K = p(u) λ K + q(u) µ K (K = λR , µR ) , (2.49)
∂λ ∂µ
and u = ~λ/(2µ2 ). In the dot model, the sums of dotted diagrams in
Eqns.(2.7,2.9) yields closed forms for the renormalised parameters :
   
R 3 R uc1 (s) 2 2 2
λ = λ −2 + , µ =µ 1− + u c2 (s) .
1 + uc2 (s) 1 + uc2 (s) 3
(2.50)
Applying Eq.(2.49) to λR then tells us that
−3u c02 (s) = p(u) 1 − 4u c2 (s) + 2u2 c2 (s)2 + 6q(u) u c2 (s) .


(2.51)
It is easy to see that p(u) = αu + O (u2 ), with some constant α, and
q(u) = O (u). The terms proportional to u1 in Eq.(2.51) demand that
c2 (s) = −α(s − s0 )/3 (with arbitrary s0 ). But in that case the right-hand
side of Eq.(2.51) cannot be made s-independent unless p(u) = q(u) = 0,
which implies that c1,2 do not depend on s at all. The dot model exhibits
a curious half-renormalisability, where the dots can be absorbed into renor-
malised parameters, but the running of these parameters cannot be made
scale-independent. For a viable, fully renormalisable theory the divergence
structure of the Feynman diagrams must be much more complicated and
subtle than simple ‘dotting’ !

2.4 Asymptotics of renormalisation in ϕ4 the-

ory
As mentioned in section 2.1.2, the perturbation expansion of the connected
Green’s functions appears to have numerically smaller coefficients than that
80 August 31, 2019

of the unrenormalised ones. We can investigate this for the ϕ4 theory sys-
tematically as follows. The bare connected Green’s functions are given by
 n
~ ~λ X (k)
C2n = (−u)n−1 t2n (u) , u = 2 , t2n (u) = t2n uk . (2.52)
µ µ k≥0

We employ a renormalisation scheme in which V2 and C4 are free of correc-

tions :
~ ~2 ~λ̂
C2 = , C4 = − 2 û , û = 2 , (2.53)
µ̂ µ̂ µ̂
where the renormalised quantities are indicated with a hat. Using
C4 u t4 (u)
û = − 2 = (2.54)
C2 t2 (u)2
we express û = u − 5u2 /2 + · · · as a series in u, which can be inverted to yield
u = û s(û) , (2.55)
and the series s(û) starts with unity. The bare and renormalised µ’s are then
related by

µ = µ̂ τ2 (û) , τ2 (û) = t2 û s(û) , (2.56)
and C2,4 are indeed free of corrections. For the higher connected Green’s
functions C2n , n ≥ 3 we then have
 n
~
C2n = (−û)n−1 τ2n (û) ,
µ̂

 n−1 X
t2n û s(û) s(û) (k)
τ2n (û) = = τ2n ûk . (2.57)
τ2 (û) τ2 (û) k≥0

All these manipulations are well suited to computer algebra. We can then
(k) (k)
compute the ‘improvement factor’ t2n /τ2n . The results for C6,8,10,12 are de-
picted in figure 2.1 as a function of k, the loop order, up to 155 loops. The
improvement is less for higher connected Green’s functions but at high order
the factors approach an asympotic value. We see that the asymptotic charac-
ter of perturbation theory is completely unchanged by renormalisation, since
neither the superexponential (n!) nor the purely exponential (cn ) or even
the polynomial (na ) behaviour of the coefficients is affected. The asymptotic
value of all improvement factors is exp(15/4) = 42.521 · · ·, which is explained
in appendix 15.15.3.
August 31, 2019 81

Figure 2.1: Improvement factor for ϕ4 theory in zero dimensions

2.5 The method of counterterms

2.5.1 Counterterms in the action
We can view the renormalization program in a slightly different way, by
insisting that the original parameters and the renormalized ones coincide,
and shifting the burden onto the action itself. That is, we include interaction
terms that are of higher order in ~, the so-called counterterms. For the ϕ4
theory this leads to
µ + δµ 2 λ + δλ 4
S(ϕ) → ϕ + ϕ ,
X2 4! X (k)
δµ = ~k δµ(k) , δλ = ~ k δλ . (2.58)
k≥1 k≥1

Inspecting the dimensions we see that we can write

λn (n) λn+1
δµ(n) = δn , δλ = ηn , (2.59)
µ2n−1 µ2n
where the δn and ηn are pure numbers. In the rest of this section we can
therefore use λ = µ = 1 without causing confusion. The counterterms (in-
cluding the δµ ) are treated as extra interactions, and we can choose them so
that they give us our favourite renormalization scheme, for instance we can
 82 August 31, 2019

again decide to have C2 and C4 free of loop corrections. Obviously it will be E16
sufficient to restrict ourselves to the 1PI diagrams. Denoting the counterterm
interactions (and their formal loop order) by dots, we start by computing the
one-loop correction to C2 :

1 1 1
+ = − − δ1 ⇒ δ1 = − . (2.60)
2 2

Similarly, for the one-loop correction to C4 we have

1 3 3
+ = − η1 ⇒ η1 = . (2.61)
2 2

And so we continue : at two loops we have

1 2
1
+ + + +
1 1 δ1 η1 7
= + + − − δ2 ⇒ δ1 = − ,
4 6 2 2 12
1 1 2
+ + + + +
3 3 3
= − − − 3 − 3δ1 − 3η1 − η2 ⇒ η2 = . (2.62)
4 2 4

E18 The resulting values for δ1,2 and η1,2 are of course precisely those we already
encountered in Eq.(2.2), only now we can compute them in a systematic,
E19
diagrammatic way.
E20

2.5.2 Return to the dot model, and a preview

If we compare our dotted-loop model with the counterterm approach, it be-
comes clear why the dot model fails : in essence, it only considers one-loop
couterterms, or in other words it assumes that all loops can only be divergent
in a single way. If (as we shall do from chapter 4 onwards) we recognize that
the loop divergences arise from momenta in the various loop reaching all the
way to infinity, then more-loop diagrams contain several loop momenta and
we will have to worry in which way they go to infinity, all together or one
after the other. In chapter 4 we shall come back to this issue.
August 31, 2019 83

2.6 Exercises
Excercise 15 Some reflection
Verify the claim made below Eq.(2.5) by showing that every diagram contain-
ing dotted loops can be assigned its place in the modified SDe of Eq.(2.10).

Excercise 16 Argue it yourself

Give the argument for restricting ourselves to the 1PI diagrams only in the
computation of the counterterms as discussed in section 2.5.

Excercise 17 Dotting in ϕ3
Show that for pure ϕ3 theory (never mind its ill-defined character) there
are only two dotted 1PI diagrams possible. There are therefore no ‘infinite
towers’ of dotted diagrams. Write the appropriate SDe in this dotted model.
Such theories, where only a finite number of diagrams diverge, are called
super-renormalizable. Show that, in this case, we can afford to also dot loops
with 3 propagators, and end up with a renormalizable theory19 .

Excercise 18 Counterterms at three loops

The three-loop 1PI contribution to C2 has 5 distinct diagrams without coun-
terterms, 6 with one 1-loop counterterm, 2 with one 2-loop counterterm,
and 2 with two 1-loop counterterms : and, of course, 1 diagram with the
3-loop counterterm. Write these out, and compute their symmetry factors
and multiplicities, and compute from them that δ3 = −1/8.
19
This would correspond with a ‘realistic’ ϕ3 theory in 6 dimensions, and this is why a
6-dimensional ϕ3 theory makes a nice playground to study renormalization.
84 August 31, 2019

Excercise 19 Counterterm at three loops (cont’d)

The three-loop 1PI contribution to C4 has 12 distinct diagrams without coun-
terterms, 14 with a single one-loop counterterm, 2 with one 2-loop countert-
erm, and 4 with two 1-loop counterterms : and, of course, 1 diagram with the
3-loop counterterm. Write these out, and compute their symmetry factors
and multiplicities, and compute from them that η3 = 11/8.

Excercise 20 Employing the counterterms

Write the one-loop diagrams for C6 , including the counterterm diagrams.
From that, compute the renormalized value of the one-loop correction to C6 ,
and compare to Eq.(2.3).
Chapter 3

More fields in zero dimensions

3.1 The action and the path integral

We now have established an overview of the quantum-field theoretic be-
haviour of a single field ϕ in zero dimensions. Generalising to a theory
with more than one field is fairly straightforward, and we shall do so in
this chapter. We shall assume that there are K distinct fields, labelled ϕj ,
j = 1, 2, . . . , K. The number K can be taken as large as we please, and even
infinite provided that the fields form a countable set1 . These fields have a
combined probability density given by
 
1
P (ϕ1 , ϕ2 , . . . , ϕK ) = N exp − S(ϕ1 , ϕ2 , . . . , ϕK ) , (3.1)
~
where we have immediately introduced ~ since we are now familiar with it.
In the special case where the action is separable, that is,
S(ϕ1 , ϕ2 , . . . , ϕK ) = S1 (ϕ1 ) + S2 (ϕ2 ) + · · · + SK (ϕK ) ,
the fields are actually independent random variables ; the theory is just so
many copies of the single-field one, and in the following we shall disregard
that uninteresting situation. In concert with our convention of having a
coupling constant λn accompanied by a factor 1/n!, we shall let the cou-
pling of several fields be accompanied by combinatorial factors for each field
separately, so the action may contain a term
λ1,3,7
ϕ1 2 ϕ3 2 ϕ7 4 ,
(2!)(2!)(4!)
1
This notion will be relaxed later on, as we move to higher dimensions.

85
86 August 31, 2019

which gives the Feynman rule −λ1,3,7 /~ for this vertex.

The Green’s functions are defined by

Gn1 ,n2 ,...,nK = hϕ1 n1 ϕ2 n2 · · · ϕK nK i (3.2)
In order to be able to keep the various fields apart, we have to assign to each
of them its own source Jj , j = 1, 2, . . . , K. The path integral therefore reads2
X J n1 · · · J nK
1 K
Z(J1 , . . . , JK ) = Gn1 ,...,nK
n1,...,K ≥0
n 1 ! · · · n K !
Z K
!!
1 X
= N exp − S(ϕ1 , . . . , ϕK ) − Jj ϕj dϕ1 · · · dϕK . (3.3)
~ j=1

The extraction of the Green’s functions is then performed as

∂ n1 ∂ nK
 
n1 +···nK
Gn1 ,...,nK = ~ ··· Z(J1 , . . . , JK ) .
(∂J1 )n1 (∂JK )nK J1 =···=JK =0
(3.4)

3.2 Connected Green’s functions and field func-

tions
The relation between the Green’s functions and their connected counterparts
is again given by straightforward generalization:


W (J1 , . . . , JK ) = log Z(J1 , . . . , JK )
X J n1 · · · J nK
1 K
= Cn1 ,...,nK (3.5)
n ≥0
n 1 ! · · · n K!
1,...,K

The precise expression of the G’s in terms of the C’s is of course now some-
what more involved : for instance, for K = 3 we have
G1,0,0 = C1,0,0 ,
G1,1,0 = C1,0,0 C0,1,0 + C1,1,0 ,
G1,1,1 = C1,0,0 C0,1,0 C0,0,1 + C1,1,0 C0,0,1
+ C1,0,1 C0,1,0 + C0,1,1 C1,0,0 + C1,1,1 . (3.6)
2
Where possible, we denote multiple integral by a single integration sign. This usually
does not lead to confusion.
August 31, 2019 87

We now have K field functions, one for each field ; they are given by
∂
φj (J1 , . . . , JK ) = ~ W (J1 , . . . , JK ) . (3.7)
∂Jj
An important thing to note is that, since the field functions are derivatives,
∂ ∂
φj (J1 , . . . , JK ) = φk (J1 , . . . , JK ) . (3.8)
∂Jk ∂Jj

3.3 The Schwinger-Dyson equation

The SDe for the path integral can be summarized as follows :
 
∂
S(ϕ1 , . . . , ϕK ) Z(J1 , . . . , JK ) = Jk Z(J1 , . . . , JK ) , (3.9)
∂ϕk ϕj =~∂/∂Jj

as can easily be verified. For the field functions, the SDe is best illustrated
with an example. Suppose that we have the following action for K = 2 :
1 1 λ
S(ϕ1 , ϕ2 ) = µ1 ϕ1 2 + µ2 ϕ2 2 + ϕ1 2 ϕ2 2 . (3.10)
2 2 4
This time, the coupling constant λ carries a factor 1/(2!)/(2!) since there
are not four identical fields ‘meeting’ at the vertex, but rather two pairs of
identical fields, as mentioned above. We indicate the field type with either
‘1’ or ‘2’. The Feynman rules for this case are
~ ~ 1 2 −λ
1
↔ , 2
↔ , ↔ ,
µ1 µ2 1 2 ~
1 J1 2
J2
↔ , ↔ . (3.11)
~ ~
There are two coupled Schwinger-Dyson equations, one for each field :
1111
0000
0000
1111
0000
1111
0000
1111
0000
1111
1111
0000
0000
1111
1 0000
1111
0000
1111
0000
1111 0000
1111
0000
1111
1 0000
1111
0000
1111 1 1 2
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
= + 00000
11111
00000
11111
00000
11111
00000
11111
0000
1111 00000
11111
0000
1111 0000
1111
00000
11111
2 0000
1111
00000
11111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
000000
111111
000000
111111 000000
111111
111111
000000 000000
111111
1 111111
1 1 111111
000000
000000
111111
000000
111111 1 2 000000
111111
000000
111111
000000
111111
000000
000000
111111
111111
000000
000000
111111
111111
000000 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111 000000
111111 1 2 000000
111111
000000
111111
+ 2
00000
11111
+ 1
00000
11111
+ 2
000000
111111
000000
111111
000000
111111
000000
111111
,
11111
00000
00000
11111 11111
00000
00000
11111
2 00000
11111
00000
11111
00000
11111
2 00000
11111
00000
11111
00000
11111
00000
11111
00000
11111 00000
11111
00000
11111
00000
11111
00000
11111 00000
11111
00000
11111
 88 August 31, 2019
0000
1111
1111
0000
0000
1111
0000
1111
2 0000
1111
1111
0000 0000
1111
0000
1111
0000
1111 0000
1111
0000
1111 0000
1111
0000
1111
0000
1111 2 2
2 0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
= + 00000
11111
00000
11111
1
00000
11111
00000
11111
0000
1111 00000
11111
0000
1111 1 0000
1111
00000
11111
1111
0000
00000
11111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
000000
111111 000000
111111
111111
000000 000000
111111
2 000000
111111
000000
111111
000000
111111
1 111111
000000 1 111111
000000
000000
111111
000000
111111
000000
111111 000000
111111 000000
111111
2 000000
111111
000000
111111
000000
111111
2 000000
111111
111111
000000
000000
111111
000000
111111
1 111111
000000
000000
111111
2 000000
111111
+ 1
00000
11111
+ 2
00000
11111
+ 000000
111111
2111111
000000
000000
111111
000000
111111
,
11111
00000 11111
00000
00000
11111 00000
11111
00000
11111
00000
11111 00000
11111
00000
11111
1 00000
11111
00000
11111
1 00000
11111
00000
11111
00000
11111 00000
11111
00000
11111 00000
11111
00000
11111 00000
11111

(3.12)

E21 with the following analytical representation for the field functions φj =
φj (J1 , J2 ) (j = 1, 2) :
2
 
J1 λ 2 ∂ ∂ 2 ∂
φ1 = − φ1 φ2 + ~φ1 φ2 + 2~φ2 φ1 + ~ φ1 ,
µ1 2µ1 ∂J2 ∂J2 (∂J2 )2
2
 
J2 λ 2 ∂ ∂ 2 ∂
φ2 = − φ2 φ1 + ~φ2 φ1 + 2~φ1 φ2 + ~ φ2 .
µ2 2µ2 ∂J1 ∂J1 (∂J1 )2
(3.13)

The effective action must of course be a two-variable function Γ(φ1 , φ2 ) such

that
∂
Γ(φ1 , φ2 ) = Jj , j = 1, 2 . (3.14)
∂φj
E22 This effective action is also concave. The two-field case can, obviously, be
extended to the case of arbitrarily many fields, provided the couplings are
E23
unambiguously defined.

3.4 The sum rules revisited

Chapter 1 discusses the diagrammatic sum rules for a theory with only one
field (derived in appendix 15.3). Of course, we have to extend this also. Let
us therefore assume various fields, labelled by an index j. Then, E (j) and
I (j) are the number of external and internal lines of type j, respectively, and
(j)
by Vq we denote the number of vertices that have precisely q legs of type j
sticking out. The generalization of Eq.(15.21) then reads
X
qVq(j) = 2I (j) + E (j) ∀ j , (3.15)
q
August 31, 2019 89

and that of Eq.(15.22) takes the form

X X
Vq(j) = I (j) + P − L . (3.16)
j,q j

As before, these sum rules are valid in any nonzero dimension as well.

3.5 A zero-dimensional toy for QED

3.5.1 Fields and sources
We consider the following action for three fields, including sources :
1 ¯ − ϕ̄J − HB .
S(ϕ, ϕ̄, B) = µB 2 + mϕϕ̄ + eϕ̄Bϕ − Jϕ (3.17)
2
The field ϕ̄ is not automatically 3 the ‘conjugate’ of ϕ in any sense : it is just
different from ϕ. Note the absence of symmetry factors since all the fields in
the three-point vertex are distinct. Also the two-point interaction term mϕ̄ϕ
carries no factor of 1/2. Such an action can stand for an extremely primitive
model for QED, the theory of electrons and photons. The action has three
partial derivatives :
∂
S(ϕ, ϕ̄, B) = mϕ̄ + eϕ̄B − J¯ ,
∂ϕ
∂
S(ϕ, ϕ̄, B) = mϕ + eBϕ − J ,
∂ ϕ̄
∂
S(ϕ, ϕ̄, B) = µB + eϕ̄ϕ − H . (3.18)
∂B
The SDe’s for the path integral are therefore
2
 
∂ 2 ∂ ¯ ¯ J, H) = 0 ,
~m + e~ − J Z(J,
∂J ∂J∂H
2
 
∂ 2 ∂ ¯ J, H) = 0 ,
~m ¯ + e~ ¯ − J Z(J,
∂J ∂ J∂H
2
 
∂ 2 ∂ ¯ J, H) = 0 .
~µ + e~ ¯ − H Z(J, (3.19)
∂H ∂ J∂J
3
See, however, section 3.5.2.
 90 August 31, 2019

The field-generating functions (the ‘field functions’) are, of course, each a

function of J, J¯ and H, and are given by
∂ ∂ ∂
ψ = ~ ¯ log Z , ψ̄ = ~ log Z , A=~ log Z , (3.20)
∂J ∂J ∂H
so that
∂ ∂ ∂
~ ¯Z = ψ Z , ~ Z = ψ̄ Z , ~ Z = AZ , (3.21)
∂J ∂J ∂H
and Eq.(3.18) can be written as
 
1 e ∂
ψ = J− Aψ + ~ ψ ,
m m ∂H
 
1 ¯ e ∂
ψ̄ = J− ψ̄ A + ~ ψ̄ ,
m m ∂H
 
1 e ∂
A = H− ψ̄ ψ + ~ ψ . (3.22)
µ µ ∂J

We may rewrite these SDe’s since

∂ ∂ ∂ ∂ ∂ ∂
ψ = ¯A , ψ̄ = H , ψ = ¯ψ̄ . (3.23)
∂H ∂J ∂H ∂J ∂J ∂J
The Feynman rules are, for this action, as follows :
~
ψ ψ ↔ ,
m
~
↔ ,
µ
e
↔ − . (3.24)
~

A few things are of interest here. In the first place, we have here the first
instance of an important concept : that of oriented lines. The propagator is
oriented, it runs from ϕ to ϕ̄. In the second place, in the action we find the
two terms Jϕ ¯ and ϕ̄J, which would suggest that J is the source in the SDe
¯
of ψ̄, and J is the source in the SDe for ψ ; but it is actually the other way
E24 around ! What is the source for a given field function is seen by taking the
derivative of the action, and inspecting which field then occurs as a linear
term, and which source term is left by itself after the differentiation.
August 31, 2019 91

3.5.2 Bald, furry and quenched toys

We can try to recast the above ‘QED’ in the form of a path integral. We
then immediately run into problems ! The integral over B is doable, but the
integral of exp(−mϕ̄ϕ/~) diverges if we simply take ϕ and ϕ̄ to be unrelated
real fields. The way out is to say that, after all, ϕ̄ is the complex conjugate
of ϕ so that ϕ̄ϕ is nonnegative. The action can then be written as
 2 2
µ 2 J  − |J|
 
S(ϕ, ϕ̄, B) = B + (m + eB) ϕ − − HB
2 e + mB  m + eB
µ 2 2 |J|2
= B + (m + eB)r − − HB , (3.25)
2 m + eB
where we have shified ϕ to ϕ0 = ϕ − J/(m + eB) as integration variable, then
wrote dϕ dϕ̄0 ∝ r dr with r = |ϕ|. We can now integrate over the ϕ, ϕ̄ fields :
Z   Z  
2 1 1
Z ∼ dB d(r ) exp − S(ϕ ϕ̄, B) ∼ dB exp − S(B) ,
~ ~
2
 
µ |J| eB
S(B) = B 2 − HB − + ~ log 1 + . (3.26)
2 m + eB m
Here we have dropped overal constant factors at will, and reabsorbed the
result of the r2 integral into the exponent. For reasons that shall appear
below, we can call this Bald QED.
How can we interpret the term ~ log(1 + eB/m) ? Let us consider the set
of all diagrams with a single closed ϕ loop and n B legs attached, with the
B propagators removed : E25

+ + + ··· (3.27)

The diagram with n B legs evaluates to (n − 1)!(−e/m)n . Suppose that we

wish to avoid these ϕ loops altogether. This can be done by introducing
counterterms as described in chapter 2 : these counterterms are n-point B
self-interactions vertices, that can be summed :
X  e n B n 
eB

~ − = −~ log 1 + . (3.28)
n≥1
m n m

We see that the logarithmic term in S(B) simply embodies the closed ϕ
loops ! As we shall see in section 10.2.7, in ‘grown-up’ QED we have Furry’s
92 August 31, 2019

theorem that states that ϕ loops with an odd number of B legs must vanish
when summed correctly. We can implement this by using counterterms for E26
only these odd-n loops :
 2k+1     
X ~ eB ~ eB eB
− = − log 1 + − log 1 − . (3.29)
k≥0
2k + 1 m 2 m m

Thus we arrive at three formulations for ‘zero-dimensional QED’ : Bald, with

no modifications ; Furry, with Furry’s theorem implemented ; and Quenched,
where we suppress all closed ϕ loops4 . Their respective S(B) actions are :
|J|2
 
µ 2 eB
Bald : S(B) = B − HB − + ~ log 1 + ,
2 m + eB m
|J|2 e2 B 2
 
µ 2 ~
Furry : S(B) = B − HB − + log 1 − ,
2 m + eB 2 m2
µ |J|2
Quenched : S(B) = B 2 − HB − . (3.30)
2 m + eB

3.6 Exercises
Excercise 21 Two-field action
Consider the two-field action, now with the sources included :
µ
λ
S(ϕ1 , ϕ2 ) = ϕ1 2 + ϕ2 2 + ϕ1 2 ϕ2 2 − J1 ϕ1 − J2 ϕ2
2 4
1. Determine the 2 SDe’s for the path integral.
2. From these, determine the 2 SDe’d for the field functions φ1 and φ2 .
3. Verify that this is agreement with the diagrammatically obtained result.

Excercise 22 A O(N ) symmetric theory

Consider a theory containing N fields ϕ1,2,...,N , with the following action :
µ 2 λ
S(~
ϕ) = |~
ϕ| + |~ϕ|4 ,
2 24
and with sources J1,2,...,N . Since the action is invariant under any real rotation
in the space of vectors (ϕ1 , ϕ2 , . . . , ϕN ) it is called O(N ) symmetric.
4
Leaving plenty of other loops !
August 31, 2019 93

1. Show by using diagrams that the SDe can be written as

 
λ 2 ~ ~ ∂ ~2 2~ 2
µφk = Jk − φk |~
ϕ| + ~φk (∇ · φ) + ~ |φ| + ~ ∇ φk
6 ∂Jk
Notice that to obtain this form one may have to use the fact that
∂φk /∂Jm = ∂φm /∂Jk .
2. Show that the same result can be obtained algebraically, by working
out de SDe in the following form :
 
Sk φ ~ + ~∇
~ e = Jk

where Sk stands for the partial derivative of the action S(~

ϕ) to ϕk , and
∇k stands for ∂/∂Jk .
~ alone. We can therefore write
3. By symmetry, W must depend on |J|
 
~ 2 /2
φk = Jk F |J|

for some function F . Show that this function obeys


λ
2xF (x)3 + ~ (N + 2)F (x)2 + 6xF (x)F 0 (x)


µF (x) = 1 −
6

2 0 00


+ ~ (N + 2)F (x) + 2xF (x)

Excercise 23 The 123 theory

We consider a theory containing not 1 but 3 fields, labeld ϕ1,2,3 . The action,
including the sources, is given by
µ
ϕ1 2 + ϕ2 2 + ϕ3 2 + gϕ1 ϕ2 ϕ3 − J1 ϕ1 − J2 ϕ2 − J3 ϕ3


S(ϕ1 , ϕ2 , ϕ3 ) =
2
In the following we shall concentrate on Feynman diagrams and will not
worry about the convergence of the path integral.
1. Prove that there are now 3 different SDe’s, of the form
∂ ∂ ∂
µ~ Z + g~2 Z = Ji Z
∂Ji ∂Jj ∂Jk
where i, j, k is a permutation of 1,2,3.
94 August 31, 2019

2. There are now of course also 3 field functions φ1,2,3 . Prove that

∂ ∂
φj = φi , i, j ∈ (1, 2, 3)
∂Ji ∂Jj

3. Give the SDe for the field functions using diagrams.

4. Prove that for any diagram with nj external lines of type j (j = 1, 2, 3)

the following must hold: the nj are either all even, or all odd.

5. By Cij we denote the connected Green’s function with two external

legs, one of type i and one of type j. Prove that

Cij = 0 , i 6= j ,

and furthermore that there are no tadpole diagrams.

Excercise 24 Symmetry factors in QED

For the model of section 3.5, write the three SDe’s in diagrammatic form.
Show that all symmetry factors are equal to unity for diagrams with at
least one external leg. Show that this implies the same for actual Quantum
Electrodynamics, that is, the same model but extended to Minkowski space
and with more complicated propagators and vertices.

Excercise 25 Throwing in the loops in QED

Compute the diagrams in Eq.(3.27), and prove the result in the text. Keep
track of the symmetry and multiplicity factors ! Use this to show the cor-
rectness of the counterterms in Eq.(3.28).

Excercise 26 Furry’s theorem in higher loops

Prove the following : if Furry’s theorem holds at one-loop order, it holds to
all loop orders.
Chapter 4

QFT in Euclidean spaces

4.1 Introduction
The main characteristic of a space(-time) of more than zero dimensions is
the fact that the quantum field is defined at more than one point ; in fact,
at an infinity of points. The possibility of sending signals from one point to
another one requires the existence of correlations between the field values
at different points. The nature of this correlation, and its reflection in the
appropriate Feynman rules, is our subject now.

4.2 One-dimensional discrete theory

4.2.1 An infinite number of fields
We shall consider a theory of a countably infinite set of fields in zero dimen-
sions. We denote by {ϕ} the set of all these fields :

{ϕ} = . . . , ϕ−3 , ϕ−2 , ϕ−1 , ϕ0 , ϕ1 , ϕ2 , ϕ3 , . . .

where the field labels run from −∞ to +∞. The collection of all their
corresponding sources is denoted by {J}. We shall, as a working example,
consider a theory where the interaction consists of four fields with the same
label meeting at one point. Moreover, we shall assume the kinetic terms to

95
96 August 31, 2019

be uniform in the field labels. Thus, the action will be1 :

X 1 λ4 4

2
S({ϕ}, {J}) = µϕn − γϕn ϕn+1 + ϕn − Jn ϕn , (4.1)
n
2 4!

where we include the sources in the action2 . If γ were zero, the action would
be separable and the theory would be an uninteresting series of replicas of the
zero-dimensional action for a single field. We shall consider positive values of
γ ; in that case, the action tends to minimize if ϕn and ϕn+1 carry the same
sign : a positive correlation between ‘neighbour’ fields is the result. Note,
moreover, that the action is, by choice, invariant under the relabelling of n
by n + K with any fixed K : this is called translation invariance, in this case
translation by a fixed increment in labelling3 . The model is also invariant
under the relabelling of n by −n : this is called parity invariance.
The Feynman rules are easily derived from the action of Eq.(4.1) :

n ~
↔
µ

n m γ

↔ + δm,n+1 + δm,n−1
~

n n λ4
n n ↔ −
~

Jn
n ↔ +
~
Feynman rules, version 4.1 (4.2)

1
If not indicated explicitly otherwise, sums will run from −∞ to +∞.
2
Both µ and γ are independent of n simply because I choose them so. I might choose
differently, only I don’t want to.
3
This will lead to momentum conservation later on. Note however that, as indicated
above, momentum conservation is a consequence of our choice, or in practice of our belief
in the translation invariance of our physical laws. Other models are possible and not a
priori wrong : they are simply much more complicated.
August 31, 2019 97

The identity of the field is indicated by its label. Alternatively, the four-
vertex and the source vertex may be labelled. The SDe now takes the fol-
lowing form, for any n :
111111
000000
000000
111111 111111
000000 111111
000000
000000
111111 000000
111111
000000
111111 000000
111111
000000
111111
000000
111111
000000
111111 000000
111111 000000
111111
n 000000
111111 n 000000
111111
000000
111111 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
= + n n+1
000000
111111
000000
111111
000000
111111
000000
111111
+ n n−1
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111 000000
111111
000000
111111
000000
111111
000000
111111
11111
00000 111111
000000
000000
111111 1111
0000
0000
1111
00000
11111 000000
111111 0000
1111
00000
11111
00000
11111 000000
111111 0000
1111
n 00000
11111 n 000000
111111
000000
111111 0000
1111
0000
1111
+ 0000
1111
00000
11111
+ 000000
111111
000
111
000
111
+ n
0000
1111
0000
1111
0000
1111
, (4.3)
1111
0000 000
111 0000
1111
0000
1111
0000
1111
0000
1111
000
111
000
111
0000
1111
0000
1111
0000
1111 000
111
0000
1111
0000
1111 000
111
000
111
0000
1111
0000
1111
0000
1111 000
111
0000
1111
0000
1111 000
111
000
111

or, in terms of the field functions φn ({J}), that depend on all sources :
Jn γ
φn = + (φn−1 + φn+1 )
µ µ
∂2
 
λ4 3 ∂ 2
− φn + 3~φn φn + ~ φn . (4.4)
6µ ∂Jn (∂Jn )2

4.2.2 Introducing the propagator

The Schwinger-Dyson equation (4.3) can be cast in another, more useful
form. Consider the fact that, upon entering the field function via its exter-
nal leg, one must encounter either zero or more two-point functions before
encountering a source vertex or a four-vertex. Let us denote by

n m
Πm,n ≡ (4.5)
the total set of diagrams that contain only two-point vertices (or no vertices),
and have fields n and m at its external legs4 . The SDe can then be rewritten
as follows :

000000
111111
000000
111111
000000
111111
000000
111111 n
n 000000
111111
000000
111111
X k k
000000
111111
000000
111111
000000
111111
= +
000000
111111 n

000000
111111
000000
111111
k


n k 
+ + n  ,

(4.6)
k 

4
To go from ϕn to ϕm one needs, of course, at least |n − m| vertices, but more vertices
are also possible.
98 August 31, 2019

where we must sum over the label of the field exiting the Π. Therefore, we
have
X
φn = Πn,k ×
k
∂2
 
Jk λ 3 ∂ 2
− φk + 3~φk φk + ~ φk . (4.7)
µ 6µ ∂Jk (∂Jk )2

The object Πn,k describes to what extent the fields ϕn and ϕk influence one
another : it will be called the propagator from now on.

4.2.3 Computing the propagator

From the translation and parity invariance of the model we have discussed,
we can infer that Πn,k can actually only depend on |n − k|, so that we can
restrict ourselves to Π0,n ; we denote this by Π(n). For Π(n), we have a very
simple Schwinger-Dyson equation :

0 n = 0 n + n + n , (4.8)
0 1 0 −1

or  
~ γ
Π(n) = δ0,n + Π(n + 1) + Π(n − 1) . (4.9)
µ µ
The easiest way to solve this set of equations is by Fourier transform. We
define5 X
R(z) = Π(n) e−inz , (4.10)
n
6
from which the propagator may be recovered using

Z+π
1
Π(n) = e+inz R(z) dz . (4.11)
2π
−π

5
We are forced to use exp(−inz), with absolute value unity ; any other absolute value
would make the generating function divergent either as n → ∞ or n → −∞.
6
We choose e−inz rather than e+inz in Eq.(4.10) by convention. Although this may
not be completely, glaringly obvious at this point, this convention is ultimately related to
the fact that, in nonrelativistic quantum mechanics, the Schrödinger equation has been
chosen to read i~∂|ψi/∂t = Ĥ|ψi rather than −i~∂|ψi/∂t = Ĥ|ψi.
August 31, 2019 99

Multiplying both sides of Eq.(4.9) by exp(−inz) and summing over n leads

to
~ γ
+ R(z) eiz + e−iz


R(z) =
µ µ
~ ~u
= iz −iz
= (4.12)
µ − γ (e + e ) µu − γ(u2 + 1)

where we have introduced u = eiz . This allows us to write the integral (4.11)
as
un
I
~
Πn = − du , (4.13)
2iπγ (u − u+ )(u − u− )
|u|=1

where u± are the two roots of the quadratic form µu − γ(u2 + 1) :

1  p 
u± = µ ± µ2 − 4γ 2 . (4.14)
2γ
Provided that µ exceeds 2γ, the two poles of the integrand are real, and
0 < u− < 1 < u+ . We can then contract the contour around the point
u = u− , upon which we find
u− n
Π(n) = ~ , n≥0 . (4.15)
γ(u+ − u− )

The general solution for the propagator is therefore7

~
Π(n) = p u− |n| . (4.16)
µ2 − 4γ 2

Unsurprisingly, the propagator falls off exponentially with |n|. Note that if E27
γ were negative, then u− would also be negative, and the propagator would
oscillate between positive and negative correlations. Moreover if µ were 2γ
or smaller the poles of the integrand would lie on the unit circle |u| = 1,
making the integral ill-defined.
With this explicit form of the propagator, we can now switch to a new
set of Feynman rules :
7
This derivation is valid for n ≥ 0. For negative n, Cauchy’s theorem on which it is
based does not hold immediately : but in that case we can perform the variable transform
from u to 1/u and obtain the result.
100 August 31, 2019

m
n ↔ Π(m − n)

n n λ4
n n ↔ −
~

n Jn
↔ +
~
Feynman rules, version 4.2 (4.17)

The difference with the previous set of rules is that now the line denotes a
propagator running between n and m. The SDe is now very similar to that
of the zero-dimensional ϕ4 theory :

1111
0000
0000
1111
0000
1111
000000
111111 0000
1111
0000
1111
111111
000000
000000
111111
000000
111111
X n m n m 0000
1111
0000
1111
000000
111111 000
111
0000
1111
= +
n 000000
111111
000000
111111
 000
111
000000
111111
000000
111111
 000
111
000
111
000000
111111
000000
111111
000000
111111 000
111
000
111
0000
1111
m
 000
111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111

1111
0000
0000
1111 1111
0000

n 0000
1111 n m 0000
1111
0000
1111
+ 0000
1111 + 0000
1111
 . (4.18)

m 0000
1111 0000
1111
0000
1111
0000
1111 0000
1111
0000
1111
000
111 0000
1111
000
111
000
111 0000
1111 
000
111 0000
1111
0000
1111
000
111 0000
1111
000
111
000
111 0000
1111
000
111
000
111

4.2.4 Figments of the imagination : a sermon

The concept of an infinite number of fields all huddling together at a single
point simply cries out for a better visualization. The most useful picture is
that of each field occupying its own point. Indicating by a line those fields
that have a direct coupling, we arrive at a picture like

ϕ ϕ ϕ ϕ ϕ3
−1 0 1 2

where we now need to introduce a new notion, that of distance. In our

sensorial experience, distances are, in their essence, measured by the sending
and receiving of signals, and the weaker the signal from one point to another,
August 31, 2019 101

the further those points are deemed to be apart ; in the language of these
notes, the smaller Π(m − n), the larger the ‘distance’ between n and m. We
can therefore dress up our picture by introducing a fundamental distance ∆,
subsequent field locations being separated by this distance,

∆ ∆ ∆ ∆ ∆ ∆
}
}
}
}
}
}
ϕ ϕ ϕ ϕ ϕ3
−1 0 1 2

where the distances between the successive points are all equal since the
couplings γ are all equal. We have, as it were, constructed a one-dimensional
universe. It may come as a surprise that the concept of space is here presented
as a visualization device. If we reflect, however, on how someone who (like
a new-born infant) has no a priori concept of spacelike separations would
have to envisage the workings of the physical world, we shall conclude that
that person had better invent space in order not to go insane pretty quickly.
In its essence, space, like so much else in the world around us, is simply a
mental construction that allows us to come to grips with, and control, our
environment8 .
After all this has been said, we must acknowledge the empirical fact
that to our knowledge space seems not to be made up from single points9 .
Therefore we have to assume that ∆ must be much smaller than the smallest
distances that can, at present, be resolved10 . We therefore introduce the
continuum limit : we assume that the theories we consider are such that the
limit ∆ → 0 can be taken in a sensible manner, yielding sensible results.
This sidesteps the interesting question of whether ∆ is really zero or not.
Indeed, we do not know. Any theoretical result that depends sensitively on
whether ∆ = 0 or ∆ 6= 0 would be extremely important since experimental
information about it would allow us a look at the fundamental structure of
space ; but for us it is safer to construct theories the predictions of which
do not hinge on this unknown. As we shall see, this can be made to work ;
8
See also Peter L. Berger and Thomas Luckmann, The Social Construction of Reality :
A Treatise in the Sociology of Knowledge (Garden City, New York: Anchor Books, 1966).
Additionally, the thought of all of us sharing a single point with every particle of our
bodies clashes with my sense of personal space to say the least.
9
Nor does it appear to be one-dimensional – but that is easily repaired, as we shall see.
10
About 10−19 meter.
102 August 31, 2019

as an added bonus, we can feel free from misgivings about the mathematical
rigour of taking the continuum limit — we may not be at the limit anyway.

4.3 One-dimensional continuum theory

4.3.1 The continuum limit for the propagator
Having identified the positions occupied by the various fields with points in
space (or time), we define the distance between points m and n by
x = (n − m)∆ . (4.19)
The dimension of x is that of ∆, that is, a length L. The continuum limit
is, then, that where ∆ → 0 and |n − m| → ∞ while x remains fixed. The
propagator is now a function of x, so we redefine it as
Π(x) ← Π(x/∆) .
This means that
Z+π
~ exp(ixz/∆)
Π(x) = dz , (4.20)
2π µ − 2γ cos(z)
−π

where µ > 2γ as before. A corresponding change in the integration variable

z is now in order : we write
z = k∆ , (4.21)
The dimension of k is therefore L−1 . The propagator becomes
+π/∆
Z
~∆ exp(ixk)
Π(x) = dk
2π µ − 2γ cos(k∆)
−π/∆
Z
~∆ exp(ixk)
≈ dk . (4.22)
2π (µ − 2γ) + γ∆2 k 2
In the last line, we have taken ∆ to be very small indeed. Note that the
approximation cos(z) ≈ 1 − k 2 ∆2 /2 is, of course only justified as long as
k is finite ; but for very large k the integrand is extremely oscillatory and
contributes essentially nothing11 . Now, in order to avoid a propagator that
11
This handwaving argument is justified by the fact that we get the right propagator in
the continuum limit.
August 31, 2019 103

either blows up or vanishes, we must define the ∆-dependence of µ and γ

such that γ ∼ 1/∆ and µ − 2γ ∼ ∆. We shall take
1 m2 ∆ 2 m2 ∆
γ→ − , µ→ + , (4.23)
∆ 4 ∆ 2
with m2 a positive number (remember that we need µ > 2γ). We shall also
take m itself to be positive. We then find the exact results
p 1 − m∆/2
µ − 2γ = m2 ∆ , µ2 − 4γ 2 = 2m , u− = . (4.24)
1 + m∆/2
The propagator takes the form12
eixk
Z
~ ~
Π(x) = dk 2 = exp(−m|x|) . (4.25)
2π k + m2 2m
To check that this result is indeed the correct one, we can consider the
continuum limit directly for the propagator result (4.16) :
 |x/∆|
~ 1 − m∆/2 ~
Π(n) → → exp(−m|x|) , (4.26)
2m 1 + m∆/2 2m
as desired.

4.3.2 The continuum limit for the action

In the action (4.1), we shall want to replace the sum over n by an integral
over x : X Z
∆ → dx .
n
It is therefore necessary that every term in the action acquires a factor ∆.
Now, the action depends on the quantum fields ϕn . As we let the distance
between the points shrink to zero, the collection of values {ϕ} turns into a
function ϕ(x). The precise correspondence between {ϕ} and ϕ(x) is some-
thing that, in the end, we have to decide for ourselves. Out of the several
possibilities we shall adopt the following :
   
1 0 1
ϕ(x) = ϕn+1 + ϕn , ϕ (x) = ϕn+1 − ϕn . (4.27)
2 ∆
12
To obtain this result we can close the integral in the complex-k plane, circling around
either the pole at k = im or that at k = −im.
104 August 31, 2019

This assignment is called the Weyl ordering. Its converse reads, of course,
∆ 0 ∆ 0
ϕn = ϕ(x) − ϕ (x) , ϕn+1 = ϕ(x) + ϕ (x) . (4.28)
2 2
In a sense, the field value ϕ(x) is sitting ‘in between’ the points ϕn and
ϕn+1 . Other assignments can be proposed, for instance ϕn = ϕ(x). However,
these are less attractive13 . Upon careful application of Weyl ordering and the
assumed continuum limits for µ and γ, the kinetic part of the action (4.1)
has the following continuum limit :
i X 1 ∆2
X hµ 
2 2 0 2
ϕn − γϕn ϕn+1 = (µ − 2γ)ϕ(x) + (µ + 2γ)ϕ (x) =
n
2 n
2 8
X 1 1 0 2
 Z 
1 2 1 0 2

2 2 2
m ϕ(x) + ϕ (x) ∆ = m ϕ(x) + ϕ (x) dx . (4.29)
n
2 2 2 2

The interaction and source terms in the path integral do not have a factor
∆ coming out naturally, but we may simply define the continuum limits by
redefining the objects in the action :
λ4 → ∆λ4 , Jn → ∆J(x) , (4.30)
so that the continuum limit of the full action, including this time also the
sources, becomes14
Z  
1 2 2 1 0 2 λ4 4
S[ϕ, J] = m ϕ(x) + ϕ (x) + ϕ(x) − J(x)ϕ(x) dx . (4.31)
2 2 4!
Note the notation with square brackets: the action is now no longer a num-
ber depending on (a countably infinite set of) numbers, but rather on the
functions ϕ(x) and J(x) ; this is called a functional.
13
For example, consider a function ϕ(x) that vanishes for x → ±∞. The integral
2ϕ(x)ϕ0 (x) dx then vanishes upon partial integration. Weyl ordering tells us that
R

2ϕ(x)ϕ0 (x) = (ϕn+1 2 − ϕn 2 )/∆, leading to the correspondence

Z X
2ϕ(x)ϕ0 (x) dx ↔ ∆ (ϕn+1 2 − ϕn 2 ) ,
n

where the sum also vanishes explicitly after relabelling. For the alternative assignment
ϕn = ϕ(x) the vanishing cannot be proven.
14
Strictly speaking, the Weyl ordering requires the replacement of Jn not by ∆J(x) but
by ∆J(x) + ∆2 J 0 (x)/2. The additional term, however, vanishes in the continuum limit as
∆ → 0, as do the higher powers of ∆ involved in the ϕn 4 term.
August 31, 2019 105

4.3.3 The continuum limit of the classical equation

For the discrete action, there is an obvious classical equation :
∂
S({ϕ}) = 0 for every n , (4.32)
∂ϕn
where, again, the source terms have been subsumed into the action. For the
ϕ4 model of Eq.(4.1), the classical equation is therefore
∆λ4 3
µϕn − γ(ϕn+1 + ϕn−1 ) + ϕ = ∆Jn (4.33)
3! n
for all n, and the extra factor ∆ in the coupling constant and the sources
have been taken into account. The Weyl prescription leads us to write
µϕn − γ(ϕn+1 + ϕn−1 ) ≈ m2 ∆ϕ(x) − ∆ϕ00 (x) , (4.34)
so that the continuum limit of the classical field equation takes the form
λ4
m2 ϕ(x) − ϕ00 (x) + ϕ(x)3 = J(x) . (4.35)
3!
This is precisely the Euler-Lagrange equation, that may also be obtained
immediately from the continuum form of the action by taking functional
derivatives ; we shall not really have a use for them in these notes, but for
good measure they are discussed in appendix 15.7.

4.3.4 The continuum Feynman rules and SDe

Let us have a look again at the SDe for the discrete model, for simplicity
taking the ϕ4 model again :
X
φn = Π(n − m)
m
∂2
  
λ ∂
× Jm − φ3m + 3~φm φm + ~2 φm . (4.36)
6 ∂Jm (∂Jm )2
Going over to the continuum limit entails, as we have seen, the following
substitutions :
φn , φm → φ(x), φ(y) , Jm → ∆J(y) , Π(n − m) → Π(x − y) ,
Z
X 1 ∂ δ
λ4 → ∆ λ4 , → dy , → . (4.37)
m
∆ ∂Jm δJ(y)
106 August 31, 2019

With this, the SDe becomes

Z  
λ4
φ(x) = dy Π(x − y) J(y) − φ(y)3
6
δ2

δ 2
+3~φ(y) φ(y) + ~ φ(y) . (4.38)
δJ(y) (δJ(y))2

This version of the SDe is not meant to be solved, rather it tells us the
Feynman rules in the continuum limit :

↔ Π(x − y)
x y
λ4
↔ −
~
J(x)
x ↔ +
Z ~
dx for a vertex at position x

Feynman rules, version 4.3 (4.39)

This comes with the understanding that the positions of all vertices are to
be integrated over, and that the field function φ is now a functional of the
source J. For a free theory there are no interactions, and we find
Z
φ(x) = dy Π(x − y) J(y) . (4.40)

We see that the free field is the sum of its responses to the source, weighted
by the correlation between the position where the field is measured and that
of the strength of the source at all points. It is this property that establishes
the propagator as the ‘differential-equation’ Green’s function; but this cor-
respondence is only valid for non-interacting theories (see also footnote 3 in
Chapter 1).

4.3.5 Field configurations in one dimension

Before entering spaces of more dimensions, we may have a look at the field
variables. The zero-dimensional variable ϕ, with its integration element, is
August 31, 2019 107

in the discrete one-dimensional formulation replaced by the whole set ϕ, for

which the path integration element reads, of course,
Y
Dϕ = dϕn
n

The continuum limit of this object is defined to be the continuum-formulation

path integration element, however badly defined this may be. The assigning
a functional value S[ϕ] to a given field ϕ(x) is not problematic ; rather it is
the prescription of how all field configurations are to be summed over that
makes it so hard to define path integrals rigorously15 . It is instructive to
consider the nature of the dominant contributions. Consider the part of the
path integrand that governs the point-to-point variation of the paths: it is
 
1 2
K∆ (ϕn , ϕ+1 ) ≡ exp − (ϕn+1 − ϕn ) .
2~∆

It is clear that the majority of values (ϕn+1 − ϕn )2 will be of order O (~∆),

as usual for Gaussian distributions. This means that ϕn+1 and ϕn must ap-
proach each other as ∆ → 0, so the contributing fields are continuous. On the
other hand, the approach is not too fast, since by ϕn+1 − ϕn ≈ ∆ϕ0 (x) we see
that the derivative ϕ0 (x) diverges as ∆−1/2 , hence the contributing functions
are nowhere differentiable. This is not to say that differentiable fields are not
allowed : rather, the nondifferentiable ones are the overwhelming majority.
Two conclusions follow. In the first place, the use of continuum-formulation
objects like ϕ0 (x) or ϕ00 (x) in the action are to be treated as highly sym-
bolic, almost purely mnemonic, concepts. In the second place, the classical
solution, which is typically almost everywhere differentiable, is itself not the
dominant contribution to the path integral ; rather, it is the bundle of fields
close to the classical one that constitutes the lowest-order approximation to
the behaviour of the theory.

To gain some insight in the structure of a typical path (field configura-

tion), let us consider the interrelation of three consecutive fields : it is given
by K∆ (ϕ0 , ϕ1 )K∆ (ϕ1 , ϕ2 ). For simplicity, we neglect the rest of the action.
The positions of these three fields are separated by ∆. The ‘typical’ jumps
15
In fact, the mathematical definition of continuum path integrals relies on the discrete
formulation !
108 August 31, 2019

Figure 4.1: Zooming out from a fractal path

√
in field values are of order ∆, as mentioned above. Now imagine ‘zoom-
ing out’, that is, disregarding the value of ϕ1 , and inspecting only ϕ0 and
ϕ2 , which
R are now separated by 2∆. This is obtained by integrating over
ϕ1 : dϕ1 K∆ (ϕ0 , ϕ1 )K∆ (ϕ1 , ϕ2 ) = K2∆ (ϕ0 , ϕ2 ) , where the proportionality
constant is absorbed in the normalization
√ of the path integral. The typical
jump from ϕ0 to ϕ2 is now of order 2∆. We conclude that, if we resolve
the continuum path down to√a scale ∆, the typical fluctuations over this
scale will always be of order ∆. The typical path has a fractal structure.
Such behaviour, with zigs and zags at every length scale, is encountered in
Brownian motion – and in the behaviour of the stock market16 . In figure
4.1 we plot a typical fractal path running over 10,000 points separated by a
distance of 0.01, with ∆ = 1. The first plot shows all points ; in the second,
only every 10th point is used, and in the third plot only every 100th point is
used. The qualitative form of the three paths remains the same, as expected
for a fractal path. The average absolute value of the point-to-point jumps
are 0.80, 2.49, and 6.97, √
respectively : the ratios between these numbers are
indeed roughly equal to 10.

16
Note that this qualitative picture holds only for one-dimensional theories (and, luckily,
the price of stocks, bonds, futures etc is expressed in one-dimensional currency). In more
dimensions, the paths’ behaviour is even more wild.
August 31, 2019 109

4.4 The momentum representation

4.4.1 Fourier transforming the SDe
We are now ready to make an important technical change. So far, we have
considered the fields and their expectation values as functions of position. It
will turn out to more practical to consider them as functions of momentum
or, in the one-dimensional case, of wave number 17 . There are a number of
good reasons for doing so. In the first place, in the free theory the various
momentum modes are independent of one another, in contrast to the fields
at different space points18 : propagators are simpler in momentum language
than in position language. In the second place, there is a law of momen-
tum conservation operative in the universe, and not a law of conservation
of position. In the third place, momenta are more directly the physical
characteristics that are controlled and measured in actual particle physics
experiments. E28
The transition from position to momentum is nothing but applying Fourier
transforms : we already had, from Eq.(4.25),
Z
dk ~
Π(x − y) = exp(ik(x − y)) , (4.41)
2π k + m2
2

and we now introduce19

Z Z
dk dk
φ(x) = φ(k) exp(ikx) , J(x) = J(k) exp(ikx) . (4.42)
2π 2π

We now have to figure out what the correct Feynman rules are in this
new language. To do so we use (what else ? ) the SDe. It suffices to restrict
ourselves to ϕ3 theory at the tree level, where it reads
Z  
1 λ3 2
φ(x) = dy Π(x − y) J(y) − φ(y) . (4.43)
~ 2~
17
Recall the discussion on loose terminology in Chapter 0.
18
Indeed, the more-dimensional theories have been constructed expressly to make fields
at different points correlate to one another!
19
We use the same notation for the position-dependent quantities and their momentum-
dependent Fourier transforms. This will not lead to confusion since we shall soon drop
the position-dependent ones anyway.
110 August 31, 2019

Inserting the Fourier representation leads to

Z Z Z
dk ikx dk ik(x−y) ~
φ(k) e = dy e ×
2π 2π k + m2
2
Z Z 
dk3 J(k3 ) ik3 y λ3 dk1 dk2 i(k1 +k2 )y
e − φ(k1 ) φ(k2 ) e . (4.44)
2π ~ 2~ 2π 2π
We find the following form of the SDe in momentum language :

~ J(k)
φ(k) = 2 2
−
k +m ~
Z 
λ3 dk1 dk2
φ(k1 ) φ(k2 ) (2π) δ(k − k1 − k2 ) . (4.45)
2~ 2π 2π

4.5 Doing it in momentum space

4.5.1 The Feynman rules
On the basis of the above we can now formulate the Feynman rules for our
theory in momentum space (for the example of ϕ4 theory) :

~
k ↔
k 2 + m2
k1 !
k2 λ X
↔ − (2π) δ kj
k4 ~ j
k3
k q
1
↔ + J(q) (2π) δ (q + k)
~
Z∞
dk
for every momentum k
2π
−∞

Feynman rules, version 4.4 (4.46)

Where before we had to integrate over the position of every vertex, we

now have to integrate over every momentum. It is of course possible (and
this is in fact the most common situation) that the source contains only a
August 31, 2019 111

single momentum mode. In that case the external legs in a diagram carry a
single momentum ; but all momenta of the internal lines have to be integrated
over. Also note that the vertices now carry Dirac deltas imposing momentum
conservation. This is a direct consequence of our choosing the vertices of the
theory to be position-independent20 . In addition, it has becomes necessary
to indicate how the momenta involved in the vertices are to be counted. It
is usual to count all the momenta either incoming or outgoing. The precise
convention is unimportant, but it is important that you use it consistently.

4.5.2 Some example diagrams

Here we present some diagrams, evaluated according to the rules we have
formulated so far. The first one,

q k
(4.47)

with a source on only one endpoint, evaluates to

Z
dk 1 ~ 1
J(q) (2π) δ(q − k) 2 2
= J(q) 2 . (4.48)
2π ~ k +m q + m2
When we add another source,
q1 q2
(4.49)

this gives us
1 1
J(q1 ) 2 J(q2 ) (2π) δ(q1 + q2 ) . (4.50)
~ q1 + m2
we see here an important fact : every connected diagram contains one Dirac
delta informing us that overall momentum must be conserved. The diagram

q1

q2 q3
(4.51)
20
This means that the homogeneity of space(-time) can be investigated by very carefully
checking momentum(-energy) conservation in interactions. Of course, if vertices take on
different values very far away in space or time these effects may be undetectable.
 112 August 31, 2019

bears this out : it reads

Z
dk1 dk2 dk3 J(q1 ) J(q2 ) J(q3 )
2π 2π 2π ~ ~ ~
~ ~ ~
× 2 2 2
k1 + m k2 + m k3 + m2
2 2

×(2π)δ(q1 − k1 )(2π)δ(q2 − k2 )(2π)δ(q3 − k3 )

−λ3
× (2π)δ(k1 + k2 + k3 )
~
1 1 1
= J(q1 )J(q2 )J(q3 ) 2
q1 + m2 q2 2 + m2 q3 2 + m2
−λ3
× (2π)δ(q1 + q2 + q3 ) . (4.52)
~
Next, we consider the one-loop diagram
k2
q1 k1 q2

k4
k3 (4.53)

for which we have to write down (including the symmetry factor ! )

Z
dk1 dk2 dk3 dk4 J(q1 ) J(q2 )
2π 2π 2π 2π ~ ~
~ ~ ~ ~ 1
× 2 2 2 2 ×
k1 + m2 k2 + m2 k3 + m2 k4 + m2 2
×(2π) δ(q1 − k1 ) (2π) δ(q2 − k4 )
   
−λ3 −λ3
× (2π) δ(k1 − k2 − k3 ) (2π) δ(k2 + k3 − k4 )
~ ~
λ3 2 1 1
= J(q1 )J(q2 ) 2 (2π) δ(q1 − q2 )
2 q1 + m q2 + m2
2 2
Z
dk2 1 1
× 2 (4.54)
2π k2 + m (q1 − k2 )2 + m2
2

In addition to the overall momentum conservation delta, we see here the

other significant fact : every closed loop involves a momentum that is not
E29 fixed by conservation, and that demands an integral over that momentum.
August 31, 2019 113

4.6 More-dimensional theories

4.6.1 The more-dimensional continuum
Our choosing a labelling of fields with a single integer index is, of course,
arbitrary. We can consider an alternative in which the fields are labelled by
D integer indices :

ϕn → ϕ~n , ~n = (n1 , n2 , . . . , nD ) .

In the interest of brevity, we introduce the notation

~n ± k = (n1 , . . . , nk−1 , nk ± 1, nk+1 , . . . , nD ) . (4.55)

An appropriate action for this choice would be

" D
#
X 1 X λ4
S({f }) = µϕ~n 2 − γ ϕ~n ϕ~n+k + ϕ~n 4 − J~n ϕ~n . (4.56)
2 k=1
4!
~
n

The obvious visualization for this choice is that of a space rather than
a line, covered with a regular square grid of fields, each connected to 2D
nearest neighbors: the corresponding continuum picture, therefore, is that of
a theory in D equivalent dimensions. The propagator of this theory obeys,
of course, the SDe
D D  
~Y γX
Π(~n) = δn ,0 + Π(~n + k) + Π(~n − k) , (4.57)
µ k=1 k µ k=1

with the solution

Z+π
~ exp(i(n1 z1 + · · · + nD zD ))
Π(~n) = dD z , (4.58)
(2π)D µ − 2γ cos(z1 ) · · · − 2γ cos(zD )
−π

so that, now, µ must exceed 2Dγ. The continuum limit takes a different
form than in the one-dimensional case. We define

~x = (x1 , x2 , . . . , xD ) , xj = nj ∆ ,
~k = (k 1 , k 2 , . . . , k D ) , k j = zj /∆ , (4.59)
114 August 31, 2019

The simplest nontrivial choice is then to approach the continuum as follows :

γ → ∆D−2 , µ → 2Dγ + m2 ∆D , λ4 → ∆D λ4 ,
ϕn1 ,n2 ,...,nD → ϕ(~x) , Jn1 ,n2 ,...,nD → ∆D J(~x) . (4.60)

The propagator21 takes the continuum form

 
~
Z exp i~x · ~k
Π(~x) = dD k . (4.61)
(2π)D ~k · ~k + m2

The continuum form of the action is

Z  
1 2 1 ~ λ4
S[ϕ, J] = m ϕ(~x) + (∇ϕ(~x)) + ϕ(~x) − J(~x)ϕ(~x) dD x ,
2 2 4
2 2 4!
(4.62)
The Feynman rules are seen to be

↔ Π(~x − ~y )
x y
λ4
↔ −
~
J(~x)
↔ +
x ~
Feynman rules, version 4.5 (4.63)

and also the SDe is a straightforward generalization of the one-dimensional

case :
Z 
D
φ(~x) = d y Π(~x − ~y ) × J(~y )

δ2
 
λ4 3 δ 2
− φ(~y ) + 3~φ(~y ) φ(~y ) + ~ φ(~y ) . (4.64)
6 δJ(~y ) (δJ(~y ))2
21
The propagator is still a smooth function ; but, as you can easily check by comparing
to the discussion in section 4.3.5, for D = 2 the field function ϕ is no longer continuous.
For D ≥ 3, the discontinuities are not even guaranteed to be finite, and only the interaction
terms can keep the field from jumping infinitely wildly !
August 31, 2019 115

The classical field equation,

~ 2 ϕ(~x) + λ4 ϕ(~x)3 = J(~x) ,

m2 ϕ(~x) − ∇ (4.65)
3!

can be obtained directly from the continuum action by the functional Euler-
Lagrange equation

!
δ ~ · δ
S[ϕ, J] − ∇ S[ϕ, J] =0 . (4.66)
δϕ(~x) ~ x)
δ ∇ϕ(~

The propagator only depends on |~x| and is therefore rotationally invariant :

this is a larger symmetry22 than that of the original lattice, that only allows
rotations over multiples of π/2. The mutual influence of things happening
at a point ~x and at a nearby point ~x + d~x is governed by ds(d~x), the real
distance between points, on which the physics depends. In this case,

D
X
2 2
ds(d~x) = |d~x| = (dxj )2 , (4.67)
j=1

the Euclidean distance between the points ; this type of quantum field theory
is therefore said to be Euclidean.

As before, it will turn out to be more useful to go over to a momentum

formulation of the theory. This is performed by a completely straightforward
generalization of what we did in the one-dimensional case, and we can give
the Feynman rules (in D dimensions) without more ado :

22
The increase in symmetry depends on an interplay between the lattice action and the
form of the continuum limit ; it is possible to construct actions in which the continuum
symmetry is not larger than that of the lattice theory. But I prefer to choose actions that
(will come to) look like the physics of the real world.s
116 August 31, 2019

~
k ↔
~
|k| + m2
2

k1 !
k2 λ X
↔ − (2π)D δ ~kj
k4 ~ j
k3
k q
1 D

~

↔ + J(q) (2π) δ ~q + k
~
Z∞
dD k
for every momentum k
(2π)D
−∞

Feynman rules, version 4.6 (4.68)

4.6.2 The propagator, explicitly

It is possible to express the Euclidean propagator Π(~x) in terms of known
functions, using a Gaussian representation :
Z∞ Z  
~
Π(~x) = dt dD k exp i~x · ~k − t~k · ~k − tm2
(2π)D
0
Z∞ D Z
~ −m2 t
Y
dk j exp −z(k j )2 + ik j xj


= dt e
(2π)D j=1
0
Z∞ D  
(xj )2
 
~ −m2 t
Y π 1/2
= dt e exp −
(2π)D j=1
t 4t
0
Z∞
|~x|2
 
~ −D/2 2
= dt t exp −m t −
(4π)D/2 4t
0
 1−D/2
~ 2π|~x|
= K1−D/2 (m|~x|) . (4.69)
2π m

The function K is the so-called modified Bessel function of the second kind,
discussed in appendix 15.15.8 ; here we have used its integral representation
August 31, 2019 117

given by Eq.(15.296). For large m|~x|, the propagator decreases exponentially,

while for small m|~x|, we have
~

Π(~x) ≈ − log m|~x| , D = 2 ,
2π
~Γ (−1 + D/2) 2−D
Π(~x) ≈ |~x| , D≥3 . (4.70)
4 π D/2
In every dimension, the propagator is normalized in the same way :
 
Z Z Z exp i~k · ~x
~
Π(~x) dD x = dD x dD k
(2π) D
|~k|2 + m2
Z D D ~
~ D (2π) δ (k) ~
= d k = . (4.71)
(2π) D
|~k| + m
2 2 m2

4.7 Exercises
Excercise 27 Guessing the correlator
For the ‘one-dimensional discrete model’ we found for the correlator the
recursion relation
 
~ γ
Πn = δ0,n + Πn+1 + Πn−1
µ µ
Make the following Ansatz:

Πn = A B |n|

and compute A and B. Compare to the correct answers.

Excercise 28 Fourier transformation of the action and the rules

Consider the one-dimensional continuum action
1 0 2 m2
Z  
2 λ 4
S = dx ϕ (x) + ϕ(x) + ϕ(x) − ϕ(x)J(x)
2 2 4!

Here, ϕ(x) and J(x) are real fields.

1. We define the Fourier transformations by Eq.(4.42). how that ϕ(k) =
ϕ(−k), and J(k) = J(−k).
118 August 31, 2019

2. Compute the transformed form for the action.

3. Determine the Feynman rules in the transformed formulation.
Excercise 29 Being clever in one dimension
In one dimension, loop diagrams can be conveniently computed using Cauchy
integration. We shall do this in the momentum representation for amputated
diagrams, that is, the Feynman factors for external lineas are not included.
1. Consider the tadpole diagram in ϕ3/4 theory :

k
where the loop momentum is indicated.
(a) Show that the external line carries no momentum.
(b) Write down the expression for the tadpole diagram. By direct inte-
gration over k from −∞ to +∞, show that the result is −λ3 π/(2m).
(c) Redo the above calculation in another way : the integrand has
poles at k = ±im. Close the integration contour in the complex k
plane and contract it around a pole. Show that it does not matter
which pole you choose.
2. Consider the following diagram :
p−k

p
k
where the external momentum and the loop momenta are indicated.
Show, using the same contour technique as above, that it evaluates to
1/(2m)/(p2 + 4m2 ). Show that again the choice of the upper or the
lower half-plane is irrelevant.
3. Apply the same technique twice to show that

1
=
(8m2 )(p2 + 9m2 )
August 31, 2019 119

4. Show that
p2 p1 2 + p2 2 + p3 2 + 24m2
p1 =
p3 (p1 2 + 4m2 )(p2 2 + 4m2 )(p3 2 + 4m2 )

where all external momenta are counted ingoing, so that p1 +p2 +p3 = 0.
120 August 31, 2019
Chapter 5

QFT in Minkowski space

5.1 Moving to Minkowski space : making time

Since the known space in which particle physics takes place is not of a Eu-
clidean but rather of a Minkowskian nature1 , it behooves us to make the
transition to this new type of space. Essentially, this involves singling out
one of the coordinate for a special rôle.

5.1.1 Distance in Minkowski space

Whereas the real distance, that is that notion distance measure that actually
governs the relative influence of fields at different points, is in Eulidean space
given by the Euclidean square distance of Eq.(4.67), we know that in the
spacetime in which we actually live and do physics, the real distance is quite
different. In particular, one of the coordinate directions represents time.
That is, events in spacetime taking place at position ~x = (x1 , x2 , x3 ) and
time t relative to some freely chosen origin are denoted by four coordinates:
xµ = (x0 , x1 , x2 , x3 ) , x0 = ct , (5.1)
where c is the universal constant providing the exchange rate between units of
distance and units of time2 ; it is the necessary velocity of massless particles3 ,
1
We shall not involve ourselves in the complications that arise in curved space ; a consis-
tent theory of quantum gravity is not — at present — relevant to particle phenomenology.
2
See section 0.2.1.
3
As soon as we have defined what we mean by a particle. It is customary to specify
‘the speed in vacuo’ but since particles inside a medium with which they interact are no

121
122 August 31, 2019

and the real distance between two events with coordinates xµ and xµ + dxµ
is given by
3
X
2 0 2
s(dx) = (dx ) − (dxj )2 ≡ gµν dxµ dxν ≡ dxµ dxµ ≡ dx · dx , (5.2)
j=1

(summation over repeated indices implied), where gµν is the covariant metric
tensor4

 1 if µ = µ = 0
gµν = diag(1, −1, −1, −1) ≡ -1 if µ = ν ∈ {1, 2, 3} (5.3)
0 otherwise


We also have the contravariant metric tensor g µν , defined by

g µα gαν = δ µ ν , (5.4)

so that g µν is numerically equal5 to gµν . The metric tensors allow for the
raising or lowering of indices : for instance,

xµ = gµν xν : x0 = x0 , xj = −xj (j = 1, 2, 3) . (5.5)

The special rôle of time in physics is evidenced by the relative minus sign in
the metric tensor.

5.1.2 Farewell probability, hello SDe

Up to now, the action of our theory was real-valued, and the path integrand
a real probability density. In de derivation of the SDe, however, and the con-
sequent use of our Feynman diagrams, we have not used that fact anywhere ;
the only requirement for the validity of the SDe is that the path integrand
go to zero sufficiently fast at the endpoints. As long as this is guaranteed
we may generalize the parameters of the action (including the sources) to
complex values : indeed we shall take them to be purely imaginary, with one
exception.
longer massless, this may not be necessary.
4
See section 0.2.4.
5
By coincidence. Even in the flat Minkowski space, another set of coordinates (spherical
ones, for instance) would lead to a g µν quite different from gµν . However, we shall always
use the sensible (pseudo)Cartesian coordinates in these lectures.
August 31, 2019 123

We can repeat the treatment of chapter 4 in this different setting. In

what follows we shall use the labelling ~n = (n0 , n1 , . . . , nD−1 ), and use the
notation of Eq.(4.55). For the action we choose
" D−1
#
X 1 X λ4
S({ϕ}) = µϕ~n 2 − γ0 ϕ~n ϕ~n+0 − γ ϕ~n ϕ~n+k + ϕ~n 4 − J~n ϕ~n ,
2 k=1
4!
~
n
(5.6)
with a disctinction between the ‘time’ component and the ‘space’ conponents.
We ensure the validity of the Schwinger-Dyson equations by letting µ have
a real part. This can be arbitrarily small, as long as it is positive. The SDe
for the propagator now reads
D−1  
~ Y γ0
Π(~n) = δn ,0 + Π(~n + 0) + Π(~n − 0)
µ k=0 k µ
D−1  
γX
+ Π(~n + k) + Π(~n − k) . (5.7)
µ =1
For the generating function we choose
X
Π(~n) exp in0 z 0 − i(n1 z 1 + · · · nD−1 z D−1 ) .


R(~z) = (5.8)
~
n

Note the special treatment of the zeroeth component, which we again put in
by hand 6 We can now solve Eq.(5.7)
Z
1 D 0 0 1 1 D−1 D−1


Π(~n) = d z exp −in z + i(n z + · · · + n z R(z) ,
(2π)D
D−1
!−1
X
R(z) = i~ µ − 2γ0 cos(z 0 ) − 2γ cos(z k ) . (5.9)
k=1

We now proceed to the continuum limit ; we choose, as ∆ → 0,

1 µ
~n = x , xµ = (x0 , x1 , . . . , xD−1 ) , ~z = ∆ k µ , k µ = (k 0 , k 1 , . . . , k D−1 ) ,
∆
(5.10)
6
This sidesteps the question of whether we can safely move from Euclidean to
Minkowski space by some kind of ‘analytic continuation’. At any rate such an analytic
continuation must be carefully tailored to give the Minkowskian theory that we want ; we
may as well put that effort into setting up a Minkowskian theory from the start, as we do
here.
 124 August 31, 2019

and, in addition to ϕ~n → ϕ(x),

γ = −γ0 = i∆D−2 , µ = 2γ0 + 2(D − 1)γ + im2 ∆D +  ,

λ4 → i∆D λ4 , J~n → i∆D J(x) . (5.11)

The continuum propagator now has the form

e−ik·x
Z
i~ D
Π(x) = d k (5.12)
(2π)D k 2 − m2 + i

Here k · x = k 0 x0 − ~k · ~x and k 2 = (k 0 )2 − |~k|2 . By a slight redefinition7 , the

path integral can then be written as
Z  
i
Z[J] = N Dϕ exp S[ϕ] , (5.13)
~

E31 with the final, full Minkowski action for ϕ4 theory :

m2 − i
Z  
D 1 µ 2 λ4 4
S[ϕ] = d x ∂ ϕ(x)∂µ ϕ(x) − ϕ(x) − ϕ(x) + J(x)ϕ(x) .
2 2 4!
(5.14)

5.1.3 A close look at almost nothing: i and −

In the above, the introduction of the quantity  deserves attention. We have
seen that it is necessary in order to maintain the validity of the SDe’s by
making the path integrand vanish at the endpoints. Another issue is the cor-
E32 rectness of Eq.(5.9) ; in order to stay well away from singularities, we would
rather have |µ| > 2(D − 1)|γ| + 2|γ0 | rather than |µ| > 2|(D − 1)γ + γ0 |. On
the other hand, the very presence of  ensures that, in the integration over ~z,
the singularities are bypassed and the integral is well-defined. We may also
notice that (as can be seen from our treatment of the one-dimensional case in
chapter 4) the essential contributions to the integral come from z ∼ u− ∼ 1.
The choice of Eq.(5.11) is therefore justified. But, it may be argued, a new
parameter  appearing in the theory, even if it becomes infinitesimal, must
surely have a physical meaning ? Indeed it has, and we shall come back to
this point later on. In the meantime, the  is understood and almost never
written explicitly in the action ; indeed ‘real’ physics assumes that we take
7
We take a factor −i out of S.
August 31, 2019 125

 → 0 in the end of any discussion anyway.

As we have seen, the Minkowskian nature of the theory has been imposed,
rather than derived, at two separate moments : first, the distinguishing of
γ0 and γ in Eq.(5.6), and second, the definition (5.8). Of course, there is
no a priori reason why the universe would have Minkowskian rather than
Euclidean symmetry8 . So the difference between γ and γ0 can be understood
from a phenomenological point of view. It is somewhat surprising, then, that
a second ‘by hand’ intervention, having z 0 act differently from z 1,...,D−1 , is
necessary to have position x and momentum k have the right Minkowskian
product.

5.1.4 The need for quantum transition amplitudes

We now find ourselves in a new interpretational situation. Since the ex-
ponent in the path integrand is now no longer real but complex-valued a
straightforward probabilistic picture of the path integral is no longer possi-
ble. Indeed, every path gives a contribution which is a complex phase factor,
with the same absolute value, namely precisely one9 . In fact, all possible
dynamics must now arise from interference effects. The leading contribution
still comes from the bundle of paths around the classical solution (that is still
given by the Euler-Lagrange equation), because there the phases are to first
order approximation constant. Further away from the classical solution the
phases of nearby path fluctuate wildly as ~ → 0 and these paths contribute
very little10 .
As mentioned before, the complex-valued character of the action does
not prevent us from keeping the machinery of Green’s functions, connected
Green’s functions and the Feynman diagrams to compute them, since we
have ensured that the SDe’s still hold. But we shall have to reinterpret
them. In accordance with standard quantum mechanical practice, we pos-
tulate that the (connected) Green’s functions are related to the
quantum-mechanical transition amplitudes. The squared modulus of
8
Although you may wonder what the world would look like if time was indistinguishable
from space. But, even in approaches to spacetime that attempt to develop a picture of
quantum gravity, such as causal dynamical triangulations, a foliation that preconceives
something like a special rôle for time has to be put in, in fact, by hand.
9
In the limit  → 0.
10
The remarks about instantons remain valid also in Minkowski space.
126 August 31, 2019

such an amplitude is the transition probability, to be used in the computa-

tion of cross sections and decay rates. The precise nature of the Green’s
function-amplitude relation will be elucidated later.

5.1.5 Feynman rules for Minkowskian theories

Having deduced the propagator in four-dimensional Minkowski space, we can
now formulate the provisional Feynman rules for Green’s functions with fixed
external wave vectors :

k i~
↔
k · k − m2 + i
k1 k2 i
↔ − λ4 (2π)4 δ 4 (k1 + k2 + k3 + k4 )
k4 k3 ~
k1 k2 i
↔ + J(k2 )(2π)4 δ 4 (k1 + k2 )
~
Z∞
dk
for every momentum k (5.15)
2π
−∞

At the vertices the momenta must be counted

either all incoming or al outgoing.

Feynman rules, version 5.1

The vertices also pick up an additional factor i, and all vectors from now on
are assumed to be Minkowskian four-vectors.

5.1.6 The propagator, explicitly

We can find position-space expressions for the Feynman propagator by actu-
ally doing the Fourier transform :

e−ik·x
Z
i~ 4
Π(x) = d k
(2π)4 k 2 − m2 + i
0 0 ~
e−ik x eik·~x
Z
i~ 0 3
= dk d k 0 2 , (5.16)
(2π)4 (k ) − ω(k)2
August 31, 2019 127

Im k0

Re k0

Figure 5.1: Closing the contour for positive times

where q q
~ 2 2
ω(k) = |k| + m − i ≈ |~k|2 + m2 − i (5.17)

in accordance with our idea of having  infinitesimal but positive. Note

that Π(x) is Lorentz invariant : it will not change if we perform a Lorentz
transformation on xµ the Minkowski vector separating the two endpoints
of our propagator. This simplifies our work considerably, and we need to
consider only a few cases.

Timelike separation : x2 > 0

In this case we can adopt a Lorentz frame in which ~x = 0 and we denote

s = x0 , while we assume s > 0 for the moment. Our integral hence becomes

exp(−ik 0 s)
Z
i~
Π(x) = dk 0 d3 k . (5.18)
(2π)4 (k 0 − ω(k))(k 0 + ω(k))

The integrand has two simple poles, one at ω(k) and the other at −ω(k). For
s > 0 the contour must be closed in the lower half plane in order to make the
exponent exp(−ik 0 s) vanish at infinity (cf figure 5.1). We can now perform
the k 0 integral :
Z
~ 1
Π(x) = d3 k exp(−isω(k)) . (5.19)
(2π)3 2ω(k)
128 August 31, 2019

The integration over the solid angle of ~k is trivial, and we are left with a
single integral :
Z∞
~ k2
Π(x) = dk exp(−isω) , ω = ω(k) . (5.20)
(2π)2 ω
0

Using the fact that k dk = ω dω we can switch to a new variable τ = ω/m :

Z∞ √
~m2 ~m
Π(x) = dτ τ 2 − 1 exp(−imsτ ) = −i K1 (ims) , (5.21)
(2π)2 (2π)2 s
1

where we have used Eq.(15.296) and partial integration. For negative s the
same result is found, but with −s instead of s. So we can formulate the final
form as
~m
Π(x) = −i K1 (im|s|) . (5.22)
(2π)2 |s|

Spacelike separation : x2 < 0

In this case we can adopt a Lorentz frame in which x0 = 0, and we may take
|~x| = s. Now, we can choose either contour for the k 0 integral to obtain

exp(i~k · ~x)
Z
i~ 0 3
Π(x) = dk d k
(2π)4 (k 0 )2 − ω(k)2
Z∞ Z1 Z2π
~ k2
= dk dc dϕ exp(iskc)
(2π)3 2ω(k)
0 −1 0
Z∞ Z∞
−i~ k ~ k
= dk sin(ks) = − 2 dk exp(iks) . (5.23)
(2π)2 s ω(k) 8π s ω(k)
0 −∞

The integral of k over the real axis can be deformed as indicated in figure
5.2. The integrand has two cuts along the imaginary axis in conformity with
the usuals choice where the cut of z 1/2 lies along the negative real axis. Since
s > 0 we have to move around the cut in the upper half plane. The result is
~m
Π(x) = K1 (ms) . (5.24)
(2π)2 s
August 31, 2019 129

im

−im

Figure 5.2: Closing the contour for spacelike separation

It is gratifying that, in spite of the very different way in which they are arrived
at, the Minkowskian propagator for spacelike separations and the Euclidean
propagator of Eq.(4.69) (where all separations are spacelike anyway) exactly
coincide !

In conclusion, we see that the propagators connecting points with a space-

like or with a timelike separation are related to one another by an analytic
continuation11 . It may be instructive to have a look at limiting cases. For E33
small s we find the same result in both cases (using Eq.(15.297)) :
~
Π(x) ∼ − , (x · x) → 0 . (5.25)
(2π)2 (x · x)

For large values of ms, we can use the asymptotic expansion of Eq.(15.305)
to arrive at
m1/2
Π(x) ∼ exp(−im|s|) for x · x = s2 > 0 ,
|s|3/2
11
Incidentally, this can easily be obscured since K1 (iz) is sometimes written as
(2) (2)
(π/2)H1 (z), where H1 is a so-called Hankel function. The world of Bessel functions
can be bewildering...
130 August 31, 2019

Figure 5.3: Bessel functions K1 for real and imaginary argument

m1/2
Π(x) ∼ exp(−m|s|) for x · x = −s2 < 0 . (5.26)
|s|3/2

In figure 5.3 we give the two functions |K1 (ix)| and K1 (x) for real x > 0. For
imaginary argument, (i.e. in the timelike region) it shows a power-damped
oscillatory behaviour while for real argument (i.e. the spacelike region) it falls
off exponentially ; similar to the behaviour of, say, exp(x). This practically
prohibits any faster-than-light signalling by particles obeying this propagator
over distances much larger than 1/m.

(Almost) lightlike separation: x2 ≈ 0

We have seen that for small x · x of either sign, the propagators coincide
and are independent of m. We can understand this as follows. If x · x ≈ 0,
then |x0 | ≈ |~x| can be made very small indeed using an appropriate Lorentz
transformation12 . In that case the k integral will be dominated by very large
values, for which ω(k) ≈ k and hence the value of m becomes unimportant.

12
If x · x = 0 then x0 can be brought arbitrarily close to zero.
August 31, 2019 131

Let us therefore put m = 0, upon which the propagator reads

Z∞
k2
Z
~
Π(x) = dk dΩ exp(−ik · x)
(2π)3 2(k − i)
0
Z∞
~ e−ikx+ − e−ikx−
≈ dk , x± = |x0 | ± |~x| . (5.27)
(2π)2 2i|~x|
0

In the sprit of the i prescription we regulate this by a limiting procedure :

Z∞
~ e−ikx+ − e−ikx− −ηk ~ 1
Π(x) = lim dk e = lim
η↓0 (2π)2 2i|~x| η↓0 (2π)2 (η + ix+ )(η + ix− )
0
 
−~ x · x + iη ~ 1
= lim =− P + iπδ(x · x) . (5.28)
η↓0 (2π)2 (x · x)2 + η 2 (2π)2 x·x
The P stands for the principal-value distribution (see appendix 15.15.6).
We recover the result (5.25), and now in addition that for x2 = 0. Note,
however, that the propagator is defined in terms of distributions rather than
of functions : no problem since in the SDe the propagator occurs inside
integrals anyway !

5.2 Moving in Minkowski space : particles

5.2.1 The Klein-Gordon equation
For a free theory the SDe is again quite simple. In position, rather than
momentum, representation we have
Z
i
φ(x) = d4 y Π(x − y) J(y)
~
exp(−ik · (x − y))
Z
1
= − 4
d4 y d4 k J(y) . (5.29)
(2π) k · k − m2 + i
Forgetting for the moment about the i, we can apply a differential operator13
on φ(x) :
∂ µ ∂µ + m2 φ(x) =

13
Note that in Eq.(5.30), ∂ stands for a derivative to x, not to y !
132 August 31, 2019
Z µ 2 −ik·(x−y)
1 4 4 (∂ ∂µ + m ) e
= − d y d k J(y)
(2π)4 k 2 − m2
Z 2 2 −ik·(x−y)
1 4 4 (−k + m ) e
= − d y d k J(y)
(2π)4 k 2 − m2
Z Z
1 4 4 −ik·(x−y)
= dydke J(y) = d4 y δ 4 (x − y) J(y) ,(5.30)
(2π)4

since the integral over k has become trivial14 . The free-field function φ(x) is
therefore seen to obey the equation

∂ µ ∂µ + m2 φ(x) = J(x) ,


(5.31)

and this is known as the Klein-Gordon equation. In more conventional treat-

ments, this equation is the starting point for a relativistic quantum field the-
ory, being introduced as a direct relativistic adaptation of the nonrelativistic
Schrödinger equation ; for us, it is a fairly unimportant15 result following
from the Feynman rules.
You might fret about just leaving out the i in this derivation. Indeed
we can keep it, and the the Klein-Gordon equation would contain m2 − i
instead of just m2 . This more puristic approach is certainly more correct, but
in writing down the Klein-Gordon equation no-one really can be bothered.

5.2.2 Enter the particle !

Equation (5.29) leads to an equation for φ(J), which I find not very inter-
esting. What is important, however, is the light it sheds on the source J :
the natural interpretation is, indeed, for J to be a physical source, generat-
ing the field φ via Huygens’ principle. The propagator takes the rôle of the
Green’s function as used in the solution of inhomogeneous differential equa-
tions. This is the place where the notion of particle comes in ! We
picture the source influencing its environment by sending out (or receiving)
signals that carry information such as energy, (angular)momentum, charge,
colour, etcetera. Under the right circumstances these signals obey Newton’s
first law, and that is what justifies our particle picture. The field function φ
is now interpreted as the quantummechanical wave function of the particle.
14
This is of course solely due to our taking just the right differential operator.
15
Unimportant in the sense that we shall not derive any consequences from it. The same
will be seen to hold for the Dirac, Proca and Maxwell equations.
August 31, 2019 133

5.2.3 Unstable particles, i and the flow of time

We are now in a position to investigate the physical meaning of the i pre-
scription. In order to do so, let us assume that  is not infinitesimal, but
rather of fixed value γ. That is, we shall use a propagator
exp(−ik · (x − y))
Z
i~
Πγ (x − y) = 4
d4 k , γ>0 . (5.32)
(2π) k 2 − m2 + iγ
Moreover, let us choose a source that emits particles simultaneously16 at time
t = 0, all over space : we take17
J(x) ∼ δ(x0 ) . (5.33)
The response of the field can be written as
Z
i
φ(x) = d4 y Πγ (x − y) J(y 0 )
~
 
Z exp −ik · x + ik 0 0
y − i~k · ~y δ(y 0 )
1
= − d4 y d4 k
(2π)4
(k 0 )2 − |~k|2 − m2 + iγ
−1 exp(−ix0 k 0 )
Z
= dk 0 0 2 . (5.34)
2π (k ) − m2 + iγ
The integrand has poles in the complex k 0 plane at
0
p
2
 γ 
k = ± m − iγ ≈ ± m − i ,
2m
where we have assumed that γ is small compared to m2 . For times later than
t = 0, the integration contour must be closed along the lower half complex
plane (just as in 5.1.6), and we find
 γ 0
φ(x) ∝ exp −imx0 − x . (5.35)
2m
In accordance with the quantum-mechanical interpretation of our theory,
|φ(x)|2 must be (related to) the probability of finding particles. In the present
case, we have
 γ   
2 0 t 1 γc
|φ(x)| ∝ exp − x = exp − , ≡ . (5.36)
m τ τ m
16
Simultaneity is an ambiguous concept in Minkowski space : here, we mean simultane-
ous in our frame : a ‘Little Bang’.
17
We do not worry about normalization issues here.
134 August 31, 2019

That is, the probability of finding particles anywhere decreases exponentially

as time goes on. This is what one expects for unstable particles with a mean
lifetime equal to τ . We shall write γ = mΓ, where Γ is called the total decay
width of the particle, so that 1/τ = c Γ18 . We see that a Feynman rule is
now available for unstable particles :

k 1
↔ i~
k · k − m2 + imΓ
The propagator for an unstable particle with mean life-
time 1/(Γc).
Feynman rules, version 5.1 (addendum) (5.37)

The i prescription is seen to just mean that we should treat stable particles
as the infinitely-long-lifetime limit of unstable particles.
Equation (5.35) describes particles at rest since there is no space depen-
dence in the wave function. The lifetime/width is therefore that of particles
at rest, which is indeed the usual definition. For particles in motion, time
dilatation indicates that the lifetime be increased by a factor p0 /m ; we shall
encounter this situation in the next chapter.
An issue that appears resolved is the direction of time flow. Whereas
Minkowski space itself, being essentially static, does not assign any preferred
direction associated with the time coordinate, the direction of time flow is
now defined to be that direction in which unstable particles disappear , rather
than appear19 .
Another point to be noted is the following. The unstable propagator by
itself is seen to lead to a decreasing overall probability : that contradicts
the normal unitary evolution of quantum mechanics. But it is not the whole
18
Remember that Γ has the dimension of inverse length, not kilograms ! In practice, it
is customary to give total decay widths, like particle masses, in GeV, and then τ = ~/Γ.
19
Attractive as the above argument appears, a drawback comes from the case x0 < 0.
In that case, the contour integral must be closed along the upper half plane, so that the
pole k 0 = −m + iγ/(2m) becomes the significant one. We find φ(x) ∝ exp(−|t|/τ ), which
is to be interpreted as a particle density that starts out as zero at t = −∞, and grows to a
crescendo at t = 0 ; this lacks an obvious interpretation. We ascribe this to the use of the
simple form (5.33). A better source, needed for a more rigorous treatment, can be simply
constructed. Notice that this really means that the direction of time is governed by the
sources !
August 31, 2019 135

story : for a particle to be unstable it must be able to go over into other

particles, that is, there must be interactions. These have been left out of our
discussion. In a more complete treatment, we shall of course see that, as the
unstable particles disappear, the density of other particles will increase, and
total probability will be preserved. In other words, the decay width must be
consistently computable from the interactions present in the theory.
The assumption that γ is considerably smaller than m2 implies that Γ is
small compared to m. Indeed, if we assume that Γ becomes nonzero due to
interactions, the very spirit of perturbation theory argues that Γ is relatively
small. Rigorous upper limits on the width of any given particle cannot easily
be given ; but let us imagine a particle of mass M (in kilograms, not inverse
meters !). Its natural ‘size’ is given by its Compton wavelength λc = ~/(M c).
If Γ (a quantity with the dimension of inverse length) were larger than 1/λc ,
this would mean that such a particle would, upon production, decay even
before a lightlike signal could have crossed its diameter : it is as if the par-
ticle would vanish before it was even aware that it existed. In general, the
situation Γ > m is held to signal a breakdown of the concept of a particle as
a more or less identifiable entity.

5.2.4 The Yukawa potential

For another application, we can consider a static point-like source :

J(x) ∼ δ 3 (~x) . (5.38)

The response of the field is then

e−ik·x
Z Z
4 i~ 4 i 3
φ(x) = dy 4
d k 2
δ (~y )
(2π) k · k − m + i ~
~
eik·~x
Z
1 3~
= dk . (5.39)
(2π)3 |~k|2 + m2
The i term in the denominator can safely be neglected here. Writing |~x| ≡
r ≥ 0 and k ≡ |~k|, and going over to polar coordinates for ~k, we have
Z∞ Z2π Z1
1 2 eikr cos θ
φ(x) = dk k dϕ d cos θ
(2π)3 k 2 + m2
0 0 −1
 136 August 31, 2019

Z∞
eikr e−ikr
 
1
= dk k −
4iπ 2 r k 2 + m2 k 2 + m2
Z0
1 k
= dk eikr . (5.40)
4iπ 2 r k2 + m2
For r > 0 we can close the integration contour in the upper half of the
complex-k plane, and we find
1 exp(−mr)
φ(x) = . (5.41)
4π r
This is the so-called Yukawa potential, introduced in the 1930’s as a model for
E34 the strong nucleon-nucleon force, with m the mass of the pion. The Compton
wavelength of the pion is, indeed, roughly the range of the nuclear forces. If
we take m → 0 we find the Coulomb potential of a static electric source ;
but the real propagator of the photon field, responsible for the Coulomb
interaction, is more complicated, so that the above derivation is more or less
just handwaving in the case of electromagnetism. A fuller treatment is given
in section 10.5.

5.2.5 Kinematics and Newton’s First Law

Let us see to what extent the picture of the source as an object that, in
a sense, emits particles can be reconciled with standard ideas in classical
relativistic mechanics. That is, we want to measure positions and times, as
well as energies, velocities and momenta, as well as possible. To this end, we
shall choose the source to be
|x0 | |~x|2
  
i 0 0
J(x) ∼ exp − − 2− p x − ~x · p~ . (5.42)
σ0 4σ ~
That is, the source is active for a period σ0 /c around t = 0, and in a region
of volume σ 3 around the spatial origin. Its Fourier transform,
"   #−1
0 2
 2 !
1 p p
~
J(k) ∼ + k0 − exp −σ 2 ~k − , (5.43)
σ0 2 ~ ~

shows that it emits particles with all kinds of wave vectors k µ = (k 0 , ~k),
centered around values pµ /~, with pµ = (p0 , p~). For a bridge to non-quantum
August 31, 2019 137

Im k0

0
Re k

Figure 5.4: Closing the contour for Newton’s first

physics to be built, both the position and wave representation of the source
should be adequately localized ; σ0 and σ should be neither too large nor too
small. For now, we do not assume any particular relation between p0 and p~.
We want to study the response of the field to this source for positive times.
We have  
Z 0 0
exp −ik x + ik · ~x~
φ(x) ∼ d4 k J(k) . (5.44)
(k 0 )2 − |~k|2 − m2 + i
In figure 5.4 we exhibit the k 0 integral contour in the complex plane. For
x0 > 0, the contour is to be closed in the lower half complex-k 0 plane. The
integrand has simple poles at the loci

0 ~ 0 p0 i
k = ω(k) − i , k = − ,
~ σ0
p0 i
k 0 = −ω(~k) + i , k 0 = + ,
~ σ0
the latter two lying outside the contour.
The k 0 integral therefore leads to the following expression for φ(x) :
Z  2 !
p
~
φ(x) ∼ d3~k exp i~x · ~k − σ 2 ~k − A(k 0 , ~k) ,
~
138 August 31, 2019
 
exp −ix 0
ω(~k)
1
A(k 0 , ~k) = 2
2ω(~k) p0 /~ − ω(~k) + 1/σ0 2


iσ0 exp(−ix0 p0 /~ − x0 /σ0 )

+ . (5.45)
2 (p0 /~ − i/σ0 )2 − ω(~k)2 + i

The second term decays exponentially at the same rate as the source. Since
we are interested in the behaviour of the field when it is free, i.e. unaffected
by any interactions, we can only study that behaviour once the source has
died out, and then so has this term20 . The first term describes Fourier modes
of the field that obey the dispersion relation k 0 = ω(~k), together with the
resonance condition that tells us that the field can only be appreciable if
both p0 /~ ≈ ω(~k) and p~/~ ≈ ~k. We therefore expect any fruitful resonance
in the field, which can allow for the transmission of signals over macroscopic
distances, only if
p0
 
p~
≈ω . (5.46)
~ ~
If we relate the zero component p0 (with dimension kg m/s) to an energy E
by writing
p0 = E/c , (5.47)
we find that the only particle modes emitted by the source that have a chance
of propagating over distances much further than σ must satisfy

Mc
q
E≈ |~p|2 c2 + M 2 c4 , m= . (5.48)
~
This is the mass shell condition, which prescribes the relation between the
energy E (in Joule), momentum p~ (in kg m/s), and mechanical mass M (in
kg) of a particle moving freely through spacetime. We recognize the quantity
m that we have been using so far as the inverse Compton wavelength of the
particle. A particle is called on-shell if its momentum pµ satisfies Eq.(5.48) ;
if not, it is called off-shell. Off-shell particles are not exotic or improbable ;
they are just not visible as the result of an experiment since they cannot
20
This is comparable with what you would do classically: studying the trajectory of a
thrown ball to see whether Newton’s laws are obeyed only makes sense once the ball has
definitively left your hand.
August 31, 2019 139

propagate well21 . Note the occurrence of ~ in the denominator : interpreted

strictly, this implies that a truly classical limit leads to m → ∞. The Euler-
Lagrange equations are therefore not classical at all except for the Maxwell
equations holding for massless photons.
Given that the particle is emitted on its mass shell, the integral φ(x) is
not yet automatically appreciable. The complex phase in Eq.(5.44) will lead
to extremely rapid oscillatory behaviour of the integrand, and an essentially
vanishing result, except for those regions where the phase of the integrand is
stationary. This happens if

∂
 
∂
  ~k
x k − ~x · ~k =
0 0
x ω(~k) − ~x · ~k =
0
x0 − ~x = 0 . (5.49)
∂k~ ~
∂k ω(k)~

That is, φ(x) is appreciable on a line in spacetime given by

c p~ p~
~x = t 0
= ~v t , ~v = c 0 : (5.50)
p p
the particle moves along a straight line, with constant velocity ~v . This is
Newton’s First Law.
One might envisage other time-dependences of the source. Two additional
cases are discussed in Appendix 15.9 ; the conclusions (although based on
slightly different mathematics) are the same.
From this simple investigation we may conclude that (a) free particles
moving over macroscopic distances must be on-shell ; and that (b) the motion
of such particles follows Newton’s first law.

5.2.6 Antimatter
We again consider the free SDe :

dk 0 d3~k exp(−ik · x)
Z Z
0
φ(x , ~x) ∼ J(k 0 , ~k) ,
2π 3 ~
(2π) (k ) − ω(k) + i
0 2 2
q
ω(~k) = |~k|2 + m2 . (5.51)
21
In popular literature, off-shell particles are often dicussed with a lot of mumbling
about ‘uncertainty relations’, ‘borrowing energy from the vacuum’, and so on. Do not
allow yourself to be misled ! When a theorist starts invoking the uncertainty principle as
a reason for something, keep your hand on your wallet. The ‘uncertainty principle’ is not
a principle, not a reason, but a result.
140 August 31, 2019

x x

B B

E>0 E<0

A A
t t

Figure 5.5: Moving forward ; moving backward

If x0 > 0, the integration contour can be closed through the lower half of the
complex k 0 plane :

d3~k
Z  
0
φ(x , ~x) ∼ exp −i(x0 ω(~k) − ~x · ~k) J(ω(~k), ~k) . (5.52)
(2π)3 2ω(~k)

If, on the other hand, x0 < 0, the closure must be over the upper half of the
plane, and then

d3~k
Z  
0
φ(x , ~x) ∼ exp −i(−x ω(k) − ~x · k) J(−ω(~k), ~k) . (5.53)
0 ~ ~
(2π)3 2ω(~k)

We see that the propagator essentially describes plane waves, with the fol-
lowing characteristic: positive energies travel towards the future, and negative
energies travel towards the past.
While the concept of particles with positive kinetic energy, moving from
past to future, conforms to our everyday experience, the idea of negative
(kinetic) energies and movement backwards in time is not only æsthetically
repellent but may lead to splitting headaches in the verbal description of
physical processes. When, however, we consider more closely how such a sit-
uation will appear, it becomes clear that negative energies moving backwards
in time are indistinguishable from positive energies moving forward. Some
simple bookkeeping with the help of the two diagrams in figure 5.5 comes
in handy here. Consider two loci in space, denoted by A and B. In the
August 31, 2019 141

first diagram a particle moves forward in time, with positive energy, from
A to B. As a result the energy at A decreases, and that at B increases.
In the second diagram, a particle with negative kinetic energy starts at B,
and moves backwards in time to A. The net effect on the energies at A and
B is exactly the same ! The two situations are indistinguishable from the
point of view of the energy balance22 . There may still be a difference, of
course ; if the particles have additional properties such as electric charge,
the backwards-moving particles will appear with the opposite charge. For
instance, a negatively charged electron moving backwards will appear as a
positively charged positron moving forward, as can be seen from the two di-
agrams above. Such re-interpreted time-reversed particles are called
antiparticles 23 . Every particulate object whose propagator contains the
denominator of Eq.(5.51) is seen to contain both the regular particles and
their antiparticles. Moreover, we find the fundamental result that particles
and their antiparticles must have exactly the same mass and lifetime. The
antiparticle interpretation is just the way we surrender to a prejudice about
motion in time24 . Particles and their antiparticles may be identical, the pho-
ton being an example. Such particles must, of course, be electrically neutral.
On the other hand, not all neutral particles are their own antiparticles ; neu-
trons and antineutrons are distinct from one another25 . We have thus found
the following result for free particles : if we (a) replace all particles by their
antiparticles and vice versa, the so-called charge conjugation operation C,
(b) inverse all space directions26 , the so-called parity transformation P, and
(c) invert the direction of time, the so-called time reversal operation T, then
the world will look exactly the same ! This is (a restricted form of) the CPT
theorem, valid for the propagation of free particles. The more interesting,

22
At this point you may wonder about lots of the particles that constitute you and
the world actually moving backwards in time. Try to formulate, say, your own digestive
processes in this light — it doesn’t bear thinking about.
23
A commonly made error is to say that antimatter moves backwards in time. This is
wrong. Antimatter is invented precisely to make everything move forward in time !
24
Physicists from some alien civilization might have less problems with the other inter-
pretation.
25
Once the neutron is seen to be a collection of charged quarks, the distinction be-
comes obvious. On the other hand, neutrinos, while electrically neutral, are not equal to
antineutrinos, and are yet believed to be elementary.
26
Since, as can be seen from our diagrams, inverting the direction of the motion through
time will simultaneously change motion towards the left (say) into motion towards the
right, and so on.
142 August 31, 2019

A
t
t0

Figure 5.6: An annihilating spacetime path

real CPT theorem, valid also for interacting particles, needs more tools than
we have at our disposal right now : its proof is referred to Appendix 15.14.
Let us now consider the (classically depicted) path a particle tracks out
in spacetime, as given by the space-time diagram given in figure 5.6. In one
description, the particle starts at A and moves to B, where at time t0 it
reverses its time direction, and moves backwards in time to C. In the alter-
native description, a particle starts at A and its antiparticle starts at C, and
the pair collides at B at time t0 . For times later than t0 , the particle and/or
its antiparticle have disappeared ; but because of momentum conservation
their combined energy has to be transferred onto one or more other parti-
cles (not depicted). The two descriptions are completely equivalent, but the
second one conforms much better to the way we tend to view the world. At
the ‘collision/reversal-point’ B the particle coming from A must dump its
energy, and even an additional amount since its energy must become nega-
tive for it to start moving backwards to C. Therefore, particle-antiparticle
annihilation releases energy, often in the form of photons27 . For instance,
27
It is sometimes stated that particles can only annihilate with their own antiparticle.
This is a somewhat restricted point of view, since for instance electrons can annihilate
with anti-neutrinos into W particles, as we shall see. It may be more appropriate to
say that it needs particles with their own antiparticles to annihilate into something that
has quantum numbers (electric charge, fermion number, etcetera) equal to those of the
vacuum. Neutrinos and their antineutrinos cannot easily annihilate into photons, being
electrically neutral : but they can annihilate into one or more Z bosons.
August 31, 2019 143

when positrons meet electrons, the usual result28 is e− e+ → γ γ. We also

see that nothing forbids the opposite process, in which available energy turns
into particle-antiparticle pairs : γ γ → e− e+ .

5.2.7 Counting states : the phase-space integration el-

ement
The treatment of the previous section is also useful in that it provides a hint
on how to count the wave-vector states. For on-shell particles of mass m we
use the integration element

d3~k
q
1
, ω(k) = ~k 2 + m2 .
~
(2π)3 2ω(~k)

This object has dimension L−2 . It is not explicitly Lorentz-covariant, but we

can write it also in a more attractive form : E36
1 d3~k 1 4 2 2
θ(k 0 ) .


= d k δ k − m (5.54)
(2π)3 2ω(~k) (2π)3

Note that if k 0 is positive for an on-shell particle in any given inertial frame,
it is positive in all intertial frames that can be reached by Lorentz transfor-
mations from the first one. This ensures that the step function θ(k 0 ) always
has the same value, irrespective of any Lorentz boosts we may care to make.
Lorentz covariance of the phase space integration element is thus guaran-
teed. We shall use the density of states (5.54) for all on-shell particles in the
calculation of cross sections and lifetimes.
If, for a given scattering process, the final state contains N particles
with masses mj , j = 1, 2, . . . , N , and momenta pµ1 , pµ2 , . . . , pµN , the combined
phase-space integration element is

dV (P ; p1 , p2 , . . . , pN ) ≡
N
! N
!
Y 1 X
3
d4 pj δ(pj 2 − mj 2 ) (2π)4 δ 4 P − pj , (5.55)
j=1
(2π) j=1

28
Note that the simpler-seeming process e− e+ → γ is kinematically impossible if the
resulting photon is to be on its mass shell. On the other hand, a single off-shell photon
can be produced, but such a photon must immediately decay again, in for instance a
particle-antiparticle pair of some kind.
144 August 31, 2019

where P µ is the total momentum of the scattering system. The four-dimensional

Dirac delta forces the overall conservation of wavevector29 . The condition
θ(p0 > 0) imposing positive energy for the outgoing particles is, here and in
the following, always understood.

5.3 Exercises
Excercise 30 Dangerous 
Explain why, in Eq.(5.11), it is important that  > 0.

Excercise 31 Zero Minkowski dimensions

Once we have discarded the purely probabilistic interpretation of the path
integral, we may as well return to zero dimensions, with an action

m2 − i 2 λ 4
S(ϕ) = ϕ − ϕ − Jϕ
2 4!
1. Using the i prescription, show that
Z∞  r
m2 2

2π~ −iπ/4
dϕ exp −i ϕ = e
2~ m2
−∞

2. Find the Feynman rules for the propagator and the vertices.

3. Write down the SDe for this model.

Excercise 32 Useful 
Show that the integral Z
I0 = dx exp(ixy)

where x runs over the whole real axis and y is an arbitrary real number, is
ill-defined. Show that the integral
Z   
I = dx exp ixy − x2
2
29
Conservation of total energy and momentum.
August 31, 2019 145

where  is an infinitesimally small but positive real number, is well-defined.

Compute I , and show that
lim ↓0 I = 2π δ(y) .
This handling of the integral is precisely what we also do by including an 
in the action.
Excercise 33 careful, careful...
Perform the steps leading to Eq.(5.22) and to Eq.(5.24). Especially in the
latter case you have to be quite careful in the complex plane !
Excercise 34 Starting out at Yukawa
Consider the Yukawa potential in three space dimensions:
1 −mr
V (~x) = e , r = |~x|
4π r
Compute its (three-dimensional) Fourier transform explicitly, to show that
Z
~ 1
d3 x V (~x) ei~x·k =
|~k| + m2
2

Excercise 35 Short-range weak forces

From our discussion of the Yukawa potential, estimate the effective range of
the static weak-interaction potential, mediated by particles of masses 80 to
90 GeV/c2 .
Excercise 36 Counting states
Prove Eq.(5.54).
Excercise 37 Auxiliary fields
In Eq.(5.14) the action for ϕ4 theory was given. Let us consider a slightly
different one, involving two fields :
Z
S = d4 x L(ϕ, ψ) ,
1 µ m2 − i µ2 − i
L(ϕ, ψ) = ∂ ϕ(x)∂µ ϕ(x) − ϕ(x)2 − ψ(x)2
2 2 2
g 2
− ψ(x)ϕ(x) + J(x)ϕ(x) .
2
The field ψ does not occur with a derivative term : such fields are called
auxiliary fields. This exercise examines some of their effects.
146 August 31, 2019

1. Show that the path integral over ψ(x) can be performed independently
at every point in Minkowski space.

2. Compute the result of the ψ integration for the path integral. This is
called integrating out the field ψ.

3. Show that after having integrated out ψ we are left with the ‘original’
ϕ4 theory, provided we choose λ to be a combination of g and µ. Find
this combination.

4. Write out the coupled SDe’s for the ϕ−ψ theory in configuration space,
using the formalism of functional derivatives. Insert the SDe for ψ into
that for ϕ and show that the correct ϕ4 -SDe for ϕ is again obtained.

5. Show that, if we insist that ψ obeys its classical (Euler-Largange) equa-

tion, the same result is obtained.

6. Show that the method of auxiliary fields does not work in Euclidean
space since there the path integral is not defined (it is easiest to do this
in zero dimensions).
Chapter 6

Scattering processes

6.1 Introduction
In this chapter we turn our attention to the bread-and-butter subject of parti-
cle phenomenology : the description of scattering processes. We shall discuss
the way in which Feynman diagrams and their evaluation are postulated to
predict the probability for finding specified final states given specified initial
states. We also investigate the consequences of the claim that our approach
describes quantum physics and is therefore of a probabilistic nature : that
is, we can only compute probabilities, which are necessarily bounded1 . This
leads to the notion of unitarity and the use (and usefulness) of cutting rules.

6.2 Incursion into the scattering process

6.2.1 Diagrammatic picture of scattering
To a large extent, particle phenomenology can be viewed as the study of
scattering processes, in which some initial state is prepared and allowed to
time-evolve, and finally an observation is made in which the system is seen
to have resulted in some final state. A useful example is provided by the
current practice in high-energy colliders : here the initial state is prepared
by machine physicists operating the collider, and it consists of two (beams
of) particles with more or less well-defined momenta coming out of the beam
pipes. The interesting part of the time-evolution of the system is that during
1
After all, the probability of a certain scattering process occurring cannot exceed 100%.

147
148 August 31, 2019

which the initial-state particles approach one another and meet, hopefully2 ,
in the interaction point, where the dynamics takes place. The final state is
observed by the detector operated by the particle physicists.
Since not only the scattering itself but also the initial-state preparation
and the final-state observation are quantum processes, all these parts of the
process must, according to our assumptions, be described by Feynman di-
agrams in a manner still to be established. The diagrammatic form of the
complete process will then look as follows :

and here (and in the following) we adopt the convention that the initial state
appears on the left-hand side of the diagrams, and the final state on the right-
hand side3 . This does not imply any spatial or timelike relation between any
of the vertices in the diagram: indeed, they are supposed to be integrated
over all of spacetime4 . Another observation on the above diagram is also
relevant : the initial-state preparation and the final-state observation should
contain physics that is better understood than the scattering part, and there
should be a clear notion of precisely which particles constitute the initial
and final states. This is indicated by the identifiable propagators connecting
the various ingredients of the process. We therefore adopt the idealization
that the only relevant part of the scattering should reside in the central, or

2
In the idealised setting in which particles with perfectly well-defined momenta form
plane waves of infinite spatial extent, they can hardly avoid meeting. In practice, the
momenta and spatial extensions of the particles’ wave packets are of course more limited.
3
This convention is based on not much more than the fact that in Latin writing you
move from left to right. Indeed a not uncommon alternative is to move upwards, so that
the incoming state is at the bottom. I do not know of any script that moves like that. It
could be worse : fortunately no-one employs boustrophedonic diagrams !
4
Of course, if there is any justice the contribution from paths in which a vertex is very
far out ought to be small.
August 31, 2019 149

scattering part, in this case

We now have to confront the two following questions. In the first place,
which Feynman diagrams should occur in the scattering part ? And sec-
ondly, in actual experiments the initial- and final-state particles travel over
many meters between preparation, scattering, and detection. These particles
should therefore be on their mass shell, but isn’t this precisely the case in
which their propagators blow up ? The situation obviously calls for some
reinterpretation and additional Feynman rules, to which we shall come.
Before finishing this section, let us remark that also initial states consist-
ing of only a single particle occur :

In this case, we simply study the decay properties of the particle, such as its
total or partial decay width.

6.2.2 The argument for connectedness

Let us consider the set of all Feynman diagrams describing a decay process.
As discussed in section 1.3.7, we omit all vacuum bubbles. The set can then
be split up into its connected pieces, for instance
1111111
0000000
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111 0000000
1111111
= 1111111
0000000
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
+ 0000000
1111111
0000000
1111111
11111
00000
+ · · · (6.1)
0000000
1111111 00000
11111
0000000
1111111 00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111

where as before the shading indicates connected diagrams. Now, recall that
every vertex in any diagram contributes a Dirac delta imposing energy-
momentum conservation. Therefore, every connected diagram has an overall
 150 August 31, 2019

Dirac delta imposing overall energy conservation. That, however, implies

that a satellite diagram such as
00000
11111
11111
00000
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111 (6.2)

asks for particles carrying positive energy to originate (by some interactions)
from the vacuum. Such contributions therefore vanish by energy conserva-
tion, and the only contributing diagrams are contained in the totally con-
nected blob. Next, consider two-particle scattering. Since we forbid satellite
diagrams the only possible contributions are given by
0000000
1111111
1111111
0000000
0000000
1111111
1111111
0000000 0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111
0000000
1111111
= 0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
+ 0000000
1111111

0000000
1111111
+ · · · (6.3)
0000000
1111111 1111111
0000000
0000000
1111111 0000000
1111111
0000000
1111111
0000000
1111111

Now, the second term here is in principle possible but only if a) the two
E38 incoming particles are inherently unstable5 and b) the outgoing particles ar-
range themselves in precisely two groups according to the indicated decay
patterns. Leaving aside such special cases, we conclude that the scattering
amplitude is given by the connected Feynman diagrams. Note that
the restriction to connected diagrams only arises here from simple energy
considerations, and not from any deep inherent superiority of connected di-
agrams over disconnected ones : in essentially all cases of interest, the result
of the disconnected diagrams vanish anyway6 .
We may imagine situations where particles can be created from the vac-
uum. This is the cases in ‘field theories at high temperature’ where processes
take place in a heat bath which can deliver energy to create particles, so en-
ergy conservation is relaxed. In such a picture the heat bath is the ‘vacuum’
of the theory, and diagrams such as that of Eq.(6.2) are not automatically
zero. Another delicate situation is that of more incoming particles : for
instance, we might consider four particles scattering into four, in which we
5
This makes the notion of particles ‘coming in from infinity’ conceptually dubious in
this scattering.
6
There are plans on the table for a muon collider. Muons are unstable particles, but
their lifetime (especially at high energies) is extremely long as particle lifetimes go, so we
can treat them as essentially stable in most processes. However, it is possible that final
states are produced where one of the muons actually decays ; then we have to take more
care (see also exercise 38).
August 31, 2019 151

might recognize two groups of two particles scattering into two :

00000000
11111111
11111111
00000000
0000000
1111111
1111111
0000000 00000000
11111111
0000000
1111111 00000000
11111111
00000000
11111111
0000000
1111111 00000000
11111111
0000000
1111111
0000000
1111111 00000000
11111111
0000000
1111111
= 0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111 + 0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
+ · · · (6.4)
0000000
1111111

In this case, the only argument to disregard the disconnected diagrams is an

appeal to the special kinematics.

6.3 Building predictions

6.3.1 General formulæ for decay widhts and cross sec-
tions
Consider a ‘slightly unstable’ particle of mass m at rest, with momentum
P µ . We shall adopt the following prescription for its differential decay width
into n particles with momenta pµ1 , pµ2 , . . . , pµn :

dΓ = ΦΓ h|M|2 i dV (P ; p1 , p2 , . . . , pn ) Fsymm . (6.5)

Here, M stands for the transition amplitude, which we still have to establish.
The symbol h|M|2 i indicates that in accordance with quantum-mechanical
practice we have to square the absolute value of M in order to arrive at
a probability, and the brackets indicate summation and/or averaging over
degrees of freedom other than the momenta : at present such degrees of
freedom are not in our theory yet, but they will arrive ! The symbol ΦΓ
denotes the collection of factors that must be included to account for the
density of states for the incoming particle, etcetera. That such a factor must
be present is clear from the fact that dΓ must have dimension 1/L. dV
denotes the phase-space integration element discussed in section 5.2.7. The
total decay width in this particular decay channel is obtained by integrating
over the available phase space7 . The momentum P µ is that of the incoming
particle at rest. The symmetry factor Fsymm is included to handle identical
particles in the final state. In quantum mechanics, the statement that two
particles are identical means that an interchange of these particles leads to
the physically identical final state, so that an unconstrained summation over
7
This can become very challenging !
152 August 31, 2019

their momenta (and other quantum numbers) would lead to over-counting.

We therefore prescribe that Fsymm contains a factor 1/k! for every group of
precisely k indentical particles in the final state8 . For example, a final state
containing precisely 2 photons, 3 electrons and 1 positron leads to Fsymm =
1/(2!)(3!)(1!) = 1/12. Note that the decay width is inversely proportional
to the particle’s lifetime. This means that for a moving particle the decay
width must decrease by a factor m/P 0 to account for time dilatation.
In the case of two stable incoming particles with momenta pµa and pµb , we
rather talk about the transition rate per unit flux, that is, the cross section
for their scattering. It has dimension L2 , and must be given by a formula of
the form

dσ = Φσ h|M|2 i dV (pa + pb ; p1 , p2 , . . . , pn ) Fsymm , (6.6)

with its own flux factor, also still to be determined. We see that, in order to
get the formulae (6.5) and (6.6) to actually work, we have to establish

• the flux factors ΦΓ and Φσ ;

• the algorithm to derive from the connected Green’s function the am-
plitude. In particular this calls for a special treatment of the external
lines.

We shall solve these issues in the next section.

6.3.2 The truncation bootstrap

We have come to one of the centrally important steps in our treatment of
scattering. We shall solve the questions of the previous section by assuming
that we have solved them and drawing consequences : a bootstrap mechanism
Consider the process in which two particles with momenta pa and pb scatter
and yield j + n stable particles in the final state, whose momenta we label by
k1 , k2 , . . . , kj and q1 , q2 , . . . , qn . The distinction between these groups lies in
the fact that, whereas the k’s emerge ‘directly’ from the scattering, the q’s are
in fact the decay products of an unstable particle that was ‘directly’ produced
8
√
Some authors choose to include a factor 1/ k! in the transition amplitude M. I
am opposed to this since such a prescription introduces a distinction between particles in
the initial and those in the final state, which may destroy the crossing symmetry of the
amplitude.
August 31, 2019 153

together with the k’s. Nevertheless, the complete final state consists of both
the k’s and the q’s. The relevant diagrams thus look like

pa 111111
000000
000000
111111
000000
111111
000000
111111
}k 1,2,...

M=
000000
111111
000000
111111
A
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111 p
1111111
0000000
0000000
1111111
0000000
1111111
0000000
1111111
B
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
}q 1,2,...
(6.7)
pb
Note that the connected blobs may themselves contain many different in-
dividual diagrams. By separating the blobs A and B we indicate that the
unstable particles is actually quite long-lived so that the place where it is
produced and that where it decays tend to be clearly separated.
Now, we shall assume that we have, somehow solved the problem of how
to go from connected Green’s function to amplitude, and that we have applied
this procedure to the above process, in particular to the external lines. We
then have for the amplitude the form
i~
M=A B , (6.8)
p2 − m2 + imΓ
where p = q1 + · · · + qn is the momentum of the (internal !) line correspond-
ing to the unstable particle, and p2 = p · p. The unstable particle’s mass
is m, and its total decay width is Γ. The symbols A and B stand for the
‘processed’ connected Green’s functions for the ‘production’ process A and
the ‘decay’ process B, but with the Feynman factors for the unstable inter-
mediate particle removed. Assuming, for simplicity, that Fsymm = 1, we then
have for the differential cross section the form
~2
dσ = Φσ |A|2 |B|2 dV (P ; k1 , . . . , kj , q1 , . . . , qn ) , (6.9)
(p2 − m2 )2 + m2 Γ2
where P = pa + pb . In order to emphasize that p is the sum of the q’s, we
may write this also as

dσ = Φσ |A|2 |B|2 dV (P ; k1 , . . . , kj , q1 , . . . , qn )
~2 d4 p
(2π)4 δ 4 (p − Σq) , (6.10)
(p2 − m2 )2 + m2 Γ2 (2π)4
with obvious notation for the sum over the wavevectors q.
 154 August 31, 2019

Now, we let the unstable particle approach stability, so that the location
where it decays becomes widely separated from that where it is produced.
That is, we examine the case that Γ becomes very, very small, and we may
E39 approximate
1 π
→ δ(p2 − m2 ) . (6.11)
(p2 − m2 )2 + m2 Γ2 mΓ

We can then use this to rewrite

dV (P ; k1 , . . . , kj , q1 , . . . , qn ) d4 p
2 4
(2π)4 δ 4 (p − Σq) (6.12)
2 2
(p − m ) + m Γ 2 2 (2π)
as
1 d4 p δ(p2 − m2 )
dV (P ; k1 , . . . , kj , q1 , . . . , qn ) (2π)4 δ 4 (p − Σq)
2mΓ (2π)3
1
= dV (P ; k1 , . . . , kj , p) dV (p; q1 , . . . , qn ) . (6.13)
2mΓ
Inserting this in Eq.(6.10) we see that the cross section now takes the form

dσ = ~ |A|2 dV (P ; k1 , . . . , kj , p)

1 1
~ |B|2 dV (p; q1 , . . . , qn ) .


(6.14)
Γ 2m
Now, step back and consider what it is we are actually computing here : it
is the cross section for producing an almost-stable particle p, together with
the k’s in a specified configuration, followed by the decay of the particle p
into a specified configuration of q’s. Under the usual ideas of conditional
probability, this is the same as first computing the cross section for the
production of p and the k’s, followed by the conditional probability that,
given p, we see it decay into the q’s. This conditional probability, called the
(differential) branching ratio, is the partial decay width for p to go into the
q’s (computed in the p rest frame), divided by the total decay width, in this
case Γ (also defined in the rest frame). We conclude that
• ~ |A|2 is h|M|2 i for the process pa + pb → k1 + · · · + kj + p ;

• ~ |B|2 is h|M|2 i for the process p → q1 + · · · + qn ;

• ΦΓ is given by 1/(2m).
August 31, 2019 155

In a sense, we have managed to cut through the p line, and interpret the
process rather as it would be given by the diagrams

pa 000000
111111
111111
000000
000000
111111
000000
111111
}k 1,2,...

M →
000000
111111
000000
111111
A
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
p p 1111111
0000000
0000000
1111111
0000000
1111111
0000000
1111111
B
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
}q 1,2,...
(6.15)
pb

A point to be noted here has been somewhat hidden so far. The connected
Green’s functions contain factors (2π)4 δ 4 (· · ·) for overall momentum conser-
vation. This conservation has been imposed already, however, in our choice
of the phase space integration elements dV . We therefore have to remove
these factors as well in the transition from connected Green’s function to M.

What about the treatment of the external lines ? In the above discussion
we started with p as an internally ocurring unstable particle, carrying its
own propagator. As we let it become stable, the propagator has disappeared
into the phase space counting, leaving only a residue of a factor ~2 . At the
end of the story the particle p has become a stable particle occuring as an
external line in the blob A. This, therefore, must be the prescription for
the external lines ! This is called truncation or amputation of external lines.
An external line√must apparently carry, instead of its undefined propagator,
simply a factor ~. We arrive at the following, expanded set of rules for the
calculation of scattering amplitudes M (as opposed to Green’s functions) :
156 August 31, 2019

k i~
↔ internal
k · k − m2 + i
111111111
000000000
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111 k √
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
↔ ~ external
000000000
111111111
000000000
111111111
000000000
111111111

k1 k2 i
↔ − λ4 (2π)4 δ 4 (k1 + k2 + k3 + k4 )
k4 k3 ~
k1 k2 i
↔ + J(k2 )(2π)4 δ 4 (k1 + k2 )
~
Z∞
dk
for every momentum k (6.16)
2π
−∞

At the vertices the momenta must be counted either all

incoming or al outgoing. Drop the 2π 4 δ(· · ·) for overall
momentum conservation.

Feynman rules, version 6.1

Let us now turn to the two-particle initial state. The flux factor ΦΓ for
particle decay has been found to be 1/(2m). It is related to how we count
the density of states of the incoming particle9 . We can directly translate this
to the case of two-particle scattering. Let us work in the Lorentz frame in
which particle b is at rest while particle a impinges upon it10 . Keeping in
mind the effect of Lorentz transformations on the density of states we see
that whereas mb remains, ma has to be replaced by p0a , in accordance with
the discussion in section 5.2.3. The density-of-states factor for the two-body
initial state is therefore 1/4p0a mb . Since, however, we are asking for a cross
section rather than a transition rate, we have to divide this by the velocity
of particle a in b’s rest frame, that is, by a factor |~pa |/p0a . The flux factor
therefore becomes
Φσ = (4mb |~pa |)−1 .
This expression, being given in a specific Lorentz frame, is not very attractive.
9
If we fix the particle to be at rest, we must count the states as d3 (~k) d3~k/2ω(~k) =
R

1/2m.
10
Not the most practical frame, to be sure : at the LHC it is a frame in which the whole
ATLAS detector, say, moves at 99.99999882% of the speed of light.
August 31, 2019 157

We can, however, write it in an explicitly Lorentz-invariant form :

1
−1/2
Φσ = λ (pa + pb )2 , p2a , p2b , (6.17)
2
where we have introduced the Källén function

λ(x, y, z) ≡ x2 + y 2 + z 2 − 2xy − 2xz − 2yz = (x − y − z)2 − 4yz . (6.18)

It often happens that the colliding particles have masses that are negligible
compared to their combined invariant mass, the square of which is commonly
denoted by the Mandelstam variable s. In that case, we may write
1
Φσ ≈ , s ≡ (pa + pb )2 . (6.19)
2s
This finishes our bootstrap treatment of the relation between connected
Green’s functions and scattering amplitudes, or matrix elements.

6.3.3 A check on dimensionalities

It is instructive to check that the widths and cross section expressions that
we have derived do, indeed, have the correct dimensionality. By dim[ ] we
shall denote the dimensionality of objects, and L will denote
R a length. In the
first place, from the fact that the action S, that contains d4 x (∂ϕ)2 , must
have the same dimension as ~, we can immediately derive the dimensionality
of the fields11 :
dim[ϕ] = dim[φ] = dim ~1/2 /L .
 
(6.20)
A connected Green’s function with n external lines is the expectation value
of ϕn and must therefore have dimension12

dim[Cn ] = dim ~n/2 /Ln .

 
(6.21)

The Dirac delta function imposing wavevector conservation has dimension-

ality
dim δ 4 (k) = dim k −4 = dim L4 .
     
(6.22)
11
In
 four spacetime
 dimensions! In d dimensions it would read dim[ϕ] =
dim ~1/2 L1−d/2 .
12
Higher-order contributions to Green’s functions contain, of course, additional powers
of ~: but these must occur only in dimensionless combinations with the coupling constants
of the theory.
158 August 31, 2019

To go from the connected Green’s function Cn to the n-point matrix ele-

ment Mn , we have to extract the external propagators as well as the overall
wavevector conservation delta function, and assign a factor ~1/2 to each ex-
ternal line: therefore,
 
Cn n/2
= dim Ln−4 .
 
dim[Mn ] = dim n 4
~ (6.23)
(C2 ) δ (k)

The n-particle phase-space integration element dVn has dimensionality L4−2n

as we have seen. Taking into account that the flux factor ΦΓ = 1/2m must
have the dimensionality of 1/m, that is, L, the dimensionality of the decay
width of a single particle into n particles is given by

dim[Γ(1 → n)] = dim (1/m)|Mn+1 |2 dVn = dim L−1 ,

   
(6.24)

as required. Similarly, for the cross section of two particles going into n
particles we have

dim[σ(2 → n)] = dim (1/m2 )|Mn+2 |2 dVn = dim L2 ,

   
(6.25)

again as required. Note that the above analysis is kept simple because we
have restricted ourselves to the use of wavevectors rather than mechanical
momenta, which would introduce additional factors of ~ in the calculation.
The other natural constant, c, need not enter here.

6.3.4 Crossing symmetry

In our treatment of antimatter in the previous chapter we have seen that
the production (absorption) of a particle is, in a sense, æquivalent to the
absorption (production) of its antiparticle. We can make this even more
specific as a relation between various scattering amplitudes : this goes by the
name of crossing symmetry. Consider a generic 2 → 2 scattering process :

a(p1 ) + b(p2 ) → c(q1 ) + d(q2 )

where we have indicated the momenta of the particles. Let us write the
corresponding amplitude as M(p1 , p2 , q1 , q2 ). By moving particles from the
initial to the final state13 , or vice versa, we can then find the amplitudes for
13
You can visualize this by taking an outgoing particle, say, and dragging its external
leg from the final to the initial state.
August 31, 2019 159

the crossing-related processes, for example :

a+b→c+d : M(p1 , p2 , q1 , q2 ) ,
a + c̄ → b̄ + d : M(p1 , −p2 , −q1 , q2 ) ,
a + d¯ → b̄ + c : M(p1 , −p2 , q1 , −q2 ) ,
c̄ + d¯ → ā + b̄ : M(−p1 , −p2 , −q1 , −q2 ) . (6.26)
Since the momenta of all (anti)particles have positive energy, the minus signs
yield momenta with negative energy. Depending on the type of the particle14 ,
this may involve an analytic continuation of the amplitude function M.

6.4 Unitarity issues

6.4.1 Unitarity of the S matrix
If M is to be a correct form of the scattering amplitude for a given initial
state to be observed, after time evolution, as a given final state, it must obey
the constraints of unitarity which we shall now discuss. In a more traditional
quantum-mechanical parlance, the initial state is given to us at some time
in the far past, where the incoming particles are supposed to be so widely
separated that they are essentially free : the state of the system is then
|in, t = −∞i
We now let nature take its course : the incoming particles approach one
another, the interaction is ‘switched on’, and the system evolves into some,
possibly very complicated, superposition of free-particle states :
|in, t = −∞i → |in, t = +∞i
Finally, the final state is observed to be a particular free-particle state (as-
suming the final-state particles have been able to move very far away from
one another), that is,
|out, t = +∞i .
The probability amplitude for this to happen is of course
M = hout, t = +∞|ins, t = +∞i ≡ hout, t = +∞| S |in, t = −∞i ,
(6.27)
14
Especially for Dirac particles.
 160 August 31, 2019

where S is the matrix describing the time evolution of the incoming state from
t = −∞ to t = +∞. An important consideration here is the conservation of
probability. That is, any initial state |ii must go to some final state |f i with
100% probability15 ; of course |f i = |ii may also be one of the possibilities.
Writing this using the S matrix we have
X X
  
  

i S †  f f S  i = i S † S  i .
 
1= | hf |S| ii |2 = (6.28)
f f

Conversely, any final state |f i must have come from some initial state |ii
with 100% probability, so that
X X
  
  

f S  i i S †  f = f SS †  f .
 
1= | hf |S| ii |2 = (6.29)
i i

E40 Since this must hold for all states |i, f i we have

S † S = SS † = 1 , (6.30)

and the S matrix is unitary. Note that we have had to assume that the set of
initial states |ii and final states |f i are complete16 . The free-particle states
are natural choices for complete orthonormal bases, and we see that M is
simply a matrix element of the S matrix. We shall investigate this in some
more detail.
For simplicity, let us assume that we can label the initial states with a
discrete label i, and the final states by a similar discrete label f . We can
then write the S matrix element as

S f i = δ f i + Mf i , (6.31)

where the Kronecker delta embodies what would happen if there were no in-
teractions : the only possible observed final state would in that case be iden-
tical to the initial state (two particles, say, continuing on their way without
having interacted). The remainder Mf i is the object described by Eq.(6.27) ;
it is the result of the interactions of the theory, and is described by the Feyn-
man diagrams. Note that Mii 6= 0 is quite possible ; it corresponds to the
case where the final state happens to reproduce the initial state, so to speak
15
Try to imagine a world in which this does not hold ! Conservation of probability is a
dogma — but a reasonable one.
16
Again, try to imagine what things would look like if that were not the case. . .
August 31, 2019 161

in spite of the interactions. This is called the forward scattering amplitude.

Now, the unitarity of the S matrix is expressed17 as SS † = S † S = 1, or
X
∗
Skf Ski = δf i , (6.32)
k

or, in terms of M :
X
Mf i + M∗if + M∗kf Mki = 0 . (6.33)
k

As a special case, we can consider f = i : we then have the optical theorem,

X
2 Re (Mii ) + |Mki |2 = 0 , (6.34)
k

which immediately shows that the forward scattering amplitude must have
negative real part18 . In Eq.(6.34), the first term is called the dispersion term,
and the second one the absorptive term, terms which originate in the study
of waves moving through a medium. Also, in our ‘discrete’ formulation,

|Sf i | ≤ 1 for all i, f , (6.35)

which implies that Mf i can not be arbitrarily large ; we shall employ this
idea extensively later on. In appendix 15.10 we derive a more precise bound,
called partial-wave unitarity, but that calls for Feynman rules we still have
to develop.
A an elementary illustration of the optical theorem we consider the fol-
lowing physical process. We start with an empty initial state i (that is, a
state containing no particles). At some moment a source kicks in, producing
an unstable particle with wavevector p, mass m and total width Γ. Sometime
later, the same source absorbs the particle, and at the end the final state f
is empty again. The simple Feynman diagram describing this is
J
p J
17
Since S may be an infinite matrix, both conditions are necessary, whereas for a finite
matrix one would suffice.
18
A word of caution : in much of the literature, the statement reads that the amplitude
must have positive imaginary part. This is simply due to the fact that in those texts, the
S matrix element is written not δ + M but δ + iM. I do not see any particular virtue in
this.
162 August 31, 2019

and since the initial and final state coincide (f = i) this is a forward scattering
amplitude ; it must obey the optical theorem. We shall now verify this. The
matrix element is given by19
   
J i~ J
Mii = i 2 2
i , (6.36)
~ p − m + imΓ ~
so that
J2 mΓ
Re (Mii ) = − , (6.37)
~ (p2 − m2 )2 + m2 Γ2
which is indeed negative. Now, we consider the matrix elements Mki as used
in Eq.(6.34). These describe the initial state i going over in any final state
k, that is, they describe the decay of the particle after it has been produced
by the source :
111
000
J 000
111
000
111
000
111
000
111
000
111
p 000
111
000
111
000
111

and we shall denote them by

J ~
Mki = −i D , (6.38)
~ p − m2 + imΓ
2

where iD is the contribution of the ‘decay blob’. We then have

X J2 X
|Mki |2 = 2 |D|2 . (6.39)
2 2 2 2
(p − m ) + m Γ k
k

The optical theorem (6.34) will therefore be satisfied if

1 X
Γ= ~ |D|2 . (6.40)
2m k

But this is, of course, precisely the prescription for the decay width of the
particle, if we realize that the final state k indicates not only all possible
final states, but also that the summation over k should include the phase-
space integration. This short excercise illustrates both the optical theorem
and the computational prescriptions arrived at before. Note that the factor
~ corresponds √ precisely with the Feynman rule that an external line should
carry a factor ~.
19
You might opt for the ‘absorbing’ source to be J ∗ rather than J. The conclusions are
the same.
August 31, 2019 163

6.4.2 The cutting rules

We shall now consider how the unitarity relation (6.33) can be made useful
in the language of Feynman diagrams. To start, we realize that this equation
contains, in addition to the ‘standard’ matrix element Mf i for initial state i
and final state f , also M∗if which describes the (complex conjugate) matrix
element for initial state f going over into final state i, that is, the time-
reversed process. We shall embody this in a useful manner by introducing a
cutting line. A cutting line cuts across diagrams separating them into a ‘left’
and ‘right’ piece. Any diagram to the left of a cutting line is interpreted in
the usual manner ; any diagram to the right of a cutting line is interpreted to
be the complex conjugate of the time-reversed version of the diagram. That
is,
 ∗
0000000
1111111 1111111
0000000 0000000
1111111 111111
000000
1111111
0000000 1111111
0000000
0000000
1111111 0000000
1111111 0000000
1111111 000000
111111
0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111 0000000
1111111 000000
111111
000000
111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
= 0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
, 0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
= 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111

0000000
1111111 0000000
1111111 0000000
1111111
0000000
1111111 000000
111111
0000000
1111111
0000000
1111111 0000000
1111111 0000000
1111111 000000
111111
0000000
1111111 0000000
1111111 000000
111111
0000000
1111111
0000000
1111111 000000
111111
000000
111111

If the diagram contains oriented lines, the time-reversal also inverts the ori-
entation of those lines (if the orientation is indicated by an arrow, we reverse
the arrow). We can write Eq.(6.33) diagrammatically as

000000
111111
111111
000000
000000
111111
111111
000000
000000
111111 111111
000000 X 111111
000000 000000
111111
000000
111111 000000
111111 000000
111111 000000
111111
000000
111111 000000
111111 000000
111111 000000
111111
i 000000
111111
000000
111111
000000
111111 f + i 000000
111111
000000
111111
000000
111111
f + i 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111 =0 . (6.41)
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
k k 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
f
000000
111111 k 000000
111111 000000
111111
000000
111111

It is possible to sharpen this equation to make it more useful. In the first

place, Eq.(6.41) holds for whole matrix elements, evaluated to all orders
in perturbation theory. This implies that it must also hold for each or-
der separately20 . However, even at some fixed order, Mf i can contain very
many diagrams. Consider a somewhat-complicated Feynman diagram in ϕ3
theory :

(6.42)
20
If Eq.(6.41) were not to hold order-by-order, this would imply subtle relations between
coupling constants, ~, and the like. We would then be in a position to actually compute
coupling constants from first principles, which would be good — too good to be true, in
fact, if we believe that we are allowed some freedom in choosing the values of the coupling
constants while retaining a consistent theory.
 164 August 31, 2019

The corresponding Lagrangian reads

1 1 1
L = (∂ µ ϕ)(∂µ ϕ) − m2 ϕ2 − λϕ3 . (6.43)
2 2 6
The unitarity structure of the above Feynman diagram is not immediately
E41 obvious since there are, at this order of perturbation theory, no less than 58
diagrams that contribute to this amplitude. We can, however, employ the
following trick. Let us assign a different label to each line in the diagram, in
an arbitrary manner, for instance
2 5
6
1 4 (6.44)
7
3
9 8
and let us pretend that each line corresponds to a different field. This diagram
can then be interpreted as coming from a theory with 9 distinct fields (with
identical mass) and Lagrangian
9  
X 1 µ 1 2 2
L = (∂ ϕn )(∂µ ϕn ) − m ϕn − V ,
n=1
2 2
V = λ123 ϕ1 ϕ2 ϕ3 + λ245 ϕ2 ϕ4 ϕ5 + λ349 ϕ3 ϕ4 ϕ9
+λ567 ϕ5 ϕ6 ϕ7 + λ789 ϕ7 ϕ8 ϕ9 . (6.45)
Nothing forbids us to assign to the various ϕϕϕ couplings precisely the nu-
merical value λ later on. Now, it is easily seen that, in order λ123 λ245 λ349 λ567
λ789 , the diagram (6.44) is the only diagram that can contribute in this the-
ory21 ! We can do even more : by inspection of all possibilities, we can simply
realize that the only final states k in the unitarity condition (6.41) must be
precisely k = {2, 3}, {5, 9}, {2, 4, 9} or {3, 4, 5}, if we want to end up with
the right order in perturbation theory22 . In other words,

+ +

21
The secret resides in the fact that in V the external fields 1,6 and 8 occur precisely
once, and the other fields precisely twice.
22
Note that, for instance, the choice k = {5, 7, 8} would result in the right-hand half of
the diagram being disconnected ; the choice k = {2, 4, 7} is inconsistent since both 6 and
8 are in the final state.
August 31, 2019 165

+ + + = 0 , (6.46)

where we have omitted the line labellings : indeed, the same identity must
hold for the original diagram (6.42). This establishes the so-called cutting
rules (also called the Cutkosky rules), which can be most simply expressed in
words : take a diagram and move the cutting line through it from right to left
in all possible manners, making sure that the two halves in which the diagram
is cut remain connected and that neither the inital state or the final state
is dissected. The particles described by internal lines through which the cut
runs must be assumed to be on their mass shell23 . The sum of all the possible
contributions then vanishes24 . It must be stressed, however, that we have not
proven the Cutkosky rules in this section : rather we have established that a
theory must satisfy them if it is to be a good, probability-conserving theory.

6.4.3 Infrared cancellations in QED

As an illustration of how the cutting rules may be applied we shall make a
slight jump ahead and consider quantum electrodynamics, that is the theory
of photons and electrons. Their Feynman rules will be dicussed later ; for
now it is sufficient to know that the only interaction vertex in the theory is
the three-point vertex

where the oriented lines stand for electrons and positrons, and the wavy
line denotes the photon. Let us consider the 1PI two-loop corrections to the
photon propagator. These are given by
111111111
000000000
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
= + +
000000000
111111111
000000000
111111111
000000000
111111111

23
This may mean that the situation thus described fails to meet the restrictions of
momentum/energy conservation ; then, that contribution vanishes.
24
You might object that in a theory with many different fields the symmetry factors of
the diagrams will, in general, be different from those of a theory with only a single field,
and this is true : however, in the summation over the ‘intermediate states’ k we must of
course also include the ‘indentical-particle’ symmetry factor Fsymm , which precisely repairs
the correspondence — another illustration of the crucial rôle of the symmetry factors !
166 August 31, 2019

By applying the cutting rules we can investigate the real part of this two-loop
contribution:
000000000
111111111
111111111
000000000
000000000
111111111
000000000
111111111
000000000
111111111
 ∗
000000000
111111111
111111111
000000000
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111 000000000
111111111

+
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
↔
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111 000000000
111111111
000000000
111111111 000000000
111111111
000000000
111111111 000000000
111111111

↔ + + + +

+ + +

+ + . (6.47)

This set of 10 cut diagrams is, as we can see, equal to25

  ∗
+ + + (c.c.)
 2
 
+  +  ,
 (6.48)

where integration over the final state is implied. As we shall see, the presence
of a photon in the final state leads to a so-called infrared (IR) divergence
arising from the fact that the probability of emitting an on-shell photon goes
to infinity as the photon energy goes to zero. The process described by the
last two diagrams has therefore an infrared divergence. This divergence is
neatly cancelled by a compensating divergence in the diagrams with a virtual
photon in the first line. This is a well-known fact26 ; but it is instructive to
see that the statement about the cancellation of the infrared divergences can
be replaced by the simpler statement that the photon propagator is free from
infrared divergences27 . This is one example of a useful rule of thumb : when
25
Recall the remark about oriented lines just above Eq.(6.41).
26
And a fortunate one.
27
Two remarks are in order here. In the first place, the virtual-photon diagrams do
contain divergences related to the loop momentum going to infinity : these are ultraviolet
(UV) divergences. ß The photon propagator is therefore still ultraviolet divergent, and
this is cured in the usual manner by renormalization. In the second place, the cancellation
of IR divergences takes place even when we restrict the phase space for the outgoing
particles, provided that zero-energy photons are admitted. This last point is, of course,
not proven by our argument ; but if it were not true you would see really infinite physical
cross sections by making suitable cuts on the final state ! Nature, fortunately, is more
friendly than that.
August 31, 2019 167

you encounter loop diagrams, try to envisage the physics that is described by
cutting them. In fact, the cancellation can be pinpointed further ; the single
statement that the single diagram

is IR-finite means that the IR divergences in

 2 ! !∗
 
 and + (c.c.)
 

 

must cancel between them.

6.5 Some example calculations

6.5.1 The FEE model
As an example of an application of what we have learned so far, we shall
investigate at theory that contains two particle types, one of mass m, denoted
by E, and another denoted by F , of mass M . The Lagrangian density of this
theory is given by

1 µ m2 2
L = (∂ ϕE ) (∂µ ϕE ) − ϕE
2 2
1 M 2 2 mλ
+ (∂ µ ϕF ) (∂µ ϕF ) − ϕF − ϕF ϕE 2 . (6.49)
2 2 2
There exists a single coupling between two E’s and one F . Note that the
Feynman rule for the vertex is given28 by −imλ/~ ; we have introduced a
factor m in order to ensure that

dim[λ] = dim ~−1/2

 

with no length scale.

28
It is customary to leave out the (2π)4 δ 4 () of momentum conservation, since it is present
in all vertex Feynman rules for translation-invariant interactions.
168 August 31, 2019

6.5.2 Two-body phase space

Since we shall consider processes ending in a two-body final state, it is ex-
pedient first to work out the corresponding two-body phase space. For the
sake of generality we shall do this for a final state containing two momenta
q1,2 µ with general masses m1,2 . Furthermore we shall write

P µ = q1 µ + q2 µ , s = P µ Pµ . (6.50)

The phase space (and with it widths and cross sections) is often most easily
evaluated in the rest frame of P µ , in which ~q1 = −~q2 . The phase space
integration element is given by29
1
dV (P ; q1 , q2 ) = d4 q1 δ(q1 2 − m1 2 ) d4 q2 δ(q2 2 − m2 2 ) δ 4 (P − q1 − q2 ) .
(2π)2
(6.51)
As a first step, we cancel d4 q2 against the four-dimensional Dirac delta, and
write the q1 integration in its not-explicitly-covariant form :
1 d3 ~q1 2 2


dV (P ; q1 , q2 ) = δ (P − q 1 ) − m2 . (6.52)
(2π)2 2q1 0
Now, the q1 integration element can be expressed in polar coordinates as
d3 ~q1 |~q1 |2 d|~q1 | dΩ 1
0
= 0
= |~q1 | dq1 0 dΩ , (6.53)
2q1 2q1 2
where we denote the ~q1 solid angle by

dΩ = d cos θ dφ (6.54)

with a polar angle θ running from 0 to π and an azimuthal angle φ running

from 0 to 2π, and use the fact that

|~q1 | d|~q1 | = q1 0 dq1 0 . (6.55)

The Dirac delta imposing the mass shell condition on q2 can be written as
√

δ (P − q1 )2 − m2 2 = δ s + m1 2 − m2 2 − 2q1 0 s

s + m1 2 − m2 2
 
1 0
= √ δ √ − q1 , (6.56)
2 s 2 s
29
It is usual not to include the step functions that require the energies to be positive.
August 31, 2019 169

where the rest frame of P has been used. We immediately find that

s + m1 2 − m2 2 s + m2 2 − m1 2
q1 0 = √ , q2 0 = √ , (6.57)
2 s 2 s

and
1
|~q1 | = |~q2 | = √ λ(s, m1 2 , m2 2 )1/2 , (6.58)
2 s
where the Källén function crops up again. In the P µ rest frame, the phase
space integration element is therefore given by
1/2
m1 2 m2 2

1
dV (P ; q1 , q2 ) = λ 1, , dΩ . (6.59)
32π 2 s s

6.5.3 A decay process

As a first application, we shall assume that M > 2m so that the F particle
can decay into a pair of E’s:

F (P ) → E(q1 ) E(q2 ) .

In lowest order, its single Feynman graph is given by

q1
P q2

The corresponding matrix element is quite simple :

mλ √ 3 √
M = −i ~ = −imλ ~ , (6.60)
~
so that it has dimensionality dim[1/L] as it should. The decay width is
therefore
1 1
dΓ(F → EE) = |M|2 dV (P ; q1 , q2 )
2M r 2!
2 2 2
mλ~ 4m
= 2
1− dΩ . (6.61)
128π M M2
170 August 31, 2019

Note the occurrence of the symmetry factor 1/2 arising from the fact that the
two final-state E particles are indistinguishable. The angular integration is
of course trivial in this simple case, and we immediately find the total width
r
m2 λ2 ~ 4m2
Γ(F → EE) = 1− , (6.62)
32πM M2
with the correct dimensionality dim[Γ] = dim[1/L].

6.5.4 A scattering process

As a second application, we F to be massless, M = 0. We now have an
extremely primitive picture of the electron-photon system, where E is the
electron and F the photon. We consider the process of ‘Compton scattering’ :

E(p1 ) F (p2 ) → E(q1 ) F (q2 )

which, to lowest order, is given by two Feynman diagrams:

p1 q2 p1 q2
M= q1 + q1 (6.63)
p2 p2

We start by defining the various momenta. It is usually best to work

p in the
centre-of-mass frame. If we denote the total scattering energy by (s), we
can write the momenta as follows, using Eqs.(6.57,6.58) :

pµ1 = (p, q~ep ) , pµ2 = (q, −q~ep ) , q1µ = (p, q~eq ) , q2µ = (q, −q~eq ) . (6.64)

Here ~ep,q are the

√ unit vectors2 that
√ give the directions of p~1 and ~q1 , and
2
p = (s + m )/2 s, q = (s − m )/2 s. The tree-level matrix element is then
 
2 2 1 1
M = −i~λ m +
(p1 + p2 )2 − m2 (p1 − q2 )2 − m2
~λ2 m2 (1 − cos(θ))
= =i , (6.65)
s + m2 + (s − m2 ) cos(θ)

where a = 2s/(s + m2 ), b = (s − m2 )/(s + m2 ), and θ is the angle between

~ep and ~eq . At large energies, b is close to one and the angular distribution
is strongly peaked in the backwards direction, θ ∼ π ; at lower energies if
August 31, 2019 171

flattens out. We can also see that in the strictly forward direction, θ = 0, the
amplitude vanishes. The flux factor and the phase space integration element
read

1 1 s − m2
Φσ = , dV (p 1 + p ; q
2 1 2, q ) = d cos(θ) dϕ , (6.66)
2(s − m2 ) 32π 2 s

with ϕ the azimuthal angle, on which M does not depend. The differential
cross section thus becomes
2
~2 λ4 m4

1 − cos(θ)
dσ = d cos(θ) dϕ . (6.67)
64π 2 s (s + m2 ) + (s − m2 ) cos(θ)

The angular integral readily performed, and we find the total cross section

~2 λ4 m4
 
2s  s  s
σ(s) = 1− log + 2 . (6.68)
16πs(s − m2 )2 s − m2 m2 m

This expression would appear to diverge at the threshold, s = m2 , but by E42

carefully expanding in powers of s − m2 we find

~2 λ4 m4
σ(m2 ) = , (6.69)
48πm2

which can be considered the cross section for static scattering. Let us consider
this expression more closely. In the first place, the factor λ4 and consequently
the factor ~2 could have been foreseen from the start. The fact that the cross
section must have dim[σ] = dim[L2 ] implies that at the threshold, where
m is the only length scale in the problem, there must also be an overall
factor 1/m2 . Moreover, n body phase space contains a power π 4−3n from its
definition ; and also it contains n − 1 solid angles to be integrated over, each
giving rise to30 a factor π. This means that the total cross section for an E43
n-body final state will contain a factor π 3−2n . In this way, almost the whole
E44
cross section formula is determined, and all the calculational effort is only
used to find the numerical factor 1/48. E45

30
This is to say that the angular integral does not necessarily evaluate to π, but rather
that a factor π invariably arises in the result of a solid-angle integral.
172 August 31, 2019

6.6 The one-loop cookbook

6.6.1 The one-loop calculation
In the computation of loop corrections, many more technicalities and tricks
are required ; here we present a first introduction to them. As a working
example we consider the diagram in ϕ3 theory :

k p
2
p
L(p ) = (6.70)
p+k

The momentum flowing through the external propagators is pµ . The diagram,

which excludes the external propagators, is given by
Z
2 1 1
L(p ) = 4
d4 k 2 . (6.71)
2(2π) (k − m + i)((k + p)2 − m2 + i)
2

We have dropped the coupling constants, but included the symmetry factor
1/2. We start by combining the two denominators by using the so-called
Feynman trick, which we shall now discuss.

The Feynman trick

Consider n positive real numbers aj and real numbers νj (j = 1..n). We can
write
n Z∞ Y n
!
dzj zj νj −1
 
Y 1 n
= exp − Σj=1 zj aj . (6.72)
a νj
j=1 j j=1
Γ(νj )
0

In this integral, we may define σ as the sum of the z’s, and define xj as zj /σ,
as follows:
n Z∞ Y n
!
dzj zj νj −1
 
Y 1
νj
= exp − Σj zj aj
j=1
a j j=1
Γ(ν j )
0
  Y !
 zj 
dσ δ σ − Σj zj dxj δ xj − . (6.73)
j
σ
August 31, 2019 173

We can now eliminate the z’s in favor of the x’s, and integrate out σ :
n Z∞ Yn 
dxj xj νj −1
  
Y 1
= δ 1 − Σj xj
a νj
j=1 j j=1
Γ(νj )
0
Z∞ P
 
−1+ j νj
× dσ σ exp − σ Σj xj aj
0
(dxj xj νj −1 ) δ(1 −
P  Q P
Γ Z1 xj )
j νj
j
j
= Q Pj νj . (6.74)
j Γ(νj )

0 x 1 a1 + x 2 a2 + · · · + x n an
In our example we have n = 2 and ν1 = ν2 = 1 :
Z1
1 dx
=
2 . (6.75)
a1 a2 xa1 + (1 − x)a2
0

Incidentally we now also have a proof of the so-called Euler’s integral :

Z1 Yn Q
νj −1


j Γ(νj )
dxj xj δ − 1 + Σ j xj = P , (6.76)
j=1
Γ( j ν j )
0

by simply putting a1 = a2 = · · · = an = 1.

Loop momentum shift

After the Feynman trick, the loop integral (6.71) is given by
Z1 Z
1
−2
L(p2 ) = dx d4 k k 2 + 2x(pk) + xp2 − m2 + iη , (6.77)
2(2π)4
0

where we have switched the notation for the infinitesimal width from  to η.
Replacing the loop integral momentum k by k − xp we obtain
Z1 Z
1
−2
L(s) = dx d4 k k 2 + x(1 − x)s − m2 + iη , s = p2 ,
2(2π)4
0
(6.78)
and we see explicitly that the loop integral in this case depends only on s
(and m).
174 August 31, 2019

Dimensional regularization

The k integral, containing as many positive powers of k as negative ones, is

logarithmically divergent, and we have to regularize it somehow. The most
popular method of doing so is nowadays that of dimensional regularization31 .
We change the number of dimensions of spacetime from 4 to D = 4 − 2
where we keep  finite up to the end of the calculation, and only then (after
renormalization, say) take the limit  → 0. The k integral is then no longer
d4 k but rather dD k. However, since this changes the engineering dimension
of the diagram this would lead to loop-corrected amplitudes with inconsistent
units ; therefore we introduce a quantity µ with the dimension of k 32 , and
write

Z1
µ2
Z

−2
L(s) = dx dD k k 2 + x(1 − x)s − m2 + iη (6.79)
2(2π)4−2
0

It is important to realise the flexibility that we allow ourselves. If we interpret

 as a real number then choosing  > 0 will make the above integral finite. In
other situations (cf section 11.1)  > 0 mihgt be more ‘natural’. The power of
dimensional regularization is that we may assign these properties to  in any
way we like ; dimensional regularization treats the integral’s dimensionality
as a complex number, and the integral amenable to analytic continuation. A
particularly striking instance of dimensional doublethink is found in exercise
46.

The Wick rotation

Now we let the iη come into its own. The k integral can be written as
dD k = dk 0 dD−1~k, and in terms of k 0 the integrand skirts along a singularity
if (k 0 )2 ∼ ~k 2 − sx(1 − x) + m2 . we see that we can change the integration
contour from s(−∞, +∞) to (−i∞, +i∞) by a counterclockwise rotation.
That is, we may write dk 0 = idk 4 and let k 4 run over the real axis. This
(a) gives an overall factor of +i and (b) changes every occurrence of k 2 into

31
This is mainly due to the fact that the dimensionality of a theory usually has only a
small influence on properties such as causality or current conservation.
32
That is, inverse length. Loosely speaking, an energy scale.
August 31, 2019 175

−kE2 with kE2 = (k 4 )2 + |~k|2 . Our integral therefore becomes

Z1
iµ2
Z

−2
L(s) = dx dD k kE2 + m2 − x(1 − x)s − iη . (6.80)
2(2π)4−2
0

The t-shell formula

The integrand in Eq.(6.80) depends on k µ only through kE2 . It is therefore
sensible to first compute
Z
R(D; t) = dD k δ(kE2 − t) . (6.81)

Since we want to do this for possibly non-integer D this is on necessity

somewhat symbolic. We write
Z∞
R(D; t) = 2D dk 1 dk 2 · · · dk D δ (k 1 )2 + (k 2 )2 + · · · + (k D )2 − t

0
Z∞
= t D/2−1
dy1 · · · dyD y1 −1/2 · · · yD −1/2 δ(y1 + · · · + yD − 1)
0
Γ(1/2)D
= tD/2−1 . (6.82)
Γ(D/2)
√
To get this result we have written k j = yj t and used Euler’s formula,
Eq.(6.76). Our loop integral thus becomes, with D = 4 − 2 :
Z1 Z∞
iµ2 t1−
L(s) = dx dt . (6.83)
2(4π)2− Γ(2 − ) (t + m2 − x(1 − x)s − iη)2
0 0

The t integral
By writing t = a(1 − y)/y and using again Eq.(6.76) we can compute the t
integral :
Z∞ Z1
tα Γ(β − α − 1)Γ(α + 1)
dt = a1+α−β dy y β−α−2 (1−y)α = a1+α−β .
(t + a)β Γ(β)
0 0
(6.84)
 176 August 31, 2019

Thus we get, in our example,

Z1
i µ2 Γ()
−
L(s) = dx m2 − x(1 − x)s − iη . (6.85)
2(4π)2−
0

Incidentally, when a vanishes so does the t integral : we say that

Z∞
dt t1− tn = 0 , (6.86)
0

simply because the integral has a dimensionality but there is no quantity

E46 available to carry the appropriate power.

The  expansion
Keeping in mind that at some point we want to take  → 0 it is sensible to
expand in relevant powers of . First, we use Eq.(15.274),
1
Γ() = (1 − γE  + O 2 ) ,


(6.87)

where γE ≈ 0.5772 is Euler’s constant33 , to see that our integral diverges as
−1 whern  vanishes. This implies that in all other factors we must keep the
first-order terms but may discard higher orders34 . For the other factors we
may use
A = exp( log(A)) = 1 +  log(A) + O 2 ,


(6.88)
and putting everything together we arrive at the final, intermediate result :
i 2


L(s) = R  − R(m , s) + O () ,
2(4π)2
1
R = − γE + log(4π) + log(µ2 ) ,

Z1
R(m2 , s) = dx log(m2 − x(1 − x)s − iη) . (6.89)
0

33
Lots and lots of mathematical things are called after Euler. To spread the credit
somewhat, it is also called the Euler-Mascheroni constant. See also appendix 15.15.4
34
Higher-loop integrals typically diverge with higher powers of 1/, and figuring out
what to keep becomes accordingly more complicated.
August 31, 2019 177

So far the computation has been quite straightforward, in that many of its
steps are the same for all one-loop calculations. In particular we can see that
1/ always occurs with −γE + log(4π) + log(µ2 ) in the combination R .

The x integral : the hard work

In the evaluation of the x integral, it of course becomes important to keep
careful track of the logarithm’s singularity structure. We can distinguish
three cases, depending on the value of s = p2 .
1. s < 0. In this case m2 − sx(1
p − x) is always positive. Writing x =
(1 + y)/2 we have, with β = 1 − 4m2 /s (so that β > 1) :
Z1
R(m2 , s) = dx log(m2 + |s|x − |s|x2 )
0
Z1
1
= dy log(m2 + |s|/4 − |s|y 2 /4)
2
−1
Z1    
|s| 2 2
= dy log + log(β − y )
4
0
 
2 β+1
= log(m ) − 2 + β log . (6.90)
β−1
Limiting cases are R(m2 , 0) = log(m2 ) and R(0, s) = −2 + log |s|.
s < 4m2 . Proceeding in a similar way as above, but now with
2. 0 < p
η = 4m2 /s − 1, we can write
Z1
2
R(m , s) = log(s/4) + dy log(y 2 + η 2 )
0
= log(m2 ) − 2 + 2η arctan(1/η) . (6.91)

3. s > 4m2 . This is the more tricky case since the argument crosses zero
twice as xpmoves between 0 and 1. The two roots are given as x0,1 ,
with β = 1 − 4m2 /s (but now 0 < β < 1) :
x1 = (1 + β)/2 + iη , x0 = 1 − x1 = (1 − β)/2 − iη . (6.92)
178 August 31, 2019

Figure 6.1: The function R(m2 , s) of Eq.(6.89)

The iη becomes important when taking logarithms :

log(−x1 ) = log(x1 ) − iπ , log(−x0 ) = log(x0 ) + iπ . (6.93)

We evaluate the integral as follows :

Z1
R(m2 , s) = log(s) +


dx log(x − x0 ) + log(x − x1 )
0
 1
= log(s) + (x − x0 ) log(x − x0 ) + (x − x1 ) log(x − x1 ) − 2x


0
= log(s) + x1 log(x1 ) + log(−x1 ) + x0 log(x0 ) + log(−x0 ) − 2
= log(s) + 2x1 log(x1 ) + 2x0 log(x0 ) + iπ(x0 − x1 )
   
2 1+β
= log(m ) − 2 + β log − iπ . (6.94)
1−β

The form of R(m2 , s) in the three different regions can be seen as analytic
continuations of each other, but of course precisely how to analytically con-
tinue is the tricky point here ! In figure 6.1 we plot the function R(m2 , s) for
August 31, 2019 179

m = 3.14. It is continuous both at s = 0 and s = 4m2 . The real part of R

is positive (since log(3.142 ) > 2), its imaginary part is negative. At s = 4m2
the imaginary part ‘switches on’ suddenly. This indicates that R(s) has a cut
along the real axis starting at that value ; and of course the ‘kink’ in the real
part tells us the same. We also see that the real part of L(p2 ) only develops
above the ‘threshold’ value p2 = 4m2 , and is positive there, as required by
unitarity.

6.6.2 Dispersion relations

Kramers and Kronig rule!
We can place the above treatment into context by considering it as part of an
amplitude. Since the momentum p cannot possibly be on-shell for all values
of p2 , we consider the diagram of the FEE model

Mf = (6.95)

This diagram is therefore given by35

~2 λ2 iλ2
R − R(m2 , s) ,


Mf = Σ(s) , Σ(s) = (6.96)
(s − M 2 )2 16π 2

where we recognize M as the F mass, and m that of the E particle. We take

the final state to be identical to the initial state36 . In that case, R(m2 , s)
has negative imaginary part, and the real part of M is negative as required
by the optical theorem of Eq.(6.34). Next, we concentrate on the imaginary
part of R(m2 , s), for which we had found
p
=R(m2 , s) = −π β θ(s > 4m2 ) , β = 1 − 4m2 /s . (6.97)
By the Kramers-Kronig relations of appendix 15.15.9 we can relate this to
its real part by an integral :
Z∞ Z∞ p
2 1 =R(m2 , z) 1 − 4m2 /z
<R(s, m ) = P dz = −P dz , (6.98)
π z−s z−s
−∞ 4m2
35
Note that ~2 occurs, rather than ~.
36
In this FEE model the value of M does not depend on this, but in more evolved
models it does, and we want to study the forward scattering amplitude.
 180 August 31, 2019

where the P reminds us that the integral

p is a principal value integral (cf
appendix 15.15.6). By going over to v = 1 − 4m2 /z as integration variable,
we can write
Z1 
β2 4m2

2 1 2
<R(m , s) = 2 P dv − , β = 1 − . (6.99)
β 2 − v2 1 − v2 s
0

The second term is obviously divergent. But we should realize that what we
are trying to do here is to relate the imaginary part of an amplitude to its real
part ; and that real part is indeed divergent, as the R shows. The divergent
part of the v integral is therefore just a reflection of the loop divergence, and
since we have already taken that into account we may write
Z1
2 β2
<R(m , s) = 2 P dv +C , (6.100)
β 2 − v2
0

where the constant C must be fixed by additional information, for which we

may choose the condition R(m2 , 0) = log(m2 ) that we had already found.
E47 For s < 0 and s > 4m2 , β 2 is positive and we find
 v=1  
2 β+1
<R(m , s) = 2β (log |β + v| − log |β − v|) = β log ;
v=0 |β − 1|
(6.101)
2 2 2 2
for 0 < s < 4m , β < 0 and we use η = −β to arrive at
Z1  v=1
η2
  
2 v 1
<R(m , s) = 2 dv 2 2
= 2η arctan = 2η arctan .
η +v η v=0 η
0
(6.102)
We also find that our constant must read C = log(m2 ) − 2, and so we arrive
at precisely the result found in section 6.6.1. Such methods, by which real
and imaginary parts of amplitudes are related to one another, is called the
technique of dispersion relations.

Checking the optical theorem

As we have seen, for the diagram 6.95 we found its imaginary part to be
~2 λ4 β
=Mf = − . (6.103)
(s − M 2 )2
August 31, 2019 181

As a check, we can look at the lower-order FEE amplitude EE → EE,

restricted to the single diagram : its amplitude is given by

~λ2
Mee = −i : (6.104)
s − M2

its absolute value squared, summed over all final states (that is, the cross
section without the flux factor) is given by

~2 λ4
Z
X 2 1 β 1
|Mee | = |Mee |2 dV (p; q1 , q2 ) Fsymm = (4π) ,
(s − M 2 )2 (2π)2 8 2
(6.105)
and we find precisely
X
=Mf + |Mee |2 = 0 . (6.106)

Note the crucial rôle of the symmetry factor Fsymm !

6.7 Exercises
Excercise 38 Stability implies safety
Consider a process in which two stable particles collide and produce a number
of final-state particles. One possible diagram is

The propagator indicated with ‘p’ reads, of course,

i~
p2 − m2 + i

and this would blow up if p2 = m2 , making the cross section infinite. Show
that this cannot happen.
182 August 31, 2019

Excercise 39 Dirac is the limit

Consider the function
 1
f (x) =
π x2 + 2
Show the following :
R∞
1. f (x) ≥ 0 for all x, and f (x) dx = 1
−∞

2. lim→0 f (x) = 0 if x 6= 0, and lim→0 f (0) = ∞.

This proves that f (x) approaches the Dirac delta function δ(x) as  → 0.

Excercise 40 Brushing up on your quantum mechanics

Consider quantum states that are all normalised to unity.

1. By considering |ci = |ai − |bi hb| ai, prove that | ha| bi |2 ≤ 1, and that
| ha| bi | = 1 implies that |bi = eiφ |ai for some complex phase eiφ .

2. Let G be an operator on states, and let |bi = G |ai. Show that if

ha| bi = 1 for all |ai, then G must be the unit operator.

Excercise 41 Find’em !
In section 6.4.2, it is mentioned that in ϕ3 theory the three-point amplitude
contains 58 diagrams at two loops. Find these diagrams, and show that their
symmetry factors add up to 175/8. Hint : also have a look at appendix
15.2.1.

Excercise 42 At the threshold

Consider the scattering process of section 6.5.4. Evaluate the matrix element
at s ↓ m2 . Compute the cross section directly at the threshold, and recover
Eq.(6.69).

Excercise 43 A real calculation ?

We shall now do another simple calculation, again using the FEE model,
with the mass of the E particles equal to zero, and everything at the tree
level. We make use of the formula for massless two-body phase space:
1
d4 p1 δ(p1 2 ) d4 p2 δ(p2 2 ) δ 4 (P − p1 − p2 ) = dΩ
32π 2
August 31, 2019 183

where Ω is the solid angle in the rest frame of P . There is one vertex :

E F mλ
↔ −i
~
E

Here λ is a dimensionless coupling constant.

1. Compute the decay width Γ(F → EE).

2. Give the three tree-level diagrams for E(q1 )E(q2 ) → E(p1 )E(p2 ).

3. Write the scattering matrix element M. NB 1: The F particle is un-

stable, as we have just proven ! NB 2: do not square it, because that
becomes ugly.

4. Show that, if we write M = iT , then =(T ) is positive, as required by

unitarity.

5. Assume λ to be a quite small number. For s (the centre-of-mass energy

squared) equal to m2 , find the approximate form of |M|2 , and compute
the cross section.

Excercise 44 Yet another one

Compute the total cross section for the process EE → F F in the FEE model,
again assuming the E particles to be massless.

Excercise 45 Infrared divergence

We consider again the FEE theory containing but now with massive E and
massless F. This is a very crude example of QED (Quantum Electrodynam-
ics), the theory of charged spin-1/2 particles and photons.

1. Consider a general process (perhaps involving other particles) in which

an E with momentum p is emitted. Draw its blob.

2. Conider he same process but now with the additional emission of also
an F of momentum k. Consider the diagram where the F comes from
the external E line, and draw it.

3. Show that, if we neglect the i term, the resulting amplitude contains

a factor 1/(p · k) .
184 August 31, 2019

4. Show that the cross section is proportional to 1/(k 0 )2 .

5. Show that the phase space integration element is proportional to k 0 dk 0 .

6. Show that the total cross section for this process of F emission must
be divergent.

Excercise 46 Why is zero equal to zero ?

To understand why dimensional regularization uses Eq.(6.86) can be obtained
a follows.

1. Consider the integral

Z∞
Y = dt t1− tn , n = −2
0

Split the integration region in two, from 0 to a, say, and from a to

∞. Assume now that  always has the ‘right’ sign, that is  < 0 in
0 < t < a, and  > 0 in t > a. Show that Y = 0.

2. Show that Y = 0 can be obtained for other values of n, by choosing 

appropriately.

Excercise 47 Principal value in practice

Let us consider a function f (x) that has a series expansion :

f (x) = f0 + f1 x + f2 x2 + f3 x3 + · · · , f0 6= 0

and has no singularities in an interval a < 0 < b. Show that computing the
principal-value integral
Zb
f (x)
P dx
x
a

is equivalent to using the integration rule (used to obtain Eq.(6.101))

Z
1
dx = log |x|
x
Chapter 7

Dirac particles

7.1 Pimp my propagator

So far we have been studying particles that can carry only a limited amount of
information : such a particle is completely specified once we have determined
its identity and its momentum. In this chapter we shall start increasing the
number of properties that particles can carry, by examining how the Feynman
propagator can be modified. Since the pole structure of the propagator is
closely related to the causality1 of the theory, and must be used to derive
Newton’s first law in the approximation of propagation over macroscopic
distances, we will not mess with the denominator of the propagator. The
generalisations we shall propose therefore involve the numerator, and are of
the form

1 T
i~ → i~ , (7.1)
p2 − m2 + i p2 − m2 + i

where T is some object that informs us that the particle propagating is not
as simple as we have seen so far, but has additional properties. What those
properties are depends, of course, on the choice of T , and in the following
chapters we shall investigate some of the possible choices.

1
Positive energies moving to the future, and all that.

185
186 August 31, 2019

7.1.1 Down with dyads !

The particle propagator never occurs in isolation, but always between two
vertices, where the particle is ‘produced’ and where it is ‘absorbed’2 . This
implies that as long as we have not committed ourselves to particular vertices
a change in the propagator may be compensated to some extent by a change
in the vertices. For instance, suppose that T is a simple scalar3 : then the
predictions of the theory will remain unchanged if we opt to multiply the
vertices by T −1/2 . Therefore, T must be more complicated than a single
number, i.e. it must have some matrix form. In which kind of space such
matrices live is at the moment of course not yet determined, we just assume
that the matrix has some indices. Now, consider the case where the matrix
is actually a dyad, a tensor product of a column ‘vector’ and a row ‘vector’ :
for instance, if the matrix space is three-dimensional, it might read
   
a1   a1 b 1 a1 b 2 a1 b 3
T =  a2  b 1 b 2 b 3 =  a2 b 1 a2 b 2 a2 b 3  .
a3 a3 b 1 a3 b 2 a3 b 3

In such a case, the row ‘vector’ could be assigned to one of the vertices, and
the column ‘vector’ to the other vertex, and the remaining propagator would
again be trivial. We therefore assume that the numerator of the propagator
has a matrix structure that is not a simple dyad, but rather a sum of dyads4 :
X  
 U (n) W (n)  .


T = (7.2)
n

In terms of explicit components this reads

X
a
T ab = U(p)(n) W(p)(n) b ,


(7.3)
n

where a, b are some indices living in some linear space. They may be Lorentz
indices5 , but not necessarily. The sum over n must contain at least two terms.
The vertices of the theory must, of course, contain corresponding indices a, b
2
This also holds true also in the case of external lines, if we keep in mind that they are
defined in the square of the matrix element.
3
A number, or a function of p2 .
4
Any matrix r × s can be written as the sum of at most rs dyads.
5
As in the case of spin-1 particles, see later on.
August 31, 2019 187

with which those of the propagator are contracted, otherwise the matrix ele-
ment could not be a scalar. The objects | Ui are ‘columns’ indicated by upper
indices, and the hW| are ‘rows’ indicated by lower indices. It is tempting to
think of the hW| as ‘hermitean conjugates’ of the | Ui but this is not neces-
sarily true6 .

Going back to the bootstrap argument of section 6.3.2, we see that we

can redo it with the more complicated propagator. Again the denominator
contribution will end up in the phase space, but the numerator will be left.
We can remedy this for instance by assigning the factor Wb to the production
matrix element, and U a to the decay amplitude, with the understanding that
this only holds if the particle is on-shell 7 . A pimped propagator naturally
leads to new Feynman rules for the external lines as well ; and we shall in-
vestigate some of those extensions.

7.1.2 The spin interpretation

As we have seen, particles with generalized propagators will carry factors
| Ui or hW| when they occur as external lines in Feynman diagrams. Such
particles therefore carry, by definition, additional information which is some-
how embodied in the label (n). Adhering to good quantum practice, we shall
assume that particles with different values of (n) are physically distinct from
one another even if their momentum is the same. That is, for p2 = m2 we
require

(n)  (`)  X
a
W (n) a U (`) = K δn,` ,


W  U = (7.4)
a

with K some constant (that is, the external-line factors are (multiples of)
the elements of an orthonormal set). This implies that

T2 ∝ T , (7.5)

6
For Dirac particles, the hW| are the Dirac conjugates of the | Ui. More about this
later.
7
Where the U and the W are assigned depends, of course, on which vertex their re-
spective indices are coupled to.
188 August 31, 2019

In other words, T must have properties of a projecting operator8 . Later on,

we shall see that the ‘additional property’ can be interpreted as the spin
 (n)  of
the particle. Simple counting suggests that the external factors U  are
members of a k/2-spin multiplet if the label (n) runs over 1 + 2k values :
but the more careful treatment is to first see how the | Ui transform under
rotations in the rest frame of pµ , and only then to assign them a spin inter-
pretation9 . We shall do this explicitly for various particle types.

We have noted that the sum over dyads is actually a sum over distinct
additional properties of the particle that propagates. Now, we should like the
different ‘versions’ of the particle to propagate in the same manner, otherwise
by just sitting and waiting we would see the ‘additional’ properties of the
particle change10 . We shall therefore require that T depends only on the
momentum (and possibly on the mass) of the particle : T = T (p, m). The
particles we have studied in the previous chapter, whose propagator has the
trivial numerator T = 1, of course transform trivially (i.e. not at all) under
rotations : such particles are therefore scalars, or spin-0 particles.

7.2 The Dirac algebra

7.2.1 The Dirac matrices
The simplest nontrivial momentum dependence of T is presumably linear.
But we run into trouble immediately : for the momentum carries a Lorentz
index. Now we do not want to contract this index with a corresponding index
in one of the vertices since this would simply amount to a redefinition of the
vertices. On the other hand, we cannot afford to have the Lorentz index
floating loose, which would destroy the Lorentz invariance of the theory. We
therefore choose
T (p) = pµ γµ + K m 1 . (7.6)
8
Here we use the term projecting to indicate that T 2 is proportional to T . A true
projection operator has T 2 = T .
9
This becomes particularly important in the case of massless particles.
10
Suppose an electron with ‘spin-up’ propagates less than an electron with ‘spin-down’ ;
after some time, we would see almost exclusively ‘spin-down’ electrons coming by. This is
not how nature works, apparently ; and that is why we use only a single, `-independent
constant K in Eq.(7.4).
August 31, 2019 189

Here, γµ (µ = 0, 1, 2, 3) is a set of four matrices since as we have argued

the propagator’s numerator must be of matrix form11 . The symbol 1 stands
for the unit matrix of whatever space the γ matrices live in, and the term
Km1 has been added since there is no clear reason to forbid it . Of course,
simply prescribing the γ matrices would again destroy the Lorentz invari-
ance of the theory since any matrix element would have γ’s all over the
place. We therefore demand that in the final form of the matrix element
all reference to the specific choice of these matrices can be removed in a
Lorentz-invariance-respecting manner. That is, the γ matrices must be en-
dowed with a property that allows us to remove them from the final answer.
The momenta with which they are contracted should then end up in ordinary
Minkowski products. That is, there must be a requirement of the form
(algebraic combination of γ µ , γ ν , γ ρ , . . . , γ σ ) = (tensor)µνρ···σ . (7.7)
This requirement had better be as simple as possible, otherwise we might
not be able to eliminate the Dirac matrices from very simple amplitudes.
A moment’s reflection will tell us that essentially the only possible such
property is12
γ µ γ ν + γ ν γ µ = 2g µν 1 . (7.8)
Note that this is a matrix equation: in its full glory it would read
X 
µ a ν c ν a µ c
(γ ) c (γ ) b + (γ ) c (γ ) b = 2 g µν δ a b ,
c

but, as is conventional, we shall not explicitly write out the Dirac indices
unless it is unavoidable. Note also that Eq.(7.8) immediately confirms that
the Dirac objects γ cannot be simple numbers13 . Dirac matrices with different
indices anti-commute, while
(γ 0 )2 = 1 , (γ k )2 = −1 (k = 1, 2, 3) . (7.9)
11
Another argument against the γ’s being simple numbers is that, in that case, they
would define a preferential vector γ µ . This would destroy the assumed isotropy of
Minkowski space, and a frame in which ~γ vanishes would deserve to be equated with
Newton’s absolute reference frame.
12
The anticommutation is necessary because of the symmetry of g µν in its indices.
Another possibility might read something like γ µ γ ν γ α γ β = µναβ 1 but this woud not
allow us to remove fewer than 4 Dirac matrices in any matrix element. The factor 2 in
Eq.(7.8) is simply conventional.
13
Because in that case the fact that g 01 = 0 would imply that γ 0 or γ 1 , or both, vanish:
and that would clash with g 00 = −g 11 = 1.
190 August 31, 2019

We also find immediately that14

γ µ γµ = 4 . (7.10)

From Eq.(7.9) we see that the eigenvalues of γ 0 are either 1 or -1 ; and

those of γ 1,2,3 are either i or −i. We therefore have the following Hermiticity
properties for the Dirac matrices :

(γ 0 )† = γ 0 , (γ k )† = −γ k (k = 1, 2, 3) . (7.11)

For the rest of these notes, the eigenvalues of the Dirac matrices are ac-
tually unimportant. Any choice of Dirac matrices satisfying Eqs.(7.8) is
acceptable ; many have been proposed in the literature15 . That none of
them possesses a physical advantage over the others follows from the ‘funda-
mental theorem of Dirac matrices’ which shows that any two representations
of the Dirac algebra (7.8) can be transformed into each other16 . This again
strengthens our conviction that any result involving Dirac particles should
be derivable without any reference whatsoever to the particular form of the
Dirac matrices, and we shall endeavour to adhere to this. Note that, at this
point, we have not specified the dimensionality of the Dirac matrices. In or-
der to avoid confusion with Lorentz indices, the Dirac indices will be called
spinor indices, and the objects | Ui and hW| for this propagator will be called
spinors. Spinors carry a single spinor index.
Before finishing this section, let us introduce the Feynman slash notation :
if aµ is a Lorentz vector, we shall mean by a / its contraction with Dirac
matrices:
/ ≡ aµ γµ .
a (7.12)

14
At least in precisely four spacetime dimensions. In so-called dimensional regularization
this may change.
15
You may argue that we must show that solutions to Eq.(7.8) actually exist ! That is
correct, and so I am forced to exhibit at least one solution. Here goes, the Pauli-Dirac
representation in block notation :
     
1 0 0 σk 0 1
γ0 = , γk = , γ5 =
0 -1 −σk 0 1 0

where σk are the three Pauli matrices (γ 5 is defined later on). Can we continue now ?
16
We defer the proof of this theorem to Appendix 15.11.
August 31, 2019 191

The Dirac equation (7.8) can therefore also be written as17

a / = 2 (a · b) for all aµ , bν , a
/b/ + b/a /a/ = a2 . (7.13)

The vector object aµ and the matrix a

/ encode exactly the same information ;
further on we shall see how the vector can be recovered once the matrix is
given. A few simple results, which can be checked by repeated application
of the anticommutation rule, are18 E48

γµ a / , γµ a
/ γµ = −2 a / b/ γµ = 4(a · b) , γ µ a
/ b/ c/ γµ = −2 c/ b/ a
/ . (7.14)

7.2.2 The Clifford algebra

By the anticommutation relation (7.8), any product of more than four Dirac
matrices can be reduced to a smaller number. Let us define the enormously
useful object19
γ5 ≡ i γ0 γ1 γ2 γ3 , (7.15)
for which we can immediately derive that

γ 5 γ µ = −γ µ γ 5 , (γ 5 )2 = 1 . (7.16)

Also we can define the commutator of Dirac matrices as E49

i µ ν i
σ µν ≡ [γ , γ ] = (γ µ γ ν − γ ν γ µ ) , (7.17)
2 2
whence
γ α γ β = g αβ − i σ αβ . (7.18)
Obviously there are 6 independent σ matrices. The most general object that E50
can be constructed using Dirac matrices is therefore

Γ = S 1 + Vµ γ µ + Tµν σ µν + Aµ γ 5 γ µ + P γ 5 , (7.19)
17
It is customary to leave the unit matrix 1 out of the notation. Its presence can always
be inferred where necessary.
18
Again, in four spacetime dimensions.
19
In some texts the definition of γ 5 is slightly different, for instance it may lack the
factor i, therefore take care when comparing different texts. The reason why it is called
γ 5 and not γ 4 is that in some older treatments the Minkowski indices were assumed to
run from 1 to 4, with the 4th index playing the rôle of our 0th one.
192 August 31, 2019

and these objects form the Clifford algebra. We see that T (p) must be an ele-
ment of the Clifford algebra. The various coefficients are called, respectively,
the scalar (S), vector (Vµ ), tensor (Tµν ), axial-vector (Aµ ) and pseudo-scalar
(P ) coefficients. Since the tensor coefficient may be taken antisymmetric,
there are in total 1+4+6+4+1=16 coefficients. This suggests (but does not
prove) that the Dirac matrices are 4 × 4 matrices. Given an element Γ in
the Clifford algebra, we can recover its coefficients using the trace identities
that we shall discuss below. Finally, we can define the two handy Clifford
elements
1
1 ± γ5 .


ω± = (7.20)
2
They are mutually exclusive complete projection operators, that is
2
ω± = ω± , ω+ ω− = ω− ω+ = 0 , ω+ + ω− = 1 . (7.21)

7.2.3 Trace identities

In the world of Dirac, an important rôle is played by traces of Dirac matrices
or Clifford elements. To start, we have of course
Tr (1) = N , (7.22)
where N is the (as yet unknown) dimensionality of the Dirac matrices20 .
Using γ 5 and the cyclicity property of the trace operation, we see that
Tr (γ µ ) = Tr γ µ γ 5 γ 5 = Tr γ 5 γ µ γ 5 = −Tr γ µ γ 5 γ 5 = −Tr (γ µ ) (7.23)

so that the trace of a single Dirac matrix vanishes ; and by the same method
we see that the trace of a product of an odd number of Dirac matrices is also
zero, in particular
Tr γ 5 γ µ = 0 .


(7.24)
For two Dirac matrices we have
1
Tr (γ µ γ ν ) = Tr (γ µ γ ν + γ ν γ µ ) = N g µν , (7.25)
2
from which we see that the trace of a σ matrix must vanish21 . To continue,
Tr γ 5 = Tr γ 5 γ 0 γ 0 = Tr γ 0 γ 5 γ 0 = −Tr γ 5 γ 0 γ 0 = −Tr γ 5 , (7.26)

20
We shall prove later on that N = 4.
21
Generally, all commutators are always traceless, by the cyclicity property of traces.
August 31, 2019 193

so that also this trace evaluates to zero. The trace of 4 Dirac matrices requires
a bit more work, anticommuting γ m u to the right results in

Tr γ µ γ ν γ α γ β = Tr 2g µν γ α γ β − 2g µα γ ν γ β + 2g µβ γ ν γ α − γ ν γ α γ β γ µ ,

(7.27)
so that, by cyclicity,

Tr γ µ γ ν γ α γ β = N g µν g αβ − g µα g νβ + g µβ g να ;



(7.28)

and the same method may be used to arrive at the 15 terms for a trace
of 6 Dirac matices, the 105 terms for a trace of 8 matrices, and so on22 .
Furthermore, since the anticommutation operations used in Eq.(7.27) might
as well have moved to the left inside the trace instead of to the right, we
immediately find that23

Tr (γ µ1 γ µ2 γ µ3 · · · γ µn−1 γ µn ) = Tr (γ µn γ µn−1 · · · γ µ3 γ µ2 γ µ1 ) . (7.29)

Since γ 5 is the product of all four different Dirac matrices, the product γ 5 γ µ γ ν
(with µ 6= ν) is actually a product of two different Dirac matrices, and
therefore
Tr γ 5 γ µ γ ν = 0 .


(7.30)
Finally, it is immediately24 seen that

Tr γ 5 γ µ γ ν γ α γ β = iN µναβ .


(7.31)

Returning to the general Clifford algebra element Γ, we can straightforwardly

derive the following results :

Tr (Γ) = N S , Tr (Γ γ µ ) = N V µ , Tr (Γ σ µν ) = 2N T µν ,
Tr Γ γ 5 γ µ = −N Aµ , Tr Γ γ 5 = N P ,



(7.32)

which we can also write as the so-called Fierz identity :

1
Γ = {Tr (Γ) + Tr (Γγµ ) γ µ + Tr (Γσµν ) σ µν /2
N
−Tr Γγ 5 γµ γ 5 γ µ + Tr Γγ 5 γ 5 .



(7.33)
22
It is clear that such trace evaluations are best performed by computer algebra.
23
For even n, we have the proof here ; for odd n it is trivial since 0 = 0.
24
By writing out γ 5 , and noting that of the 105 terms in the trace only one survives.
 194 August 31, 2019

This shows that we can indeed recover all coefficients from a given Γ25 . It
also leads to the following useful insight : if all the above five traces vanish,
then Γ itself must be identically zero. The above method of computing the
Clifford coefficients from the algebra element is also called Fierzing.
A final, important remark : we have shown that the trace identities,
which have been obtained using only Eq.(7.8), evaluate to expressions con-
taining only the metric and the Levi-Civita symbol, which are Lorentz ten-
sors. Therefore, if we can show that all matrix elements (or, at a pinch, their
absolute squares) can be written as traces, we have realized our goal : the
particular representation of the Dirac matrices is irrelevant, and all possible
E51 choices will lead unambiguously to a unique result.
Below, we shall prove that we may take N = 4 : for this choice, we here
summarize the more important trace identities.

Tr (1) = 4
Tr (γ α1 γ α2 · · · γ α2n+1 ) = 0
Tr γ α γ β = 4 g αβ

Tr γ α γ β γ µ γ ν = 4 g αβ g µν − g αµ g βν + g αν g βµ

Tr γ 5 = 0

Tr γ 5 γ α γ β = 0

Tr γ 5 γ α γ β γ µ γ ν = 4i αβµν

Tr (γ α1 γ α2 · · · γ αn ) = Tr (γ αn · · · γ α2 γ α1 ) (7.34)

7.2.4 Pauli and Chisholm, their identities

Consider the two trace results

Tr (/ab/c/γ µ ) = N (ab)cµ − (ac)bµ + (bc)aµ

,
/b/c/γ 5 γ µ = iN µ (a, b, c) .


Tr a (7.35)

From these we immediately see that

/b/c/ = (ab)/c − (ac)/b + (bc)/a − iγ 5 γµ µ (a, b, c) ,

a (7.36)
25
aγ µ ) = N aµ we can indeed find aµ from given /a as promised
In particular, since Tr (/
in sect.7.2.1 once we know what N is.
August 31, 2019 195

which is known as the Pauli identity.

Next, let Γ be a Clifford algebra element that consists of only an odd
number of γ matrices. In that case it has the decomposition

Γ = Vµ γ µ + Aµ γ 5 γ µ . (7.37)

Let us define the reverse ΓR as the result of writing all the Dirac matrices
involved in the reverse order26 . By the reflection property of Eq.(7.29), this
means that
Tr ΓR = Tr (Γ) ,


(7.38)
for all elements of the Clifford algebra. In the present case, we have

ΓR = Vµ γ µ − Aµ γ 5 γ µ . (7.39)

Therefore,
ΓR + Γ = 2Vµ γ µ . (7.40)
We immediately arrive at the so-called Chisholm identity :
N
γµ Tr (Γ γ µ ) = Γ + ΓR .


(7.41)
2
In any computation it is always wise to try and get rid of repeated Lorentz
indices, and the Chisholm identity is one of the tools to do so.

7.2.5 Hermite no, Dirac yes

Any complex linear space has the notion of Hermitean conjugation, but as
we see from Eq.(7.11) objects like a
/ do not conjugate in a simple way. The
Clifford algebra can be endowed with another attractive notion of conjuga-
tion, called Dirac conjugation, which we shall now construct. Denoting the
Dirac conjugation by an over-bar, we require that a/=a / for real aµ , so the
Dirac matrices must be all self-conjugate :

γ µ = γ µ , µ = 0, 1, 2, 3 . (7.42)

We look for a definition of the form

Γ = Ω Γ† Ω−1 (7.43)
26
Note that, fortunately, (γ 5 )R = γ 5 , so that (γ 5 γ µ )R = −γ 5 γ µ .
196 August 31, 2019

for any Clifford element Γ ; such a form ensures the reasonable property
Γ1 Γ2 = Γ2 Γ1 (7.44)
for two Clifford elements. Double conjugation should be equal to the identity :

†
Γ = Γ = Ω Ω−1 Γ Ω† Ω−1 = B −1 Γ B , B = Ω† Ω−1 .


(7.45)
The element B must therefore commute with all Clifford elements, which
implies that B is a multiple of the unit element27 . Without loss of generality
we may therefore take B = 1, so that Ω is Hermitean. The straightforward
choice is therefore Ω = γ 0 , and the Dirac conjugate is defined as
Γ = γ 0 Γ† γ 0 . (7.46)
For a spinor ξ (which carries an upper spinor index) we have
ξ = ξ† γ0 , (7.47)
which is seen to carry a lower spinor index. A conjugate spinor η, which
carries a lower index, obeys
(η) = η , (7.48)
which has an upper index. A spinor sandwich 28 is an object of the form
ηΓξ ,
and it carries no spinor indices as can be seen ; reasonably, we have

∗
η Γ ξ = ξ Γ η = ηΓξ . (7.49)
Further conjugacy properties follow immediately from Eq.(7.42) :
σ µν = σ µν , γ 5 γ µ = γ 5 γ µ , γ 5 = −γ 5 , ω± = ω∓ . (7.50)
In order for a general Clifford element of the form (7.19) to be self-conjugate,
the coefficients S, V µ , T µν and Aµ must be real, and P imaginary.
The standard Dirac spinors which we shall investigate are defined such
that W = U, although as we have already mentioned this is not an unavoid-
able choice to make. Note that the Dirac choice implies that
T (p) = T (p) . (7.51)
27 5 5
From Bγ = γ B, B must
consist of
an even number of Dirac matrices. The require-
ment Tr (Bγ α − γ α B)γ β ∼ Tr Bσ αβ = 0 then shows that B can only contain 1 and
γ 5 , and γ 5 does not commute with γ µ .
28
Named after John Montagu, 4th Earl of Sandwich, PC, FRS (13 November 1718 - 30
April 1792).
August 31, 2019 197

Figure 7.1: A Mental Flip : one vase or two faces ?

7.2.6 Flipping in and flipping out

Consider a spinor sandwich in terms of explicit indices :
X
ηΓξ= (η)a (Γ)a b ξ b . (7.52)
a,b

Once we realise that the individual terms in this double sum are, in fact,
simple numbers, it is clear that we may also write
X
ηΓξ= ξ b (η)a (Γ)a b = Tr (ξ η Γ) , (7.53)
a,b

where ξη is seen as a (dyadic) matrix. This mental flip, whereby we may

suddenly reinterpret the combination spinor-conjugate spinor as a matrix,
frequently comes in very handy indeed ! In the same spirit, there are two
ways to look at the following object :

(η̄1 Γ1 ξ1 ) (η̄2 Γ2 ξ2 ) = η̄1 (Γ1 ξ1 η̄2 Γ2 ) ξ2 (7.54)

On the left-hand side we have two sandwiches, on the right-hand side only
one. This is in particular useful if the dyad ξ1 η̄2 has some simple form.
198 August 31, 2019

7.2.7 A Fierz identity

As an application of what we have learned of the Clifford algebra, we shall
prove the Fierz identity. This deals with the object

F (1, 2, 3, 4) = ξ 1 ω+ γ µ ξ2 ξ 3 ω+ γµ ξ4 , (7.55)

where the ξ’s are arbitrary spinors. Obviously, F (1, 2, 3, 4) = F (3, 4, 1, 2).
Now, as F stands denoted above, it appears to be the (Minkowski) product
of two spinor sandwiches, but we may also (by the ‘mental flip’ mentioned
above) see it as the single sandwich

F (1, 2, 3, 4) = ξ 1 ω+ γ µ ξ2 ξ 3 ω+ γµ ξ4 .


(7.56)

The dyad ξ2 ξ 3 is an element of the Clifford algebra, so

ξ2 ξ 3 = S + Vα γ α + Tαβ σ αβ + Aa γ 5 γ α + P γ 5 . (7.57)

The contraction over the indices µ is then possible :

ω+ γ µ ξ2 ξ 3 ω+ γ µ = ω+ γ µ Vα γ α + Aα γ 5 γ α ω+ γµ

= ω+ γ µ Vα γ α + Aα γ 5 γ α γµ = −2 ω+ (Vα γ α − Aα γ α ) , (7.58)

where we have used the fact that ω+ Γω+ = 0 if Γ contains an odd number of
Dirac matrices29 . We can therefore write

F (1, 2, 3, 4) = −2 ξ 1 ω+ (Vα − Aα ) γ α ξ4 . (7.59)

Now, we also know that, whatever the spinors ξ2 and ξ3 are,

1
1
Vα = Tr ξ2 ξ 3 γα = ξ 3 γα ξ2 ,
N N
1 1
= − Tr ξ2 ξ 3 γ 5 γα = − ξ 3 γ 5 γα ξ2 .


Aα (7.60)
N N
This leads us to the alternative form
 
2
F (1, 2, 3, 4) = − ξ 1 ω+ ξ 3 1 + γ γα ξ2 γ α ξ4
5


N
2 4
= − ξ 1 ω+ γ α ξ4 ξ 3 1 + γ 5 γα ξ2 = − F (1, 4, 3, 2) . (7.61)


N N
29
So that S, T , and P drop out.
August 31, 2019 199

As we have already mentioned, we shall show that N = 4 and the Fierz

identity then becomes

F (1, 2, 3, 4) = −F (1, 4, 3, 2) . (7.62)

In words, the spinors ξ2 and ξ4 may be interchanged at the price of a minus

sign30 .

7.3 Dirac spinors

7.3.1 Chirality spinors
So far we have seen spinors only in the abstract, as the ‘vectors’ in the linear
space on which the Clifford algebra lives ; it is now time to flesh them out a
bit31 . In the first place, we shall define chirality spinors to be those spinors
that are eigenspinors of γ 5 . Since (γ 5 )2 = 1, they must have eigenvalues 1 or
0 under ω± , and for now we write

u+ = ω+ u+ , u− = ω− u− (7.63)

for the positive-chirality and negative-chirality spinors, respectively. Using

the fact that u− u− must be an element of the Clifford algebra and can there-
fore be Fierzed, we have

u− u− = ω− u− u− ω+
= ω− S + Vµ γ µ + Tµν σ µν + Aµ γ 5 γ µ + P γ 5 ω+

= ω− Vµ γ µ + Aµ γ 5 γ µ ω+ = ω− (Vµ − Aµ ) γ µ ,


(7.64)

and similar for u+ . We see that there must exist some four-vector q, called
the spinor’s momentum, such that32

u− u− = ω− q/ . (7.65)
30
This is very suggestive, once we are convinced that the Dirac system describes
fermions. However, the Fierz identity holds only for this particular sandwich, and re-
lies heavily on the presence of the ω± . On the other hand again, it is eminently suited to
resolve a potential problem in the Fermi model of muon decay, that we shall discuss later
on.
31
Considerably, in fact.
32
In this sense, spinors are ‘square roots of vectors’.
200 August 31, 2019

We shall therefore denote the chirality spinors by u− (q) and u+ (q). The
momentum q can be recovered from the spinor components once we know N ,
since
N µ
u± (q) γ µ u± (q) = Tr (u± (q)u± (q)γ µ ) = Tr (ω± q/γ µ ) = q . (7.66)
2
This also tells us immediately that q µ is real-valued and finite, and that
ua± (q)2 (7.67)
X
2q 0 = Tr ω− q/γ 0 = u± (q) γ 0 u± (q) = u†± (q)u± (q) =



so that q 0 is positive 33 .

7.3.2 Chirality spinors are massless

We can obtain more information on chirality-spinor momenta, and on the
dimensionality N of the spinor space as well, by the following interesting
approach. Let us first define the dyadic product

Θ ≡ u− (q)u†− (q) = ω− q/γ 0 . (7.68)

Since γ 0 q/γ 0 = 2q 0 γ 0 − q/ the higher powers of Θ are related :

Θn = θ(n = 1)Θ + θ(n = 2) 2q 0 Θ − q 2 ω− + θ(n ≥ 3) 2q 0 Θn−1 − q 2 Θn−2

(7.69)
for n ≥ 1. This allows us to build a generating function, defined34 for suffi-
ciently small z :
X
f (z) = z n Θn = z Θ − z 2 q 2 ω− + 2q 0 z f (z) − z 2 q 2 f (z)
n≥1

z Θ − z 2 q 2 ω−
= . (7.70)
1 − 2q 0 z + q 2 z 2
Now, Tr (Θ) = N q 0 /2 and Tr (ω− ) = N/2 so that
!
X zN −2q 0 + 2zq 2
Tr (f (z)) = Tr z n Θn = − 0z + q2z2
, (7.71)
n≥1
4 1 − 2q
33
Note that in the sum 7.67 we still have not specified how many indices are to be
summed over.
34
This is always possible since both q 0 and q 2 are finite.
August 31, 2019 201

and by cleverly integrating we find

Zx X xn
!
−1
dz Tr (f (z)) = Tr − Θn = Tr (log (1 − x Θ))
z n≥1
n
0
N
= log(1 − 2q 0 x + q 2 x2 ) . (7.72)
4
The Cayley-Hamilton theorem (discussed in more detail in appendix 15.15.13),

log(det (1 − xΘ)) = Tr (log(1 − xΘ)) , (7.73)

then allows us to write

N/4 Y
det (1 − x Θ) = 1 − 2q 0 x + q 2 x2 = (1 − xλj ) , (7.74)
j

where the product runs over the N eigenvalues λj of Θ. Now, a dyad can
have at most one nonzero eigenvalue35 . The determinant in Eq.(7.74) must
therefore have the form 1 − xλ, where λ is the single nonzero eigenvalue.
Two conclusions fall into our lap : N , the dimension of Dirac space, must
be equal to 4 (and we shall use that everywhere from now on), and q 2 = 0.
We conclude that the momentum of any chirality spinor is finite,
real-valued and massless36 , with positive time component37 .

7.3.3 Phase conventions, spinor products and Weyl

The chirality spinors are defined only up to a complex phase : for actual
calcuations it is often useful to adopt a phase convention. This is of course
somewhat arbitrary ; a particularly useful one will be introduced in sect.8.1.
35
The proof is simple : if |aiha| bi = λ |bi, then either |bi = |ai and λ = ha| ai, or λ = 0.
36
Strictly speaking, this is not the only possible solution. We might take N = 2 and
q · q = (q 0 )2 to find det (1 − x Θ) = 1 − 2xq 0 . That, however, implies ~q = 0 and is therefore
only valid for Dirac particles at rest. Another possibility it to take q 2 = 0 but N = 4k
with k ≥ 2, which results in det (1 − x Θ) = (1 − 2xq 0 )k . There is still a single nonzero
eigenvalue but it is degenerate. In this case we have reducible representations of the Dirac
α
matrices:
 αif γ is the N = 4 form for the Dirac matrix, then the N = 8 case has the block
γ 0
form , and so on. Nothing is to be gained by doing this.
0 γα
37
We might have tried this using Θ = uu rather than Θ = uu† . Unfortunately, under
that choice we would have Θ2 = 0 and all eigenvalues would vanish, which rather spoils
the fun.
202 August 31, 2019

Having adopted a particular phase convention, we can then define a prod-

uct of spinors, in analogy to an inner product of vectors. In these notes, we
shall use the definition
s+ (q1 , q2 ) = u+ (q1 )u− (q2 ) , s− (q1 , q2 ) = u− (q1 )u+ (q2 ) . (7.75)
The other possible products, u+ u+ and u− u− , both vanish38 . We also have
s− (q1 , q2 ) = s+ (q2 , q1 )∗ ,
|s+ (q1 , q2 )|2 = u+ (q1 )u− (q2 ) u− (q2 )u+ (q1 )
= Tr (ω+ q/1 ω− q/2 ) = 2(q1 · q2 ) . (7.76)
These spinor products turn out to be extremely handy in actual calcula-
tions39 .

7.3.4 General spinors and their dyads

Let us now consider a general nonzero spinor η, and see what we can say
about η η̄. In the first place, the previous sections tell us that there must be
two real-valued, massless momenta q1,2 with positive energy, and two complex
phases z1,2 such that
η = ω+ η + ω− η = z1 u+ (q1 ) + z2 u− (q2 ) . (7.77)
The phases z1,2 take into account whatever phase convention for the spinors
we have adopted. We can define the real nonnegative number w by
2w = |s± (q1 , q2 )| , (q1 · q2 ) = 2w2 , (7.78)
38
Since u+ u+ = (u+ ω− )(ω+ u+ ).
39
Here follows a policy statement ! In much of the literature the chirality spinors are
introduced in another way. In the so-called Weyl representation of the Dirac matrices
the matrix γ 5 is diagonal, of the form diag(1,1,-1,-1), so that, say, ω+ =diag(1,1,0,0) and
chirality spinors have only two nonzero components. To distinguish between u+ and u−
authors sometimes go as far as to use ‘dotted’ and ‘undotted’ indices, and a spinor u+ (p) is
then denoted by pȦ or suchlike. Likewise, s± (p, q) is then written as [pq] or hpqi, depending
on the + or -. This approach is, in my view, somewhat misleading in that it suggests that
‘helicity techniques’, as they are called, are somehow tied to the exclusive use of the Weyl
representation. Since I maintain that all results involving Dirac particles ought
to be derivable independent of representation, I endeavour to do that consistently.
The Weyl-representation approach (also called the use of Weyl-van der Waerden spinors)
appears to have the advantage of allowing for momenta that are complex-valued or have
negative energy, but that is indeed only apparent, as sect.8.1.6 shows.
August 31, 2019 203

and two other important vectors :

1 1
k= (q1 + q2 ) , s= (q1 − q2 ) , (7.79)
2w 2w
so that
k·k =1 , s · s = −1 , k·s=0 . (7.80)
All this allows us to write

η η̄ = u+ (q1 )u+ (q1 ) + u− (q2 )u− (q2 )

+ z1 z2∗ u+ (q1 )u− (q2 ) + z1∗ z2 u− (q2 )u+ (q1 )
z1 z2∗ z1∗ z2
= ω+ q/1 + ω− q/2 + ω+ q/1 q/2 + ω− q/2 q/1
s+ (q1 , q2 ) s− (q2 , q1 )
= w k/ + γ 5 s/ + cos(θ) 1 + k/ γ 5 s/ + i sin(θ) γ 5 − k/ s/



,
2w 1
exp(iθ) = z1 z2∗ = η̄ ω+ η . (7.81)
s+ (q1 , q2 ) 2w

Let us now introduce the following matrix :

Σ = Σ(φ) = cosh(φ) + i sinh(φ) γ 5 k/ , (7.82)

so that ΣΣ = 1. The operation ΣΓΣ is therefore a similarity transformation

on the general Clifford element Γ. We find

Σ η η̄ Σ = w cos(θ) 1 + k/ γ 5 s/ + (cosh(2φ) − sinh(2φ) sin(θ)) k/ + γ 5 s/



+i(cosh(2φ) sin(θ) − sinh(2φ)) γ 5 − k/ s/


. (7.83)

We can cleverly choose

1
φ = arctanh(sin(θ)) , (7.84)
2
so that
1 sin(θ)
cosh(2φ) = , sinh(2φ) = , (7.85)
| cos(θ)| | cos(θ)|
and finally arrive at

1 + σγ 5 s/



Σ η η̄ Σ = w| cos(θ)| k/ + σ , σ = sign(cos(θ)) . (7.86)
 204 August 31, 2019

7.3.5 General spinors and Dirac spinors

Let us define a mass m by m = 2w| cos(θ)|, and a momentum pµ by pµ = m k µ
(so that p · p = m2 and p · s = 0). We can then rewrite Eq.(7.86) as follows :

η = Σ u(p, s) for σ = +1 , η = Σ v(p, −s) for σ = −1 , (7.87)

where we have introduced the Dirac spinor and Dirac antispinor by the
definition of their dyads :
1
p/ + m 1 + γ 5 s/ ,



u(p, s) u(p, s) = U(p, s) =
2
1
p/ − m 1 + γ 5 s/ .



v(p, s) v(p, s) = V(p, s) = (7.88)
2
Note that the overall complex phase of the spinors is not determined40 . It
E52 can easily be checked that the U’s and V’s are mutually exclusive projecting
operators, and span the whole Clifford space :

U(p, s)2 = mU(p, s) , U(p, s)U(p, −s) = 0 ,

V(p, s)2 = mV(p, s) , V(p, s)V(p, −s) = 0 ,
U(p, s)V(p, s0 ) = 0 ,
U(p, s) + U(p, −s) − V(p, s) − V(p, −s) = 2m . (7.89)

Noticing that the k/ in Σ can be absorbed into the spinors41 , we finally arrive
at 
 (cosh(φ) + i sinh(φ) γ 5 ) u(p, s) for σ = +1
η= . (7.90)
5
(cosh(φ) − i sinh(φ) γ ) v(p, −s) for σ = −1


In this sense, a Dirac spinor is hidden in a general spinor.

The rewriting that we have done is only sensible if we can actually find
E53 the various vectors and phases ! Fortunately that is possible : from Eq.(7.81)
we see that

η̄ γ µ η = 4w k µ , η̄ γ 5 γ µ η = −4w sµ , η̄ ω+ η = 2w eiθ (7.91)

40
We have already stressed that any result about Dirac matrices should (ultimately)
rely on the relation (7.8) alone. In the same spirit, any result about Dirac spinors should
(ultimately) rely on Eq.(7.88) alone !
41
k u(p, s) = u(p, s) and /k v(p, s) = −v(p, s).
Because /
August 31, 2019 205

which gives us everything we need : w, k µ , sµ , m and σ. Therefore the spinor

u(p, s) (or v(p, −s), as the case may be) is fully determined by the elements
of η alone42 . In particular, |η̄η| = 2m.

7.3.6 Particular general spinors

Our reasoning in the previous two sections will only work as long as w and
cos(θ) are nonzero. What happens in these special cases, for which m (or
η̄η) vanishes ? Let us first concentrate on the case w = 0 : then q1µ and q2µ
must in fact be dependent, and we can write Eq.(7.77) in another form :
η = z1 u+ (q) + az2 u− (q) , (7.92)
with a some real number43 , and q µ is massless as before. Let us now pick
two more vectors, pµ and sµ , with p2 = 0, s2 = −1, and p · s = q · s = 0. The
dyad can be written as
η η̄ = ω+ q/ + a2 ω− q/ + aB ,
z1 z2∗ z1∗ z2
B = ω+ q/p/s/q/ + ω− q/s/p/q/ . (7.93)
u+ (q)/ps/u− (q) u− (q)/sp/u+ (q)
Naturally, the object B claims our attention. Since
|u+ (q)/ps/u− (q)|2 = Tr (ω+ q/p/s/q/s/p/) = 4(p · q)2 , (7.94)
we see that there is a complex phase ψ such that
B = eiψ ω+ s/q/ − e−iψ ω− s/q/ (7.95)
(note the interchange of q/ and s/ in the second term). Now for some nifty
Dirac algebra ! We can introduce yet another vector tµ by44 the Pauli idenitiy
for this situation :
q/p/s/ − s/p/q/ = 2i(pq) γ 5 t/ . (7.96)
42
Always count your degrees of freedom ! The spinor η has 4 complex elements, and
therefore 8 degrees of freedom. The two vectors k and s, with their 3 constraints, contain
5 degrees of freedom. With m and σ we come to 7 degrees of freedom. You may worry
about this until you realise that the missing degree of freedom is the remaining complex
overall phase, which is not determined.
43
We tacitly assume that η 6= ω− η. But if η = ω− η we are finished anyway since then
the u+ (q) term is simply absent.
44
That the right-hand side of Eq.(7.96) is correct can be seen from the fact that the
left-hand side is completely antisymmetric in interchange of q, p and s.
206 August 31, 2019

It is immediately clear that t2 = −1 and that t · p = t · q = t · s = 0.

Multiplying Eq.(7.96) from the right by q/ we then find

(/qp/s/ − s/p/q/)/q = 2(pq)/sq/ = 2i(pq)γ 5 t/q/ , (7.97)

in other words, s/q/ = iγ 5 t/q/. This allows us to rewrite B :

1
B = γ 5 eiψ (i/t + s/) − e−iψ (i/t − s/) q/ = γ 5 s/⊥ q/ ,


(7.98)
2
where
sµ⊥ = cos(ψ) sµ − sin(ψ) tµ . (7.99)
Putting everything together, we find for our dyad45 :
1
1 + a5 γ 5 + a⊥ γ 5 s/⊥ k/ ,


η η̄ =
2
1 − a2 2a
k µ = (1 + a2 )q µ , a5 = 2
, a⊥ = . (7.100)
1+a 1 + a2
As before, we have to verify that k µ , sµ , a5 and a⊥ can be found from the
elements of η. We have

η̄ g µ η = 2k µ , η̄ γ 5 γ µ η = −2a5 k µ , (7.101)
p
which gives us k µ and a5 , and also a⊥ = 1 − a25 : we can always take
a⊥ > 0 without loss of generality. To find out about sµ⊥ , we pick any vector
rµ (except r = k) and construct

1 (s⊥ · r) α
yα = η̄ γ 5 r/γ α η = sα⊥ − k . (7.102)
2a⊥ (k · r) (k · r)

We have some freedom in determining s⊥ since we can always replace it by

s0⊥ = s⊥ + x k, with x arbitrary : after all, s0⊥ · s0⊥ = −1 and k · s0⊥ = 0 as well,
and the extra term does not change the dyad. We can then, for instance,
insist that s0⊥ = 0, and find s⊥ via

y0 µ
sµ⊥ = y µ − k . (7.103)
k0
45
This is a curious-looking expression, that you might not think to be dyad at all at
first sight. The secret is that a25 + a2⊥ = 1.
August 31, 2019 207

The remaining particular case is that of w 6= 0, cos(θ) → 0. Then the

matrix Σ(φ) is divergent, hence not so useful. Our dyad now reads

η η̄ = w k/ + γ 5 s/ + iλ γ 5 − k/ s/


, λ = sin(θ) = ±1 . (7.104)

For this case there exists another similarity transform :

1
Σ0 = Σ0 (ρ) = √ (1 + iρ/k ) , ρ = ±1 , (7.105)
2
that gives us
0
Σ η η̄ Σ0 = w 1 − λργ 5



k/ − λρ/s . (7.106)
The result is therefore surprisingly simple :

 (1 + iρ/k ) u− (q2 ) for λρ = 1
η= . (7.107)
(1 + iρ/k ) u+ q1 ) for λρ = −1


We can actually choose the one we like best ! The phases z1,2 of Eq.(7.77)
and the spinor product s+ (q1 , q2 ) have conspired to make this possible.

It has to be noted (see sect.7.4.6) that the forms (7.100) and (7.107) are
also dyads of Dirac spinors, taken in the limit m → 0. The conclusion of
all these exercises is the following : a general spinor can always be brought
in the form of a Dirac spinor, using similarity transformations that depend
only on k. This will help us in choosing a propagator later on.

7.4 Dirac particles

The stage is now set to apply all the above mathematical machinery to choose
a propagator for Dirac particles, and find out about external lines for these
particles.

7.4.1 Casimir does tricks

In the last section we saw that u-spinors with the same momentum p and
opposite spin vectors are orthogonal. Could there be other spin vector choices
also yielding an orthogonal state ? To this end we can consider u(p, s)u(p, s0 )
208 August 31, 2019

where sµ and s0µ are spin vectors. If the spinors refer to orthogonal quantum
states, then the absolute square of their product must vanish. We shall now
compute this exactly, by turning the product into a trace using the so-called
Casimir trick, an example of the ‘mental flip’ :
X
a
b
|u(p, s)u(p, s0 )|2 = u(p, s) a u(p, s0 ) u(p, s0 ) b u(p, s)



a,b
X
b
a
u(p, s) a u(p, s0 ) u(p, s0 ) b



= u(p, s)
a,b
X
b
a
= u(p, s)u(p, s) a u(p, s0 )u(p, s0 ) b
a,b

= Tr u(p, s)u(p, s) u(p, s0 )u(p, s0 ) .

(7.108)
For any correctly constructed amplitude involving Dirac particles, its abso-
lute square is always amenable to the Casimir trick : traditionally, therefore,
the evaluation of such amplitudes is done in this way46 . This establishes
the last requirement for the uniqueness (up to a phase) of matrix elements
involving Dirac particles (cf. section 7.2.3). We can evaluate the trace by
standard operations. For didactic purposes we give them here in excruciating
detail :
Tr u(p, s)u(p, s) u(p, s0 )u(p, s0 ) =

1
Tr (/p + m)(1 + γ 5 s/)(/p + m)(1 + γ 5 s/0 )


=
4
1
Tr (/p + m)2 (1 + γ 5 s/)(1 + γ 5 s/0 )


=
4
m
Tr (/p + m)(1 + γ 5 s/)(1 + γ 5 s/0 )


=
2
m
Tr p/ + m + p/γ 5 s/ + mγ 5 s/ + p/γ 5 s/0 + mγ 5 s/0 + p/γ 5 s/γ 5 s/0 + mγ 5 s/γ 5 s/0


=
2
m
m
Tr m + mγ 5 s/γ 5 s/0 = Tr (m − m/ss/0 ) = 2m2 1 − (s · s0 ) (. 7.109)


=
2 2
46
Note that there is a price: the length of the expressions is doubled by the squaring, and
if the amplitude contains many diagrams the algebra can become very cumbersome indeed.
A lot of computational shortcuts have been proposed, the most useful of which appears to
be not to bother with squaring at all but rather to evaluate the spinor products themselves
directly as complex numbers, by so-called spinor techniques. On the other hand, the
existence of the Casimir trick ensures that, as required, one can completely get rid of
the Dirac matrices in the prediction of cross sections using only their anticommutation
properties.
August 31, 2019 209

Note that only two out of the eight terms contain combinations of Dirac
matrices that survive the trace. Since we can work in the pµ rest frame,
where the spin vectors must be spatial unit vectors, we conclude that, in
that frame
|u(p, s)u(p, s0 )|2 = 2m2 1 + ~s · ~s0 .


(7.110)
The states are only strictly orthogonal if ~s0 = −~s. E54
We can use the Casimir trick also to simply prove some properties of
E55
Dirac (anti)spinors :
E56
u(p, s)u(p, −s) = 0 , u(p, s)v(p, s0 ) = 0 ,
u(p, s)u(p, s) = 2m , v(p, s)v(p, s) = −2m , E57
u(p, s)γ µ u(p, s) = 2pµ , v(p, s)γ µ v(p, s) = 2pµ ,
u(p, s)γ 5 γ µ u(p, s) = −2msµ , v(p, s)γ 5 γ µ v(p, s) = 2msµ ,
u(p, s)γ 5 u(p, s) = 0 , v(p, s)γ 5 v(p, s) = 0 ,
u(p, s)γ 5 v(p, s) = 0 , |u(p, s)γ 5 v(p, −s)| = 2m . (7.111)

7.4.2 The Dirac propagator, and a convention

We have now arrived at a possible choice for the Dirac propagator. In the
first place, we can realize that since we can turn any spinor into a Dirac
spinor by similarity transformations, that can be absorbed in the vertices47 .
Therefore, we shall only consider Dirac spinors48 .
We have started with assuming that T (p) = p/ + Km. Also, we have
argued that on the mass shell this should be a projecting operator, in other
words
K2 + 1
 
2 2 2
T (p) = p/p/ + 2mK/p + m K = 2mK p/ + m = 2mK T (p) ;
2K
(7.112)
hence K = 1 or −1. By universal convention we adopt

T (p) = p/ + m , (7.113)
47
Here it is important that the similarity transformations depend only on k, i.e. the
momentum, discussed in sect.7.3.4 Vertices are perfectly allowed to depend on the momen-
tum of a leg (these are called derivative couplings), but we would surely run into trouble
if they would also depend on, say, s.
48
Note that this implies that a funny, non-Dirac propagator can always be mimicked by
a Dirac propagator, where the particle has peculiar interactions !
210 August 31, 2019

and adopt this choice also off the mass shell (where it is actually used). The
Dirac propagator therefore takes the form
p/ + m
i~ .
p2 − m2 + i
The fact that the numerator is linear in p means that the propagator is ori-
ented, in contrast to what we have used so far. To indicate this we define the
orientation with an arrow, and adhere to the convention that the momentum
is counted in the direction of the arrow, irrespective of the sign of the energy
component. The first Dirac Feynman rule therefore becomes

k/ + m
↔ i~
k k · k − m2 + i
Feynman rules, version 7.1
(7.114)
In writing out Feynman diagrams containing Dirac particles, we of course
have to keep track of the Dirac indices resident in propagator and vertices.
This may lead to incredibly cumbersome notation, that may however be
greatly simplified if we adopt the following writing convention : write out
the Dirac-index carrying factors in order, moving against the ori-
entation of the line. Then, all these factors are contracted together using
the usual rules for matrix multiplication, and one hardly ever needs to write
the Dirac indices explicitly. This convention is really to be urged on anyone
contemplating any calculation involving Dirac particles49 !
A final word on notation : since

p/ + m p/ − m = p2 − m2 ,



(7.115)

the Dirac propagator might be written as

p/ + m i~
i~ = . (7.116)
p2 2
− m + i p/ − m + i
In instances were the i can be neglected, this is certainly allowed ; however in
more delicate situations (such as inside loops) the first alternative is probably
to be preferred. Nevertheless we shall occasionally also use Eq.(7.116).
49
Try it out for yourself ; after ten minutes (max) you will be convinced.
August 31, 2019 211

7.4.3 Truncating Dirac particles : external Dirac lines

Let us now return to the truncation argument that gave us the Feynman rule
for external lines in chapter 6. We shall redo this for Dirac particles moving
between production and decay. As a first case, let the ‘p’-line connecting
production and decay be oriented from production to decay, as indicated in
the following diagram :

pa }k 1,2,...

M= A
p B }q 1,2,...
(7.117)
pb
According to the convention described above we then have for the amplitude
i~(/p + m)
M=B A . (7.118)
p2 − m2 + imΓ
Note that, in this amplitude, the factor A must carry the upper Dirac index
of a spinor, and B the lower index of a conjugate spinor. Obviously, pµ
carries positive energy. As we let Γ vanish and pµ approaches the mass shell,
we may then write X
p/ + m = u(p, s) u(p, s) , (7.119)
s
where the sum over s runs over two values, sµ and −sµ . Following the
truncation argument, we readily see that the spinor u(p, s) must then be
included in the decay amplitude, and u(p, s) in the production amplitude.
In the alternative case, where the line is oriented against the flow of
energy, the amplitude is given by

pa }k 1,2,...

M= A
p B }q 1,2,...
(7.120)
pb
and reads (again with our convention !)
i~(−/p + m)
M=A B . (7.121)
p2 − m2 + imΓ
212 August 31, 2019

Note that it is now A that is the conjugate spinor, and B the regular one.
Of course, they describe a physical process different from the first case ! We
are now forced by the negativity of the energy to write
X
−/p + m = − v(p, s) v(p, s) . (7.122)
s

The sign flip in the projecting operator is of course precisely that which
turns a particle description (with negative energy, moving backwards in time
along the orientation of the propagator) into the antiparticle description, with
positive energy. The truncation argument then tells us that v(p, s) must be
the factor associated with the production, and v(p, s) must be associated with
the annihilation, of the antiparticle. There remains the question of where to
put the left-over Fermi minus sign. Consistently, we may decide to keep it
with the v, in which case we arrive at the following Dirac Feynman rules :

k/ + m
↔ i~
k k · k − m2 + i
√
↔ ~ u(p, s)
p,s
√
↔ ~ u(p, s)
p,s
√
↔ ~ v(p, s)
p,s
√
↔ − ~ v(p, s)
p,s
Feynman rules, version 7.2
(7.123)
The awkward-looking minus sign is usually subjected to the argument that
any matrix element containing an incoming antiparticle will have the factor
−v in each of its diagrams, and since we are interested in absolute values
squared anyway, there would appear to be little harm in deleting this overall
minus sign from the Feynman rules : and this is what is commonly done. A
little reflexion, though, will remind us that the sign of the amplitude’s real
part is fixed by unitarity, and now we have changed it ! Clearly, the minus
sign will be back to haunt us later on50 .
50
The minus sign becomes expecially urgent when we use amplitudes to describe inter-
action potentials, see section 10.5.
August 31, 2019 213

7.4.4 Lorentz transformations in Dirac space

We have to determine the spin of Dirac particles. Although the fact that they
have two orthonormal spin states strongly suggests that they have spin-1/2,
a real proof must rest on the way they form a representation of the rotation
group. The rotation group is, of course, a subgroup of the Lorentz group.
Now, we have argued that the vector pµ and the matrix p/ contain exactly the
same information, for any vector pµ . Therefore, we must be able to find how
p/ transforms under a Lorentz transformation. Let us define by Λ(p; q) the
minimal Lorentz transformation51 , that is it makes pµ go over in q µ while
keeping any vector rµ unchanged for which p · r = q · r = 0. Rotations are
an example : in that case p0 = q 0 = 0, |~p| = |~q|, and ~r · p~ = ~r · ~q = 0. Since p/
is a matrix, the effect of a Lorentz transformation must be represented by a
matrix transformation, that is

Λ(p; q) : p/ → Σ1 p/ Σ2 . (7.124)

Since we must ensure that Dirac conjugation commutes with Lorentz trans-
formation, we must have Σ2 = Σ1 ; and in order to have matrix multiplication
commute with Lorentz transformations as well52 we must have Σ2 Σ1 = 1. We
conlude that
Λ(p; q) : p/ → Σ p/ Σ , Σ Σ = 1 . (7.125)
The explicit form of Σ, the minimal Lorentz transformation reads53

p2
 
q/p/
Σ=C 1+ 2 , |C|2 = . (7.126)
p (p + q)2

You can simply check that this is indeed correct :

2(pq) p2 q 2
   
2 q/p/ + p/q/ q/p/p/q/ 2
Σ Σ = |C| 1 + + 4 = |C| 1 + 2 + 4 =1 ,
p2 p p p
(7.127)
51
Not to be confused with an infinitesimal Lorentz transformation. A product of two
minimal transformations is, in general, not itself minimal.
52
So that we can either first mutiply /p1 and /p2 , and then Lorentz-transform them, or
do the Lorentz transform first and the multiplication afterwards.
53
This form tacitly assumes that under minimal Lorentz transforms the sign of p2 and
(p + q)2 are the same. If p2 < 0 this is not always true ; however, for boosts and spatial
rotations it does hold.
214 August 31, 2019

and
   
2 q/p/p/ + p/p/q/ q/p/p/p/q/ 2 q/p/q/
Σ p/Σ = |C| p/ + + = |C| p/ + 2/q + 2 = q/ ,
p2 p4 p
(7.128)
where we have used the anticommutation result q/p/q/ = 2(pq)/q − p/q 2 . The
other requirements, Σ/qΣ = p/ and Σ/rΣ = r/, are proven trivially. For general
Clifford elements Γ, we have now also ensured that

Γ → ΣΓΣ (7.129)

under (minimal) Lorentz transformations. It is somewhat surprising to see

that the form of this Lorentz transformation in Clifford space is quite simple.
Since all spinorial dyads ξη are Clifford elements, we find from the above that
the transformation rules are

ξ → Σξ , ξ → ξΣ . (7.130)

7.4.5 The spin of Dirac particles

Let us now select the spinor of a particle in its rest frame, and consider
rotations of the space axes. By tµ , xµ , y µ and z µ we shall mean the four-
dimensional extensions of the spatial unit vectors in the zeroeth (time), x-,
y- and z-directions, respectively. A rotation Σz (θ) over an angle θ from x
towards y around the z axis54 is then determined by choosing the rotation

xµ → x̃µ = cos(θ)xµ + sin(θ)y µ , (7.131)

so that (x + x̃)2 = −4 cos(θ/2)2 and x/

/̃ x = − cos(θ) − sin(θ)/xy/. The rotation
matrix is therefore

Σz (θ) = cos(θ/2) + sin(θ/2) x

/ y/ . (7.132)

The generators of the rotation group must therefore be

Tx = β/yz/ , Ty = β/zx
/ , Tz = β/xy/ , (7.133)
54
Here the confusing active-passive distinction rears its ugly head. We shall not worry
about it since the rotation algebra is the same in each case.
August 31, 2019 215

where we have used cyclicity, but not specified the constant β. This constant
can be determined from the rotation group algebra requirement:

[Tx , Ty ] = Tx Ty − Ty Tx = i~Tz , (7.134)

which for the Dirac system is seen to read

[Tx , Ty ] = β 2 y/z/z/x / y/z/ = 2β 2 x

/ − z/x / y/ = 2β Tz , (7.135)

from which we see that β = i~/2. Noticing also that55

~2
Tz 2 = β 2 x / y/ = −β 2 x2 y 2 =
/ y/x = Tx 2 = Ty 2 , (7.136)
4
we conclude that the total-spin operator comes to
 
2 2 2 1 1
Tx + Ty + Tz = + 1 ~2 . (7.137)
2 2

The spinors are, therefore, representatives of a spin-1/2 system.

An observation is in order here. If we rotate the vector xµ towards y µ but

simply keep on rotating, we shall get xµ back after 360 degrees ; but Σz (2π)
is minus unity. To restore all things to their original forms, we must turn
around twice, not once !

7.4.6 Massless Dirac particles ; helicity states

In the projecting operators u(p, s)u(p, s) and v(p, s)v(p, s) as we have defined
them, the limit m → 0 appears unproblematic. There is, however, a subtlety.
Let us take a Dirac particle with definite helicity : in that case, the spin vector
is parallel to the direction of motion56 . Let us write p~ = p~e, with ~e the unit
vector in the direction of p~. The requirements s2 = −1, (ps) = 0 determine
that
1
pµ = p0 , p ~e , sµk = p, p0 ~e .



(7.138)
m
55
The fact that the square of any of the generators is proportional to the unit matrix is
more or less a coincidence ; for systems with higher spins it no longer holds.
56
This is, obviously, not a Lorentz-invariant notion. As the particle’s velocity approaches
c, however, it becomes Lorentz-invariant.
216 August 31, 2019

As m → 0, the spin vector diverges, and the massless limit is not so obvious.
We may, however, write for this case
1 0
p0 − p 1
sµk = 1, −~e = pµ + O m/p0 ,



p , p ~e + (7.139)
m m m
since (p0 −p)/m = m/(p0 +p). The projecting operator can then be evaluated
by
1
1 + γ 5 s/k (/p + m)


u(p, s)u(p, s) =
2 
1 1 5 0


= 1 + γ p/ + O m/p (/p + m)
2 m
1
= (1 + γ 5 )(/p + m) + O (m)
2
≈ ω+ p/ , (7.140)

and we recover a chirality state. Of course, for sµ antiparallel to the velocity,

we find
u(p, s)u(p, s) ≈ ω− p/ . (7.141)
The massless limit therefore reads

lim u(p, ±sk ) = u± (p) , lim v(p, ±sk ) = u∓ (p) . (7.142)

m→0 m→0

But states without pure helicity are also possible : we can consider the case
where p~ and ~s make a fixed angle θ. In that case the spin vector reads
m cos θ sµk + p0 sin θ sµ⊥
µ
s = p , sµ⊥ = (0, ~e⊥ ) , (7.143)
(p0 )2 − p2 cos2 θ
with ~e · ~e⊥ = 0. If we now let m → 0 so that p → p0 , while keeping φ fixed,
then the limit of the projecting operator becomes
1
u(p, s)u(p, s) ≈ (1 + γ 5 s/⊥ )/p , (7.144)
2
and we see that this limit is indistinguishable from a masless, transversely
polarised Dirac particle. We may also decide to let θ also approach zero with
θ/m fixed, and then we will end up with
1
1 + PL γ 5 + P⊥ γ 5 s/⊥ p/ ,


u(p, s)u(p, s) ≈ (7.145)
2
August 31, 2019 217

with PL2 + P⊥2 = 1 : precisely the special case discussed in sect.7.3.6 ! The
message is that the massless limit is always defined, but must be taken with
some care57 .

7.4.7 The parity transform

An interesting excercise is the following. Let ξ be an arbitrary spinor. The
object
u(p, s) = C (/p + m)(1 + γ 5 s/)ξ (7.146)
is then exactly the spinor for a Dirac particle with momentum pµ and spin
vector sµ , provided that C is chosen appropriately58 . Now, let us consider

γ 0 u(p, s) = C γ 0 (/p + m)(1 + γ 5 s/)ξ . (7.147)

By anticommuting the γ 0 to the right, we can arrive at

γ 0 u(p, s) = C (p/̂ + m)(1 + γ 5 s)

/̂ γ 0 ξ . (7.148)

Here, the vectors with and without hats are related as follows :

pµ = (p0 , p~ ) , p̂µ = (p0 , −~p ) , sµ = (s0 , ~s ) , ŝµ = (−s0 , ~s ) . (7.149)

Since γ 0 ξ is also an arbitrary spinor, the object γ 0 u(p, s) is exactly59 the

spinor u(p̂, ŝ) for a Dirac particle with momentum p̂µ and spin vector ŝµ .
What state is this, precisely ? The spatial momentum of the particle has
been reversed : this is called the parity transform. The spin vector, however,
retains its spatial part while its time-part has now been flipped. The spin
vector is, therefore, a four-vector of a different type from the more regular
vector pµ : such four-vectors are called axial vectors60 . We conclude that
multiplying a spinor by γ 0 induces its parity transform. For antiparticles the
treatment is completely identical.
57
It is also clear that to produce, say, beams of ultrahigh-energy electrons with given
helicity, one needs to be able to align the spin vector very precisely with the momentum, to
angles of order m/p0 . Nevertheless, this is feasible in practice, as the LEP/SLC colliders
have proven.
58
This idea lies at the basis of the spinor techniques, to be discussed below.
59
You may worry about the normalisation factor C, which is determined by insisting
that u(p, s)u(p, s) = 2m. Inserting the γ 0 we see that the normalisation condition now
becomes (u(p, s)γ 0 )(γ 0 u(p, s)) = 2m so that the normalisation remains correct.
60
This explains the term ‘axial-vector’ coefficient we used in the Clifford algebra.
218 August 31, 2019

7.5 The Feynman rules for Dirac particles

7.5.1 Dirac loops. . .
As mentioned above, there is a natural tendency in formulating the Feynman
rules to leave out the Fermi minus sign in the rules for external particles. Let
us suppose that we choose to do that. Now, consider the following cutting
rule :

q q q

+ + = 0 (7.150)
p p p

Here, a scalar particle has a three-point coupling to a pair of Dirac particles61 .

We shall not evaluate the whole diagram, but rather concentrate on the two
Dirac propagators. In the third, cut-through diagram, they occur as external
lines, giving rise to a factor (viz. section 6.4.2)

u(p)Γ1 v(q) v(q)Γ2 u(p) ,

where Γ1,2 represent the rest of the diagrams. The momenta p and q are
assumed to run from left to right. We have not indicated the spins since
anyway we have to sum over them. Therefore we would have to evaluate the
trace  
Tr (/p + mp )Γ1 (/q − mq )Γ2 ,

where we have indicated that the two Dirac particles are not necessarily of
the same type. Let us now shift our attention to the first diagram, say. A
closed loop of Dirac particles is automatically also a trace: this diagram,
then, requires the analogous trace
 
Tr (/p + mp )Γ1 (−/q + mq )Γ2 ,

61
The requirement that amplitudes do not contain uncontracted indices essentially forces
us to use Feynman rules in which the orientation of Dirac lines is conserved at every vertex.
For so-called Majorana fermions this is not true : Majorana fermions, therefore, have no
distinction between particle and antiparticle.
August 31, 2019 219

since the momentum q is running against the orientation62 . The second

trace has the opposite sign of the first one ! To solve this problem (and save
unitarity of the S matrix !) we therefore have to introduce an additional
Feynman rule for Dirac particles :

k/ + m
↔ i~
k k · k − m2 + i
√
↔ ~ u(p, s)
p,s
√
↔ ~ u(p, s)
p,s
√
↔ ~ v(p, s)
p,s
√
↔ ~ v(p, s)
p,s
-1 for every closed Dirac loop
Feynman rules, version 7.3
(7.151)

7.5.2 . . . and Dirac loops only

In the above we have not yet explained why the minus sign must be assigned
only to those closed loops that contain only Dirac particles. The reason for
this is crossing symmetry. Consider a (cut) diagram like this one

where the lines without arrows have no Dirac but scalar propagators. The
cut crosses two Dirac lines, and we might conclude that a minus sign is called
62
We disregard the denominators of the Dirac propagators since they do not influence
our argument.
220 August 31, 2019

for. However, by crossing symmetry this diagram is related to

where now the cut crosses one Dirac line and one line without an arrow.
Since the propagator in that line is even in its momentum, we can always
choose the loop momentum to run in the ‘correct’ direction for the Dirac
line, and no minus sign is needed. Therefore, the first diagram also takes no
extra minus sign, since crossing symmetry forbids an amplitude to suddenly
pick up an extra minus sign under crossing. It is only when a closed loop
consists of only Dirac particles that no crossing can be found for which the
loop momentum can be chosen to run in the ‘correct’ direction. Therefore,
only for such loops is a minus sign unavoidable63 .

7.5.3 Interchange signs

Consider the following two diagrams, that can both contribute to the decay
of a scalar into a Dirac-antiDirac pair at the one-loop level,

where the first diagram contains a fermion loop and hence carries an overall
minus sign ; the second one does not. Now consider the cut versions of these
diagrams :

and notice that the left-hand sides of the cut-through diagrams are identi-
cal. The right-hand sides differ in the way that the in-going fermions are
connected to the out-going ones ; the ingoing ones are interchanged in the
second diagram with respect to the first one. This, then, must correspond to
63
This holds true later on, where we also introduce vector particles, the propagator of
which is also even in the momentum.
August 31, 2019 221

a minus sign associated with the interchange of external lines in a diagram,

and we arrive at the final form of the Feynman rules for Dirac particles :

k/ + m
↔ i~
k k · k − m2 + i
√
↔ ~ u(p, s)
p,s
√
↔ ~ u(p, s)
p,s
√
↔ ~ v(p, s)
p,s
√
↔ ~ v(p, s)
p,s
-1 for every closed Dirac loop
-1 for every Dirac particle interchange
Feynman rules, version 7.4
(7.152)

Note that the interchange rule only determines the relative sign between two
Feynman diagrams. How the interchange sign can be determined is best
illustrated by an example. Consider, for instance, a process with 6 external
fermions. Three of them must then be oriented outward from the diagram,
carrying a u or v, and the other three must be oriented inward and carry a u
or a v. Let us assume that there are three Feynman diagrams, schematically
given by64

diagram 1 : u1 Γ1 u2 v 3 Γ2 u4 u5 Γ3 v6
diagram 2 : u1 Γ4 u2 v 3 Γ5 v6 u5 Γ6 u4
diagram 3 : u1 Γ7 u4 v 3 Γ8 v6 u5 Γ9 u2 ,

Clearly, we have left out an enormous amount of detail here, and the Γ’s
can be anything. Note that we have written the three diagrams in such a
way that the conjugate spinors u1 , v 3 and u5 are in the same order in each
diagram : this is always possible. Now, we see that to go from diagram 1 to
diagram 2, the positions of u4 and v6 must be interhanged, whereas one can
64
The process e− e− e+ → e− e− e+ is an example.
 222 August 31, 2019

go from diagram 1 to diagram 3 by, say, interchanging first u2 and u4 , and

then u2 and v6 . Therefore, diagram 1 and 3 have no relative minus sign, and
diagram 2 has a minus sign with respect to 1 and 3. In actual practice, the
determination of the relative signs can be made even easier ; simply decide
on some preferred ordering of all your u’s, v’s, u’s and v’s , and compare
the ordering in your given diagram with your preferred one. Note that, since
spinor sandwiches always contain two spinors, spinor sandwiches may be in-
terchanged at will without destroying this simple rule.

Before finishing this section we want to make an important observation.

The loop and interchange minus signs as we have discussed them depend on
E59 the structure of the diagrams, and not on the type of the Dirac particles ;
even if a neutrino and a top quark were interchanged, the minus sign would
E60
crop up65 . The minus signs depend only on the fact that they are Dirac
particles, that is, spin-1/2 fermions. No notion of ‘indistinguishability’ is
relevant here.

7.5.4 The Pauli principle

Let us consider a possible experiment in which we attempt to produce two
Dirac particles of the same type (two electrons, say), with exactly the same
momentum and spin. Any such process is, in principle, described by Feynman
diagrams. We can say immediately that the number of diagrams must be
even, since for every diagram there must be a corresponding one in which the
two electrons are interchanged. Now, if the momenta and the spins of the two
electrons are precisely the same, they will be described by identical conjugate
spinors, and in fact the two diagrams of the pair will have exactly the same
value — apart from the relative minus sign ! The total amplitude is therefore
identically zero. We conclude that it is not possible to produce two Dirac
particles in exactly the same state. By considering incoming electrons, we
can also conclude that it is not possible to observe two Dirac particles if they
are in exactly the same state, since the observation process is also describable
(presumaby !) by Feynman diagrams. This is the Pauli exclusion principle66 .
65
Of course, the interactions in the theory may be such that no such interchange is
possible : but this is beside the point.
66
Note that I do not comment on the possibility that electrons in identical states might
simply exist : they would not be observable by any process describable by Feynman di-
agrams. Such a Fermi-Dirac condensate would be the most stable and invisible form
August 31, 2019 223

7.6 The Dirac equation

7.6.1 The classical limit
So far we have not mentioned the Dirac equation, nor have we had need for
it. As an illustration, we shall show how it can be obtained. To this end, we
need to provide a few Feynman rules in position, rather than in momentum
space. The Dirac propagator, oriented from spacetime point x to spacetime
point y, is
Z
i~ k/ + m
↔ 4
d4 k e−ik·(y−x) 2 , (7.153)
x y (2π) k − m2
where we have dropped the i for simplicity. The Dirac particles are created
by a spinorial source J(x), and absorbed by a conjugate-spinorial source
¯
J(x), with the rules
i
↔
− J(x) ,
~
i ¯
↔ − J(x) . (7.154)
~
If we forget about any other couplings, the Dirac field is free, and its SDe is
exactly its own classical limit. Now, consider the following form of it :

x
x = . (7.155)

With the field function of the Dirac field denoted by ψ(x), this SDe reads
Z
1 k/ + m
ψ(x) = 4
d4 y d4 k e−ik·(x−y) 2 J(y) , (7.156)
(2π) k − m2
where matrix multiplication is implied as usual. We can now apply a well-
chosen differential operator :
Z
1 k/ + m
(i/∂ − m) ψ(x) = 4
d4 y d4 k e−ik·(x−y) (/k − m) 2 J(y)
(2π) k − m2
of matter. Its only connection to the rest of the universe would arise through some
non-diagrammatic process, involving possibly gravity since that appears not to be very
amenable to diagrammatics. Of course, classical quantum mechanics finds that the com-
bined wave function for identical-state electrons vanishes identically, but again quantum
and gravity do not see completely eye to eye.
224 August 31, 2019
Z
1
= 4
d4 y d4 k e−ik·(x−y) J(y)
(2π)
Z
= d4 y δ 4 (x − y) J(y) = J(x) , (7.157)

which is the classical Dirac equation :

(i/∂ − m) ψ(x) = J(x) . (7.158)

We can also consider the ‘Dirac-conjugate’ SDe :

x
x = , (7.159)

which is written as
Z
1 ¯ k/ + m −ik·(y−x)
ψ̄(x) = d4 y d4 k J(y) e . (7.160)
(2π)4 k 2 − m2

By the same simple manipulation as above, we can then show that the con-
jugate Dirac equation reads
←
ψ̄(x)(−i ∂ ¯
/ −m) = J(x) , (7.161)

where the leftward arrow indicates that the derivative must be taken towards
the left 67 .

7.6.2 The free Dirac action

We can cast the above in the form of the – possibly more familiar – La-
grangian treatment. The action for the free Dirac field including sources is
then given by Z
¯
S[ψ, ψ̄, J, J] = d4 x L(x) , (7.162)

where the Dirac Lagrangian is given by

¯
L(x) = ψ̄(x) (i/∂ − m) ψ(x) − J(x)ψ(x) − ψ̄(x)J(x) . (7.163)
67
The operator i/ ∂ is self-conjugate, rather the minus sign in front of it comes from the
fact that the direction of the derivative is now also reversed.
August 31, 2019 225

This Lagrangian does not contain a derivative of ψ̄: the Euler-Lagrange

equation is therefore simply
Z
δS δL(y)
= d4 y = (i/∂ − m)ψ(x) − J(x) = 0 . (7.164)
δ ψ̄(x) δ ψ̄(x)
By partial integration we can see that the same action can also be obtained
from the Lagrangian
←
L̂(x) = ψ̄(x)(−i ∂ ¯
/ −m)ψ(x) − J(x)ψ(x) − ψ̄(x)J(x) , (7.165)
which is now independent of any derivative of ψ. The Euler-Lagrange equa-
tion for ψ, then gives us precisely Eq.(7.161). Finally, it is easily seen that
the dimensionality of the field ψ is given by
dim ψ = dim ~1/2 /L3/2 .
   
(7.166)

7.7 Exercises
Excercise 48 Lorentz-contracted gamma matrices
Prove the results given in Eq.(7.14).
Excercise 49 Five spacetime dimensions
Let γ 4 be defined as γ 4 = iγ 5 . Show that this γ 4 has just the right properties
so that the relations
γ µ γ ν + γ ν γ µ = 2 g µν , γ µ = γ µ
are valid in a five-dimensional spacetime (µ, ν = 0, 1, 2, 3, 4) with metric
γµν = diag(+1, −1, −1, −1, −1).
Excercise 50 Cyclic sigmas
Prove, by explicit calculation, that
γ 5 γ 0 γ j = σ kn , γ 5 γ j γ k = σ n0
where (j, k, n) is a cyclic permutation of (1, 2, 3).
Excercise 51 Another relation between σ’s
Use trace identities and Fierzing to prove that
i
γ 5 σ αβ = − αβµν σµν
2
226 August 31, 2019

Excercise 52 Projections playing around

Prove the results given in Eq.(7.89).

Excercise 53 Delving into spinor dyads

Prove Eq.(7.91).

Excercise 54 Finding mass, spin and momentum

Prove the results in Eq.(7.111).

Excercise 55 Dirac spinors: positive energy, on the mass shell

Let u = u(p, s). Using the results from the previous exercises about finding
mass, momentum and spin from given spinors, prove that p0 > 0, p · p = m2 ,
p · s = 0 and s · s = −1.

Excercise 56 Building a Dirac spinor

Let ξ be an arbitrary spinorial object (i.e. a four-component column). Show
that
(/p + m)(1 + γ 5 s/)ξ
is proportional to u(p, s).

Excercise 57 Spin basis for given momentum

Consider the two spinors u(p, s) and u(p, −s). Show that any other spinor
with momentum p can be written as

u(p, s0 ) = a+ u(p, s) + a− u(p, −s)

for some coefficients a± , and write a± as spinor products.

Excercise 58 Rotating over 60 degrees and so on

p µ = xµ , q µ = cos(θ) xµ + sin(θ) y µ
Compute Σ(p → q) for θ = π/3, and from there, by taking powers, for
θ = 2π/3 and finally for θ = 2π.

Excercise 59 Dirac FEE model

We reconsider the FEE model,but this time the E are Dirac particles. This
August 31, 2019 227

means that we have not only E particles but also Ē particles. There is one
vertex :
E
i
E F ↔ ~λ

where λ is a dimensionless coupling constant. As before, F has mass M and

E has mass m.

1. There are now 3 diagrammatic SDe’s for this model. Write them out.

2. Assume M > 2m. Compute the decay width Γ(F → EE) at the tree
level.

3. Assume 0 < M < 2m.

(a) Write the tree diagrams for the process E(p1 ) Ē(p2 ) → E(q1 ) Ē(q2 )
by F exchange. We have indicated the momenta of the E parti-
cles.
(b) Compute h|M|2 i for this process, where we sum over the final-
state spins and average over the initial-state spins of the E’s.
(c) Compute the total cross section σ(E Ē → E Ē), as a function of
the total invariant mass-squared s ≡ (p1 + p2 )2 . This is best done
in the centre-of-mass frame.

4. In the above process, assume now M = m = 0 (and still Γ = 0).

Determine h|M|2 i, using momentum conservation to simplify the result
as much as possible.

5. Assume M = 0 and m > 0.

(a) Write down the 2 diagrams for E(p1 ) F (k1 ) → E(p2 ) F (k2 ) at
the tree level.
(b) Work out h|M|2 i for this process.
(c) Compute the total cross section as a function of s = (p1 + k1 )2 .
This is best done in the centre-of-mass frame.
(d) Compute the total cross section in the limit s → m2 .

6. Make no assumption on m or M .
228 August 31, 2019

(a) Write down the 24 diagrams for EF → EF at one loop.

(b) Assume that there are also F F F and F F F F vertices.
i. Write the 3 diagrams for EF → EF at the tree level.
ii. Write the 13 diagrams for F → EE at the one-loop level.
iii. Determine the number of Feynman diagrams at the tree level
for the following processes:
A. E Ē → E ĒF
B. E Ē → E ĒF F
C. E Ē → EE Ē Ē
D. E Ē → EEE Ē Ē Ē

Excercise 60 Dirac FED model

We consider a variation on the FEE model: there are now 2 types of Dirac
particles, E and D, with masses mE and mD , respectively. The F particle is
massless. There are now 2 vertices :
D E

E F F
D

where both vertices have Feynman rule iλ/~ as before. We shall assume
mE > mD .

1. Compute the decay width Γ(E → F D) at the tree level.

2. Compute the cross section σ(F F → E Ē) at the tree level, as a function
of the total invariant mass-squared s.
Chapter 8

Helicity techniques for Dirac

particles

8.1 The standard form for spinors

In order to perform calculations involving Dirac particles we need to develop
computational techniques that are better1 than the Casimir trick. This chap-
ter discusses how we can compute things like amplitudes without invoking ex-
plicit representations of Dirac matrices. This goes by the general appellation
of helicity methods, although chirality methods might be more appropriate2 .

8.1.1 Opting for helicities, opting for antisymmetry

The case of massless Dirac particles is of special interest, both since calcula-
tions typically become easier if the masses vanish and because in high-energy
interactions the masses can often be neglected as a good approximation3 .
Therefore, helicity states are the theorist’s favoured choice in such calcu-
lations. As we have seen before, the complex phase of spinors is not well
defined, and we can make use of this freedom to endow spinor products with
special symmetry properties. Let us start by realizing that, for massless p and
q, we can write the helicity spinors (possibly with a suitable normalization

1
Or at least faster !
2
I think that this really is because ‘chirality’ pronounces less fluenty than ‘helicity’.
3
At the LHC, electrons are almost perfectly massless.

229
230 August 31, 2019

factor)
u± (p) ∼ p/ ξ∓ , ξ∓ = ω∓ ξ∓ , (8.1)
where the helicity spinors ξ∓ can be chosen almost4 arbitrarily. Note that
this is a phase convention ! Then, the spinor product can be written in the
form
s+ (p, q) = u+ (p) u− (q) ∼ ξ¯− p/q/ ξ+ . (8.2)
You might try to find a spinor product that is symmetric in p ↔ q, but that
is useless since it would give
1 1
s+ (p, q) = (s+ (p, q) + s+ (q, p)) ∼ ξ¯− (/pq/ + q/p/) ξ+ = (p · q) ξ¯− ξ+ , (8.3)
2 2
and such a spinor product is simply the vector product. The alternative is
to choose the spinor product antisymmetric, so that s+ (p, q) + s+ (q, p) = 0,
and we see that this requires ξ¯− ξ+ = 0 : the helicity spinors ξ± must be
orthogonal. Then we shall have

s+ (p, q) = −s+ (q, p) = −s− (p, q)∗ = s− (q, p)∗ . (8.4)

8.1.2 The standard form for helicity spinors

Let us flesh out the discussion of the previous section by introducing what we
shall call the standard form for massless spinors. Let pµ be the momentum
of the spinor, so that p2 = 0. We now choose two basis vectors k0µ and k1µ ,
which satisfy
k0 · k0 = k0 · k1 = 0 , k1 · k1 = −1 . (8.5)
Furthermore we require that k0 ·p is nonzero5 for any massless momentum pµ
encountered in the problem at hand ; this is usually not difficult to arrange.
Since k 0 is massless, it may serve to define the basis spinor

u0 ≡ u− (k0 ) ⇒ u0 u0 = ω− k/0 . (8.6)

We can then also define6

u+ (k0 ) = k/1 u0 , (8.7)
4
With extremely bad luck you might find, say, a ξ+ that will give a vanishing product.
If so, simply pick another one !
5
And preferably not too small, for your favourite definition of ‘too small’.
6
k1 u− (k0 ) = u+ (k0 ) but /k1 u+ (k0 ) = −u− (k0 ).
Be careful here ! /
August 31, 2019 231

and easily verify that u− (k0 )u+ (k0 ) = 0. Reversal ‘flips’ the helicity :

(u− (k0 )u− (k0 ))R = (ω− k/0 )R = k/0 ω− = ω+ k/0 = u+ (k0 )u+ (k0 ) . (8.8)

Using the basis spinor, we now define all other massless spinors by
1
u± (p) = √ p/ u∓ (k0 ) . (8.9)
2p · k0
This is by construction a valid form since
1 1
u± (p)u± (p) = p/ ω∓ k/0 p/ = ω± p/k/0 p/ = ω± p/ . (8.10)
2p · k0 2p · k0
This choice is at the basis of the so-called spinor techniques, or helicity meth-
ods. we shall use Eq.(8.9) extensively to good effect.

8.1.3 Some useful identities

At this point we prove a few results that often
turn out to be useful. In the
first place, from the property Tr (Γ) = Tr ΓR , we can see that

u+ (p1 ) γ µ u+ (p2 ) = K u0 p/1 γ µ p/2 u0 = K Tr (u0 u0 p/1 γ µ p/2 )

= K Tr p/2 γ µ p/1 (u0 u0 )R = K Tr (/p2 γ µ p/1 k/1 u0 u0 k/1 )

= K u0 k/1 p/2 γ µ p/1 k/1 u0 , (8.11)

with K = (4p1 · k0 p2 · k0 )−1/2 , which leads to the useful spinor reversal :

u+ (p1 )γ µ u+ (p2 ) = u− (p2 )γ µ u− (p1 ) . (8.12)

In fact this can easily be generalized to

uλ1 Γ uλ2 (q) = λ1 λ2 u−λ2 (q) ΓR u−λ1 (p) , λ1,2 = ± . (8.13)

The Chisholm identity of Eq.(7.41) can be applied to simple spinor sand-

wiches so as to yield
   
µ
u± (p1 )γ u± (p2 ) γµ = 2 u± (p2 )u± (p1 ) + u∓ (p1 )u∓ (p2 ) . (8.14)

Finally, we have the so-called Schouten identity. It is based on the truism

u+ (p1 )/p2 p/3 u− (p4 ) + u+ (p1 )/p3 p/2 u− (p4 ) = 2(p2 · p3 ) u+ (p1 )u− (p4 ) . (8.15)
 232 August 31, 2019

Writing this out in terms of spinor products, we have

s+ (p1 , p2 )s− (p2 , p3 )s+ (p3 , p4 ) + s+ (p1 , p3 )s− (p3 , p2 )s+ (p2 , p4 )
= s+ (p2 , p3 )s− (p3 , p2 )s+ (p1 , p4 ) . (8.16)

Using the antisymmetry property of s, and dividing out the factor s− (p2 , p3 ),
we obtain the so-called Schouten identity :

s+ (p1 , p2 )s+ (p3 , p4 ) + s+ (p1 , p3 )s+ (p4 , p2 ) + s+ (p1 , p4 )s+ (p2 , p3 ) = 0 . (8.17)

E61 Note the cyclicity in p2,3,4 . Obviously, the identity holds for s− as well. At
the end of this chapter we give a summary table of all these spinor techniques.

8.1.4 How to compute spinor products

To evaluate amplitudes numerically we need a recipe to compute spinor prod-
ucts numerically, just like the recipe to compute vector products. For spinors
in the standard form we can use the Casimir trick :

s+ (p, q) = (4(p · k0 )(q · k0 ))−1/2 u0 p/ q/ k/1 u0

= (4(p · k0 )(q · k0 ))−1/2 Tr (ω− k/0 p/q/k/1 )

1
= p (p · k0 )(q · k1 ) − (p · k1 )(q · k0 )
(p · k0 )(q · k0 )

µ ν α β
− iµναβ k0 k1 p q . (8.18)

You can easily check the spinor products’ properties by explicit inspection.
From |sλ (p, q)|2 = 2(pq) we can see that spinor products are square roots
of vector products. Actually, in practical applications it pays to compute
the spinor products of massless momenta in your problem anyway. As an
example, if two massless momenta pµ and q µ make a small angle θ, their
vector product reads (pq) = pp q 0 (1 − cos θ) ∼ θ2 . For their spinor product,
on the other hand, we have |s+ (p, q)| ∼ θ so the numerical cancellations are
much less severe in te spinor product than in the vector one. The vector
product can then be obtained as a simple square.
µ
We may consider an explicit choice for the vectors k0,1 :

k0 µ = (1, 1, 0, 0) , k1 µ = (0, 0, 1, 0) : (8.19)

August 31, 2019 233

this gives the explicit form for the spinor product

s s
q 0 − q1 p0 − p1
s+ (p, q) = p2 + ip3 2 3



− q + iq , (8.20)
p0 − p1 q0 − q1

which is very useful for actual numerical applications. Note that this choice E62
presupposes that none of the light-like vectors in the problem is oriented
E63
exactly along the x-axis. Since the ‘special’ direction in many problems is
traditionally chosen to be the z-axis, this is usually safe. E64
E65
8.1.5 The standard form for massive particles
The standard form for Dirac spinors given in Eq.(8.9) can be generalised to
the case of massive particles. Let pµ be the momentum of such a particle,
and let m be its mass. We then simply define
1
u± (p) = √ (/p + m)u∓ (k0 ) ,
2p · k0
1
v± (p) = √ (/p − m)u∓ (k0 ) . (8.21)
2p · k0

We can find out the spin vector for these two cases : writing u± (p) =
u(p, ±s0 ) we obtain

1
s0 µ = − u+ (p) γ 5 γ µ u+ (p)
2m
1
Tr ω− k/0 (/p + m)γ 5 γ µ (/p + m)


= −
4m p · k0
1 µ m
= p − k0 µ , (8.22)
m (pk0 )

which is indeed the only vector built from p and k0 that can have the right
properties s0 2 = −1 and (ps0 ) = 0. Note that for small(ish) m and generally
positioned k0 , ~s0 points in the general direction of p~. Therefore we call u+ (p)
a right-handed spinor, and u− (p) a left-handed spinor. In addition, from the
fact that, for the antispinor v± (p),

1
v + (p) γ 5 γ µ v+ (p) = −s0 µ , (8.23)
2m
234 August 31, 2019

we see that v+ (p) is a left-handed antispinor and v− (p) is a right-handed

antispinor7 .

8.1.6 The standard form for complex momenta

We have seen that for a helicity spinor u+ there is a momentum q such that
u+ u+ = ω+ q/, and that this momentum is real and massless, with positive en-
ergy. This is, of course, just what the actual momentum of a massless particle
is supposed to look like. But in the course of actually computing amplitudes
it frequently happens that we would like to have spinors for momenta p that
have negative energy, or are complex-valued8 . The question is then : can we
find two spinors u and w such that uw = ω+ p/ ? Obviously, now w 6= u but
that doesn’t matter as long as we don’t need to suddenly conjugate a spinor
inside an amplitude9 . Note that uw† = ω+ p/γ 0 has precisely the form of the
dyad Θ of sect.7.3.2, and therefore the restriction p2 = 0 holds.
The solution to the task we set ourselves above is actually almost trivial :
we use the same basis spinor u0 as before, and the same (real-valued !) vectors
k0 and k1 . We then simply define

u+ (p) = N p/u0 , w+ (p) = N u0 p/ , (8.24)

p
with N = 1/ 2(p · k0 ). Note that N is now no longer purely real ; but as
long as we take it equal in the definition of u and w it is directly seen that
they do the trick. Of course, for the other helicity we can use

u− (p) = N p/k/1 u0 , w− (p) = N u0 k/1 p/ . (8.25)

7
Remember the ‘not too small’ in
sect.8.1.2 ? In a frame where ~k0 and p~ are parallel the
dot product (k0 p) is of order O m2 and the ‘subleading’ term with k0 in Eq.(8.22) is no
longer subleading. Whereas normally the spinor u+ (p) looks a lot like a positive-chirality
spinor u+ for small m, it then looks more like u− : a continuous source of worry especially
in large computer programs that deal with many fermions numerically.
8
In cases like this, you will usually hear your lecturer going ‘mumble mumble analytic
continuation mumble mumble’ but that is unhelpful unless you know how to analytically
continue. This section, in fact, explains how.
9
I have never encountered this need. Of course, the complex conjugate amplitude occurs
in the cross section, but that can be accounted for by replacing every u by w and vice
versa, if you really insist on computing the complex conjugate amplitude separately (why
would you ?).
August 31, 2019 235

We have now found the dyadic forms we wanted. In what way do they differ
from the Dirac case ? The spinor products

s+ (p, q) = w+ (p)u− (q) , s− (p, q) = w− (p)u+ (q) (8.26)

are still antisymmetric, but we no longer have s− = −(s+ )∗ . On the other

hand, it is still true that

s+ (p, q)s− (q, p) = 2(p · q) ; (8.27)

and, in fact, it is easily checked that all the other relations such as the reversal
formula, the Chisholm identity, and the Schouten identity still hold.

8.2 Summary of tools for spinor techniques

Here we present the definitions and identities useful for performing calcula-
tions with massless spinors.

• Dirac equations :

p/ uλ (p) = 0 , ωλ uλ (p) = uλ (p) , ω−λ uλ (p) = 0 (8.28)

• projecting operator :
uλ (p) uλ (p) = ωλ p/ (8.29)

• Spinor products :

uλ (p) uλ (q) = 0
sλ (p, q) = uλ (p) u−λ (q) = −sλ (q, p) = −s−λ (p, q)∗
|sλ (p, q)|2 = sλ (p, q) s−λ (q, p) = 2(p · q) (8.30)

• Elimination of repeated indices10 :

γα uλ (p) uλ (q) γ α = −2 u−λ (q) u−λ (p)

γα uλ (p) u−λ (q) γ α = 2 s−λ (q, p) ω−λ (8.31)
10
This is also valid for p = q.
236 August 31, 2019

• Chisholm identity :

uλ (p) γα uλ (q) γ α = 2 uλ (q) uλ (p) + u−λ (p) u−λ (q)

   
(8.32)

• Reversal :
uλ1 (p) Γ uλ2 (q) = λ1 λ2 u−λ2 (q) ΓR u−λ1 (p) (8.33)

• Schouten identity :

sλ (q, p1 ) sλ (p2 , p3 ) + sλ (q, p2 ) sλ (p3 , p1 ) + sλ (q, p3 ) sλ (p1 , p2 ) = 0 (8.34)

8.3 Fermionic decays : the Fermi model

After all the mathematics it is high time to do some actual phenomenology.
Up to now all interacting theories have been toys to acquaint ourselves with
Feynman rules and the like, but our goal ought of course to be to be able to
say something about real, observable processes. Fortunately, there are a few
processes that can be formulated (almost) purely in terms of Dirac particles.

8.3.1 The amplitude for muon decay

An example of an actually occurring process involving only Dirac particles
is provided by muon decay in the Fermi model. The process is11

µ− (p) → e− (q) νµ (k1 ) ν e (k2 )

and is pictured by the single Feynman diagram

p k1
M= q (8.35)
k2

Here, a muon at rest undergoes a three-particle decay into an electron, a

muon neutrino and an electron antineutrino. We shall assume the neutrinos
11
In this section, the vector k1 is a momentum, and has nothing to do with the auxiliary
vector of section 8.1.
August 31, 2019 237

to be massless. The Fermi amplitude introduced to describe the phenomenol-

ogy of this process contains only a single pointlike
√ vertex where four fermions
meet with a coupling constant called GF / 2, and is given by
GF ~
M=i √ u(q) (1 + γ 5 )γα v(k2 ) u(k1 ) (1 + γ 5 )γ α u(p) . (8.36)
2
The decision to ‘hook up’ the muon and the muon neutrino is in principle
arbitrary12 , but as we have seen in section 7.2.7 we may easily interchange
the muon neutrino and the electron, and end up with the matrix element in
the charge retention form :
GF ~
M = −i √ u(q) (1 + γ 5 )γα u(p) u(k1 ) (1 + γ 5 )γ α v(k2 ) . (8.37)
2
The amplitude (8.36) implies that the neutrinos must have negative helic-
ity13 : we can write
4GF ~
M=i √ u(q) γα v− (k2 ) u− (k1 ) γ α u(p) . (8.38)
2
By eliminating the repeated indices (cf sect.8.2) we can arrive at the very
compact form
8GF ~
M = −i √ u(q) u+ (k1 ) v + (k2 ) u(p) . (8.39)
2
The transition rate can now easily be computed with a few simple traces :
1 X
|M|2 = |M|2 = 16 GF 2 ~2 Tr ((/q + me )ω+ k/1 ) Tr ((/p + mµ )ω+ k/2 )


2
= 64 G2F ~2 (q · k1 ) (p · k2 ) . (8.40)
Here, the sum is over the electron and muon spins. It is practical to evaluate
this in the muon rest frame. We shall write E1,2 for k1,2 0 in this frame. Then
(p · k2 ) is equal to mµ E2 , and by momentum conservation we find
1
1
(q + k1 )2 − me 2 = (P − k2 )2 − me 2 = mµ (K − E2 ) ,


(q · k1 ) =
2 2
(8.41)
12
Unless lepton flavour number is invoked.
13
In the standard form of spinors, the helicity for antispinors is reversed. The antineu-
trino therefore actually has positive chirality.
238 August 31, 2019

where
mµ 2 − me 2
K= . (8.42)
2mµ
The transition rate then takes the form
|M|2 = 64 GF 2 ~2 mµ 2 E2 (K − E2 ) ,


(8.43)
and for the partial decay width we find
dΓ(µ → eνµ ν e ) = 32 GF 2 ~2 mµ E2 (K − E2 ) dV (p; q, k1 , k2 ) . (8.44)

8.3.2 Three-body phase space

The phase space for the muon decay process reads
1
dV (p; q, k1 , k2 ) = d4 q d4 k1 d4 k2 δ 4 (p − q − k1 − k2 )
(2π)5
δ(q 2 − me 2 ) δ(k1 2 ) δ(k2 2 ) . (8.45)
Since the rate depends only on E2 , we shall implicitly integrate over all other
phase space variables. By cancelling the q integration against the Dirac delta
for momentum conservation, we arrive at
1 E1 E2
dV (p; q, k1 , k2 ) = 5
dE1 dE2 dΩ1 dΩ2 δ((p − k1 − k2 )2 − me 2 ) .
(2π) 4
(8.46)
The Dirac delta function can be written as
δ mµ 2 − me 2 − 2mµ E1 − 2mµ E2 + 2E1 E2 − 2E1 E2 cos θ ,

where θ is the angle between the neutrino momenta. Hence we can integrate
trivially over the other polar, and the two azimuthal, angles (leading to a
factor 8π 2 ), and the integral over θ is resolved by the delta function. The
result is
π2
dV (p; q, k1 , k2 ) = dE1 dE2 . (8.47)
(2π)5
In terms of these variables, the phase space is perfectly flat14 . Since |cos θ|
cannot exceed unity, we also have the restrictions
mµ 2 − me 2 − 2mµ E1 − 2mµ E2 ≤ 0 ,
14
This flatness does not depend on the masslessness of the neutrinos. For massive
neutrinos the same phase space density is found, only the boundaries of the phase space
become (horribly) complicated.
August 31, 2019 239

Figure 8.1: Elecrton-mass dependence of the muon decay width

mµ 2 − me 2 − 2mµ E1 − 2mµ E2 + 4E1 E2 ≥ 0 , (8.48)

which we can work into bounds on E1 :

mµ 2 − me 2 − 2mµ E2
K − E2 ≤ E1 ≤ K̂(E2 ) ≡ , (8.49)
2(mµ − 2E2 )

while E2 is seen to run from 0 to K.

8.3.3 The muon decay width

After the simple integration over E1 , we have the muon partial decay width
d
Γ(µ → eνµ ν e ) = π 2 GF 2 ~2 mµ E2 (K − E2 )(K̂(E2 ) + E2 − K) . (8.50)
dE2
The remaining integral over E2 can now be performed, and the final result is

GF 2 ~2 mµ 5
Γ(µ → eνµ ν e ) = F (me 2 /mµ 2 ) ,
192 π 3
F (x) = 1 − 8x + 8x3 − x4 − 12x2 log(x) . (8.51)

The function F (x), shown in figure 8.1, is strictly decreasing since with
increasing me /mµ the available phase space shrinks. For the realistic values
 240 August 31, 2019

of me and mµ F (x) is smaller than 1 by about 2 × 10−4 . The effects of nonzero

electron mass are therefore completely negligible, certainly if we realize that
we have not included any loop diagrams, the contribution of which is very
considerably larger than this. Before finishing, it is instructive to inspect the
muon width formula
GF 2 ~2 mµ 5
Γ(µ → eνµ ν e ) =
192 π 3
from the point of view of dimensional analysis. In the first place, the ma-
trix element M, being of 2 → 2 type, must be strictly dimensionless (see
Eq.(6.23)). Since every spinor carries half a power of momentum15 , the
Fermi coupling constant GF must carry dimension momentum−2 . Since de-
cay widths carry the dimension of momentum, as do masses like mµ , and
the only mass scale in the problem is mµ if we neglect the electron mass,
the width is necessarily proportional to GF 2 mµ 5 . The discussion at the end
of section 6.5.4 shows that the factor 1/π 3 was also to be expected. It is a
E67 somewhat sobering thought that all the work of this section amounts to no
more than computing the number 1/192 !

8.3.4 Observable distributions in muon decay

We can look more closely into the behavior of the electron in muon de-
cay, in particular its angular and energy distribution : in fact, these are
the only quantities of interest since the neutrinos can essentially never be
detected. If the sample of decaying muons is unpolarised, no particular di-
rection is favored in its rest frame, and the electrons may be expected to
emerge isotropically. Let us therefore assume that the muons are fully po-
larised, with polarisation vector sµ = (0, ~s) in their rest frame. Since we are
interested in distributions rather than the overall decay rate, we can afford to
be sloppy with pre-factors in this computation. Additionally we shall assume
me = 0. The matrix element is now given by
M ∼ u− (q) γµ v− (k2 ) u− (k1 ) γ µ u(p, s)
∼ u− (q)u+ (k1 ) u+ (k2 )u(p, s) . (8.52)
Squaring this, we obtain
|M|2 ∼ Tr (ω− q/k/1 ) Tr ω+ k/2 (1 + γ 5 s/)(/p + mµ )

15
Since the spin sum of uu contains /p.
August 31, 2019 241


∼ (q · k1 ) (k2 · p) − mµ (k2 · s)
∼ k20 (mµ − 2k20 ) (1 + cos(θ) cos(θ2 ) + sin(θ) sin(θ2 ) cos(φ2 )) ,
(8.53)

where θ is the angle between ~q and ~s, and θ2 , φ2 are the polar and azimuthal
angles of ~k2 with respect to ~q. As we have seen above, the phase space
integration element is, up to overall factors,

dq 0 dk20 d cos(θ) dφ dφ2 ,

where φ is ~q ’s azimuthal angle around ~s, while

mµ (q 0 + k20 − mµ /2)
cos(θ2 ) = 1 − . (8.54)
q 0 k20

Performing the integral over the (unobservable) angle φ2 , and disregarding

the trivial φ dependence, we therefore have

mµ (q 0 + k20 − mµ /2)
  
0 0
dΓ ∝ k2 (mµ −2k2 ) 1 + cos(θ) 1 − dq 0 dk20 d cos(θ) .
q 0 k20
(8.55)
The integral over k20 from mµ /2 − q 0 to mµ /2 is straightforward, leading to
the following, properly normalised distribution :

1 d2 Γ 2q 0
= y 2 3 − 2y + cos(θ)(1 − 2y) ,


y= . (8.56)
Γ dy d cos(θ) mµ

The overall angular distribution is

1 dΓ 1

= 3 − cos(θ) , (8.57)
Γ d cos(θ) 6

and the overall distribution of the electron energy reads

1 dΓ
= 2y 2 (3 − 2y) . (8.58)
Γ dy

Let us now consider the parity properties of the decay. Under the parity
transform, Eq.(7.149) tells us that ~q → −~q and ~s → ~s, so that effectively
cos(θ) → − cos(θ). We see that the muon decay amplitude, which is not
242 August 31, 2019

symmetric in cos(θ), displays parity violation. This is most prominent when

the electron has maximal energy :
d2 Γ

1
= 1 − cos(θ) . (8.59)
Γ dy d cos(θ) y=1

Qualitatively, we can understand this as follows. At y = 1, the two neutrinos

recoil in parallel against the electron. Since νµ is left-handed and ν̄e is right-
handed (because of the factors 1 + γ 5 in the coupling), their spins cancel, and
the spin of the electron must therefore be in the direction of the muon spin
by conservation of angular momentum. If the electron were to be emitted
in the direction of the muon spin, it would therefore be right-handed, which
is again forbidden by the coupling16 . Hence the amplitude must vanish at
θ = 0.

8.3.5 Charged pion decay: helicity suppression

Another example of a process that we can now compute is that of pion decay.
The charged pion π − is a bound state of a d quark and a u antiquark, which
are fermions. These annihilate into a charged lepton ` (e− or µ− ) and the
corresponding anti-neutrino, assumed to be massless as before. The process
therefore reads
π − (p) → `− (q1 ) ν ` (q2 )
where we have indicated the observable momenta. We assume that the in-
teraction vertex is again of Fermi type, so that the lepton and antineutrino
couple with the familiar17 (1 + γ 5 )γ µ coupling. But for the quarks it is dif-
ferent : they are certainly not free particles, since they form a bound state,
and the best we can do is to assume that they conspire to form some object
with a Lorentz index. The only possibility (since the pion is spinless) is that
somehow they form something proportional to the pion momentum pµ . The
phenomenological amplitude therefore reads
√ GF
M = i ~ √ fπ pµ u(q1 )(1 + γ 5 )γ µ v(q2 ) . (8.60)
2
Here the totality of our lack of knowledge about the quarks’ goings-on is
embodied in the pion decay constant fπ . We can use the spinors’ properties
16
Remember that we took me = 0.
17
By now, I hope.
August 31, 2019 243

to simplify

u(q1 ) (1 + γ 5 ) p/ v(q2 ) = u(q1 ) (1 + γ 5 ) (/q1 + q/2 ) v(q2 ) = m` u(q1 ) (1 − γ 5 ) v(q2 ) .

(8.61)
We therefore find

|M|2 = 2~ G2F fπ2 m2` m2π − m2` ,



(8.62)

and for the 2-body phase space we have

1 λ1/2 (m2π , m2` , 0) dΩ m2π − m2`
dV (p; q1 , q2 ) = = dΩ . (8.63)
(2π)2 m2π 8 32π 2 m2π
Consequently,
2
− − ~ G2F fπ2 m2` (m2π − m2` )
Γ(π → ` ν ` ) = . (8.64)
8π m3π
This tells us that a muon is much more likely to decay into a muon and its
antneutrino than into an electron and its antineutrino :
2
Γ(π − → µ− ν µ ) m2µ m2µ

≈ 2 1− 2 ∼ 7.8 × 103 (8.65)
Γ(π − → e− ν e ) me mπ
The smallness of the electronic decay mode is thus assigned to the mechanism
of helicity suppression. Since the pion has no spin, the spins of the charged
fermion and the antineutrino must be opposite. The antineutrino is right-
handed and therefore the charged lepton has right-handed helicity. On the
other hand, the weak interactions couple only to the left-handed chirality
part of the charged lepton. For massless leptons helicity and chirality are
the same, and the decay would be impossible. The more massive the lepton E68
is, the larger is the overlap between the right-handed helicity and the left-
handed chirality components.

8.4 Exercises
Excercise 61 Helicity operator for massless spinors
Show that, for a general timelike or lightlike momentum pµ , the operator
1 5
Σp = cos(θ/2) + i sin(θ/2) γ [/p, γ 0 ]
2|~p|
244 August 31, 2019

corresponds to a rotation around the p~ axis over an angle θ. For massless

pµ , prove that it acts as a helicity probe :
Σp uλ (p) = exp(iλθ/2) uλ (p)
Excercise 62 Spinor products and vector products
Using the form of Eq.(8.20) for s+ (p, q) to prove, by direct computation, that
|s+ (p, q)|2 = 2(p · q)
Excercise 63 Long Dirac strings
Let pµj , (j = 1, 2, 3, . . .) be massless momenta. Write the expression
A = u+ (p1 )/p2 p/3 p/4 p/5 u− (p6 )
in terms of spinor products, and compute |A|2 .
Excercise 64 Why is it zero ?
for massless momenta (as in the previous exercise), let
A1 = u(p1 )/p2 p/3 u(p7 ) , A2 = u(p1 )/p4 p/5 p/6 u(p7 )
where you can chooce the helicities of p1,7 yourself. Show that in all cases
A1 A2 ∗ = 0
Excercise 65 Levi-Civita with spinor currents
Let pµ1,2,3,4 be massless vectors. Prove that
µναβ u+ (p1 )γα u+ (p2 ) u+ (p3 )γβ u+ (p4 )
= i u+ (p1 )γ µ u+ (p4 ) u+ (p3 )γ ν u+ (p2 ) − (µ ↔ ν)

Excercise 66 Towards a real theory prediction

We consider the process
e+ (p1 ) e− (p2 ) → e+ (q1 ) e− (q2 )
where the electrons/positrons are massless spin-1/2 particles. In Quantum
Electrodynamics (QED), the amplitude for this process is given by
M(λ1 , λ2 , ρ1 , ρ2 ) =

2 1
i~Qe 2
uλ1 (p1 )γ µ uλ2 (p2 ) uρ2 (q2 )γµ uρ1 (q1 )
(p1 + p2 )

1 µ
− uλ (p1 )γ uρ1 (q1 ) uρ2 (q2 )γµ uλ2 (p2 )
(p1 − q1 )2 1
August 31, 2019 245

where (as usual for massless fermions) we have disregarded the distinction
between spinors and antispinors. The quantity Qe is a fixed coupling con-
stant. The helicities λ1,2 and ρ1,2 are explicitly indicated. There are, in total,
16 helicity combinations.
1. Using spinor techniques, compute M for the sixteen helicity cases.
2. Assume that we are in the centre-of-mass frame, where p~1 + p~2 = 0.
Let θ be the angle between p~1 and ~q1 , and E = p1 0 . Give the simplest
form you can find for h|M|2 i.
3. Show that h|M|2 i is really divergent for θ = 0, and that this has nothing
to do with neglecting the electron mass.

Excercise 67 Muon decay revisited

We reconsider muon decay, again with vanishing neutrino and electron mass.
We shall investigate the result of leaving out the factors (1 + γ 5 ) in the
amplitude, so that we take it to have the form

M = i~ G u(k1 ) γ α u(p) u(q) γα v(k2 )

and all fermion can be in two spin states.

1. Compute h|M|2 i, summed over the outgoing fermions’ helicities and
averaged over the muon spin.
2. Use the results for three-body kinematics of section 8.3.2 to write the
above expressions in terms of q 0 , k2 0 , and mµ .
3. After the angular integrations, as explained in section 8.3.2 there re-
mains the energy integral
m
Zµ /2 m
Zµ /2
dq 0
dk2 0
0 mµ /2−q 0

The electron energy q 0 is observable, but the antineutrino energy k2 0 is

not. Integrate over k2 0 to get the normalised electron energy spectrum
(dΓ/dq 0 )/Γ..
4. Compare this spectrum to that of Eq.(8.58) and show how one can
experimentally infer the presence of the 1 + γ 5 .
246 August 31, 2019

Excercise 68 It’s different for us who are vectors

Let us reexamine pion decay of section 8.3.5. If the decaying particle is not
a scalar but a vector particle (as we shall see, that corresponds to unit spin
instead of zero spin), the part of the decay amplitude involving the leptons
has a vectorial character, for instance

u(q1 )(1 + γ 5 )γ µ v(q2 )

Show that in that case there is no helicity suppression.

Chapter 9

Vector particles

9.1 Massive vector particles

9.1.1 The propagator
In the last chapter we have studied the consequences of embellishing the
scalar propagator by endowing it with a numerator linear in the momentum.
The next obvious generalization is to let T (p) depend on two powers of the
momentum. That is, we assume it to be of the form
T (p) → T (p)µν = Ag µν + Bpµ pν , B 6= 0 ,
for some A and B that may depend on p2 . The numerator now carries two
Lorentz indices, one of each to be contracted with a corresponding index in
the vertices between which the propagator runs. The discussion of the last
chapter leads us to require that for momenta on the mass shell T (p) must
be a projecting operator :
T (p)µα T (p)α ν = kT (p)µν if p2 = m2 , (9.1)
for some k 6= 0, in other words
A2 = kA , B 2 m2 + 2AB = kB . (9.2)
We might choose the solution A = 0, but then the resulting form T (p)µν ∼
pµ pν would be factorizable. It follows that A must equal −m2 B, and therefore
we shall use1
1
T (p)µν = −g µν + 2 pµ pν . (9.3)
m
1
This choice of T is called the unitary gauge.

247
 248 August 31, 2019

The first Feynman rule for these particles, that we call vector particles since
they carry a Lorentz index, is therefore established :

µ −g µν + pµ pν /m2 internal lines

p ν ↔ i~
p2 − m2 + i

Feynman rules, version 9.1

Note that this propagator is even in p and therefore has no orientation2 .

Another thing, to which we shall refer to often enough, is that the propagator
of a vector particle, in contrast to that of a Dirac particle, can to some extent
be chosen out of several possibilities. These are denoted gauges. The above
choice is called the unitary gauge.

9.1.2 The Feynman rules for external vector particles

From the form of T (p) we must be able to derive the form of the external-line
factors. Indeed, let us assume pµ to be in its rest frame, pµ = (m, ~0). There,
we have
T (p)µν = −g µν + g 0µ g 0ν = diag(0, 1, 1, 1) , (9.4)
that is, the unit tensor in the spatial sector of Minkowski space. We see that
we can write  
T (p)µν = − xµ xν + y µ y ν + z µ z ν , (9.5)

which means that, for the objects U, W three mutually orthogonal choices
can be made, for instance U (1) = x, U (2) = y, and U (3) = z. Of course,
complex linear combinations of these are also possible : in general, we can
say that there can be found three polarisation vectors µλ , with λ = −1, 0, 1,
such that
1
µ µν
X ν
(λ ) ( )µ = −δ
λ0 λ,λ0 , T (p) = (λ )µ (λ ) . (9.6)
λ=−1

E69 If pµ is not in the rest frame, we simply boost both p and the various  to
2
That is, its spacetime part is unoriented. There may of course be other properties
such as charge that do impose a distinction between production and decay of the particle.
August 31, 2019 249

the actual Lorentz frame. We can now go once more through the truncation
argument of chapter 6, with the obvious result that the polarisation vectors
are to be assigned to the external lines, and we immediately arrive at the full
set of Feynman rules for massive vector particles :

µ −g µν + pµ pν /m2 internal lines

p ν ↔ i~
p2 − m2 + i
000000
111111
111111
000000
000000
111111
000000
111111
000000
111111
√ incoming lines
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
↔ ~ λ µ
000000
111111

000000
111111
111111
000000
000000
111111
000000
111111
000000
111111
000000
111111
√ outgoing lines
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
↔ ~ λ µ
000000
111111

Feynman rules, version 9.2 (9.7)

Owing to the lack of orientation, the rules for the external lines are quite
simple, and fortunately no Dirac indices appear, nor do any curious and
cumbersome minus signs. The fact that T (p)µν corresponds precisely to sum
of the polarization vectors’ dyads, with no terms left out or left over, justifies
the name unitary gauge.

9.1.3 The spin of vector particles

To ascertain the spin of vector particles3 , we need to establish the form
of the Lorentz transformation in the space of the polarisation vectors, i.e.
Minkowski space. We can do this conveniently using the transform in Clifford
space, as follows. Let us denote by Λ(p; q)µ ν the representation of the minimal
Lorentz transformation between pµ and q µ in Minkowski space : that is, if
an arbitrary vector aµ is transformed into bµ , we have
Λ(p; q)µ ν aν = bµ . (9.8)
/ and b/ encode exactly the same information as do aµ and bµ , consis-
Since a
tency requires that
b/ = Λ(p; q)µ ν aν γµ = Σ/aΣ = Σ aν γν Σ , (9.9)
3
The fact that there are three polarisation vectors of course suggests that the spin is
unity, but again suspicion does not equal proof.
 250 August 31, 2019

with Σ as defined in section 7.4.5 ; since this must hold for arbitrary a, we
have the relation
Λ(p; q)µ ν γµ = Σ γν Σ , (9.10)
By multiplying with γ α on both sides and taking the trace, we immediately
E70 find the form of Λ(p; q) in Minkowski space :
1
1
Λ(p; q)α ν = Tr Λ(p; q)µ ν γµ γ α = Tr Σ γν Σ γ α


4  4
2
  
p q/p/ p/q/
= Tr 1 + 2 γν 1 + 2 γ α
4(p + q)2 p p
2 2
= δαν − 2
(p + q)α (p + q)ν + 2 q α pν . (9.11)
(p + q) p

Let us now specialize to the case of rotations from pµ = xµ , to q µ =

E71 cos(θ)xµ + sin(θ)y µ , as in section 7.4.5 : the rotation matrix now reads

Λµ ν = (tµ tν − z µ zν ) − cos(θ) (xµ xν + y µ yν ) + sin(θ) (xµ yν − y µ xν ) . (9.12)

In the case of Dirac particles, a rotation over 2π resulted in a sign change :

E72 here, such a full rotation is seen to be simply the identity. For infinitesimal
θ this reads

Λ(p; q)µ ν = δ µ ν + θ (xµ yν − y µ xν ) + O θ2 ,

(9.13)

so that the generators of the rotation group must in this case have the form

(Tz )µ ν = β(xµ yν − y µ xν ) and cyclic permutations , (9.14)

with the constant β again to be determined from the commutation algebra :

[Tx , Ty ]µ ν = β 2 (xµ yν − y µ xν ) = β(Tz )µ ν . (9.15)

We conclude that β = i~ in Minkowski space. We find

(Tx 2 )µ ν = −~2 (y µ yν + z µ zν ) , (9.16)

etcetera, so that the total-spin operator takes the form

 
~ ) = −2~ (x xν + y yν + z zν ) = 2~
2 µ 2 µ µ µ 2 µ 1 µ
(L ν −δ ν + 2 p pν . (9.17)
m
August 31, 2019 251

we conclude that the spin is indeed unity. The total spin operator contains,
as it must, the projection of all vectors on the spatial subspace. In words:
to be a good polarisation vector, µ must be properly normalised and satisfy
the Lorenz condition4 :

 · ∗ = −1 , ·p=0 . (9.18)

Any part of a polarisation vector that is parallel to pµ does, of course, not

transform under rotations in the space orthogonal to pµ (in our case, the
spatial part of Minkowski space since pµ is at rest). That part, therefore,
corresponds to a scalar degree of freedom. Returning to T (p) we may inter-
pret the form
1
T (p)µν = −g µν + 2 pµ pν (9.19)
m
as a propagator in which a priori four degrees of freedom propagate (the
g µν part), and where the scalar part (the pµ pν term) is carefully excised.
The pµ pν term is sometimes loosely called the longitudinal part of the prop-
agator, but this is incorrect ; we should do better by calling it the scalar part.

9.1.4 Polarisation vectors for helicity states

As usual, the helicity of a state refers to its spin as measured along the
direction of its motion. For definitiveness, let us assume that our massive
vector particle moves along the z direction. If we boost carefully5 back to the
rest frame, p~ of course vanishes, but we shall remember that to go back to
the original situation we must boost along the z direction. The operator for
the helicity is therefore Tz in this case. Good polarisation vectors for helicity E73
1,0 and -1 are then
1 1
+ µ = √ (xµ + iy µ ) , 0 µ = z µ , − µ = − √ (xµ − iy µ ) , (9.20)
2 2
The vectors ± are said to describe transverse polarisation, and the vector 0 is
called longitudinal. If we now perform the boost back to the original system
4
Note the spelling ! This does not refer to the famous Dutchman Hendrik Antoon
Lorentz (1853-1928) of transformation fame, but to the Dane Ludvig Valentin Lorenz
(1829-1891), quite another person. A relation between the polarizability and the refractive
index of a medium goes by the funky name of the Lorentz-Lorenz equation.
5
And minimally !
252 August 31, 2019

in which pµ is moving along the z direction, the transverse polarisations

remain unaffected, while the longitudinal one takes the form6
   
µ |~p| µ m
0 → p + zµ . (9.21)
mp0 p0

Very fast-moving particles, for which m  p0 ≈ |~p|, have longitudinal polar-

isation vector  
µ 1 µ m
0 → p + O . (9.22)
m p0
This is, of course, precisely what we also found for Dirac particles, since the
algebraic conditions (9.18) are the same as those of Eq.(7.80).

9.1.5 The Proca equation

Massive vector particles have their own ‘classical’ equation, which we shall
now uncover. The coupling of a massive vector particle to a source is given
by the following Feynman rule for position space :

µ i
↔ − J µ (x) (9.23)
~
The SDe for a free vector particle’s field function V µ is then again very
simple : 000000
111111 111111
000000
000000
111111
000000
111111
000000
111111 x
x 000000
111111
000000
111111
000000
111111
000000
111111
= , (9.24)
000000
111111
000000
111111
000000
111111

or, more explicitly,

e−ik·(x−y)
Z  
µ 1 4 4 µν 1
V (x) = dydk 2 −g + 2 kµkν Jν (y) . (9.25)
(2π)4 k − m2 m

We can then form the following derivative operator acting on V µ :

∂ α ∂α V µ (x) − ∂ µ ∂α V α (x) + m2 V µ (x) =

Z −ik·(x−y)
1 4 4 e
= d y d k W µν Jν (y) , (9.26)
(2π)4 k 2 − m2
6
In a somewhat simpler notation, if pµ = (p0 , p~), with p = |~
p| and ~e = p~/p, then the
longitudinal polarisation vector reads 0 µ = (p, p0~e)/m.
August 31, 2019 253

where W µν can be evaluated as

   
µν 2 2 µν 1 µ ν µ αν 1 α ν
W = (−k + m ) −g + 2 k k + k kα −g + 2 k k
m m
2 2 µν
= (k − m )g . (9.27)

The remaining integrals over y and k now lead immediately to the so-called
Proca equation for V µ :

∂ · ∂ V µ − ∂ µ ∂ · V + m2 V µ = J . (9.28)

This is the ‘Maxwell equation’ for massive vector fields. It is instructive to

examine this equation in empty space, that is, for J = 0. Multipying it by ∂µ ,
we find that the first two terms cancel, and we recover the Lorenz condition
∂ · V = 0 : all physical polarisations must be orthogonal to the momentum,
as we had already found. Reinserting this condition in Eq.(9.28), we are left
with the Klein-Gordon equation (∂ ·∂ +m2 )V µ = 0, which essentially requires
the particles to be on the mass shell. Note that this nicely compact way of
enforcing the Lorenz condition only works for m 6= 0 : for massless vector
particles, it must be put in by hand.
We can also write down the Lagrangian corresponding to the Proca equa-
tion, that is, that Lagrangian that has the Proca equation as its Euler-
Lagrange equation. It reads
1 1 1
L = (∂µ Vν )(∂ µ V ν ) − (∂µ Vν )(∂ ν V µ ) + m2 V µ Vµ
2 2 2
1 µν 1 2 µ
= F Fµν + m V Vµ , (9.29)
4 2
where the field strength tensor is defined as

F µν = ∂ µ V ν − ∂ ν V µ . (9.30)

9.2 The spin-statistics theorem

9.2.1 Spinorial form of vector polarisations
Although there may not (now !) appear to be a special need for it, we
can define the polarisation vectors for a massive vector particle using Dirac
254 August 31, 2019

spinors. Let the momentum of the vector particle be pµ and its mass m. We
can choose two massless momenta k1µ and k2µ that sum to pµ :

pµ = k1µ + k2µ , k1,2 2 = 0 , 2 (k1 · k2 ) = m2 . (9.31)

Three orthonormal states can now be constructed by standard-form spinors

as follows :
1
+ µ = √ u+ (k1 )γ µ u+ (k2 ) ,
m 2
1
0 µ u+ (k1 )γ µ u+ (k1 ) − u+ (k2 )γ µ u+ (k2 ) ,


=
2m
1
− µ = √ u− (k1 )γ µ u− (k2 ) . (9.32)
m 2
In fact, the longitudinal polarisation 0 µ can (by the Casimir trick, as usual)
be seen to be nothing else than
1
0 µ = (k1 − k2 )µ . (9.33)
m
This polarisation, then, is properly normalized and orthogonal to ± µ . Fur-
thermore, from Eq.(8.31) it follows that
1
+ · − = u+ (k1 )γ µ u+ (k2 ) u− (k2 )γµ k− (p1 ) = 0 . (9.34)
2m2
The polarisation vectors are therefore all orthogonal to each other. To check
the normalization of + , we write
1 1
+ · + = 2
u+ (k1 )γ µ u+ (k2 ) u+ (k2 )γµ u(k1 ) = u(k1 )γ µ k/2 γµ u+ (k1 )
2m 2m2
1 2(k1 · k2 )
= − 2 u+ (k1 )/k2 u+ (k1 ) = − = −1 . (9.35)
m m2
We can now have a look at the helicities of these polarisation states. For ±
we can compute
1 1  0 0 ~ ~ 
|± 0 |2 = 0 0


Tr ω k
/
± 1 γ k
/ 2 γ = k k + k1 · k2 , (9.36)
2m2 m2 1 2
and this only vanishes in a Lorentz frame in which the ~k1,2 are antiparallel,
so these states can only have pure helicity in frames where ~k1,2 are aligned
August 31, 2019 255

with ~q. Similarly, ~0 is only parallel to p~ if ~k1,2 are antiparallel, therefore in
general 0 is not a purely longitudinal state. Nevertheless, the representations
(9.32) form a perfectly acceptable orthonormal set of polarisations, and in
µ
particular (as we shall see) the freedom in choosing k1,2 can often be used to
good effect.
In a frame in which ~k1,2 are antiparallel we can find the helicities as
follows. Using the result of exercise 72, we can write the rotation over an
angle θ around the common direction as
uλ (k1 ) γ µ uλ (k1 ) →
 
µν 1 − cos θ µ ν µ ν

sin θ µν
cos θg + k1 k2 + k2 k1 −  (k1 , k2 ) uλ (k1 ) γν uλ (k2 )
k1 · k2 k1 · k2
(9.37)
and since we can write
i
µν (k1 , k2 )uλ (k1 ) γν uλ (k2 ) = − Tr γ 5 γ µ γ ν k/1 k/2 uλ (k1 ) γν uλ (k2 )


4
i
= − Tr γ γ u−λ (k1 )u−λ (k2 ) k/1 k/2 = −i(k1 k2 ) u−λ (k2 )γ 5 γ µ u−λ (k1 )
5 µ


2
= −iλ (k1 k2 ) uλ (k1 )γ µ uλ (k2 ) , (9.38)
this rotation results in just a phase factor exp(iλθ). E74
Before finishing this section we point out that also the (trivial) external-
E75
line Feynman factor for scalar particles can be written in terms of spinors.
µ
For a massive scalar with momentum q µ , the same choice of k1,2 is of course
possible. We simply note that √ we can always find a complex phase eiϕ such
that the external-line factor h can be cast in a form containing two spinors :
√ √ eiϕ
~ → ~ √ s+ (k1 , k2 ) . (9.39)
2 p1 · p2
It should not come as a surprise that an external integer-spin particle can
conventiently be represented by a spinor-antispinor pair. After all, this is
precisely the way in which particles like the W and Z (and, to a lesser extent,
the Higgs as well) are most often seen in experiment : namely, through their
decay into a fermion-antifermion pair.

9.2.2 Proof of the spin-statistics theorem

The treatment of the previous section may appear somewhat academic, but
it has an interesting consequence. Integer-spin particles (scalars and vectors)
256 August 31, 2019

can be represented in their external lines with an even number of spinors,

that is an even number of Dirac particles. Particles with half-integer spin
are represented by an odd number of Dirac particles. This persists : spin-3/2
particles can be formulated using 3 spinors, spin-2 particles by 4 spinors,
and so on. This implies that the interchange of two external half-integer-
spin particles involves the interchange of an odd number of Dirac particles,
and will therefore lead to a minus sign. The interchange of two external
integer-spin particles involves the interchange of an even number of Dirac
particles, and hence no minus sign. These particles, therefore, obey op-
posite statistics : integer-spin particles are bosons, half-integer spin
particles are fermions7 .

9.3 Massless vector particles

9.3.1 Polarisations of massless vector particles

Let us reconsider the helicity states of Eq.(9.20). These are defined in the
rest frame of the particle, with the understanding that we have to boost
back to the frame in which the particle moves, in our case along the z axis.
Under this boost the longitudinal polarisation takes the form of Eq.(9.21).
Let us now imagine that the particle approaches masslessness, that is, we
let m/p0 decrease towards zero. The boost necessary to reach the original
frame then becomes enormous, and the longitudinal polarisation will go to
infinity when the particle becomes massless. The only way to avoid matrix
elements becoming arbitrarily large, and hence violating unitarity sooner or
later, is to arrange the interactions of the theory in such a way that the effect
of longitudinal polarisation are suppressed by a factor of order O (m/p0 ) :
we shall use this extensively later on. In the strictly massless case, the
longitudinal polarisation vector must decouple completely, and we arrive at
the result that for massless particles, only the two states of maximal
helicity are physical8 .

7
Traditionally, the spin-statistics theorem, like the CPT theorem, is considered to be
very deep and difficult. Make up your mind.
8
This can also be proven for particles of higher spin. There is a single exception, the
so-called Kalb-Ramond state. See Appendix 15.12.
August 31, 2019 257

9.3.2 Current conservation from the polarisation

A photon is a vector particle ; as far as we know it is massless9 . Its po-
larisation vectors must therefore be transverse. For a photon moving in
the z direction,
√ any possible √
polarisation vector must be a superposition of
(x + iy) / 2 and (x − iy) / 2. If k µ is the photon momentum, and µ its
µ µ

polarisation, we must therefore have not only k ·  but also

0 = 0 , ~k · ~ = 0 . (9.40)

However, a problem immediately arises: for the above equations are not
invariant under Lorentz boosts. If we boost k µ and µ to a generically other
frame, they no longer hold. Let us assume that we are in such a frame ; there
we have the Lorentz-invariant conditions

(k 0 )2 = |~k|2 , (0 )2 − |~|2 = −1 , k 0 0 = ~k · ~ . (9.41)

We can decompose ~ into a parallel and a perpendicular part :

~ = ~k + ~⊥ , ~k // ~k , ~⊥ · ~k = 0 . (9.42)

Inserting this into the last equation of Eq.(9.41), we find immediately that
0 = |~k |, and the second equation then gives |~⊥ | = 1. We see that, whatever
the value of µ , we can always write
 0
µ µ 
 = ⊥ + kµ , (9.43)
k0
where ⊥ µ does satisfy Eq.(9.40). We can therefore have a consistent and uni-
tary theory of massless vector particles, provided that the k µ term decouples
from the physics. Now, any matrix element involving an external massless
vector particle with momentum k µ and polarisation vector µ will be of the
form
M = J (k)µ µ , (9.44)
where J µ (k) stands for the rest of the amplitude. Note that J µ does not
carry any information about µ , but it does know what k µ is, by momentum
conservation. Our requirement then is that the interactions of the theory be
such that
J µ (k) kµ = 0 . (9.45)
9
The current limit is about 10−36 GeV/c2 .
258 August 31, 2019

That is, if we replace the polarisation vector by the momentum, the ampli-
tude must vanish.

Diagrammatically, we may indicate the replacing of polarisation by mo-

mentum by attaching a handlebar 10 to the external line, so that we may
write 1111111
0000000
0000000
1111111 1111111
0000000
0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
=M , = Mc→k .
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
(9.46)
0000000
1111111 0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111 0000000
1111111

We shall use the convention that the momentum under the handlebar is
counted outgoing. The requirement for strictly massless external vector par-
ticles then becomes 0000000
1111111
1111111
0000000
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
=0 . (9.47)
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111

What, finally, is the physical content of the requirement ? This is simply

answered if we let our massless vector particle be a photon. The object J µ
is then seen as a source of photons, that is, an electromagnetic current11 . If
we now briefly return from a momentum-language formulation to a position-
language one, we see that the Fourier transform of the requirement (9.45) is
written as
∂µ J (x)µ = 0 . (9.48)
We see that our requirement is nothing but current conservation in the case
of electromagnetism ! The fact that electric charge is conserved ensures that
longitudinally polarised photons are safely absent from our experience12 .

A similar message can be gotten from the propagator. After all, the
massive-vector propagator
−g µν + k µ k ν /m2
i~
k 2 − m2
clearly becomes horribly singular at m = 0. The solution, as before, is to
require that in our theory the k µ k ν term should drop out. There is a catch,
10
The term ‘handlebar’ is my own, biker’s privilege. You might use the term ‘taking
the divergence’ but that obscures the very mechanical way by which we investigate the
suitability of amplitudes.
11
One may for instance have the source J represent a charge whose momentum changes,
thereby emitting radiation.
12
Whether they exist is another question ; at any rate we cannot produce them, nor
observe them.
August 31, 2019 259

however: whereas external vector particles must be on the mass shell, the
momentum of internal lines is off the mass shell. We therefore arrive at the
sharper requirement that Eq.(9.47) must hold even if the particle is off-shell.

9.3.3 Handlebar condition for massive vector particles

Let us examine the situation where a vector particle does have a mass, but
the mass m is very small compared to the vector particle’s energy E or its
momentum. Clearly, it would be unacceptable13 if the limit m → 0 would be
singular while the case m = 0 is not14 . We shall therefore require that, for
massive vector particles partaking in a process at high energy, the handlebar
condition (9.47) holds in a milder form :
1111111
0000000
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
= O (m) . (9.49)
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111

The meaning of this condition is the following. The longitudinal polarisation

vector of a massive vector boson has energy behaviour different from its two
transverse ones : it grows at high energy E  m with an extra power of E.
If for transverse polarisation the amplitude is well-behaved at high energy it
may not be so for longitudinal polarisation. The requirement implied by the
handlebar condition is that the extra power E inserted into the expression
because of longitudinal polarisation is softened, by cancellations over at least
one order of magnitude in terms of E/m. We shall presently see that this
condition is sufficiently strict to determine, to a large extent, the possible
couplings of a theory containing such particles.

9.3.4 Helicity states for massless vectors

The spinor-based helicity states for massive vector particles of section 9.2.1
are apparently not well suited to the massless case. Note, however, that we
may generalize the method of Eq.(9.31) as follows :

q µ = p1 + α p 2 , p1,2 2 = 0 , m2 = 2α (p1 · p2 ) . (9.50)

13
Or at least embarassing — after all, we do not know for certain if the mass of the
photon is strictly zero or just a paltry 10−137 kilograms.
14
Note that we do not even insist that m → 0 gives the same result as m = 0, only that
the limit is nonsingular.
260 August 31, 2019

Using the fact that the spinors of massless particles are homogeneous of
degree 1/2 in the argument :
√
u± (αp2 ) = α u± (p2 ) , (9.51)

we see that (for instance) the polarisation vector + can be written, in analogy
to Eq.(9.32), as
1
+ µ = √ u+ (p1 ) γ µ u+ (p2 ) . (9.52)
2 p1 · p2
Since α does not occur in the polarisation vector, we may consider the limit
α → 0. In that case, q = p1 is massless, and the only condition on the
massless vector p2 is that (p1 · p2 ) must not vanish. By a judicious choice of
overall complex phase, this leads us to propose, for a massless vector particle
with momentum k µ , states of definite helicity as follows, where the spinors
are again in the standard form :

uλ (k)γ µ uλ (r)
λ µ = √ , λ=± . (9.53)
λ 2 s−λ (k, r)

Here, the vector rµ is an arbitrarily chosen massless vector not parallel to

k µ ; it is called the gauge vector. We can ascertain that
1
+ · − = u+ (k)γ µ u+ (r) u− (r)γµ u− (k) = 0 , (9.54)
4k · r
Also,
1 −1
+ · + = u+ (k)γ µ u+ (r) u+ (r)γµ u+ (k) = u+ (k) r/ u+ (k) = −1 .
4k · r 2k · r
(9.55)
These, then, are acceptable helicity states.
A few useful properties of these polarisation vectors are
√ √
λ 2 λ 2
ωλ /λ = uλ (r)uλ (k) , ω−λ /λ = u−λ (r)u−λ (k) (9.56)
s−λ (k, r) s−λ (k, r)

and √
k/ /λ = λ 2 u−λ (k) uλ (k) , (9.57)
and this object is explicitly gauge-invariant.
August 31, 2019 261

9.3.5 The massless propagator : the axial gauge

We can perform the sum over the physical polarisation states of a massless
vector from the helicity states :
X X 1
λ µ λ ν = uλ (k)γ µ uλ (r) uλ (r)γ ν uλ (k)
λ=± λ=±
4k · r
X 1 1
= uλ (k)γ µ r/γ ν uλ (k) = Tr (/k γ µ r/ γ ν )
λ=±
4k · r 4k · r
1
= −g µν + kµ rν + rµ kν .


(9.58)
k·r
The form of the massless vector propagator in which only physical degrees of
freedom propagate is therefore given by the following Feynman rule :

µν µ ν µ ν
µ k ν ↔ i~ −g + (k r + r k )/(k · r) massless internal lines
k 2 + i

Feynman rules, version 9.3 (9.59)

Note the appearance of the arbitrary vector r. This way of writing the
propagator is called the axial gauge. The propagator is explicitly orthogonal
to rµ whatever the value of k : The vector r acts as an ‘axis’ with respect
to which the field is always orthogonal, hence the name. The fact that
the vector r is arbitrary is of course bothersome, in the same way that the
arbitrariness of the representation chosen for the Dirac matrices in the case
of Dirac particles is bothersome. We solve it in the same way, by insisting
that we ought to be able to remove r from the final expressions for matrix
elements. This can of course not be by virtue of any property of r itself, but
must come from the handlebar condition, since every term containing r also
contains k. Two things are worthy of remark here. In the first place, the
propagator is homogeneous of degree zero in r, so any result cannot depend
on the length of r anyway. In the second place, in contrast to the propagator
proposed before, with pµ pν /m2 , the propagator in the axial gauge does not
diverge.
262 August 31, 2019

9.3.6 Gauge vector shift

Let us consider helicity states for massless vector particles as defined in
sect.9.3.4. We shall denote these by λ µ (k, r). If we change the gauge vector r
from one value into another, another perfectly acceptable helicity state is ob-
tained. What is the relation between these states ? To answer this we simply
compute the difference between the states with different gauge vector :

λ µ (k, r1 ) − λ µ (k, r2 ) =
λ uλ (k)γ µ uλ (r1 ) uλ (k)γ µ uλ (r2 )
 
√ −
2 s−λ (k, r1 ) s−λ (k, r2 )


λ u−λ (r1 )γ u−λ (k)u−λ (k)uλ (r2 ) + u−λ (r1 )uλ (k)uλ (k)γ µ uλ (r2 )
µ
= −√
2 s−λ (r1 , k)s−λ (k, r2 )
µ µ
λ u−λ (r1 )(γ k/ + k/ γ )uλ (r2 ) √ s−λ (r1 , r2 )
= √ =λ 2 k µ . (9.60)
2 s−λ (k, r1 ) s−λ (k, r2 ) s−λ (k, r1 ) s−λ (k, r2 )
we see that the two states differ only by the vector particle’s momentum. In
any current-conserving set of diagrams we may therefore choose the gauge
vector at will ; there is no risk of picking up a phase difference if two different
gauge vectors are used for two different current-conserving sets of diagrams.
As an illustration of how the gauge vector can disappear from a current-
conserving object, let us consider
 
p q
λ · − ,
2k · p 2k · q
with p and q two massless momenta. The form of section 9.3.4 turns this
into
 
λ sλ (k, p)s−λ (p, r) sλ (k, q)s−λ (q, r)
√ −
2 s−λ (k, r) 2k · p 2k · q
 
λ s−λ (p, r) s−λ (q, r)
= √ −
2 s−λ (k, r) s−λ (p, k) s−λ (q, k)
λ s−λ (p, r)s−λ (q, k) − s−λ (q, r)s−λ (p, k)
= √ (9.61)
2 s−λ (k, r)s−λ (p, k)s−λ (q, k)
Now, the Schouten identity tells us that

s−λ (p, r)s−λ (q, k) + s−λ (p, k)s−λ (r, q) = −s−λ (p, q)s−λ (k, r) (9.62)
August 31, 2019 263

so that the gauge vector indeed drops out, and

 
p q λ s−λ (p, q)
λ · − = −√ . (9.63)
2k · p 2k · q 2 s−λ (k, p)s−λ (k, q)
One can easily check that the same form is obtained without using the
Schouten identity if we choose either r = p or r = q. E76

9.4 Exercises
Excercise 69 Polarisation averages
If, for vector particles, we take not a spin sum but a spin average, we find
3
pµ p ν
 
1X µ ν 1 µν
j ¯j = −g + 2
3 j=1 3 m

for an on-shell vector particle of momentum p and mass m. We can also

consider it classically, in the rest frame where p~ = 0. In that case there is
a polarisation vector µ = (0,~) (we only consider real polarisations). Show
that, if we also average the polarisation vector over all possible orientations,
we find
pµ pν
Z  
1 µ ν 1 µν
dΩ   = −g + 2
4π 3 m
as well.

Excercise 70 A character-building calculation

Let bµ = Λ(p; q)µ ν aν for some arbitrary vector a. Use the explicit form
(9.11) of the minimal Lorentz transform to prove by explicit calculation that
b · b = a · a.

Excercise 71 Rewriting the unity

Proce that using the unit vectors t, x, y, and z we can write

δ µ ν = tµ tν − xµ xν − y µ yν − z µ zν

Excercise 72 Another minimal Lorentz transform ?

µ
Let k1,2 be two massless vectors and θ an angle, and define the tensor
1 − cos θ µ ν sin θ µν
Λµν = cos θ g µν + k1 k2 + k2µ k1ν −


 (k1 , k2 )
k1 · k2 k1 · k2
264 August 31, 2019

Show that Λµ ν is the minimal Lorentz transformation that is a rotation along

their common direction in any frame in which ~k1,2 are parallel or antiparallel.

Excercise 73 Checking helicities

For the polarisaton vectors +,0,− defined in section 9.1.4, verify that

(Tz )µ ν + ν = ~+ µ , (Tz )µ ν 0 ν = 0 , (Tz )µ ν − ν = −~− µ

Excercise 74 Another helicity check

We can check the helicity of the polarisation representation of Eq.(9.32)
another way.

1. Show that the Clifford-space operator

sin(α/2) 5 0
Σ = cos(α/2) + i γ [γ , p/]
2|~p|

satisfies ΣΣ = 1, and that it is a minimal Lorentz transformation de-

scribing a rotation around the axis p~ over an angle α (it may help to
consult Eq.(7.132)).

2. Show that this holds in any Lorentz frame.

µ
3. Let k1,2 be two massless vectors with k1µ + k2µ = pµ and let ~k1 and −~k2
point in the direction of p~. Show that

Σu+ (k1 ) = exp(−iα/2) u+ (k1 )

and
Σu+ (k2 ) = exp(+iα/2) u+ (k2 )

4. Show that, therefore, the object u+ (k1 )γ µ u+ (k2 ) acquires a phase exp(iα)
under this rotation.

Excercise 75 An orthogonality proof

For a massive vector particle with momentum pµ and mass m, choose two
µ
massless vectors k1,2 with k1 + k2 = p, and define, as in Eq.(9.32),

1 1
µ± = √ u± (k1 ) γ µ u± (k2 ) , µ0 = (k1µ − k2µ )
m 2 m
August 31, 2019 265

1. Use the Chisholm identity to prove that

p/ /+ /− − /− /+ p/ = 2 γ 5 (/k1 − k/2 )

2. Use this to show that the vectors pµ and ±,0 form an orthogonal set

Excercise 76 Current-conserving radiation factor

Check the statement below Eq.(9.63).
266 August 31, 2019
Chapter 10

Quantum Electrodynamics

10.1 Introduction
In this chapter we shall start to work our way to realistic theories about the
actual elementary particles encountered in nature1 . Almost all elementary
particles seen so far have nonzero spin2 . We shall defer the discussion of
charged spin-1 particles to a later chapter ; at this point we shall only dis-
cuss how to set up a consistent theory of spin-1/2 particles (charged leptons
and/or quarks) and photons. This is the theory of quantum electro-dynamics,
or QED.

10.2 Constructing QED

10.2.1 The QED vertex
Since the propagators of spin-1/2 particles and of the massless spin-1 photon
have already been fixed, the only ingredient which we still have to determine
is the coupling between them ; and on this coupling rests the burden of
ensuring the current-conservation requirement as embodied in Eq.(9.47). The
vertex that connects two Dirac particles must have one upper, and one lower
1
It is of course possible that the elementary particles discussed in this text are not truly
elementary and that a yet deeper level of substructure is at some moment discovered. In
that case, please insert in whatever follows the addendum (as of August 31, 2019).
2
The Higgs particle is the only known exception.

267
268 August 31, 2019

Dirac index3 : and since the photon is involved, it must also carry a Lorentz
index. The simplest, and – as we shall see – indeed the correct form of the
vertex is that of a Dirac matrix. We therefore propose the following Feynman
rule :

Q QED vertex
µ ↔ i γµ
~
Feynman rules, version 10.1 (10.1)

Here Q is the strength of the fermion-photon coupling : the charge of the

fermion4 . By dimensional analysis, we see that is has dimension

dim Q = dim ~−1/2 .

   
(10.2)

The Dirac delta function imposing momentum conservation is implied. As

is conventional, we shall employ wavy lines to indicate photons. This choice
of vertex can only been argued to be reasonable if the photon current is
conserved, so this is now our task.

10.2.2 Handlebars : a first look

Let us now start to investigate the requirements of current conservation for
our theory. One of the simplest possible processes is the decay of a photon
into a fermion-antifermion pair, shown below :
p1

q p2

Of course the photon has to be off-shell here, but that is no problem since also
off-shell photons must obey current conservation. The part of the amplitude
3
Since we cannot form constant (i.e. momentum-independent) objects with one Dirac
index, or with, say, two upper Dirac indices, we are essentially forced to have conservation
of the orientation of Dirac lines in each vertex. This is modified to some extent if Majorana
particles are present in the theory.
4
Or, rather, it is related to the charge. The precise form of this relation must, of course,
be established by investigating the coupling in a well-defined physical situation.
August 31, 2019 269

depicted is given by5

M = i Q u(p1 )γ µ v(p2 ) , (10.3)
where the index µ of the photon is coupled to a corresponding index some-
where else in the larger Feynman diagram. Let us now attach the handlebar,
so that we get
p1

q p2

With the convention, to which we shall try to adhere, that the momentum
assigned in the handlebar must be counted outgoing from the vertex, so in
this case should read −q, the handlebarred M becomes

Mc = −iQ u(p1 ) q/ v(p2 ) . (10.4)

Note that we indicate the handlebar algebraically by the symbol c. Now we

apply momentum conservation that tells us that q = p1 + p2 :


Mc = −iQ u(p1 ) p/1 + p/2 v(p2 ) . (10.5)

To the expression in the middle we add zero in a clever way :

Mc = −iQ u(p1 ) p/1 − m + p/2 + m v(p2 ) , (10.6)

where m is the mass of the fermion. Now, we know that the spinors u(p1 )
and v(p2 ) satisfy the Dirac equations

(/p1 − m)u(p1 ) = 0 and (/p2 + m)v(p2 ) = 0 (10.7)

for on-shell momenta, so that half of the expression 10.6 ‘cancels to the left’
and the other half ‘cancels to the right’, and we end up with

Mc = 0 . (10.8)

We shall see that this is the general mechanism by which unitarity and cur-
rent conservation are ensured.

The above is of course only the simplest example of current conservation

in QED, and in the following we shall in fact study all conceivable QED
5
√
The 1/~ from the vertex cancels against the ( ~)2 from the external Dirac lines.
 270 August 31, 2019

process at once, but already we can learn a few useful things. In the first
place, a possible alternative coupling, with γ 5 γ µ instead of γ µ , is ruled out
since we cannot obtain two Dirac equations :

γ 5 q/ = −/p1 γ 5 + γ 5 p/2 = −(/p1 ± m)γ 5 + γ 5 (/p2 ± m) , (10.9)

so that either the cancellation to the left would be spoiled, or that to the
right. In the second place, it is necessary that both fermions have pre-
cisely the same mass. Since all known different fermion types have different
masses, this means that the QED interaction must conserve fermion type, or
‘flavour’. Electromagnetic muon decay, µ → eγ, is therefore forbidden, not
by conservation of the electric charge (which is indeed the same for muons
E78 and electrons) but by conservation of the whole electromagnetic current6 .

10.2.3 Handlebar diagrammatics

The argument for current conservation in the previous section went through
because both fermions were on their mass shell. Since fermions in internal
lines in Feynman diagrams are not on the mass shell, we have to extend our
E77 approach to off-shell fermions. Consider an arbitrary diagram in which a
fermion of mass m propagates and couples to a photon,

where the fermion momenta p and q are indicated and for the photon mo-
mentum k we have k µ = (p − q)µ . The momenta p and q may be on-shell
(in which case the corresponding blob isn’t actually there), but any of them
may be off-shell, and hence leads into a further piece of Feynman diagram. In
that case the blobs stand for the other vertices, where the fermion is created
and absorbed7 . The part of the diagram between the blobs is
Qγ µ
    
q/ + m p/ + m
i~ 2 i i~ 2 ,
q − m2 ~ p − m2
6
Fortunately, the decay µ → eγ has never been observed, and the branching ratio is
smaller than about 10−11 .
7
Actually, the p and q lines are attached to a semi-connected graph rather than two
separate connected ones, but here the distinction is irrelevant.
August 31, 2019 271

where µ is the index belonging to the photon line ; in an actual process, µ

may be coupled to the photon’s polarisation vector if the photon is external,
or to the photon’s propagator if the photon happens to be an internal line.
In case p, say, is on-shell we have to write
Qγ µ  √ 
   
q/ + m
i~ 2 i u(p) ~ .
q − m2 ~
Let us now put the handlebar on the photon leg,

p q

and see what happens. Algebraically, we must multiply the above expression
by kµ , and then
   
iQ 2 q/ + m µ p/ + m
(i~) γ 2 kµ =
~ q 2 − m2 p − m2
 
iQ q/ + m p/ + m
= (i~)2 2 2
(/p − q/) 2
~ q −m p − m2
   
iQ 2 q /+m p/ + m
= (i~) 2 (/
p − m) − (/
q − m)
~ q − m2 p2 − m 2
   
iQ q/ + m p/ + m
= (i~)2 2 2
− 2 . (10.10)
~ q −m p − m2
We see that under the handlebar the double propagator splits up into two
single ones. Note that, for this to be possible, it is again essential that
the mass of the fermion does not change at the vertex. We may write this
operation diagrammatically as

p q = − , (10.11)

where we have introduced two new diagrammatic ingredients: a slashed

fermion line, with a trivial Feynman rule :

↔ i~ , (10.12)
272 August 31, 2019

and a new vertex, also carrying a trivial rule :

Q
↔ . i (10.13)
~
The handlebarred photon line is replaced by a dotted line which evaluates
trivially to unity, but we do not want to leave it out of the diagram since
the dashed propagator still carries momentum, so that without it momentum
conservation would not hold at the new vertex. Like the handlebar this rule
is not intended to represent some physical interaction, but serves only as a
computational device. For external Dirac lines we find even simpler rules,
since the external spinors satisfy the Dirac equation :

= 0 for on-shell external lines , (10.14)

where the external line may belong to the initial or final state, and the arrow
orientation may be also reversed. An important result follows immediately
from the triviality of our new Feynman-rule tools :

= . (10.15)

10.2.4 The Ward-Takahashi identity

We shall now prove that the Feynman rule (10.2.1) is a good one, in the
sense that a handlebar on any photon gives a zero result, both for on-shell
(external) and off-shell (internal) photon lines. This is quite a tall order,
since we have to consider a literal infinity of possible processes. We shall
base the proof on - what else ? - the SDe’s of the theory. Throughout this
section we shall use semi-connected diagrams only. Let us consider a general
Green’s function that contains r fermion lines flowing in, and s fermion lines
flowing out, together with any number (≥ 1) of photon lines, one of which
we single out :
p1 pr q1
qs

≡ (10.16)
k
The fermions’ momenta are reckoned along their respective arrows, and the
photon momentum k is counted outgoing. Note that, since we are considering
August 31, 2019 273

here a Green’s function and not an amplitude, any of the external lines may
be off-shell. It is this object that we shall submit to a handlebar operation.
The SDe’s of QED are, in our notation :

= + ,

= + ,

= + . (10.17)

We therefore have

= , (10.18)

and the handlebar operation gives us

= − . (10.19)

Each term on the right-hand side can be subjected to its own SDe, to give

qj
X pj
X
= −
j j

+ r;s − r;s . (10.20)

By virtue of Eq.(10.15) the last two terms cancel precisely, and we are left
with the Ward-Takahashi identity :

qj
X pj
X
= − . (10.21)
j j
274 August 31, 2019

We can conveniently express this in a more analytic form. Let us denote our
Green’s function by

Gµ (p1 , . . . , pr ; q1 , . . . , qs ; k) ≡ , (10.22)

where the explicit Lorentz index is that of the photon line, and the same
Green’s function, only with the special photon removed, by

G0 (p1 , . . . , pr ; q1 , . . . , qs ) ≡ . (10.23)

Taking into account the flow of the momenta and the fact that the two trivial
Feynman rules (10.12) and (10.13) together yield a factor of (iQ/~)(i~) =
−Q, we can write the diagrammatic Ward-Takahashi identity (10.21) as fol-
lows8 :

Gµ (p1 , . . . , pr ; q1 , . . . , qs ; k) kµ =
X s
Q G0 (p1 , . . . , pr ; q1 , . . . , qj + k, . . . , qs )
j=1
Xr
−Q G0 (p1 , . . . , pj − k, . . . , pr ; q1 , . . . , qs ) . (10.24)
j=1

It is this result that proves that, indeed, the choice of the vertex (10.2.1)
leads to an acceptable theory.
So far we have considered the general case, with no constraint on the
external momenta. If we now specialize to amplitudes, in which all external
momenta except perhaps for k µ , must be on their mass shell, the rule (10.14)
applies, and we find the even more attractive Ward identity :

= 0 . (10.25)
8
In many, or even most, cases of interest fermion number is conserved, which means that
in every (fundamental or effective) vertex the number of incoming and outgoing fermions
is the same ; in that case we have r = s. But since we have nowhere used this, the Ward-
Takahashi identity may be expected to hold also for processes in which fermion number
is not conserved, for example in supersymmetry where Majorana fermions occur. Note,
however, that Majorana fermions are necessarily neutral and themselves do not couple to
photons.
August 31, 2019 275

10.2.5 The charged Dirac equation

We still have to determine the precise relation between the coupling constant
Q in the Feynman rule, and the classical electric charge q of the particle. We
shall do this by establishing a relation with classical electrodynamics. The
classical (i.e. non-loop) SDe for ψ in the presence of a photon field A is given
by
11111
00000
00000
11111
00000
11111
00000
11111
00000
11111
111111
000000 00000
11111
00000
11111
000000
111111
000000
111111 00000
11111
00000
11111
000000
111111
000000
111111
000000
111111
000000
111111
= 00000
11111
00000
11111
00000
11111
, (10.26)
000000
111111
000000
111111 00000
11111
000000
111111 00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111

in other words
Z Z  
1 −ik·(x−y) k/ + m Q
ψ(x) = d4 y 4
dke i~ 2 i γµ ψ(y)Aµ (y) ,
(2π)4 k − m2 + i ~
(10.27)
whence

i/∂ − m + QA(x)
/ ψ(x) = 0 , (10.28)
which is the Dirac equation in the presence of an electromagnetic field. Let us
work this expression towards classical physics. In the first place, the deriva-
tive is, by the standard assignment rules for quantum mechanics, related to
the momentum operator :
pµ = i~ ∂ µ , (10.29)
and the mass m to the mechanical mass M by (as we have seen)
Mc
m= . (10.30)
~
The Dirac equation can therefore be written as

(pµ + ~QA(x)µ )γµ − M c ψ(x) = 0 ,

(10.31)

which is to be compared with the standard expression for the electromagnetic

momentum if a charged particle in classical electrodynamics:
q
pem µ = pµ − Aµ . (10.32)
c
where q is the classical charge of the particle and Ac the classical electromag-
netic field. In the Gaussian system of units, the charges have dimensionality
dim[q 2 ] = kg m3 /sec2 and the Coulomb field strength E therefore obeys
276 August 31, 2019

dim[E] = dim[q] /m2 . Since this is the gradient of the classical e.m. vector
potential Ac we have dim[Ac ] = dim[q] /m, and because the photon field
A has dimensionality dim[A2 ] = kg/sec, it follows that the correct relation
between the photon field and the classical e.m. field must read

Ac 2 = c A2 . (10.33)

From this it follows that the coupling Q and the charge q are related by
√
Q = −q/(~ c) , (10.34)
√
which implies the correct dimensionality dim[Q] = dim[1/ ~] ; moreover,
we find immediately that, for particles with unit electric charge,
4π
Q2 = α , (10.35)
~
where α stands for the electromagnetic fine structure constant :

α ≈ 1 / 137.036 . (10.36)

Since in QED every next loop order contains two extra powers of Q and
one (effective) power of ~, the loop expansion is in QED equivalent to an
expansion in powers of α.
What we have√done here is a pure √ dimensional analysis, it does not decide
between Q = q/~ c and Q = 2q/~ c. Later on, the discussion of Thomson
scattering will assure us that we have made the right choice. Of course, any
extra numerical factor in Q can be compensated for by a rescaling of A.

10.2.6 The Gordon decomposition

Consider a charged Dirac particle that scatters by emitting (or absorbing) a
single photon. The corresponding current reads
ie
Jµ = u(q) γ µ u(p) , (10.37)
~
where p is the incoming, and q the outgoing momentum. By the properties
of the Dirac spinors we can write this as
ie
Jµ = u(q) q/γ µ + γ µ p/ u(p) .


(10.38)
2m~
August 31, 2019 277

Since
q/γ µ = q µ − iqα σ αµ , γ µ p/ = pµ + ipα σ αµ , (10.39)
the current takes the form
ie
Jµ = u(q) (p + q)µ + i(p − q)α σ αµ u(p) .


(10.40)
2m~
This is called the Gordon decomposition : the vertex is split up into a piece
that we shall later recognize as the vertex by which a scalar particle interacts
with photons, which is called the convection term, and a tensorial part, called
the spin term. Both terms vanish individually under the handlebar operation.
The Gordon decomposition is especially useful in nonrelativistic situations,
where q µ and pµ are both dominated by their time components. In that
case u(q)u(p) ≈ 2m, and the vertex becomes simply (p + q)µ . Note that for
antispinors we find exactly the same result, namely v(p)γ µ v(q) ≈ (p + q)µ in
the nonrelativistic situation9 .

10.2.7 Furry’s theorem

An interesting observation concerns closed fermion loops in QED. Let us
consider a fermion loop that is attached by three QED vertices to the rest of
a Feynman diagram:
k ν
p2
D− = µ p3 (10.41)
p1
λ

Here, we have indicated the Lorentz indices on the photon lines, and the
momenta across the photon lines are considered incoming into the loop. In
addition to this diagram, there is also a similar diagram in which the orien-
tation of the loop is reversed :
k ν
p2
D+ = µ p3
p1
λ
9
q v(q) = −mv(q), say, which gives us a minus sign :
Since the Dirac equation reads /
on the other hand, v(p)v(q) now evaluates to approximately −2m, thereby cancelling the
minus sign.
278 August 31, 2019

Note that these graphs cannot be twisted into one another. For loops with
only one or two vertices they can be so twisted, and then do not count as
separate diagrams ; for three or more vertices, there are two distinct ones.
Without pretending to evaluate the whole loop, let us concentrate on the
Dirac structure of their numerators. The first diagram contains the trace10

D− → Tr (/k + m) γ µ (/k − p/1 + m) γ λ (/k + p/2 + m) γ ν

≡ T− , (10.42)

whereas the corresponding trace for the other diagram reads

D+ → Tr (−/k + m) γ ν (−/k − p/2 + m) γ λ (−/k + p/1 + m) γ µ

≡ T+ .
(10.43)
Note that the rest of the loops, and in particular the propagator denomina-
tors, are identical for both graphs. By using the reversibility inside traces of
Clifford algebra elements, we can write

T+ = − Tr (/k − m) γ ν (/k + p/2 − m) γ λ (/k − p/1 − m) γ µ

= − Tr (/k − m) γ µ (/k − p/1 − m) γ λ (/k + p/2 − m) γ ν

= − T− , (10.44)

since no terms with an odd power of m survives the trace. We see that
the two loops cancel each other precisely ! This can obviously be extended
to loops with more vertices, and we find Furry’s theorem : fermion loops
with an odd number of vector vertices11 and opposite orientation
cancel each other ; with an even number of vector vertices, they
are identical12 . Furry’s theorem does not hold if one or more of the ver-
tices are of axial-vector type, and so it is not generally valid for the weak
interactions. For QCD, in which the quark-gluon couplings have the Dirac-
matrix form as in QED, Furry’s theorem holds in a more restricted form :
the spacetime part of the two quark loops with even(odd) number of vertices
are equal(opposite), but the additional colour structures of the diagrams are
different. This implies, for instance, that the two quark loops with three
gluon vertices do not cancel completely. We shall come back to that case
later on.
10
By the rules of Dirac particles, closed loops automatically evaluate to traces.
11
That is, vertices consisting of a single Dirac matrix, such as in QED.
12
Furry’s theorem is usually proved by invoking the charge-conjugation matrix, discussed
in section 15.11.2. However, as we see this is not necessary.
August 31, 2019 279

Now consider a diagram (or set of diagrams) N with only one single
external line which is a photon :
p
N ≡ , (10.45)

where we have indicated the momentum of the photon. Such objects go

under the name of tadpoles 13 . By Lorentz covariance we see that, whatever
goes on inside the blob, N must always be of the form

N µ = pµ f (p2 ) . (10.46)

But . . . by momentum conservation, pµ = 0 since no momentum is coming

out on the other side of the blob. Therefore N = 0 ; photon tadpoles vanish
identically, even if Furry’s theorem does not apply in this case. It is easy to
see that the same must hold for the tadpoles of any other vector particle. For
scalar particles it does not apply, since N = f (0) has no particular reason to
vanish. Also, for particles that are described not by vectors but by symmetric
tensors, the tadpole N µν = g µν f (0) does not necessarily vanish14 .

10.3 Some QED processes

10.3.1 A classic calculation : muon pair production
We are now in a position to compute, for the first time, a realistic cross
section. We shall follow the classic steps that lead to our final result.

Description of the process with momentum assignments

The simplest calculation is that of the cross section for muon pair production
in e+ e− collisions:

e− (p1 ) e+ (p2 ) → µ− (q1 ) µ+ (q2 ) .

since it is described, at tree level, by only a single diagram15 .

13
Of course, ‘spermatozoon’ would be more appropriate. . .
14
This is not the graviton, since that is described by a symmetric traceless tensor ; it is
in fact another representation of a scalar particle.
15
It might be thought that processes involving only electrons would be simpler, but as
we shall see these contain always at least two diagrams.
280 August 31, 2019

Drawing and writing out the diagram(s)

The single lowest-order Feynman diagram is given by
p q1
1

M = (10.47)
q2
p
2

Both the electron and muon are Dirac particles. We shall denote the electron
charge by Qe , and the muon charge by Qµ , and their masses by me and mµ ,
respectively. The total invariant mass squared is conventionally denoted by
s, and of course momentum is conserved :

p1 α + p 2 α = q 1 α + q2 α , s = (p1 + p2 )2 = (q1 + q2 )2 . (10.48)

The amplitude corresponding to the Feynman diagram is

~Qe Qµ
M=i v(p2 ) γ α u(p1 ) u(q1 ) γα v(q2 ) , (10.49)
s
and is strictly dimensionless: dim[M] = dim[1], as it ought to be for a
2 → 2 process at tree order.

Squaring and averaging, eliminating Dirac stuff

In a typical muon pair production process, we shall accept muons with any
polarisation in the final state ; also, usually the beams of incoming electrons
and positrons are unpolarised. The amplitude, squared and averaged over
the incoming electron and positron spins can be evaluated using the Casimir
trick :
1 X
|M|2 = |M|2


4
spins
~2 Qe 2 Qµ 2 X
= v(p2 )γ α u(p1 ) u(p1 )γ β v(p2 )
4s2
spins
X
× u(q1 )γα v(q2 ) v(q2 )γβ u(q1 )
spins
August 31, 2019 281

~2 Qe 2 Qµ 2
Tr (/p2 − me )γ α (/p1 + me )γ β


= 2
4s

Tr (/q1 + mµ )γα (/q2 − mµ )γβ
4~2 Qe 2 Qµ 2
p2 α p1 β + p1 α p2 β − (p1 · p2 )g αβ − me 2 g αβ


=
s2
q1 α q2 β + q2 α q2 β − (q1 · q2 )gαβ − mµ 2 gαβ

4~2 Qe 2 Qµ 2

= 2(p1 · q1 )(p2 · q2 ) + 2(p1 · q2 )(p2 · q1 )
s2

2 2
+ s(me + mµ ) (10.50)

Choosing a Lorentz frame, working out dot products

We shall work in the centre-of-mass frame of the colliding electron-positron
pairs. In that frame, we shall write
   
E E
 0 
pµ1 =   , pµ2 =  0  ,
 
 0   0  (10.51)
p −p
and
   
E E
q1µ = 
q sin θ cos ϕ  µ  −q sin θ cos ϕ
   
 q sin θ sin ϕ  , q2 =  −q sin θ sin ϕ
 , (10.52)

q cos θ −q cos θ
where
s = 4E 2 , p2 = E 2 − me 2 , q 2 = E 2 − mµ 2 . (10.53)
The various vector products are therefore given by
(p1 · p2 ) = 2E 2 − me 2 , (q1 · q2 ) = 2E 2 − mµ 2 ,
(p1 · q1 ) = (p2 · q2 ) = E 2 − pq cos(θ)
(p1 · q2 ) = (p2 · q1 ) = E 2 + pq cos(θ) , (10.54)
where θ is the polar scattering angle, that is, the angle between p~1 and ~q1 .
We also use the fact that Qµ and Qe are the negative of the unit charge, so
that Qµ Qe = 4πα/~.
282 August 31, 2019

The final transition rate

We now arrive at
16π 2 α2

|M|2 = s2 (1 + cos(θ)2 ) + 4s(me 2 + mµ 2 ) sin(θ)2


s2

2 2 2
+ 16me mµ cos(θ) (10.55)

The differential cross section

Using what we have already learned about the flux factor and the two-body
phase space, we can write the differential cross section as
1/2
s − 4mµ 2

1
2

dσ = |M| dΩ
64π 2 s s − 4me 2
1/2 
α2 s − 4mµ 2 me 2 + mµ 2

2
= 2
(1 + cos(θ) ) + 4 sin(θ)2
4s s − 4me s
2 2

me mµ
+ 16 cos(θ)2 dΩ . (10.56)
s2
This cross section therefore only depends on s and the polar scattering angle:
there is, for unpolarised incoming beams, no azimuthal direction singled out
and there is therefore no azimuthal angle dependence16 .

The total cross section

The total cross section is obtained by simple angular integration, and reads
1/2
4π α2 me 2 mµ 2 s − 4mµ 2
    
σ= 1+2 1+2 . (10.57)
3s s s s − 4me 2
The cross section is only nonzero above the muon pair-production threshold,
s > 4mµ 2 . Since the muon mass mµ is much larger than the electron mass
me , we may accurately approximate by putting me ≈ 0 :
1/2
4π α2 mµ 2 mµ 2
  
σ≈ 1+2 1−4 . (10.58)
3s s s
16
This could be different, e.g. in the case of transversely polarised beams.
August 31, 2019 283

Figure 10.1: Threshold behaviour for fermions and scalars

For large s, furthermore, we have

4π α2 mµ 4
 
σ≈ 1 − 6 2 + ··· . (10.59)
3s s

By accidental cancellation of the leading mµ 2 /s terms, the large-s limit is

reached quite rapidly. Note that the cross section does not depend on ~. For
massless incoming electrons and final-state particlesp of mass m,2 the cross sec-
tion at beam energy E = xm goes as (1 + 1/(2x)) 1 − 1/x2 /x for fermions,
but as (1 − 1/x2 )3/2 /x2 for scalars in the final state. Studying the threshold
behaviour of the cross section can therefore inform us about the spin of newly
pair-produced particles : the peak cross sections differ by a factor of about
2.8. as shown in figure 10.1

10.3.2 Compton and Thomson scattering

We next consider the Compton scattering process, an elastic collision between
a photon and an electron :

e− (p) γ(k1 ) → e− (q) γ(k2 )

284 August 31, 2019

Now, there are two Feynman diagrams,

p p
q k2
M = + (10.60)
k1 k2 k1 q
The amplitude is given by

M = M1 + M2 ,
A1
M1 = −i~Qe 2 , A1 = u(q) /2 (/p + k/1 + m) /1 u(p) ,
2(p · k1 )
A2
M2 = +i~Qe 2 , A2 = u(q) /1 (/q − k/1 + m) /2 u(p) ,
2(q · k1 )
(10.61)

where 1,2 are the polarisation vectors of the respective photons. Taking into
account the averaging factor 1/4, we find17 (with m for me )
1
|A1 |2 Tr (/q + m) γ α (/p + k/1 + m) γ β (/p + m) γβ (/p + k/1 + m) γα



=
4
= 16m4 − 8(pq)m2 + 8(pk1 )(qk1 ) + 16(pk1 )m2 − 8(qk1 )m2 ,
1
|A2 |2 = Tr (/q + m) γ β (/q − k/1 + m) γ α (/p + m) γα (/q − k/1 + m) γβ



4
= 16m4 − 8(pq)m2 + 8(pk1 )(qk1 ) + 8(pk1 )m2 − 16(qk1 )m2 ,
hA1 A∗2 i = hA2 A∗1 i
1
Tr (/q + m) γ α (/p + k/1 + m) γ β (/p + m) γα (/q − k/1 + m) γβ


=
4
= 8(pq)(pk1 ) − 8(pq)(qk1 ) + 16(pq)m2 − 8(pq)2
−4(pk1 )m2 + 4(qk1 )m2 . (10.62)

We can most easily evaluate this in the photon-electron centre-of-mass frame18 .

In this frame, we have
s + m2 K
p0 = q 0 = √ , |~p| = |~q| = |~k1 | = |~k2 | = √ , (10.63)
2 s 2 s
17
Both the incoming electron and the incoming photon have 2 degrees of freedom, hence
(1/2)(1/2)=1/4.
18
In the actual experiment, the photon will of course be impingeing on the stationary
electron ; but since the cross section is invariant we may choose any frame we want.
August 31, 2019 285

where K = 2(pk1 ) = s − m2 : and the angle between ~q and ~k1 is denoted by

θ. Putting everyhting together, we find
 2

2
 2 2 8m 8m4 2m4 4m2
|M| = 16π α + 2 + −
K K (qk1 )2 (qk1 )
8m4

K 4(qk1 )
− + + . (10.64)
(qk1 )K (qk1 ) K

The phase space integration element is given by

1 1 K
dV (p + k1 ; q, k2 ) = dΩ , (10.65)
(2π)2 8 s

where Ω is the solid angle of the emitted electron. The flux factor is

1 1
2 1/2
= . (10.66)
2λ(s, m , 0) 2K

The only nontrivial quantity in the computation is

   
0 0 K 2
(qk1 ) = k1 q − |~q| cos θ = (s + m ) − K cos θ , (10.67)
4s

and we can find the angular averages

K(s + m2 )
Z
1
dΩ (qk1 ) = ,
4π 4s
Z  
1 1 2s K
dΩ = log 1 + 2 ,
4π (qk1 ) K2 m
Z
1 1 4s
dΩ = . (10.68)
4π (qk1 )2 m2 K 2

We therefore have for the transition rate, now also averaged over the scat-
tering angle :

m2 m2 s

2 2 2


|M| = 16π α 1 + + 16 2
s K
2 4
   
ms ms s K
+ −8 2 − 16 3 + 2 log 1 + 2 (10.69)
K K K m
286 August 31, 2019

The total cross section

1

|M|2

σ= . (10.70)
16πs
It is interesting19 to note that the ‘static’ limit K → 0 is well-defined :

8π α2
lim σ = . (10.71)
K→0 3 m2
This is called the Thomson cross section. It may serve as the ‘measurement’
prediction by which the electric charge of the electron is defined.

Note that, just as in the case of muon pair production, the cross section
does not depend on ~. This means that this cross section had better coincide
with the prediction from classical electromagnetism — as, indeed, it does.
Returning to the arguments that led us to make the identifications
q
Acl 2 = c A2 , Q= √ , (10.72)
~ c

we see that we have made the correct choice.

10.3.3 Electron-positron annihilation

The process
e+ (p1 ) e− (p2 ) → γ(k1 ) γ(k2 ) (10.73)
is related by crossing to Compton scattering, and is described at the tree
level by the two Feynman diagrams
p2 p2
k1 k2
M = + . (10.74)
p1 k2 p1 k1

We shall study it in the context of the way it is actually observed at high-

energy e+ e− colliders, that is, in the centre-of-mass frame with the photons
emerging an nonnegligible angles with respect to the electron and positron
beams. In that case, no invariant vector products are small, and we may
19
And comforting.
August 31, 2019 287

neglect the electron mass. We then have an example of a process in which

helicity-spinor techniques can be usefully employed. We shall do that, slowly
and explicitly. The amplitude is given by

M(ν, λ1 , λ2 ) = i~Qe 2 (A1 (ν, λ1 , λ2 ) + A2 (ν, λ1 , λ2 )) ,

(/p2 − k/1 )
A1 (ν, λ1 , λ2 ) = uν (p1 ) /λ2 (k2 ) /λ1 (k1 ) uν (p2 ) ,
2(p2 k1 )
(/p2 − k/2 )
A2 (ν, λ1 , λ2 ) = uν (p1 ) /λ1 (k1 ) /λ2 (k2 ) uν (p2 ) . (10.75)
2(p2 k2 )
Since me = 0 we may as well employ the symbol u for both the positron and
the electron. Also, the helicity of the electron fixes that of the positron, and
both are indicated by ν. The helicities of the two photons are denoted by
λ1,2 . We start by taking ν = +. Then Eq.(9.56) tells us that we then have
to use
√ √
2 − 2
/+ (kj )ω+ = u− (kj )u− (r+ ) , ω− /− (kj ) = u− (r− )u− (kj ) ,
s− (kj , r+ ) s+ (kj , r− )
(10.76)
µ
where we may choose the gauge vectors r± as we please, for each photon and
for each polarisation case individually. The best choice is then

r+ = p2 , r− = p1 (10.77)

since it makes lots of terms vanish. Indeed, if λ1 = λ2 = + both A’s contain

u− (p2 )u+ (p2 ) as a factor, and for λ1 = λ2 = − we likewise find u+ (p1 )u− (p1 )
as a common factor, and therefore

M(ν, λ, λ) = 0 (10.78)

which puts paid to 4 out of 8 helicity configurations20 . Of course the fact that E79
these amplitudes vanish cannot depend on our choice of r± , but the proof
that they vanish is now very simple ! The next case is that of λ1 = −λ2 = +.
We see that A1 still vanishes, and we can compute A2 as follows (using
(p2 k2 ) = (p1 k1 )) :
√ √
1 2 − 2
A2 =
2(p1 k1 ) s− (k1 , p2 ) s+ (k2 , p1 )
20
For ν = 0 it suffices to interchange s+ and s− , and include a minus sign for each
polarization vector.
288 August 31, 2019

× u+ (p1 )u− (k1 )u− (p2 ) (/p2 − k/2 ) u− (p1 )u− (k2 )u+ (p2 )
u+ (p1 )u− (k1 )u− (p2 ) k/2 u− (p1 )u− (k2 )u+ (p2 )
=
(p1 k1 ) s− (k1 , p2 )s− (k2 , p1 )
s+ (p1 , k1 )s− (p2 , k2 )s+ (k2 , p1 )s− (k2 , p2 )
=
(p1 k1 ) s− (k1 , p2 )s− (k2 , p1 )
s− (k2 , p2 )2
= −2 (10.79)
s− (k1 , p1 )s− (k1 , p2 )

The case λ1 = −λ2 = − amounts to simply interchanging k1 and k2 . The

four nonvanishing amplitudes are therefore

−2i~Qe 2 s−ν (k2 , p2 )2 −2i~Qe 2 s−ν (k1 , p2 )2

M(ν, ν, −ν) = , M(ν, −ν, ν) = ,
s−ν (k1 , p1 )s−ν (k1 , p2 ) s−ν (k2 , p1 )s−ν (k2 , p2 )
(10.80)
and the spin-averaged transition rate is
 

2
 2 2 k1 · p1 k2 · p1
|M| = 32π α + . (10.81)
k2 · p1 k1 · p1

The computation of the cross section is left as an excercise. We have dis-

cussed this process, rather, to show how spinor techniques may be usefully
employed to compute amplitudes for massless-particle processes in a fast
and efficient manner ; moreover, we can gain results (such as the vanishing
of the amplitude when the photons helicities are equal) that are not so easily
obtained by more traditional approaches21 .

10.3.4 Bhabha scattering

Our final 2 → 2 QED process is that of Bhabha scattering, the process

e+ (p1 ) e− (p2 ) → e+ (q1 ) e− (q2 , ) , (10.82)

21
A word of caution is in order here. The Minkowski products (pi kj ) can become small
if the photons are emitted collinearly. In that case these products are of order m2 rather
than of order s. It is therefore not adviseable to blindly put m = 0 in any process in which
photons are emitted, since then we might miss terms looking like m2 /(pi kj )2 . As can be
seen from the matrix element for Compton scattering, in this case the double-pole term
is actually suppressed by m4 rather than by m2 , and therefore at high energies we do not
have to worry about double poles for this process. For other Bremsstrahlung processes
such as e+ e− → µ+ µ− γ, the double poles are important : see section 10.3.5.
August 31, 2019 289

described by the two following Feynman graphs :

p1 q1 p1 q1
M = + (10.83)
p2 q2 p2 q2

We shall use, in addition to s, the following conventional invariants :

t = (p1 − q1 )2 = (p2 − q2 )2 , u = (p1 − q2 )2 = (p2 − q1 )2 . (10.84)

For me = 0 we have s + t + u = 0 by momentum conservation. As before, E80

we shall work in the high-energy limit so that me is neglected. The helicity-
dependent amplitude is

M(λ1 , λ2 , ρ1 , ρ2 ) = i~Qe 2 A(λ1 , λ2 , ρ1 , ρ2 ) ,

1
A(λ1 , λ2 , ρ1 , ρ2 ) = uλ (p1 ) γ µ uλ2 (p2 ) uρ2 (q2 ) γµ uρ1 (q1 )
s 1
1
− uλ (p1 ) γ µ uρ1 (q1 ) uρ2 (q2 ) γµ uλ2 (p2 ) .(10.85)
t 1
Note the relative minus sign between the two diagrams ! By the Chisholm
identity, we can now evaluate the various helicity configurations :
2 2
A(+, +, +, +) = s+ (p1 , q2 )s− (q1 , p2 ) − s+ (p1 , q2 )s− (p2 , q1 )
s  t
u2

1 1
∼ 2u + ∼ 2 ,
s t st
2 t
A(+, +, −, −) = s+ (p1 , q1 )s− (q2 , p2 ) ∼ 2 ,
s s
2 s
A(+, −, +, −) = − s+ (p1 , p2 )s− (q2 , q1 ) ∼ 2 , (10.86)
t t
where the symbol ∼ denotes our throwing away unimportant complex phases.
The other helicity configurations with λ1 = + give zero, and those with
λ1 = − follow again trivially by conjugation. We find
4 4 4
2
3 + cos2 θ

2 2 4 s +t +u 2 2


|M| = 2~ Qe = 16π α , (10.87)
s2 t2 1 − cos θ
where θ is the angle between p~1 and ~q1 in the centre-of-mass frame in which
most e+ e− scattering experiments are performed. Note that, in this case,
 290 August 31, 2019

the singularity is not due to our neglecting the electron mass ; indeed, for
nonzero mass we have
t = (p1 − q1 )2 = 2m2 − 2(p1 0 )2 + 2|~p1 |2 cos θ
= −2|~p1 |2 (1 − cos θ) . (10.88)
E81 To this order in perturbation theory, the total cross section for Bhabha
scattering is therefore indeed divergent.
E82

10.3.5 Bremsstrahlung in Mœller scattering

The nonradiative process
Mœller scattering is the mutual scattering of two electrons :
e− (p1 ) e− (p2 ) → e− (q1 ) e− (q2 )
and is just a crossed version of Bhabha scattering. The relevant expression
is therefore, for negligible electron mass,
s4 + t4 + u4
|M|2 = 2~2 Qe 4


(10.89)
s2 u 2

The radiative process

We shall now consider the so-called Bremsstrahlung22 process :
e− (p1 ) e− (p2 ) → e− (q1 ) e− (q2 ) γ(k)
At the tree level, it is described by the eight Feynman diagrams
k q1
q1 q1
q1 p1
p1 p1 p1
M = p2
k + p2 + k + p2
p2 q2
q2 q2 q2 k
k q1
q1 q1
q1 p2
p2 p2 p2
k + p
+ k + p1 ,
p1 1 p1 q2
q2 q2 q2 k
(10.90)
22
The term originated in studies of the motion of charged particles through a medium ;
they may lose energy by emitting photons, and slow down, or ‘brake’, or – in the language
of early-twentieth-century physics, which was German rather than American English –
‘Bremsen’.
August 31, 2019 291

which we may conveniently put in four groups of two diagrams each :

4
X
M = Mi ,
i=1
√
 
3 q/1 + k/ + me α /1 − k/ + me
αp
M1 = −i(Qe ~) u(q1 ) / γ −γ / u(p1 )
2q1 · k 2p1 · k
1
× u(q2 )γα u(p2 ) ,
(p2 − q2 )2
M2 = M1 cp1 ↔p2 , q1 ↔q2 ,
M3 = − M1 cp1 ↔p2 , M4 = − M2 cp1 ↔p2 . (10.91)

Note the Fermi minus sign between M1,2 and M3,4 . The four pairs of dia-
grams are separately current-conserving, i.e.

Mi c→k = 0 , i = 1, 2, 3, 4 . (10.92)

The Soft-photon approximation

Since the emitted photon is a massless particle, its energy can be arbitrarily
low. A useful result can be obtained if we take this limit, that is, the photon
energy is taken to be negligible with respect to the other particle energies.
Consider an arbitrary process in which a fermion with momentum q and mass
m is produced during a scattering :

q
M0 = A (10.93)

this amplitude can be written as

√
M0 ≡ ~ u(q) A(q) , (10.94)

where A denotes the rest of the diagram(s). The corresponding radiative pro-
cess will (amongst others) contain diagrams in which the photon is emitted
by this particular fermion :
q
Ms = A (10.95)
k
292 August 31, 2019

which evaluates to
√ q/ + k/ + m
Ms ≡ −(Qe ~) u(q) / A(q + k) . (10.96)
2q · k
Notice that the denominator q · k goes to zero as the photon energy vanishes,
and hence the diagram diverges in the limit of soft Bremsstrahlung . In
the soft-photon approximation ( and assuming that the object A does not
depend on q in too drastic a manner23 ) we have
√ q/ + m
Ms ≈ −(Qe ~) u(q) / A(q) . (10.97)
2q · k
Anticommuting / and q/, and using the property of the Dirac spinor, which
tells us that u(q)/q = mu(q), we then find
√ q·
Ms ≈ −(Qe ~) u(q)A(q) , (10.98)
q·k
that is, the diagram factorizes into the nonradiative result and an ‘infrared
factor’24 . We can repeat this procedure for those diagrams in which the
photon is emitted by the other external particles. There are, of course, also
(possibly) diagrams in which the photon is emitted from internal lines ; but,
as can easily be checked, such diagrams do not diverge as k 0 → 0. In the
soft-photon approximation, they do therefore not contribute. For radiative
Mœller scattering, we therefore have the nicely factorized form
√
M = −(Qe ~) (VIR · ) M0 ,
µ q1µ q2µ pµ1 pµ1
VIR = + − − , (10.99)
q1 · k q2 · k p1 · k p1 · k
where M0 is the amplitude for the nonradiative process ; and, using the
polarisation sum rule Σµ ν = −g µν , we find
s4 + t4 + u4
|M|2 = −2Qe 6 ~3


(VIR · VIR ) ,
t2 u2
m2e m2e m2e m2e
−VIR · VIR = − − − −
(p1 k)2 (p2 k)2 (q1 k)2 (q2 k)2
23
This assumption fails, for instance, close to a resonance. However, since every reso-
nance has a finite width, the soft-photon approximation is formally correct for infinitesimal
photon energies.
24
Since infrared light has low energy compared to visible light.
August 31, 2019 293

2(p1 p2 ) 2(q1 q2 ) 2(p1 q1 )

+ + −
(p1 k)(p2 k) (q1 k)(q2 k) (p1 k)(q1 k)
2(p1 q2 ) 2(p2 q1 ) 2(p2 q2 )
− − − . (10.100)
(p1 k)(q2 k) (p2 k)(q1 k) (p2 k)(q2 k)

As has already been intimated, the double poles are indeed suppressed by a
factor me 2 .

Hard Bremsstrahlung: massless case

Next, we consider ‘hard Bremsstrahlung’ (i.e. any photon emission which is
not soft) in the limit of vanishing electron mass. It is then most useful to
assign definite helicities to the electrons, so that the scattering process is

e− (p1 , µ1 ) e− (p2 , µ2 ) → e− (q1 , ν1 ) e− (q2 , ν2 ) γ(k, λ)

with µ1,2 , ν1,2 , λ = ±. The amplitude is then a function of the helicities, and
we write M(µ1 , µ2 ; ν1 , ν2 ; λ). We first consider M1 (+, +; +, +; +). Using
Eq.(9.56) this can be written as
√ √
(Qe ~)3 2
M1 (+, +; +, +; +) = i
2(p2 · q2 )s− (k, r)
 
q/1 + k/ α /1 − k/
αp
× u+ (q1 ) u− (k)u− (r) γ −γ u− (k)u− (r) u+ (p1 )
2k · q1 2k · p1
× u+ (q2 )γα u+ (p2 ) , (10.101)

and since M1 is current-conserving by itself we may choose r at will ; in this

case r = p1 appears to be optimal since it kills the second term. Applying
standard (hopefully, by now) spinor techniques we arrive at

M1 (+, +; +, +; +) =
√ √ s+ (q1 , k)u− (p1 )(/q1 + k/ )u− (q2 )s− (p2 , p1 )
i(Qe ~)3 8 . (10.102)
(2p2 · q2 )(2k · q1 )s− (k, p1 )

We may employ momentum conservation and masslessness for a further ma-

nipulation :

u− (p1 )(/q1 + k/ )u− (q2 ) = u− (p1 )(/q1 + k/ + q/2 )u− (q2 )

= u− (p1 )(/p1 + p/2 )u− (q2 ) = s− (p1 , p2 )s+ (p2 , q2 ) , (10.103)
294 August 31, 2019

so that

√ √ s− (p1 , p2 )2
M(+, +; +, +; +) = i(Qe ~)3 8 . (10.104)
s− (p2 , q2 )s− (k, p1 )s− (k, q1 )

Finally, we can make use of the identity of Eq.(9.63) to arrive at the form

√ s− (p1 , p2 )2
 
3 + · p1 + · q1
M1 (+, +; +, +; +) = −2i(Qe ~) − .
k · p1 k · q1
s− (p1 , q1 )s− (p2 , q2 )
(10.105)
The infrared factor also appears in this case ! Performing the appropriate
subtitutions we can write the complete amplitude as

√
M(+, +; +, +; +) = −2i(Qe ~)3 s− (p1 , p2 )2 (VIR · + )
 
1 1
× − . (10.106)
s− (p1 , q1 )s− (p2 , q2 ) s− (p1 , q2 )s− (p2 , q1 )

The minus sign in the last term is the Fermi sign ; it helps us to simplify our
expression even further using the Schouten identity, and the final form for
the amplitude is

√ s− (p1 , p2 )3 s− (q1 , q2 ) (VIR · + )

M(+, +; +, +; +) = 2i(Qe ~)3 .
s− (p1 , q1 )s− (p2 , q2 )s− (p1 , q2 )s− (p2 , q1 )
(10.107)
For the other helicity configurations, the above treatment can be repeated
straightforwardly. The resulting nonvanishing amplitudes turn out to all
have quite similar forms :

M(µ1 , µ2 ; ν1 , ν2 ; λ) =
(VIR · λ ) σ s−λ (a, b)3 s−λ (c, d)
2i~3/2 Q3 , (10.108)
s−λ (p1 , q1 )s−λ (p2 , q2 )s−λ (p1 , q2 )s−λ (p2 , q1 )

where the sign σ and the identities of the four vectors a, b, c and d are given
in the following table :
August 31, 2019 295

ν1 µ1 ν2 µ2 λ σ a b c d
+ + + + + + p1 p2 q1 q2
+ + + + – + q1 q2 p1 p2
+ + – – + + q2 p1 q1 p2
+ + – – – – q1 p2 q2 p1
+ – – + + – q2 p2 q1 p1
+ – – + – + q1 p1 q2 p2
– – – – + – q1 q2 p1 p2
– – – – – – p1 p2 q1 q2
– – + + + + q1 p2 q2 p1
– – + + – – q2 p1 q1 p2
– + + – + – q1 p1 q2 p2
– + + – – + q2 p2 q1 p1
A few things may be noted, apart from the surprising simplicity of the am-
plitudes. In each case, only s+ or s− occur. Moreover, when the electrons
have equal spins the amplitudes are explicitly antisymmetric in p1 ↔ p2 or
q1 ↔ q2 , as required by Fermi statistics25 . The spin-averaged matrix element
squared therefore has the following form in the strictly massless case :
|M|2 me =0 = −2Qe 6 ~3 (VIR · VIR )



ss0 (s2 + s02 ) + uu0 (u2 + u02 ) + tt0 (t2 + t02 )

× (10.109)
,
uu0 tt0
with s = (p1 +p2 )2 , s0 = (q1 +q2 )2 , t = (p1 −q1 )2 , t0 = (p2 −q2 )2 , u = (p1 −q2 )2 ,
and u0 = (p2 − q1 )2 . The final result is surprisingly simple. It consists
of the ‘soft-photon’ factor VIR 2 (evaluated for non-soft photon momenta),
multiplying a ‘symmetrized’ form of the nonradiative cross section.

Double-pole terms at high energy

We have already mentioned that putting me = 0 strictly may be too strict
since there are invariant products of momenta that may become equally
small. To see how this works, let us again inspect the radiation emitted from
a produced fermion, as given in figure 10.95, that can be written as
√ q/ + k/ + m
Mc ≡ −(Qe ~) u(q)/ A(q + k) (10.110)
2q · k
25
If the incoming electrons have unequal spin they are distinguishable, and no antisym-
metry is required.
296 August 31, 2019

where, as before, A stands for the rest of the diagram(s). We shall not
assume the soft-photon limit, however. Let us assume that the photon is
emitted as small angle θ with respect to the fermion momentum. We then
find, assuming the fermion energy to be large compared to its mass m :

(k · q) = k 0 q 0 − |~q| cos θ


 2 !
1 me
≈ k 0 (q 0 − |~q|) + |~q|θ2 /2 ≈ k 0 q 0 θ2 +


(10.111)
,
2 q0

where we have used the fact that q 0 − |~q| = me 2 /(q 0 + |~q|) ≈ me 2 /(2q 0 ).
we conclude that as soon as θ is of order me /q 0 or smaller26 , the product
(k · q) becomes of order m2e ; and this means that in that case the ‘single
pole’ (k · q)−1 and the ‘double pole’ me 2 (k · q)−2 are of the same order27 . The
squared matrix element (summed over fermion and photon spins) contains
of course
Qe 2 ~
|Mc |2 = −


4(k · q)2
× A(q + k)(/q + k/ + m)γ α (/q + m)γα (/q + k/ + m)A(q + k)(10.112)

Using standard Dirac algebra we can write

(/q + k/ + m)γ α (/q + m)γα (/q + k/ + m)

= 4m2 (/q + k/ + m) + 4(k · q)(m − k/ ) . (10.113)

The second term in this expression enters into the ‘massless’ result since it
will give rise only to single-pole terms, whereas the first term tells us that
the double-pole term coming from this Mc must read

2
 2 m2
|Mc | = −Qe ~ A(q + k)(/q + k/ )A(q + k) , (10.114)
(k · q)2

where we have again discarded terms of order m. The nonradiative transition

rate was given by (A)/qA(q), and in this expression we have now substituted
q+k for q. We can, by momentum conservation, always express the invariants
s, t and u in Eq.(10.89) into a form that does not contain q, and this then
26
This is sometimes called the ‘ultra-collinear case’.
27
For this reason we use the subscript c which stands for ‘collinear’.
August 31, 2019 297

gives us the double-pole terms : keeping all four collinear situations in sight,
we can write the transition rate including the double-pole terms as

2 6 2


|M| = Qe ~

ss0 (s2 + s02 ) + tt0 (t2 + t02 ) + uu0 (u2 + u02 )

 tt0 uu0
2(p1 · p2 ) 2(q1 · q2 ) 2(p1 · q1 )
× − − +
(k · p1 )(k · p2 ) (k · q1 )(k · q2 ) (k · p1 )(k · q1 )

2(p2 · q2 ) 2(p1 · q2 ) 2(p2 · q1 )
+ − −
(k · p2 )(k · q2 ) (k · p1 )(k · q2 ) (k · p2 )(k · q1 )
me 2 2(s04 + t04 + u04 ) me 2 2(s04 + t4 + u4 )
− −
(k · p1 )2 t02 u02 (k · p2 )2 t2 u2
me 2 2(s4 + t04 + u2 ) me 2 2(s4 + t4 + u02 )

− − (10.115)
(k · q1 )2 t02 u2 (k · q2 )2 t2 u02

This is our final expression for unpolarised Mœller scattering ; it is accurate

in the limit of small me even for collinear28 photon emission. E83

Radiation cones and Yennie’s crater

For every charged fermion with momentum p in a scattering process with
emission of a photon with momentum k over an angle θ between p~ and ~k
there are terms in the cross section that go as

1 1 2
= 0 0 ∼
, (10.116)
p·k k (p − |~p| cos θ) k 0 p0 (m/p0 )2 + θ2

where we have assumed that m/p0 and θ are small. This has been the reason
for statements such as ‘the radiation is essentially confined to a cone of angle
m/p0 around the fermion direction’, the so-called radiation cone. This is E84
quite wrong since the polar angular integration variable is d cos θ ∼ d(θ2 )/2
and not dθ.
Suppose that the photon is strictly parallel to the fermion, ~k//~p. Since
the photon polarisation is strictly orthogonal, ~ · ~k = 0, we must also have
~ · p~ = 0, and the soft-photon approximation 10.99 then tells us that the
diagram with denominator (p · k) vanishes in the soft-photon limit : the
298 August 31, 2019

angular distribution shows a sharp dip known as Yennie’s crater. For photons E85
that are not infinitesimally soft, the crater is less deep but still there.

10.4 Scalar electrodynamics

10.4.1 The vertices
We can also consider the possibility of interactions between photons and
charged scalar particles29 . The simplest vertex is then given by
q
µ
p

where the charge flow is indicated by the arrow. The photon index is µ. The
momenta p and q are counted along the arrow. Note that the propagator
of scalar particles may be unoriented, but the vertices do not have to be so
if there is a quantum number, such as charge, that distinguishes between
particle and antiparticle. Since no Dirac matrices occur the only quantities
in this vertex that carry a Lorentz index are the momenta p and q (and
of course the photon’s own momentum, but that is fixed by p and q). We
therefore propose a Feynman rule of the form
q Q
µ ↔ i (c1 pµ + c2 q µ ) ,
p ~
with constants c1,2 to be determined. This is simple, since we can study the
annihilation of the charged scalar-antiscalar pair into an off-shell photon :
under the handlebar operation, the amplitude becomes
p1 √
= iQ ~ (c2 p2 µ − c1 p1 µ ) kµ
p2 k √
= iQ ~ (c2 p2 µ − c1 p1 µ ) p1 µ + p2 µ

√ (c2 − c1 )
= iQ ~(p1 + p2 )2 . (10.117)
2
28
Or ultra-collinear.
29
Elementary charged scalar particles have to date not been observed, although they
are predicted in extensions of the standard model. We include them here since they will
provide indications on how to treat charged vector particles.
August 31, 2019 299

We see that c1 = c2 is required, and therefore the first Feynman rule for
scalar electrodynamics (sQED) reads

µ ↔ i Q (p + q)µ sQED vertex

q
p ~

sQED Feynman rules, version 10.1 (10.118)

Let us now consider the more complicated process of annihilation into two
on-shell photons. With the above vertex two diagrams are involved :

p1 k1 p1 k2
M = k2 + k1 (10.119)
p2 p2

The amplitude is then given, with m indicating the scalar’s mass, by

(p1 + (p1 − k1 )) · 1 ((p1 − k1 ) + (−p2 )) · 2

M = −i~Q2 + (k1 ↔ k2 )
(p1 − k1 )2 − m2
 
2 (p1 · 1 )(p2 · 2 ) (p1 · 2 )(p2 · 1 )
= −2i~Q + (10.120)
(p1 · k1 ) (p2 · k1 )

The test of current conservation now fails, since

Mc1 →k1 = −2i~Q2 ((p2 · 2 ) + (p1 · 2 )) = −2i~Q2 (k1 · 2 ) . (10.121)

The solution is to introduce a four-point vertex into the Feynman rules. For
reasons lost in the mists of time, such a vertex is called a sea-gull vertex30 .
The improved Feynman rules are now :

30
To me it does not look very gully nor even particularly birdy.
300 August 31, 2019

µ ↔ i Q (p + q)µ sQED 3-vertex

q
p ~

µ Q2 µν sQED 4-vertex
↔ 2i g
ν ~
sQED Feynman rules, version 10.2 (10.122)

Now we find immediately the desired current conservation :

+ + = 0 . (10.123)

You might worry that annihiliation into three photons will necessitate a five-
point vertex, and so on. Fortunately, the above two vertices are sufficient to
guarantee current conservation in all sQED processes, as we shall now show
using some more handlebar diagrammatics.

10.4.2 Proof of current conservation in sQED

Consider a charged scalar propagator somewhere in a Feynman diagram, and
assume a photon attached to it :
µ  
k i~ Q µ i~
q = 2 i (p + q) .
p p − m2 ~ q − m2
2

As in our proof for regular QED, none of these lines is necessarily on-shell. .
Momentum conservation again fixes the photon momentum to be k = p − q.
In analogy to regular QED we can now invent some handlebar diagrammatics
as follows :
(p − q) · (p + q)
= −iQ~
(p2 − m2 )(q 2 − m2 )
   
i~ Q Q i~
= 2 2
i (i~) + (i~) i
q −m ~ ~ p − m2
2

= − , (10.124)
August 31, 2019 301

with the trivial auxiliary rules

Q
= i~ , =i . (10.125)
~
These rules are very similar to those we adopted in regular QED : however,
in general we have
6= (10.126)

since the scalar-scalar-photon vertex still depends on the various momenta.

We now turn to the second vertex, with two photon lines. Not denoting the
two scalar propagators, we have
µ
Q2
k = 2i (p − q − k)sµ
p q ~
       
Q Q µ Q µ Q
= i (i~) i (2q + k) − i (2p − k) (i~) i ,
~ ~ ~ ~
(10.127)

in other words,

= − . (10.128)

The proof of current conservation again relies on the SDe’s for this model :

= + + ,

= + + ,

= + + ,

(10.129)

where again we have used semi-connected graphs. The handlebar operation

is now seen to lead to

= +
302 August 31, 2019

= −

+ −

(10.130)

If we now iterate the SDe cleverly for the first two of these four diagrams,
we obtain

= +

− −

+ −

= 0 , (10.131)

since we do have

= , (10.132)
owing to the simple, momentum-independent structure of the seagull vertex.
Comparing the lines of the proof for sQED with that of regular QED, the
general proof strategy becomes clear : if in a diagram a slashed propagator
occurs as one of the indicated lines of a (semi-)connected graph, we must
iterate de SDe for that line, and then we can collect the various canceling
contributions.

10.5 The Coulomb potential

The Feynman rules that we have constructed are designed to describe dy-
namical processes. To say something about the static character of QED we
return to our treatment of the Yukawa potential of section 5.2.4. There, we
used the result
~ x)
3~ exp(ik · ~
Z
1 1 exp(−mr)
d k = , r = |~x| (10.133)
(2π)3
|~k|2 + m2 4π r
August 31, 2019 303

to derive (up to an overall factor) the Yukawa potential. But we used a single
source and looked at the field’s response to it : now, we have to consider two
particles, say, scattering ever so slightly. Let us start with two scalars31 with
masses M1,2 and charge Qe . Using the sQED Feynman rules we find for the
tree amplitude :
p q i~Q2
1 1
M= k = (p1 + q1 · p2 + q2 ) . (10.134)
p2 q2 k 2 + i

Since p01 = q10 and p02 = q20 in the centre-of-mass frame we have k 0 = 0 and
we can write
i~Q2e (4M1 M2 )
M = M(~k) = − , (10.135)
|~k|2
where we have assumed the static limit (SL) in the numerator ; and this we
have to Fourier-transform to find the interaction energy V (r). The correct
convention (as we shall see) is to use
Z
i 1
V (r) = d3~k M(~k) exp(i~k · ~x) , (10.136)
4M1 M2 (2π)3
so that we find, for this scattering, V (r) = α/r, a repulsive interaction32 .
If we now change the direction of the charge flow in the (p2 , q2 ) line in
Eq.(10.134), the amplitude changes sign by the sQED Feynman rules, and
the interaction energy becomes attractive. Therefore, like charges repel, and
opposite charges attract33 . We may also consider the ‘original’ Yukawa inter-
action, mediated by a scalar particle. Its propagator has numerator i~ rather
than the −i~g µν for the photon, with no sign change in the interaction vertex
if we replace particle by antiparticle. We conclude that in the exchange of a
scalar particle rather than a vector one, the interaction is always attractive.
Let us now turn to the case of regular QED, and consider e− µ− scattering :
p q −i~Q2e
1 1
M(~k) = k = u(q1 )γ µ u(p1 ) u(q2 )γµ u(p2 ) . (10.137)
p2 q2 ~
|k|2

31
We shall assume that they are not identical so as to avoid trouble with having to deal
with two diagrams rather than one.
32
What about the factor 4M1 M2 ? It is just the wave function normalization factor for
two nonrelativistic particles.
33
If we had chosen the other convention in Eq.(10.136), with −i rather than i, we would
have found that like charges attract, contrary to experience. The point here is that the
interaction can change sign between like-charge and opposite-charge.
304 August 31, 2019

By the Gordon decomposition (see section 10.2.6) this evaluates to exactly

the result (10.135), hence like-sign fermions also repel each other. But let us
now consider fermion-antifermion scattering, e− µ+ : we would write
p q −i~Q2e
1 1
M(~k) = k = u(q1 )γ µ u(p1 ) v(p2 )γµ v(q2 ) , (10.138)
p2 q2 ~
|k|2

and find, from the Gordon decomposition, that opposite-sign fermions also
repel ! But we have forgotten the Fermi minus sign lurking backstage. The
correct Feynman rules are those of Eq.(7.123) rather than the ‘prettified’
ones of Eq.(7.151), and we obtain an extra sign that tells us that oppositely-
charged fermions attract each other just like oppositely-charged scalars34 .

10.6 Electrons in external fields : g = 2

10.6.1 The charged Klein-Gordon equation
A point of particular interest is the way in which electrons react to external
fields. Here ‘external’ is used for fields that are not of the fluctuating quantum
type, but rather applied ‘from outside’, under our experimental control ; this
is the sense in which it was used in section 10.2.5. In the absence of an explicit
source, the Dirac equation with external field A reads (cf. Eq.(10.28)) :
γµ (i ∂ µ − eAµ (x)) − m ψ(x) = 0


(10.139)
in the position representation, where e denotes the coupling. Let us compare
this to the behaviour of a scalar electron. For a charged scalar in an external
field we can write down a classical (tree-level) SDe
1111
0000
0000
1111
0000
1111
111111
000000
000000
111111 111111
000000
000000
111111 0000
1111
000000
111111 000000
111111 0000
1111
000000
111111 000000
111111 0000
1111
0000
1111
000000
111111
000000
111111 000000
111111
000000
111111
000000
111111 000000
111111
000000
111111
111111
000000
000000
111111 000000
111111 000000
111111
000000
111111 000000
111111
000000
111111 000000
111111 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
= 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
+ 000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
(10.140)
000000
111111 000000
111111 000000
111111
000000
111111 000000
111111 000000
111111

or, by explicit use of the Fourier transforms of the fields :

Z Z
1 i~
φ(x) = 4
dy 4
d4 k e−ik·(x−y) 2
(2π) k − m2 + i
34
There is a tendency among theorists to take the prettified rules as the ‘real’ ones, with
the Fermi sign occurring only in loops. As we see that is dangerous ! Note also that it
does not matter whether we assign the Fermi sign to the v or to the v, as indeed it should
not.
August 31, 2019 305
 Z
e 1
−i d4 pd4 qe−ip·y−iq·y (2p + q)µ Aµ (q) φ(p)
~ (2π)8
 2 
1 e µ
+ 2i Aµ (y)A (y)φ(y) . (10.141)
2 ~
Note the occurence of the symmetry factor 1/2 in the last line. Applying the
differential operator −∂ 2 − m2 , and using the expressions for derivatives in
Fourier representation we arrive at
− ∂ 2 − m2 φ(x) = 2ieAµ (x)∂µ φ(x) +

ie ∂µ Aµ (x) φ(x) − e2 Aµ (x)Aµ (x)φ(x)

(10.142)
or  

2 2
i∂ − eA(x) −m φ(x) = 0 . (10.143)

This is the Klein-Gordon equation for charged scalar fields. We see that the
same ‘minimal substitution rule’ pµ → pµ − eAµ as in the Dirac case is
employed to account for the presence of the e.m. field ; and we see that the
charge coupling constant e is defined in the same way for both scalar and
Dirac particles.

10.6.2 The relativistic Pauli equation

We can cast the Dirac equation for spin-1/2 electrons in a form closely re-
sembling that of the Klein-Gordon equation, by multiplying Eq.(10.139) :
  
µ µ ν ν



0 = γµ i∂ − eA (x) + m γν i∂ − eA (x) − m ψ(x)
 
2 2 2 2
= − ∂ + e A(x) − m ψ(x)

−ieγ µ γ ν ∂µ Aν (x)ψ(x) − ieγ ν γ µ Aν (x)∂µ ψ(x) . (10.144)

The second line can be rewritten as
 
µν µν
∂µ Aν (x) ψ(x) − 2ieAµ (x) ∂µ ψ(x)


−ie g − iσ (10.145)

We find, for spin-1/2 electrons, the relativistic Pauli equation, the Klein-
Gordon equation with an extra spin term, the so-called Stern-Gerlach term
added on : E86
 

2
i∂ − eA(x) − m2 − eσ µν ∂µ Aν (x) φ(x) = 0 .


(10.146)
306 August 31, 2019

10.6.3 A constant magnetic field

The spin of particles manifests itself most clearly in a magnetic field. Since
the notions of electric and magnetic fields are of dubious Lorentz covarience,
we adopt the following. Let tµ be a constant vector with t2 = 1, and let B µ
be a constant vector with (B · t) = 0. We now choose the external A field as
follows :
1
Aµ (x) = εµ (t, B, x) . (10.147)
2
~=B
In the ‘rest frame’ where ~t = 0 this reduces to A0 = 0 , A ~ × ~x/2, which
corresponds to a constant magnetic field of strength B.~
In the Klein-Gordon equation, the term linear in A is given by35

−2ieA(x) · ∂ψ(x) = −ie ε(∂, t, B, x) ψ(x) = e ε(t, B, x, i∂) ψ(x)

= −e Bµ εµ (t, x, i∂) ψ(x) . (10.148)

We can introduce the angular momentum operator

Lµ = εµ (t, x, i∂) (10.149)

~ = i~x × ∂~ = ~x × p~. In the rest frame we
since, for ~t = 0, this reduces to L
therefore have the interaction term
~ ·L
−2ieA(x) · ∂ψ(x) = e B ~ ψ(x) (10.150)

which represents the coupling between the magnetic field and the angular
momentum of the moving charge. For the Stern-Gerlach term, we have
e µν
−e σ µν ∂µ Aν (x) = σ εµναβ tα B β
2
= ieγ 5 σαβ tα B β = −eB
/ γ 5 t/ . (10.151)

We now make the approximation that the particle is moving nonrelativisti-

cally slow in the rest frame ~t = 0. In that case we may write, in momentum
language,
1
t/ u(p, s) → p/ u(p, s) = u(p, s) , (10.152)
m
so that the Stern-Gerlach interaction term in this limit reads

−e σ µν ∂µ Aν (x) ψ(x) = eBµ γ 5 γ µ ψ(x) . (10.153)

35
The term proportional to (∂ · A) vanishes because of the Lorenz condition.
August 31, 2019 307

Now, as we have seen in Eq.(7.111), the operator −γ 5 γ µ describes the spin

vector sµ , which is normalized to s2 = −1. However, as we have seen the
actual spin S is one-half this amount since the electron has total spin one
half. We therefore write
~ ·S
−e σ µν ∂µ Aν (x) ψ(x) = 2e B ~ ψ(x) , (10.154)
~ = −γ 5~γ /2 is the operator for the spin of a nonrelativistic electron.
where S
We see that the coupling of a (nonrelativistic) electron to an external
magnetic field is proportional to
~ + ge S)
(L ~ ·B
~ , ge = 2 ,
where the factor ge comes somewhat as a surprise : it is called the gyromag-
netic ratio of the electron36 . The prediction ge = 2 was one of the first and
most welcome results from Dirac’s description of the electron.

There is a way to pinpoint the gyromagnetic behaviour of an electron in a

more precise and useful manner, using the Gordon decomposition of section
10.2.6, which reads
 
µ 1 µ µ ν
u(q) γ u(p) = u(q) (p + q) + (p − q)ν [γ , γ ] u(p) . (10.155)
2m
Since the convection term (p + q)µ describes the coupling of a scalar to the
photon, it is the spin term which must be responsible for the result ge = 2
for the electron. The above implies that, by calculating loop corrections to
the electron-photon vertex, we can isolate the σ part of the loop-corrected
vertex and infer the loop corrections to ge .

10.7 Selected topics in QED

10.7.1 Three-photon production
A fine example of a quite nontrivial computation using standard-spinor (‘he-
licity’) techniques is provided by the process of three-photon annihiliation in
e+ e− collisions :
e+ (p1 ) e− (p2 ) → γ(k1 ) γ(k2 ) γ(k3 ) . (10.156)
36
Composite spin-1/2 particles can have different values for g : the proton and neutron
have about 5.586 and -3.826, respectively.
308 August 31, 2019

We shall again alculate this for massless electrons. At the tree level there are
6 Feynman diagrams, so that the amplitude reads

M(λ; λ1 , λ2 , λ3 ) = i~3/2 e3
(/k3 − p/1 ) (/p2 − k/1 )
× uλ (p1 )/λ3 (k3 ) /λ2 (k2 ) /λ (k1 )uλ (p2 )
2(k3 p1 ) 2(k1 p2 ) 1
+ (perm) . (10.157)

Here we have explicitly indicated the various helicities, and ‘perm’ stands
for the other 5 permutations of the photons. For λ = + we take the photon
polarisations as in section 10.3.3, and that immediately tells us that the
amplitude vanishes if λ1 = λ2 = λ3 (as was already anticipated in exercise
79). We see that the only helicity amplitude that we have to work hard on
is, say, M(+; − + +). Neglecting, for now, the overall factor i(2~e2 )3/2 we
can write

M(+; − + +) =
u− (k3 )u− (p2 ) (/k3 − p/1 ) u− (k2 )u− (p2 ) (/p2 − k/1 ) u− (p2 )u− (k1 )
u+ (p1 ) u+ (p2 )
s− (k3 , p2 ) 2(k3 p1 ) s− (k2 , p2 ) 2(k1 p2 ) s+ (k1 , p2 )
+ (perm) . (10.158)

Actually, only two diagrams contribute here, namely the one written down
and the one where k2 and k3 are interchanged. With

u+ (p1 )u− (k3 ) 1

= (10.159)
2(k3 p1 ) s− (k3 , p1 )

and
u− (p2 )(/p2 − k/1 )u− (p2 )
= −1 (10.160)
2(k1 p2 )
we can rewrite
u− (p2 )(/k3 − p/1 )u− (k2 ) s− (k1 , p2 )
M(+; − + +) = − + (k2 ↔ k3 ) .
s− (k3 , p1 ) s− (k3 , p2 ) s− (k2 , p2 ) s+ (k1 , p2 )
(10.161)
We can simplify further :

u− (p2 )(/k3 − p/1 )u− (k2 ) = u− (p2 )(/k2 + k/3 − p/1 − p/2 )u− (k2 )
= −u− (p2 )/k1 u− (k2 ) , (10.162)
August 31, 2019 309

Making the denominator symmetric in k2 and k3 gives us

M(+; − + +)
u− (p2 )/k1 u− (k2 ) s− (k2 , p1 ) s− (k1 , p2 ) + (k2 ↔ k3 )
= . (10.163)
s− (k3 , p1 ) s− (k3 , p2 ) s− (k2 , p1 ) s− (k2 , p2 ) s+ (k1 , p2 )
After yet another manipulation :
u− (p2 )/k1 u− (k2 ) s− (k2 , p1 ) + (k2 ↔ k3 )
= u− (p2 )/k1 k/2 u+ (p1 ) + (k2 ↔ k3 ) = u− (p2 )/k1 (/k2 + k/3 )u+ (p1 )
= u− (p2 )/k1 p/2 u+ (p1 ) = s− (p2 , k1 )s+ (k1 , p2 )s− (p2 , p1 ) , (10.164)
we arrive at
s− (k1 , p2 )2 s− (p1 , p2 )
M(+; − + +) = ; (10.165)
s− (k3 , p1 )s− (k3 , p2 )s− (k2 , p1 )s− (k2 , p2 )
putting it more symmetrically, and reinserting the overall factor,
2 3/2 s− (k1 , p2 )3 s− (k1 , p1 )
M(+; − + +) = i(2~e ) s− (p1 , p2 ) 3 . (10.166)
Q
s− (kj , p1 )s− (kj , p2 )
j=1

We can easily infer the other helicity amplitudes. The final answer is
3
(kj p1 )(kj p2 ) [(kj p1 )2 + (kj p2 )2 ]
P
j=1
|M|2 = 2~2 e4 (p1 p2 )


3
. (10.167)
Q
(kj p1 )(kj p2 )
j=1

Looking at this result, we notice two interesting things. In the first place,
it is very simple, something you would not have guessed right off, and cer-
tainly would have had to work on very hard using the classical Casimir-trick
approach. In the second place, Eq.(10.166) contains only s− and no s+ ,
as in the two-photon annihiliation case of section 10.3.3. This is a general
feature : conceding that M(+; +++) = 0 is the simplest possible amplitude,
the ‘next-simplest’, in our case M(+; − + +), is both simple and holomor-
phic in the spinor products37 . Such amplitudes are called maximal helicity
violating (MHV) and are an object of research in their own right, occurring
in many theories with massless particles such as QCD with massless quarks
and gluons, and even in gravity.
37
In the sense that s+ is the complex conjugate of s− up to a sign.
310 August 31, 2019

10.7.2 The Thomson limit : scalar vs spinor

It is interesting to note that the Thomson cross section of Eq.(10.71) is
actually a classical object, derivable within classical, nonrelativistic electro-
dynamics38 . It should therefore not depend on the fact that the electron is
a Dirac particle : the amplitudes must be the same for spin-0 and spin-1/2
electrons in the low-energy limit. We shall now investigate this in some de-
tail, at the tree level.

In the static limit the kinetic energies are negligible with respect to the
rest energies of massive particles. We shall, for our process

e(p) γ(k1 ) → e(q) γ(k2 ) , (10.168)

define the momenta as follows, in the centre-of-mass frame:

pµ = (E, 0, 0, k) , q µ = (E, 0, k sin θ, k cos θ) ,

k1 µ = (k, 0, 0, −k) , k2 µ = (k, 0, −k sin θ, −k cos θ) , (10.169)

where E 2 = k 2 + m2 for electron mass m, the polar scattering angle of the

photons is θ, and we have arbitrarily fixed the irrelevant azimuthal scattering
angle. The static limit (SL) is defined by k → 0 and E → m.
We start with the scalar case. The scattering is described by the three
diagrams

p q p
k2 p q
M = + + (10.170)
k1 k2 k1 k2
k1 q

in which the charge flow is indicated. The amplitude reads

2 2 p · 1 q · 2
M(λ1 , λ2 ) = −ie ~ −
(p · k1 )


!
2 q · 1 p · 2

− 2 1 · 2 . (10.171)
(q · k1 )
38
After all, Thomson didn’t know either about quantum physics at the time, nor about
relativity.
August 31, 2019 311

Here e is the electron charge. The photon helicities are denoted by λ1,2 , and
we shall use the representation
uλ (k1 )γ α uλ (k2 ) uλ (k2 )γ α uλ (k1 )
1 (λ)α = p , 2 (λ)α = p , (10.172)
4(k1 · k2 ) 4(k1 · k2 )

so that (j · p) = (j · q) (j = 1, 2) and 2 (λ) = 1 (λ)∗ . The first amplitude

we consider is
   
2 2 1 1
M(+, +) = −ie ~ 2|p · 1 (+)| − +2 . (10.173)
(pk1 ) (qk1 )
Using
Tr (ω+ k/1 p/k/2 p/) (pk1 )(qk1 ) m2
|p · 1 (+)|2 = = − (10.174)
4(k1 k2 ) (k1 k2 ) 2
and
1 1 (k1 k2 )
− =− (10.175)
(pk1 ) (qk1 ) (pk1 )(qk1 )
and we find the exact result
m2 (k1 k2 ) (E − k)(1 − cos θ)
M(+, +) = −ie2 ~ = −ie2 ~ . (10.176)
(pk1 )(qk1 ) E + k cos θ
In the SL this becomes

M(+, +)SL = −ie2 ~(1 − cos θ) . (10.177)

Since the amplitude must be dimensionless in energy units and m is the only
scale in the SL, it is not surprising that m drops out altogether. The second
amplitude is given by
 
2 1 1
2
M(+, −) = −2ie ~ − p · 1 (+)
(pk1 ) (qk1 )
(k1 k2 )
= 2ie2 ~ eiφ |p · 1 (+)|2
(pk1 )(qk1 )
2(pk1 )(qk1 ) − m2 (k1 k2 )
= ie2 ~ eiφ
(pk1 )(qk1 )
(E + k)(1 + cos θ)
= ie2 ~ eiφ . (10.178)
E + k cos θ
312 August 31, 2019

Here we have introduced the phase factor

p · 1 (+) u+ (k1 ) p/ u+ (k2 )
eiφ = ∗
= . (10.179)
p · 1 (+) u− (k1 ) p/ u− (k2 )
In the SL we arrive at

M(+, −)SL = −ie2 ~ eiφ (1 + cos θ) . (10.180)

We now turn to the spin-1/2 case. In section 10.3.2 we already computed

the cross section using the Casimir trick, but now we want the individual
spin amplitudes. The amplitude, now given by two diagrams, reads

2 /2 (λ2 ) (/p + k/1 + m) /1 (λ1 )
M = −ie ~ u(q)
2(pk1 )

/1 (λ2 ) (/q − k/1 + m) /1 (λ2 )
− u(p) . (10.181)
2(qk1 )
We do not bother to indicate the spins of the electrons explicitly, although
we shall have to discuss them later. As before, we start with
 ++
−ie2 ~ A++

A1 2
M(+, +) = − ,
(k1 k2 ) 2(pk1 ) 2(qk1 )
A++


1 = u(q) u+ (k1 )u+ (k2 ) + u− (k2 )u− (k1 ) (/p + k/1 + m) ×


u+ (k2 )u+ (k1 ) + u− (k1 )u− (k2 ) u(p) ,
A++


2 = u(q) u+ (k2 )u+ (k1 ) + u− (k1 )u− (k2 ) (/q − k/1 + m) ×


u+ (k1 )u+ (k2 ) + u− (k2 )u− (k1 ) u(p) . (10.182)

Using the Dirac equation and some algebra, we can rewrite

A++


1 = u(q) 2(qk2 )ω+ k/1 + 2(pk1 )ω− k/2 − mω+ k/1 k/2 − mω− k/2 k/1 u(p)


= u(q) 2(pk1 )/k2 − 2m(k1 k2 )ω+ u(p) ,
A++


2 = u(q) 2(qk1 )ω + k
/ 2 + 2(pk 2 )ω − k
/ 1 − mω − k
/ 1 k
/ 2 − mω + k
/ 2 k
/ 1 u(p)


= u(q) 2(qk1 )/k1 − 2m(k1 k2 )ω+ u(p) . (10.183)

Putting everything back together :

 
2 k/2 − k/1 m m
M(+, +) = −ie u(q) − ω+ + ω+ u(p)(10.184)
.
(k1 k2 ) (pk1 ) (qk1 )
August 31, 2019 313

Since


u(q) k/2 − k/1 u(p) = u(q) p/ − q/ u(p) = 0 , (10.185)
we arrive at the exact result
m(k1 k2 )
M(+, +) = −ie2 ~ u(q, s0 ) 1 + γ 5 u(p, s) ,


(10.186)
2(pk1 )(qk1 )
where we have indicated the spins. In the SL, we may in this expression
approximate q by p. The γ 5 term then drops out, and the amplitude vanishes
if s0 = −s. We conclude that M++ SL takes again the form (10.177), and that
the electron spin is not influenced in the SL. The second amplitude is given
by
 +−
−ie2 ~ A+−

A1 2
M(+, −) = − ,
(k1 k2 ) 2(pk1 ) 2(qk1 )
A+−


1 = u(q) u− (k1 )u − (k 2 ) + u + (k2 )u + (k1 ) (/p + k/1 + m) ×


u+ (k2 )u+ (k1 ) + u− (k1 )u− (k2 ) u(p) ,
A+−


2 = u(q) u+ (k2 )u+ (k1 ) + u− (k1 )u− (k2 ) (/q − k/1 + m) ×


u− (k1 )u− (k2 ) + u+ (k2 )u+ (k1 ) u(p) . (10.187)
Again using the various standard-form techniques we can establish that
A+−
1 = u(q)u− (k1 )u− (k2 )/pu− (k1 )u− (k2 )u(p)
+ u(q)u+ (k2 )u+ (k1 )/pu+ (k2 )u+ (k1 )u(p)
= eiφ u(q) ω− k/1 p/k/2 + ω+ k/2 p/k/1 u(p) ,


(10.188)
where eiφ is the phase factor of Eq.(10.179). For A+−
2 we find the exact same
result, and thus
ie2 ~ eiφ

M(+, −) = u(q) ω− k/1 p/k/2 + ω+ k/2 p/k/1 u(p) . (10.189)
2(pk1 )(qk1 )
In this expression, the denominator is of order O (k 2 ) which is already com-
pensated by the occurrence of k/1 and k/2 in the numerator. In the SL we can
therefore again replace q by p since q = p + O (k), and use
ω− k/1 p/k/2 + ω+ k/2 p/k/1 = (pk1 )/k2 + (pk2 )/k1 − (k1 k2 )/p − iγ α α (k1 , k2 , p) (10.190)
which yields 4(pk1 )(pk2 ) − 2m2 (k1 k2 ) when sandwiched between u(p, s) and
u(p, s), but zero when sandwiched between u(p, −s) and u(p, s). We see that
 314 August 31, 2019

also M(+, −)SL is the same as in the scalar case, while the spin again remains
unaffected by the Thomson scattering.
We have thus established that Thomson scattering will not distinguish
between scalar or Dirac electrons. It is interesting to note, however, that
the amplitudes depend on θ even in the SL, while of course in that precise
limit (zero photon energy) the value of θ becomes undetermined ! We see
that Thomson scattering (and consequently the determination of the elec-
tron charge by this process) is only meaningful if there is some momentum
transfer, no matter how small39 .

10.7.3 The Landau-Yang theorem

The photon polarisation revisited
As stated above, any good amplitude for processes in which a photon is
absorbed or produced must vanish under the handlebar operation. That
means that, provided the amplitude is acceptable, we may add to any photon
polarisation a piece of photon momentum. Let us consider a process with
several photons present, with momenta qi µ and polarisation vectors i µ . We
have, obviously, (qi · qi ) = (qi · i ) = 0 and (i · i ) = −1. From the above,
we see that, if we wish, we may employ instead of i the more complicated
object
(p · i ) µ
ηi µ = i µ − qi , (10.191)
(p · qi )
E87 where p is any vector not proportional to qi . This has the properties

(ηi · qi ) = (ηi · p) = 0 , ηi 2 = −1 . (10.192)

In numerous applications, η is actually more profitable to use than . But

we should note that, in any amplitude described by more than one Feynman
diagram, the shift from  to η simply means that parts of some Feynman dia-
grams are ‘transferred’ to other diagrams : the total result must, of course, be
the same. The most important difference between  and η is in the handlebar,
since η then vanishes :
ηi ci →qi = 0 ; (10.193)
39
This situation is, of course, the same in classical physics : to determine the elasticity
modulus of a spring you will have to stretch or compress it, no matter by how small an
amount.
August 31, 2019 315

therefore, any expression written in terms of η’s vanishes automatically under

the handlebar. On the one hand this is, of course, nice ; on the other hand,
it deprives us of a powerful check on the correctness of our diagrams, since
almost any mistake made in writing them out will show up as a failure under
the handlebar.

The Landau-Yang result

Although this may seem to fall somewhat outside the province of QED, we
can consider the decay of a spin-1 particle into photons. But even within
QED this can be envisaged, since we may have a bound state of electron and
positron (positronium ) that may, of course, have some angular momentum.
Such a positronium state can, unless we look really closely, be considered a
single particle40 . In its ground state, positronium comes in two varieties :
parapositronium in which the electron and positron’s spin are antiparallel
and hence form total spin zero, and orthopositronium in which the spins are
parallel, leading to a total spin of one.
Without knowing anything much about the bound-state structure of
ortho-positronium, let us consider the amplitude for its decay into a pair
of photons. Let us denote by P µ the ortho-positronium momentum (in its
rest frame), and by q1,2 and 1,2 the photon momenta and polarisations. We
shall define q µ = (q1µ − q2µ )/2. In addition, the positronium being a spin-1
particle, we need its polarisation vector 0 . Any amplitude for the decay must
necessarily be linear in 0 , 1 and 2 ; and current conservation is imposed by
replacing 1,2 by ηi = i − (P · i )/(P · qi )qi . Since also (P · 0 ) = 0, the three
polarisations (as well as the vector q) have no timelike component. Noting
that, in this case, (q · η1,2 ) = 0 as well, we see that to build an amplitude M
we actually have but a very few structures that we can use41 :

M = A1 (q ·0 )(η1 ·η2 )+A2 ε(P, 0 , η1 , η2 )+A3 ε(P, q, η1 , η2 )(q ·0 ) . (10.194)

The coefficients A are of course undetermined, but they can only depend on
P 2 , q 2 and (P · q). This last product is zero, and P 2 = −4q 2 = M 2 where M
is the positronium mass, so the A’s are effectively just constants. We now
come to the main observation : under interchange of the two photons we
40
Looking really closely would in this case mean splitting it up, and then the positronium
is gone.
41
You might be tempted to write down a term like ε(q, 0 , η1 , η2 ) but since all these
vectors have vanishing zeroth component, this Levi-Civita product is simply zero.
 316 August 31, 2019

have η1 ↔ η2 and q → −q. It is immediately seen that all possible terms in

M are antisymmetric under this operation, and hence cannot occur if we are
to have Bose statistics. It is obvious that this results holds to all orders of
perturbation theory, neither is it restricted to the case of ortho-positronium.
We conclude that a spin-1 particle cannot decay into two photons, which is
the Landau-Yang theorem. And so it is : para-positronium has a lifetime of
E88 1.25 × 10−10 seconds, while ortho-positronium, having to perform the much
more cumbersome decay into three photons, lives for as long as 1.39 × 10−7
seconds.
In the literature and most textbooks, the Landau-Yang theorem, espe-
cially when applied to positronium, appears to be based on fairly complicated
reasonings having to do with the charge-conjugation properties of the various
states. In our more simple-minded approach, we see that it is simply a con-
sequence of the relative paucity of building blocks available when you start
to imagine what a decay amplitude could look like. Indeed, as soon as you
envisage three-photon decay, a host of terms can be written down that re-
spect Bose symmetry, so that it is easily understood why three-photon decay
is not forbidden42 .

10.8 Exercises
Excercise 77 Compton Current Conservation
Consider the process
e− (p1 ) γ(k1 , 1 ) → e− (p2 ) γ(k2 , 2 )
where the momenta and polarisations are indicated, and write the two dia-
grams that describe it at tree level. Then, substitute 1 → k1 and show that
the amplitude, so treated, vanishes.
Excercise 78 Rare or impossible ? The process µ → e γ
The process
µ− (p) → e− (q) γ(k)
is not allowed in standard QED since it violates current conservation. Nev-
ertheless, it could be possible with a different interaction vertex coming from
some ‘new physics’. The amplitude would then read
√
M = i ~g u(q) cos θ + sin θγ 5 k/ / u(p)

42
In the words of Feynman, ‘everything that is not explicitly forbidden is allowed’.
August 31, 2019 317

where g is the coupling in the new physics, and the angle θ simply parametrizes
the relative weight of the two alternative couplings.

1. Show that this amplitude is current-conserving by applying the han-

dlebar.

2. Show that the coupling g must have the following dimensionality :

L
dim[g] = √
~
We shall therefore write
e
g=
Λ
with e the QED coupling constant and Λ the scale of the new physics.

3. Compute h|M|2 i using the Casimir trick and the trace identities.

4. Compute the total decay width Γ(µ → eγ).

5. Current limits on this process are expressed as

Γ(µ → eγ)
≤B , B ≈ 10−11 .
Γ(µ → all)

We shall assume that the decay µ → eνµ ν¯e is by far the dominant one.
Show that we can relate this to B as follows :
K
Λ≥ √ .
B
Compute K, and find the current lower limit on Λ.

Excercise 79 Multi-photon production

Consider the process e+ e− → n γ, the n-photon analogue of the process
10.73. Show the following :

1. There are, at the tree level, n! diagrams.

2. The amplitude vanishes if all photons have the same helicity.

3. If all photon helicities except one are equal, only (n − 1)! diagrams
contribute if we choose the gauge vectors cleverly.
318 August 31, 2019

Excercise 80 Relating Mandelstams

In a 2 → 2 scattering process, the momentum notation may read

p 1 + p2 → q1 + q2

We take the incoming particles to have mass m and the outgoing ones to
have mass M . From the conservation of energy and momentum,

pµ1 + pµ2 = q1µ + q2µ

and the definitions s = (p1 + p2 )2 , t = (p1 − q1 )2 , u = (p1 − q2 )2 , derive the

relation
s + t + u = 2m2 + 2M 2

Excercise 81 Bhabha scattering got wrong

In section 10.3.4 it is mentioned in a footnote how important the correct
application of the Fermi minus sign can be. Investigate this by re-computing
the cross section for Bhabha scattering using the wrong sign, and finding the
ratio between the two expressions ; then find the maximum ratio and the
scattering angle at which this is reached.

Excercise 82 Bhabha scattering the hard way

Compute h|M|2 i for Bhabha scattering at tree level, this time keeping me
nonzero. To do this, use the Casimir (trace) method.

Excercise 83 It’s a chore but someone’s got to do it

Compute the transition rates, that is, h|M|2 i, for the processes e+ e− →
µ+ µ− γ and e+ e− → e+ e− γ, taking the fermions to be massless, but keeping
the double pole terms in the radiative amplitude. This involves using spinor
techniques extensively!

Excercise 84 Measuring the radiation cone

At the LEP accelerator in its latest stages, the energy E of the incoming
electron beam was around 100 GeV, while the electron mass is me ∼ 5 × 10−4
GeV.

1. Show that the ‘canonical’ radiation cone angle derived from Eq.(10.116)
is about 5 × 10−6 radians.
August 31, 2019 319

2. The angular distribution of the photon around the electron is roughly

(1+δ−cos θ)−1 , where δ = m2e /(2E 2 ). Show that a cone that contains a
fraction f of the radiation is bounded by an angle α such that cos(α) =
1 − δ(2/δ)f .

3. Show that α ∼ 3 × 10−3 radians for f = 0.5, i.e. half the radiation is
emitted at larger angles than this.

Excercise 85 Exploring Yennie’s crater

Consider a neutral particle (the Z 0 boson or the Higgs) decaying into a
fermion-antifermion pair (with equal masses m but opposite charges) with
momenta p1,2 and a soft photon with momentum k.

1. Show that the partial decay width contains the soft-photon factor

2(p1 p2 ) m2 m2
S= − −
(p1 k)(p2 k) (p1 k)2 (p2 k)2

2. Let the angle between p~1 and ~k in the centre-of-mass frame have cosine
equal to c. Show that

2(E 2 + p2 ) 2m2 (E 2 + p2 c2 )
S= −
E 2 − p 2 c2 (E 2 − p2 c2 )2

where p01,2 = E and |~p1,2 | = p.

3. Show that S = 0 if c = ±1.

Excercise 86 Deriving the Pauli equation

Verify that Eq.(10.146) is correctly derived.

Excercise 87 Optimising the polarisation for Landau-Yang

Show, by explicit computation, that for massless p the transformation of
Eq.(10.191) is in fact precisely the adoption of p as the gauge vector. Note,
however, that in section 10.7.3 we may actually take p to be almost any
vector, so that η is actually more general.

Excercise 88 The Landau-Yang theorem in action

320 August 31, 2019

1. Consider the proces

e+ (p1 , s1 ) e− (p2 , s2 ) → γ(k1 , 1 ) γ(k2 , 2 )

where the polarisations and spins are indicated along with the mo-
menta. Write out the amplitude at the tree level, keeping the electron
mass m nonzero.

2. Choose the gauge vectors such that not only (k1 1 ) = (k2 2 ) = 0 but
also (k1 2 ) = (k2 1 ) = 0. Show that in that case 1,2 0 = 0 in the
centre-of-mass frame.

3. Take the static limit, in which p1 and p2 become equal. Show that in
this limit the amplitude is proportional to

µ (1 , k1 − k2 , 2 ) v(p1 , s1 ) γ 5 γµ u(p1 , s2 )

4. Show that the amplitude vanishes if s1 = s2 , but does not vanish if

s1 = −s2 . Hint: show that µ (1 , k1 − k2 , 2 ) ∝ p1 µ .
Chapter 11

Loop effects in QED

11.1 One-loop effects in QED

We shall discuss a number of one-loop calculations for QED. On the one hand,
these are of course phenomenologically relevant, and on the other hand they
form a nice demonstration of real-life quantum field theory at work, including
lots of gory detail.

11.2 The photon self-energy

We start with the photon self-energy due to a fermion loop :
k
p
F µν (m2 , p) = µ ν . (11.1)
p+k
The fermion mass is m, and its charge is denoted by Q.

11.2.1 Current conservation

First, we prove that the electromagnetic current is conserved in this diagram :
using what we have learned in section 10.2.3, we can write

= − ,

= = = . (11.2)

321
322 August 31, 2019

Therefore, pµ F µν (m2 , p) = 0 so that

F µν (m2 , p) = pµ pν − s g µν K(m2 , s) , s = p2 .


(11.3)

The quantity K, being a scalar, can only depend on p via p2 .

11.2.2 Using the optical theorem

We shall obtain information about K(m2 , s) by embedding it in a physical
amplitude. For this, we choose the annihilation of a massless charged scalar-
antiscalar pair (with charge Q0 ) into a fermion-antifermion pair1 , s(p1 ) s̄(p2 ) →
f (q1 ) f¯(p2 ), which is the absorptive part of the optical theorem (6.34) :
p2 q2 i~QQ0
Ma = p = u(q1 ) (/p1 − p/2 ) v(q2 ) . (11.4)
1 q1 s

Summing over the final-state spins we have

X p
|Ma |2 = 2~2 Q2 Q02 1 − β 2 (cos θ)2 ,


β= 1 − 4m2 /s , (11.5)
spins

where θ is the scattering angle between p~1 and ~q1 in the centre-of-mass
frame. The full sum over the final state of course includes the phase-space
integration :

~2 Q2 Q02
Z X
|Ma |2 dV (p1 + p2 ; q1 , q2 ) = β(3 − β 2 )θ(s > 4m2 ) . (11.6)
spins
12π

The forward scattering amplitude is that for s(p1 ) s̄(p2 ) → s(p1 ) s̄(p2 ) with
the fermion loop inserted :
p2 p2 ~2 Q02
p1 = (p1 − p2 )µ F µν (m2 , p1 + p2 )(p1 − p2 )ν
p1 s2
= ~2 Q02 K(m2 , s) . (11.7)

Note that the pµ pν term in Eq.(11.3) drops out here because of current con-
servation. The optical theorem 6.34 then tells us that

α Q2 ~
< K(m2 , s) = − β(3 − β 2 ) θ(s > 4m2 ) , α= . (11.8)
6~ 4π
 August 31, 2019 323

89 The imaginary part of K can now be computed using the Kramers-Kronig

relations of appendix 15.15.9, just as we did in section 6.6.2 :
Z∞
1 < K(m2 , z)
= K(m2 , s) = − P dz . (11.9)
π z−s
4m2

The z integral diverges since < K(m2 , z) goes to a constant for z → ∞.

Therefore we cannot make much sense of K, but we can figure out how it
depends on s by subtracting its value at s = p
0, say. This subtracted disper-
sion relation is best computed by using v = 1 − 4m2 /z as an integration
variable, and writing β 2 = 1 − 4m2 /s :
Z∞
α s v(3 − v 2 )
= K(m2 , s) − = K(m2 , 0) = P dz
6π~ z(z − s)
4m2
Z1 2
β (3 − β 2 )

α 2 2
= P dv +3−β −v
3π~ v2 − β 2
0
β(3 − β 2 )
 
α 8 2
β + 1
= −β − log 
  (β 2 > 0) ,
3π~ 3 2 β − 1
  
α 8 2 2 1
− β − |β|(3 − β ) arctan (β 2 < 0) . (11.10)
3π~ 3 |β|

11.2.3 Getting the divergence

We still have to compute K(m2 , 0). Eq.(11.1) can be written as2
−Q2 T (p, k)
Z
µν 2
F (m , p) = 4
d4 k 2 ,
(2π) (k − m + iη)((p + k)2 − m2 + iη)
2

T (p, k) = Tr (/k + m)γ µ (/k + p/ + m)γ ν .

(11.11)
We can employ the Feynman trick and the momentum shift from section
6.6.1 to arrive at
1
2 Z
−Q T (p, k − xp)
Z
µν 2
F (m , p) = 4
dx d4 k 2 . (11.12)
(2π) (k + sx(1 − x) − m2 + iη)2
0
1
Why introduce charged scalars here ? Because we can. It works equally well with
fermions in the initial state, see exercise 89.
2
The Fermi minus sign is really important here !
324 August 31, 2019

Now,

T (p, k − xp) = −8x(1 − x)pµ pν + (other terms) , (11.13)

and we see that3

Z1
Q2 8x(1 − x)
Z
2
K(m , s) = dx d4 k (11.14)
(2π)4 (k 2 + sx(1 − x) − m2 + iη)2
0

And we must compute this only for s = 0. In dimensional regularization as

described, again, in section 6.6.1 we have
Z1 Z∞
iQ2 µ2 t1−
K(m2 , 0) = dx 8x(1 − x) dt
(4π)2− Γ(2 − ) (t + m2 )2
0 0
iα
R − log m2 + O () ,


= (11.15)
3π~
where the divergent constant R was defined in Eq.(??)

11.2.4 The vacuum polarization

The photon self-energy can of course be Dyson-summed. Denoting the
dressed propagator Πµν by a thick line, we can write the SDe as

= + ,
−i~ g µν −i~ g µα
Πµν (s) = −s gαβ K(s) Παν (s)


+
s s
−i~ g µν −i~ g µν
=
= . (11.16)
s 1 − i~K(s) s (1 − Π(s))

We have dropped the pµ pν terms because of current conservation, and intro-

duced the vacuum polarization function :

β(3 − β 2 )
   
α 2 β+1 2 8
Π(s) = −R + log m + log +β − , (11.17)
3π 2 β−1 3
3
This saves us a lot of unnecessary work : the ‘other terms’ contain k µ k ν , g µν k 2 and
suchlike, with integrals more cumbersome by far.
August 31, 2019 325

Figure 11.1: The fermion-induced vacuum polarization

with β = (1 − 4(m2 − iη)/s)1/2 . For large s values, we have

   
α |s| 5 2
Π(s) − Π(0) ≈ log − − iπθ(s > 4m ) . (11.18)
3π m2 3

Note that we have defined β with the mass m2 augmented by its ‘natural’
extension into m2 −iη (η ↓ 0); the iη takes care of the correct imaginary part,
as it should. Plot 11.1 shows the behaviour of Π(s) − Π(0) as a function of
s/m2 . It is continuous both at s = 0 and at s = 4m2 , and we see the opening
of a branch cut for s ≥ 4m2 . I used the value η = 10−8 and so obviated the
need to distinguish between the three different cases s < 0, 0 < s < 4m2 and
s > 4m2 .

11.2.5 Hadronic vacuum polarization

The vacuum polarization Π(s) discussed above can reliably be computed for
charged leptons since their masses are quite accurately known, and higher-
loop QED contributions can be expected to be small. How different for
quarks ! Not only do we not know the quark masses to anything like the
same precision as the leptonic ones, but also we expect enormous higher-
326 August 31, 2019

order effects from strong interactions4 . But all is not lost ! Let us write the
hadronic self-energy of the photon as5

µ ν
= (pµ pν − s g µν ) H(s) , s = p2 . (11.19)
p

Let us now consider the forward amplitude for e+ e− → e+ e− mediated by a

hadronically corrected photon. We take the electrons as massless and denote
their helicity by λ, to write

p1 p1
Mf =
p2 p2
Q2 ~2
= 2
uλ (p1 )γµ uλ (p2 ) (pµ pν − s g µν ) H(s)uλ (p2 )γµ uλ (p1 )
s
= 2Q2 ~2 H(s) , (11.20)

which is easily computed using the helicity techniques of chapter 8. On the

other hand, we can write
2
Z p1


 
  dV = 4s σhad (s) (11.21)
p 
 2 

+ −
where σhad (s) stands for the total
√ cross section for the process e e →
hadrons at total collision energy s, and dV includes a sum over all hadronic
final states6 . The optical theorem (6.34) then tells us that

s σhad (s)
<H(s) = − : (11.22)
4πα~
4
After all, free quarks are never observed, but only their more-or-less stable bound
states. The interactions responsible must be drastic, to say the least !
5
Here we again assume current conservation in the photon-quark interactions. If that
does not hold, then the vacuum polarization is very far down the list of our problems. . .
6
The prefactor 4s arises from excising the flux factor 1/2s, realising that λ can take
two values, and that there is a factor 1/4 for the spin average. Additionally, we need the
cross section with the radiative corrections to the initial state somehow removed ; but this
is doable.
August 31, 2019 327

we can simply measure it ! We can employ a once-subtracted dispersion

relation as before and arrive at7
Z ∞
s σhad (z)
=H(s) − =H(0) = 2 P dz . (11.23)
4π α~ 4m2π z−s
√
The lower limit on z is twice the charged pion mass, since the lightest
possible hadronic system that can be formed in e+ e− collisions is the π + π −
state. The hadronic vacuum polarization is, just like above, given by Πh (s) =
i~H(s). We see that a careful measurement of σhad can help us to obtain E90
hadronic effects, valid to all orders, that would not be computable in any
other way8 !

11.3 The fermion self-energy

11.3.1 A look at gauge invariance
We now consider radiative corrections to the fermions, rather than the photon
itself. The fermion-fermion-photon vertex has three one-loop diagrams9 :

I have taken the external fermions to be on their mass shell for simplicity,
otherwise we would need even more diagrams10 . That these cannot easily
7
The first actually useful phenomenological application of this idea is in F.A. Berends
and G.J. Komen, Radiative corrections to Bhabha scattering and µ pair production from
the hadronic vacuum polarization, Phys.Lett. 63B (1976) 432. But the references therein
show that the idea is a bit older.
8
Around the turn of the century, the hadronic vacuum polarization was the dominant
uncertainty in the precision analysis of the LEP data (helping to establish bounds on the
Higgs boson mass, for instance). Not surprisingly, a small industry arose of competing
computations of Πh (s). If you wish, you can also try to extract ‘effective’ quark masses by
fitting them to the observed polarization. I have always found that of dubious relevance.
9
In order not to let the diagrams become too cumbersome I indicate the fermion flow
with just a single arrow.
10
In that case, a fermion would enter into another ‘blob’ and we would have to let the
photon also enter in there.
 328 August 31, 2019

be separated becomes clear if we check current conservation on the external

(possibly off-shell !) photon with a handlebar :

= − , = − , = .

(11.24)
By using Eq.(10.15) we see that the total set of graphs vanish. This is,
of course, just an illustration of the Ward identity of section 10.2.4. The
situation is slightly more subtle if we consider the internal photon. If we
single out the to-be-handlebarred vertex we see that we. have to consider
not 3 but the 6 diagrams

+ + + + +

E91 and then the whole lot cancels11 . The upshot is that we are allowed to use
the simple photon propagator −g µν /(k 2 + i) for the internal photon line.

11.3.2 Summing the self-energies

A fermion’s self-energy due to the electromagnetic field is, to first order, given
by the diagram
k
−ie2 γ µ (/p + k/ + m) γµ
Z
S(p) = p = d4 k .
(2π)4
((p + k)2 − m2 + iη) (k 2 − λ2 + iη)
(11.25)
The mass and charge of the fermion are m and e, respectively, and we have
introduced a very small photon mass λ, the use of which will become clear
11
There is ample room for terrible confusion here ! The photon propagator in the axial
gauge has numerator −g µν + (k µ nν + nµ k ν )/(k · n) and this suggests that, if we want to rid
ourselves of n-containing terms we have to simultaneously do the handlebar on both ends
of the internal photon line. But... on the one end the photon goes away from the fermion
line, on the other end it goes towards the fermion line, leading to diagrams with slashed
propagators with the wrong sign ! But... the factor 1/(kn) may take care of the ‘wrong’
sign of the photon momentum... or does it ? Or do we have to rely on contributions
from k and −k cancelling in the loops ? By singling out a vertex and then considering
all ways to write the loop diagram we see that, indeed, the current is conserved and no
hemicranogenic arguments are needed.
August 31, 2019 329

later on. We define

Σ(p) = i~ S(p) ≡ a(s) p/ + b(s) m , s = p2 . (11.26)

We have to distinguish between the bare mass m0 and the renormalised

(physical) mass of the fermion when including one-loop effects. The fermion
propagator with the self-energy included via Dyson summation then becomes
p/ + m i~
ı~ =
s − m2 p/ − m
i~ p/(1 − a(s)) + m0 (1 + b(s))
→ = i~ (11.27)
p/ − m0 − Σ(p) s(1 − a(s))2 − m20 (1 + b(s))2
Again, we use the on-shell renormalization scheme. This requires the denom-
inator to vanish at the physical pole p2 = m2 , giving us

1 − a(m2 )
∼ m 1 − y(m2 ) , y(m2 ) = a(m2 ) + b(m2 ) . (11.28)


m0 = m 2
1 + b(m )
Note that we keep only the one-loop corrections, so that both a and b are
formally infinitesimal. Implementing Eq.(11.28) in the summed propagator
we may Taylor-expand to find that, to one-loop accuracy, the inclusion of
the self-energy leads to
i~ i~ 1
→ + (terms with no poles) ,
p/ − m p/ − m 1 − z(m2 )
z(m2 ) = 1 − a(m2 ) − 2m2 a0 (m2 ) − 2m2 b0 (m2 ) . (11.29)

We shall restrict ourselves to on-shell (external) fermions only, so the non- E92
pole terms can be neglected.

11.3.3 The loop calculation

We now turn to the computation of Σ(p). Performing the Feynman trick and
shifting the integration variable we obtain
Z1
−i~e2 γ µ ((1 − x)/p + k/ + m) γµ
Z
Σ(p) = dx d4 k .
(2π)4 (k 2 + sx(1 − x) − m2 x − λ2 (1 − x) + iη)2
0
(11.30)
330 August 31, 2019

The k/ in the numerator vanishes upon integration, so the divergence is log-

arithmic. Under the rules and techniques of dimensional regularization, we E93
then find
Z1
α
(4πµ2 ) Γ() dx − (2 − 2)(1 − x)/p + (4 − 2)m n(x)− ,


Σ(p) =
4π
0
n(x) = m2 x − sx(1 − x) + λ2 (1 − x) − iη . (11.31)
This tells us that the functions a and b in Eq.(11.26) are given by
α
a(s) = (4πµ2 ) Γ()(2 − 2)A(s) ,
4π
α
b(s) = (4πµ2 ) Γ()(4 − 2)B(s) , (11.32)
4π
where
Z1 Z1
A(s) = dx(1 − x) n(x)− , B(s) = dx n(x)− . (11.33)
0 0

We will have to compute A(m2 ), B(m2 ), m2 A0 (m2 ) and m2 B 0 (m2 ), and this
simplifies our lives since then12 n(x) = m2 x2 + λ2 . In computing A, B(m2 )
we can let λ vanish with impunity :
Z1
1
A(m2 ) ≈ dx(1 − x) (1 −  log(m2 x2 )) = 1 − (log m2 − 3) ,


2
0
Z1
B(m2 ) ≈ dx (1 −  log(m2 x2 )) = 1 −  log m2 − 2 .


(11.34)
0

This means that a(m2 ) and b(m2 ) have an UV divergence residing in Γ().
The other two are more interesting. We may decide to keep λ nonzero or let
it vanish. In the first case we find (taking  to zero where possible) :
Z1
m2 x(1 − x)2
  2 
2 0 2  m
m A (m )λ ≈  dx 2 2 2
≈ log −3 ,
(m x + λ ) 2 λ2
0
12
Since λ only becomes relevant when x is very close to zero, we can replace λ2 (1 − x)
by just λ2 .
August 31, 2019 331

Z1
m2 x(1 − x)
  2 
2 0 2  m
m B (m )λ ≈  dx 2 2 2
≈ log − 2 . (11.35)
m x +λ 2 λ2
0

We see that the overall factor  will remove the singularity in Γ() since
Γ() ≈ 1. The UV divergence is now gone in these integrals, and instead we
find an IR divergence regularised by λ. The alternative, taking λ = 0, leads
to
Z1
0  −1
2 2
dx x−1−2 (1 − x)2 = 1 − (log m2 − 3) .


m A (m )0 =
(m2 ) 2
0
Z1
 −1
m2 B 0 (m2 )0 = dx x−1−2 (1 − x) = 1 − (log m2 − 2) .


(m2 ) 2
0
(11.36)

The IR divergence is now represented by the 1/ in Γ() with the opposite
sign provided by the −1 in Eq.(11.36). Putting everything together we finally
arrive at
α
y(m2 ) = 3R − 3 log m2 + 4 ,


4π
2 α
−3R + 3 log m2 − 4 ,


z(m )0 =
4π
2 α
−R − 2 log λ2 + 3 log m2 − 4 .


z(m )λ = (11.37)
4π

11.3.4 The curious incident of the divergences in the

nighttime
The fermion loop diagram in the previous section diverges as m → 0. But
we can also start out with strictly massless fermions and photons. The
expression (11.26) then reduces to Σ(p) = a(s)/p and we will have to evaluate
this at s = 0. If, however, we put m = λ = s = 0 in Eq.(11.30) we
find,according to Eq.(6.86) that a(0) = 0 !13 We can interpret this as follows :
there are two divergences, a UV and and IR one, and they cancel precisely.
13
The divergences do nothing : see Sir A.C. Doyle, The Adventure of Silver Blaze, in
The Memoirs of Sherlock Holmes (1894).
 332 August 31, 2019

In a sense this is to be expected : in the absence of any mass scale there

is nothing to combine with the log µ2 in R to make the argument of the
logarithm dimensionless. Therefore, log µ2 cannot occur in the result, and
therefore neither can 1/ 14 . Indeed, we can separate the two divergences
E94 explicitly by keeping λ nonzero, and then we find
 
α 2 1
a(0) = − R − log λ − . (11.38)
4π 2

In the same spirit, we can say that the photon self-energy Eq.(11.15) has
an IR divergence regularized by the fermion mass in log m2 . Indeed, if we
compute K(0, 0) directly we find again a vanishing result.

11.4 Infrared singularities in Bremsstrahlung

Since a real photon is massless its energy can be arbitrarily small, and infrared
factors such as that of Eq.(10.100) will blow up in that limit. This infrared
divergence is real, it is not an artefact of the theory ! But since it resides where
the photon momentum k µ has both k 0 = 0 and k 2 = 0 you might wonder
wether such a photon is actually a real photon — or a virtual one ? Indeed the
infrared divergences from virtual and real photon corrections tend to cancel.
To see this we have to compute the photon-emission (Bremsstrahlung) cross
section in such a way that we can compare it with what the previous section
told us. Your typical IR factor (for particles with unit electric charge e) looks
like
(p1 · p2 )
e2 ~
(p1 · k)(p2 · k)
14
The idea of discontinuous limits is not so rare as all that. For a quite homely example,
consider two points at opposite corners of a unit square. Moving from on to the other
along the square’s edges covers a distance of 2. As you can see from the picture below,
making rectangular zigzags does not change the distance covered. It is only in the ultimate
limit, where the zigs and zags vanish completely√and you go in a straight line between the
points, the distance covered drops abruptly to 2.
August 31, 2019 333

where p1,2 are two external momenta, and k is the photon’s momentum. This
we have to integrate over, with some (small !) maximum value E0 for the
energy. The integral that we have to compute is then, after the Feynman
trick,
0
(p1 · p2 )
Z
4πα 3~ θ(k < E0 )
B(p1 , p2 ) = d k
(2π) 3
2|~k| (p1 · k)(p2 · k)
Z1 0
3~ θ(k < E0 ) (p1 · p2 )
Z
4πα
= dy d k ,
(2π)3 2|~k| (py · k)2
0
py µ = y p1 µ + (1 − y)p2 µ . (11.39)
If we again give the photon a very tiny mass λ then λ < k 0 < E0 , and λ
will regulate the IR divergence15 ; but we can also let k 0 start at zero and
use dimensional regularization16 . Let us again work in 4 − 2 dimensions17 .
Using the t-shell formula (6.82) we then can cast our integration element into
the following form :

µ2 ~
3−2~ θ(0 < |k| < E0 )
Z
d k
(2π)3−2 2|~k|
2
2 ZE0 Z1
µ d cos θ
= 3/2−
dt t− , (11.40)
2(4π) Γ(3/2 − ) 2
0 −1

where we have re-introduced an integral over a polar angle θ, defined with

respect to whatever direction we like. Using results from appendix 15.15.4
15
Maybe λ is a good idea. It provides an explicit separation between virtual and real
photons since there is now a mass gap.
16
Maybe λ is a bad idea. Introducing a mass in the boson propagator would formally
require us to replace the axial propagator of Eq.(9.58) by the massive one of Eq.(9.3),
which may be ruinous in more complicated theories like QCD (see later on), altough
not in QED. You may suspect having a λ in denominators and not in numerators can
play havoc with the handlebar operation. On the other hand, as we have already seen,
dimensional regularization tends to make the distinction between UV and IR divergences
rather hazy - after all - there is only so much that a simple 1/ can do.
17
In treating the ultraviolet divergence, you would like to have D < 4 so the ‘natural’
values of  are positive ; for infrared divergences you would probably rather think of having
D larger than 4, thus  < 0. In the spirit of dimensional regularization this is unimportant
since we consider nonzero  as an analytic continuation from wherever our formulæ make
sense.
 334 August 31, 2019

we find that

2 Γ(3/2 − ) = π 1/2 1 + γE +  log 4 − 2 + O 2 .

(11.41)

Moreover18 ,
2
Z1 ZE0
d cos θ 1 1 1
0 2
= , dt t−1− = − E0 −2 . (11.42)
2 (py − |~py | cos θ) py · p y 
−1 0

Putting it all together we find, so far,

  Z1
α 1 p 1 · p2
B(p1 , p2 )0 = log(2E0 ) − R − 1 dy . (11.43)
π 2 py · py
0

The integral B is not Lorentz-invariant because we require k 0 < E0 ; but the

y integral is Lorentz-invariant, and we can evaluate it in the centre-of-mass
E95 frame. For simplicity we take the masses of p1 and p2 to be both equal19 to
m. We define
β 2 = 1 − 4m2 /s , s = (p1 + p2 )2 , (11.44)
and the y integral then reduces to

Z1 Z1
1 + β2 1 + β2
 
p1 · p2 1+β
dy = dy = log . (11.45)
py · py 1 − β 2 (2y − 1)2 2β 1−β
0 0

In the limit of very small mass it approaches log(s/m2 ). The final result, in
dimensional regularization, is therefore

1 + β2
   
α 1 1+β
B(p1 , p2 )0 = log(2E0 ) − R − 1 log ,
π 2 2β 1−β
 
α 1
B(p1 , p1 )0 = log(2E0 ) − R − 1 . (11.46)
π 2
18
Here  is ‘negative’, remember ? Therefore t− = 0 at t = 0.
19
The case of unequal masses gives us a result that is more complicated but not more
enlightening : see exercise 95.
August 31, 2019 335

Let us now turn to the λ-regularization. For simplicity we shall assume

that we are in the centre-of-mass frame of p1 and p2 . We can then write

ZE0 Z1
α (1 + β 2 ) 0 |~k|
B(p1 , p2 )λ = dk d cos θ
2π (k 0 )2 − β 2 |~k|2 cos θ2
λ −1
ZE0 !
α 1 + β2 1 0 k 0 + β|~k|
= J(β) , J(β) = dk 0 log . (11.47)
π 2β k k 0 − β|~k|
λ

The most elegant way to compute J(β) is to first compute

4E0 2 2
ZE0 Z /λ
∂ 2|~k| dk 0 1 (v − 1)2
J(β) = = dv
∂β (k 0 )2 − β 2 |~k|2 1 − β2 v(v + r)(v + 1/r)
λ 1
  2 
1 4E0 1 1−β
≈ log + log(r) , r = . (11.48)
1 − β2 λ2 β 1+β

Here we have employed the unlikely-looking but useful variable transforma-

tion from k 0 to v = (k 0 + |~k|)/(k 0 − |~k|). We can use J(0) = 0 and integrate
over β so as to obtain20
     2  
2E0 1+β 1 1+β 2β
J(β) = log log − log −Li2 . (11.49)
λ 1−β 4 1−β 1+β

The term containing log(E0 ) comes out the same, but we pick up extra terms.
For small mass, these approach −(log(s/m2 ))2 /4 − π 2 /6.

11.5 The vertex correction

We now turn our attention to the so-called vertex correction :
k
iQ u(q) V µ (p, q) u(p) = (11.50)
p q

20
For the dilogarithm function Li2 , see appendix 15.15.10.
336 August 31, 2019

Note that we have taken out a factor iQ/~ so that V µ equals γ µ at the tree
level. Q is the fermion charge which we shall take equal to the electron
charge. In addition we shall only discuss the case where the two fermions are
on-shell, p2 = q 2 = m2 . The emitted photon is therefore, generally, off-shell.
Following the Feynman rules we find, again in the Feynman gauge, and using
a photon mass λ for possible later use :

−iQ2 ~ µ
Z
µ 4 Ω
V (p, q) = dk ,
(2π)4 N
1/N = (k 2 + 2(pk) + iη)(k 2 + 2(qk) + iη)(k 2 − λ2 + iη) ,
Ωµ = γ α (/q + k/ + m) γ µ (/p + k/ + m) γα . (11.51)

The Feynman trick induces us to shift the integration variable from k to

k̂ = k − xp − yq. If we also move to 4 − 2 dimensions and perform the Wick
rotation, we find21 for the numerator Ωµ , with σ = (pq) :

Ωµ → A1 k 2 γ µ + A1 γ µ + m[/p, γ µ ] B1 + m[γ µ , q/] B2 ,

A0 = (2 − 2)2 /(4 − 2) ≈ 1 − 3/2 ,
A1 = 4σ(1 − x − y) − (2 − 2)(m2 x2 + m2 y 2 + 4m2 xy − 2xyσ) ,
B1 = 2y − (2 − 2)(x2 + xy) ,
B2 = 2x − (2 − 2)(y 2 + xy) . (11.52)

Here we have dropped all terms linear in k̂, and for the quadratic term we
used the fact that k̂ ρ k̂ ν can only integrate to something proportional to g ρν
and therefore we can replace
1
k̂ ρ k̂ ν → g ρσ k̂ 2 . (11.53)
4 − 2
The vertex correction thus becomes
(4πα) 2 µ2 2 µ µ
4−2 A0 k̂ γ − A1 γ − m[/ p − q/, γ µ ]B1
Z Z
µ
V (p, q) = δxy d k̂ ,
(2π)4−2 (k̂ 2 + D)3
D = m2 x2 + m2 y 2 + 2σxy + λ2 . (11.54)

Here we have employed the fact that D is symmetric in x ↔ y so that B2

integrates to the same as B1 ; also we have replaced λ2 (1 − x − y) by λ2 since
21
Here we also use the Dirac equation for u(q) and u(p).
August 31, 2019 337

the photon mass is only important if x = y = 0. The symbol δxy stands for
dx dy θ(0 < x < 1) θ(0 < y < 1 − x). After the t-shell formula-cum-integral
we are then left with
Z 
α
µ
V (p, q) = 2 
(4πµ ) δxy A0 γ µ Γ()(2 − ) D−
4π

µ µ −1−


− A1 γ + m[/p − q/, γ ]B1 Γ(1 + ) D . (11.55)

The A0 integral is UV divergent and responsible for charge renormalization.

The A1 integral is IR divergent, and the B1 integral is convergent.

11.6 Exercises
Excercise 89 Doing it with fermions only
Prove that Eq.(11.8) is also obtained if we start with an initial state consisting
of a fermion-antifermion pair instead of scalars. Take care not to average
over the incoming spins, as that could not possibly correspond to a forward
scattering amplitude (why ?).

Excercise 90 A tool kit for Πh (s)

The hadronic cross section σhad (s) is approximately built up from two types
of contribution : narrow resonances, and ‘plateau’ functions whenever a
new type of quark becomes kinematically accessible (essentially, the c and b
quarks, and at very high energies also the t quarks).
1. The narrow resonances are characterised by their mass M , their lep-
tonic decay width Γ` , and their total decay width Γ, which is usually
completely dominated by the hadronic decays. The cross section profile
corresponding to such a resonance reads
12π Γ` Γ
σresonance (s) = .
(s − M 2 )2 + M 2 Γ2
Assuming that Γ  M , show that the contribution to < Πh (s)−< Πh (0)
from such a resonance is given by
3 Γ` s
[< Πh (s) − < Πh (0)]resonance ≈
α M s − M2
as long as s is not too close to M 2 .
338 August 31, 2019

2. The plateau function for a quark of mass M and fractional charge q 22

is given by
4πα2 2
σplateau (s) = q Nc
3s
Here the first factor is the total cross section for muon pair production
(cf section 10.3.1), and Nc is the number of quark colours23 . Show that
its contribution reads
 s − 4M 2 
 
α
[< Πh (s) − < Πh (0)]plateau = log  
3π 4M 2 

Excercise 91 Cancellations in the eeγ loops

Show explicitly that the sum of diagrams in Eq.(11.3.1) evaluates to zero.

Excercise 92 Cavalier and careful may work

You may envisage obtaining the quantities y(m2 ) and z(m2 ) in a more ‘direct’
way. Define the denominator

N = p/ − m0 − a(s)/p − m0 b(s)

Perform a ‘Taylor expansion of N around p/ = m’. This involves derivatives

with respect to p/ for which we may write
d ∂s d d
= = 2/p
d/p ∂/p ds ds
Show that this may work provided you know exactly when to replace p/ by
m and when to let it stand. In addition, show that it is actually impossible
to have p/ = m for any pµ . You have now proven that this cavalier approach
may work if you already know the answer. . .

Excercise 93 Dirac beyond dimensions

In D dimensions we have gµν g µν = D. Use this to show, that for D = 4 − 2,

γ α p/ γα = (−2 + 2)/p , γ α p/q/γα = 4(p · q) − 2/pq/ , γ α p/q/k/ γα = −2/kq/p/ + 2/pq/k/

Excercise 94 A tale of two divergences

Explain why in the result a(0) = 0 in section 11.3.4 there can not be any
finite terms in a(0). Furthermore, compute the result (11.38).
22
They are : +2/3 for the u, c and t quarks, and −1/3 for the d, s and b quarks.
23
i.e. 3.
August 31, 2019 339

Excercise 95 Infrared factors

R1 for unequal masses
We compute the integral 0 dy(p1 · p2 )/py 2 for the case of unequal masses
m1 6= m2 .

1. In the centre-of-mass frame of p1 and p2 , show that

s + m21 − m22 s − m21 + m22 κ

p01 = √ , p02 = √ , |~p1,2 | = √
2 s 2 s 2 s

where

s = (p1 + p2 )2 , κ2 = s2 − 2s(m21 + m22 ) − δ 2 , δ = m21 − m22

2. Use this to show that

p1 · p2 s2 + κ2 − δ 2
= , z = 2y − 1
p2y (s + (κ + δ)z)(s − (κ − δ)z)

3. Show that
Z1
s2 + κ2 − δ 2 (s + κ)2 − δ 2
 
p1 · p2
dy = log
p2y 4κ s (s − κ)2 − δ 2
0

4. Show that we obtain the correct result if δ → 0 ; and that the integral
diverges if one or both masses vanish.
340 August 31, 2019
Chapter 12

Quantum Chromodynamics

12.1 Introduction: coloured quarks and glu-

ons
In chapter 10 we have studied the behaviour of electrically charged particles
and the electromagnetic field embodied by photons. Notwithstanding the
fact that particles can have different charges, all these charges are of the
same type in the sense that they can be added. For instance, atoms are
electrically neutral when studied from the ‘outside’, since the positive charge
of the nucleus is cancelled out by the negative charge of the electron cloud. It
is interesting to see what happens if we enlarge our view to the possibility of
‘different types of charge’, that cannot be meaningfully added in a simple way.
In that case, a bound state of particles with a different charge type might not
look ‘neutral’ when seen from the outside : the charges of the constituents
would show through. To avoid confusion with the electric charge we shall let
the ‘new charges’ go by the name of colours, and the dynamical theory of
their interactions is called Quantum Chromodynamics, or QCD.
We shall start our investigation with coloured fermions, called quarks1 .
The number of colours is denoted by N , where of course N ≥ 2. The quarks
are described by Dirac spinors for given momentum and spin, and also by
a colour label which we shall denote by a, b, c, . . .. These colour labels (or
1
Historically, the notion of quark predates that of colour, and the colouring of quarks
was invented to explain the possibility of the existence of particles such as the ∆++ or
the Ω− . In this chapter, we are less interested in describing the world of hadrons than in
constructing an internally consistent theory, hence the unhistorical line of reasoning.

341
342 August 31, 2019

indices) run from 1 to N . A conjugate fermion (u or v) will carry a lower, a

regular fermion (u or v) an upper index.
In addition we expect vector particles to be present, that carry the colour
force. These we call gluons. In analogy to QED, we shall assume the gluons
to be massless, but since we have different colour types there must also be
different gluon types. The gluon type will be denoted by j, k, l, m, . . ., and it
is up to us to determine2 how many gluon types occur for given N .
We now formulate a few properties that we want our world of colour to
possess:

1. Colour is conserved in interactions, just like electric charge. This must

hold for every type of colour charge separately. Particles are therefore
not allowed to change their colour ‘spontaneously’.

2. All colours are equal and none are ‘more equal than others’, which
means that particles that only differ by their colours propagate through
spacetime in the same way 3 .

12.2 Quarks and gluons : first Feynman rules

12.2.1 The propagators
Under the assumption of colour conservation, the quark propagator is defined
by
p/ + m
b
p a ↔ i~ 2 2
δab (12.1)
p − m + i
where the incoming and outgoing quark colours are denoted by a and b.
The gluon propagator requires some more care. We shall assume that it is
described by the axial propagator of chapter 9. That is, we take an axis nµ ,
and require the propagator to be orthogonal to this. Accepting that also
gluons carry colour information, we have
j k
α
↔ Παβ (p) δ jk (12.2)
p β
2
The usual approach is simply to postulate a local SU (N ) gauge symmetry, from which
the number of gluons immediately follows ; but our (or rather my) interest is to see how
we can arrive at that result from simpler, or rather physical, requirements.
3
Otherwise we would count them as different particles
August 31, 2019 343

where j and k denote the gluon colours, and

pα nλ nλ pβ
  
αβ −1 α λβ
Π (p) = i~ 2 g λ− g −
p + i (p · n) (p · n)
α β α β 2 α β
 
i~ αβ p n +n p n p p
= 2 −g + − . (12.3)
p + i (p · n) (p · n)2

The first line of Eq.(12.3) shows that, indeed, Παβ (p) nα = Παβ (p) nβ = 0.
Note that we have extended the definition of the axial-gauge propagator to
nonzero values of n2 . However, since Πα α (p) ∝ −2+n2 p2 /(p·n)2 , the number
of degrees of freedom is still equal to 2 on the mass shell, where p2 = 0.

12.2.2 The quark-gluon vertex

We start by defining the quark-gluon vertex, as a close analogue of the QED
fermion-photon vertex :
a
j i µ j a
µ ↔ ~ g γ (T ) b , (12.4)
b
where we have explicitly indicated the quark and gluon colour types. Here,
g is the coupling constant, and (T j )a b is recognized as an element of an
N × N matrix, the properties of which we still need to derive. Allowing
for complex matrices, we see that the number of different gluon colours j
cannot exceed 2N 2 . It is clear that an overall factor in the matrices T can
always be absorbed in a redefinition of g, and we shall use this to normalize
the T matrices. The above vertex, inspired by the example of QED, must
ultimately be justified by investigating the handlebar condition4 . Because
of the presence of colour we may expect that this will be more complicated
than in the simple electrodynamics case.

12.2.3 A closer look at the T matrices

We require the structure of the colour part of the interactions to take care
of colour conservation and colour democracy in the presence of interactions.
4
And, of course, by that ultimate arbiter, comparison to experiment. It is experiment
that tells us that N = 3 ; we shall stick to general N > 1 in the text.
344 August 31, 2019

Consider the following diagram :

b a
c
j

The colour part of this diagram reads

X
(T j )a b (T j )b c
j,b

and colour conservation/democracy hence demands that

X  2 a
Tj = k δab (12.5)
b
j

for some constant k. Similarly, the diagram

j b
k

a
contains the colour factor
X
(T j )a b (T k )b a = Tr T j T k


,
a,b

and using the normalization freedom we may take

1
Tr T j T k = δ jk . (12.6)
2
Since colour must be conserved, a gluon cannot lose its colour charge and
therefore gluons and photons cannot mix : in all diagrams of the form
a
j
photon
b
August 31, 2019 345

we must have b = a since the colour is conserved and the photon is colourless ;
therefore the T matrices must be traceless :
Tr T j = 0 for all gluon colours j .


(12.7)
Finally, we consider the following two-loop self-energy diagram of the photon :

Here, the fermions are quarks and the internal line labelled j is a gluon of
colour type j : of course, we have to sum over all j values. If we compare this
diagram to the corresponding QED one, we see that apart from the overall
charges (g 2 instead of Q2 ) the only difference is the colour factor, in this case
X
Tr T j T j


j

Now, if our theory is to be unitary, it must obey the Cutkosky rules, and
therefore we demand that

+ j
+ j
j

+ j
+ j
=0 . (12.8)

For the QED diagram, this indeeds holds. In the coloured case, however, the
colour structures of the diagram cut in the various ways are no longer the
same : the three lines in Eq.(12.8) are proportional to, respectively,
X   X  † †
j j†
X
j j
Tr T j T j


T = Tr T T , Tc = Tr T T , and Tcc = .
j j j

Unitarity can therefore only be safe if these three different traces are, in fact,
equal to one another. We may therefore write
 
†
X
Tr Aj Aj = 0 , Aj = i T j − T j


2Tc − T − Tcc = . (12.9)
j
346 August 31, 2019

The matrices Aj are Hermitean, so that Eq.(12.9) can also be written as

  N
j†
X
j
XX  j a 2
Tr A A = (A )  = 0 ,
b (12.10)
j j a,b=1

hence all Aj are actually identically zero, and the matrices T j must be Her-
mitean. The number of different gluon colours type is therefore N 2 − 1, and
the constant k of Eq.(12.5) is equal to (N 2 − 1)/2N .

12.2.4 The Fierz identity for T matrices

We have now zoomed in quite efficiently on the matrices T j . On the other
hand, just as in the case of Dirac particles we would prefer if predictions
for cross sections and the like dit not depend on the particular choice of
the matrices5 . We can, in fact, derive a relation between the T ’s that holds
independently of any representation : it goes under the name of the Fierz
identity6 . Any N × N matrix M can be written7 as
X
M = a0 1 + aj T j . (12.11)
j

By taking traces we can determine the coefficients :

Tr (M ) = a0 N , Tr M T k = ak /2 .


(12.12)
Therefore we have
X 1
Tr T j M T j + Tr (M ) ,


M =2 (12.13)
j
N

or, in terms of the matrix components,

X 1
M dc δcb δad = 2 M d c (T j )c d (T j )a b + M d c δ c d δ a b , (12.14)
j
N
5
In the Dirac case this was indispensable since any dependence would destroy Lorentz
invariance. In the present case one might argue that the T j could, in principle, just be
measured. Nevertheless having a representation-independent theory just feels so much
more comfortable.
6
Same Fierz.
7
You might be tempted to think that this holds only for Hermitean matrices. But since
iT j is antiHermitean we can accomodate any M provided the a’s can be complex.
August 31, 2019 347

whence the following, representation-independent identity :

 
j a j c 1 a c 1 a c
(T ) b (T ) d = δ dδ b− δ bδ d . (12.15)
2 N
Since8 the colour of quarks and gluons cannot be observed, any cross section
will involve a summation over all colours, and therefore every cross section
is expressed as (a product of) traces of strings of T matrices, in which every
matrix T k occurs exactly twice, and the index k is summed over. The Fierz
identity comes in useful here, since we can write (with summation implied)
 
j

j

1 1
Tr T A Tr T B = Tr (AB) − Tr (A) Tr (B) ,
2 N
 
j j

1 1
Tr T A T B = Tr (A) Tr (B) − Tr (AB) . (12.16)
2 N
With these trace identities we can simplify and compute any set of colour
traces without recourse to any explicit representation, especially if we recall
that Tr (1) = N, Tr (T j ) = 0 and T j T j = (N 2 − 1)/2N times unity. As an
illustration, we work out the following cases in detail :

Tr T j T k T l Tr T j T k T l



 
1 k l k l

1 k l

k l


= Tr T T T T − Tr T T Tr T T
2 N
 
1 l

l

2 l l

1 l

l


= Tr T Tr T − Tr T T + 2 Tr T Tr T
4 N N
2
1 N −1
Tr T l T l = −


= − , (12.17)
2N 4N
and

Tr T j T k T l Tr T j T l T k



 
1 k l l k

1 k l

l k


= Tr T T T T − Tr T T Tr T T
2 N
 
1 k k

2 k k

1 k

k


= Tr T T Tr (1) − Tr T T + 2 Tr T Tr T
4 N N
2 2 2
N −2 (N − 1)(N − 2)
Tr T k T k =


= . (12.18)
4N 8N
8
Empirically.
348 August 31, 2019

12.3 The three-gluon interaction

12.3.1 The need for three-gluon vertices
It is now time to investigate our theory using handlebars. In the first place, in
the process g → q q̄ the current is conserved in the same way as in QED, since
there is only a single Feynman diagram and the colour structure is therefore
irrelevant to any cancellation. The situation becomes more delicate in the
case of more complicated interactions, so let us consider the process

q̄(p1 , a) q(p2 , b) → g(q1 , 1 , j) g(q2 , 2 , k)

where we have explicitly indicated the momenta, polarisations, and colours.

We have at least the following two diagrams9 :
a j a
p1 q1 p jq
1
1

M= q2 + , (12.19)
kq
p2 k p2 2
b b
They read
q/1 − p/1 + m
M1 = −i~g 2 v(p1 )/1 /2 u(p2 ) (T j T k )a b ,
−2(q1 · p1 )
p/2 − q/1 + m
M2 = −i~g 2 v(p1 )/2 /1 u(p2 ) (T k T j )a b . (12.20)
−2(q1 · p2 )
Let us now put the handlebar on gluon 1, so that we replace µ1 by q1µ . By
the same reasoning as in chapter 10, we arrive at the handlebar rule

= − , (12.21)

with the auxiliary Feynman rules

j
b a a g
↔ i~ δ b , b a ↔ i (T j )a b . (12.22)
~
9
In the second diagram the two gluon lines form an overpass, without a vertex.
August 31, 2019 349

As before, slashed propagators vanish on external lines. Applying this to the

two Feynman diagrams of Eq.(12.19) yields10

j j
M1+2 c1 →q1 = +
k
k
j j
= − + j
− j

k k
k k
j
= − j

k
k

= −i~g 2 v(p1 )/2 u(p2 ) [T j , T k ]a b , (12.23)

where the square brackets denote, of course, the commutator of the matrices
T j and T k . Because of the colour structure we have a non-vanishing result,
and current conservation is in trouble ! The remedy must be to introduce a
third diagram, with a nontrivial ggg vertex,

a j
p1
q1
n
q2
p2
b k

and it is now our job to determine the form of the new three-gluon vertex.
We shall do this by investigating loop diagrams.

10
In calculations such as this one it is often sufficient to simply label the gluons by their
colour, thus simplifying the typography.
350 August 31, 2019

12.3.2 Furry’s failure

Consider the following Feynman diagram :

p −q 2 ρ
l
ν k q3
q2
K ∼ (12.24)
p p + q1
j q1
µ

Here three gluons are effectively coupled by a quark loop. We have explicitly
indicated the momentum flows. Note especially that the gluon momenta are
all counted flowing out of the vertex, so that we have

q1 + q2 + q3 = 0 . (12.25)

Apart from overall coupling constants and the like, the loop diagram is given
by

Tr ((/p + m)γ µ (/p + q/1 + m)γ ρ (/p − q/2 + m)γ ν )

Z
K ∼ d4 p j l k


Tr T T T .
(p2 − m2 )((p + q1 )2 − m2 )((p − q2 )2 − m2 )
(12.26)
There is also a loop diagram in which the quark runs counterclockwise instead
of clockwise. In our discussion of Furry’s theorem in sect. 10.2.7, we have
seen that the space-time part of the second diagram is exactly opposite to the
one of the first, so that in QED these two diagrams cancel. In QCD they do
not : the second diagram contains the
colour matrices in the opposite
order,
j k l j l k
that is to say it contains Tr T T T instead of Tr T T T . The sum of
the two diagrams, if we take into account the Lorentz-covariant nature of the
loop integral, and the fact that out of q1 , q2 and q3 only two momenta are
independent, must be of the form

K ∼ Y (q1 , µ; q2 , ν; q3 , ρ) Tr T j [T k , T l ] ,


(12.27)

with

Y (q1 , µ; q2 , ν; q3 , ρ) = (a1 q1 + a2 q2 )ρ g µν + (a3 q2 + a4 q3 )µ g νρ

+ (a5 q3 + a6 q1 )ν g ρµ , (12.28)
August 31, 2019 351

for some numbers a1 , . . . , a6 . For large p, each of the three propagators goes as
1/p, and the loop integral is therefore divergent. We see that indeed a three-
gluon counterterm must be allowed, and therefore a three-gluon coupling
must appear in the action, otherwise the theory would not be renormalizable ;
and the form of the three-gluon vertex must be that of Eq.(12.28).
Without evaluating the loop integral completely, we can glean all the
information we need. Consider the following transformation on K :
q1 ↔ −q2 , q3 → −q3 , µ ↔ ν . (12.29)
This transformation leaves the momentum conservation law (12.25) intact,
and also preserves the value of T (by the reversal property (7.29) of Dirac
traces). The same holds, of course, for the transformations
q1 ↔ −q3 , q2 → −q2 , µ ↔ ρ ,
q2 ↔ −q3 , q1 → −q1 , ν ↔ ρ . (12.30)
The function Y must therefore satisfy
Y (q1 , µ; q2 , ν; q3 , ρ) = Y (−q2 , ν; −q1 , µ; −q3 , ρ) =
= Y (−q3 , ρ; −q2 , ν; −q1 , µ) = Y (−q1 , µ; −q3 , ρ; −q2 , ν) ; (12.31)
and by inspection we then find that c1 = c3 = c5 = −c2 = −c4 = −c6 . We
shall therefore from now on use the Yang-Mills three-boson vertex
Y (q1 , µ; q2 , ν; q3 , ρ)
≡ (q1 − q2 )ρ g µν + (q2 − q3 )µ g νρ + (q3 − q1 )ν g ρµ . (12.32)
Note that this form is antisymmetric in the interchange of any two gluons,
and therefore invariant under a cyclic permutation11 . E100

12.3.3 The ggg vertex and its handlebar

On the basis of the previous section, we see that the only reasonable form of
the three-gluon vertex Feynman rule is
q3 j µ
ρ i
l ↔ g3 Y (q1 , µ; q2 , ν; q3 , ρ) hjkl (12.33)
q1 ~
q2
ν k
11
I have adopted the notation ‘Y ’ for this vertex since it reminds us both of the name
Yang(-Mills), and of the fact that in such a vertex three bosons meet.
352 August 31, 2019

Note that the gluon momenta are counted outgoing from the vertex. The
value of g3 must be determined, as well as the colour factor hjkl . The Y
function’s total antisymmetry strongly suggests that we take the h symbols
antisymmetric as well, and later on we shall show that this is indeed the case.
The following object will turn out to be useful :

∆(q)αβ ≡ q α q β − q 2 g αβ , (12.34)

for which
∆(q)αβ = ∆(q)βα , ∆(q)αβ qβ = 0 . (12.35)
Also,
∆(q)αβ β = q α (q · ) − q 2 α = 0 (12.36)
if  is the polarisation vector of an on-shell gluon with momentum q.
We now come to an important result. Let us consider the vertex of
Eq.(12.33), and let us put a handlebar on gluon q3 . We find, using momentum
conservation in the form q3 = −q1 − q2 ,

Y (q1 , µ; q2 , ν; q3 , q3 ) ≡ Y (q1 , µ; q2 , ν; q3 , ρ) q3 ρ
= (q1 − q2 · q3 )g µν + (q2 − q3 )µ q3 ν + (q3 − q1 )ν q3 µ
= (q2 − q1 · q2 + q1 )g µν − q2 µ (q1 + q2 )ν + q1 ν (q1 + q2 )µ
= ∆(q1 )µν − ∆(q2 )µν . (12.37)

Some algebra tells us, moreover, that in the axial gauge

Πµα (q1 ) ∆αβ (q1 ) Πβν (q2 ) = (i~g µα ) (gαβ ) Πβν (q2 ) , (12.38)

so that we find the handlebar rule

j
j j

= m + , (12.39)
m m
k k
k

with the auxiliary Feynman rules

j k ↔ i~ g µν δ jk (12.40)
µ ν
August 31, 2019 353

and
m g3 αβ jkm
↔ i g h . (12.41)
β α
~
k j
It is now time to return to the q q̄ → gg process. The new available
Feynman diagram, given by
a
p q1
1
n j

p2 q2 k
b

reads

M3 = −gg3 v(p1 )γµ u(p2 )Πµν (p1 + p2 )Y (q1 , 1 ; q2 , 2 ; −q1 − q2 , ν) hjkn (T n )a b .

(12.42)
with summation over the colour n implied. Putting the handlebar on gluon
1 as before, we get

n j n j n j
= + (12.43)
k k k

so that
M3 c = −i~gg3 v(p1 )/2 u(p2 ) hnkj (T n )a b . (12.44)
The total handlebarred amplitude now reads

a
M1+2+3 c = i~g v(p1 )/2 u(p2 ) g3 hjkn T n − g[T j , T k ] b (12.45)

The colour current will therefore be conserved if we choose

g3 = g (12.46)

and
[T j , T k ] = hjkn T n . (12.47)
354 August 31, 2019

Note that since the matrices T are Hermitean, the constants h must be purely
imaginary12 . Moreover, we can compute them, using Eq.(12.6), as

hjkl = 2 Tr T j T k T l − T l T k T j ,


(12.48)

from which we see that the h symbols must be totally antisymmetric. Since
they are related to commutators, we can use the Jacobi identity to find
relations between them13 :

0 = [[T j , T k ], T l ] + [[T k , T l ], T j ] + [[T l , T j ], T k ]

= hjk n [T n , T l ] + hkl n [T n , T j ] + hlj n [T n , T k ]
= hjk n hnl m T m + hkl n hnj m T m + hlj n hnk m T m , (12.49)

which after a few interchanges of indices leads to

hjk n hlm n + hjl n hmk n + hjm n hkl n = 0 . (12.50)

More information comes from colour conservation/democracy in the diagram

m
j k

from which we find the requirement that

X
hmnj hmnk = C δ j k , (12.51)
m,n

with some constant C. Eq.(12.51) is the gluonic equivalent of the property

(12.5) of the T matrices. It does not follow from the Jacobi identity. But
since we have already defined the h symbols by Eq.(12.48), it is not an extra
condition bur rather has to be proven : since

hmnj hmnk = 8 Tr T m T n T j Tr T m T n T k − Tr T m T k T n





, (12.52)
12
It is customary to write [T j , T k ] = i f jk n T n . The f ’s are then called the structure
constants, and the set of T matrices are then the generators of the Lie algebra of the group
SU (N ). The i is then combined with the overall i of the vertex to give a Feynman rule
without any i. This is of course a matter of taste.
13
Here and in the following, raising or lowering colour labels has no physical meaning ;
I do it only for typographical reasons.
August 31, 2019 355

and the reduction formulæ (12.16) give

1 jk
Tr T m T n T j Tr T m T n T k = −



δ ,
 4N 
m n j m k n N 1
δ jk ,



Tr T T T Tr T T T = − (12.53)
8 4N

we arrive at
hmnj hmnk = −N δ j k . (12.54)

12.3.4 On coupling quantisation

In the previous chapter we have discussed QED, characterized by the fact
that only fermion-fermion-photon couplings occur. The coupling constant
Qf for a given fermion is not constrained in any way, and there is no a priori
reason why different fermion species ought to have the same charge14 . In
QCD the situation is different. In deriving the ggg vertex we have started
with the process q q̄ → gg, without specifying the quark type, only saying
that its coupling to the gluon has coupling constant (‘colour charge’) g. We
then find that the three-gluon vertex has the same coupling constant g. But
that implies that all quark types must have the same coupling constant to
gluons, since two different quark types must lead to the same value of the
Yang-Mills coupling. We see that the presence of bosonic self-interactions
enforces a uniformity on the quark colour charges that was not there before.
In the next chapter we shall see how the presence of a W W γ vertex forces
the electron and muon charges to be, indeed, identical.

12.4 The four-gluon interaction

12.4.1 Colourful manipulations
Before proceeding it is useful to prepare some groundwork. Let us define,
first, the ‘four-colour’ object

[jklm] = hjk a hlm a = hjk a ha lm . (12.55)

14
Of course, leptons and quarks have different
√ charge ; but I rather refer to the ratio of
muon and electron charge, which could be π for all that QED cares.
356 August 31, 2019

This has the symmetries15

[jklm] = −[kjlm] = −[jkml] = [lmjk] ; (12.56)
and the Jacobi identity reads
[jklm] + [jlmk] + [jmkl] = 0 . (12.57)
In addition, we have the ‘five-colour’ object
[jklmn] = [jkla] hamn = hjk a ha l b hb mn . (12.58)
We can easily verify the symmetries
[jklmn] = −[kjlmn] = −[jklnm] = −[mnljk] (12.59)
and the two Jacobi identities
[jklmn] + [jkmnl] + [jknlm] = [jklmn] + [kljmn] + [ljkmn] = 0 . (12.60)
With such an arsenal of identities quite a few results can be derived: for
instance, by using each of the three symmetries and the two Jacobi identities
once, we can prove that
[jklmn] − [jklmn] + [lmjkn] − [lmjkn] = 0 ,
[jlkmn] + [kmjln] − [jlmkn] − [kmljn] = 0 . (12.61)

12.4.2 A purely gluonic process

We have seen that gluons, in contrast to photons, exhibit self-interactions.
We can therefore consider the process
g(q1 , 1 , j) g(q2 , 2 , k) → g(q3 , 3 , l) g(q4 , 4 , m)
which so far is given (at the tree level) by three Feynman diagrams16 :

j l j l j l
M = n + n
+ n (12.62)
k m k k m
m
15
These are precisely those of the Riemann-Christoffel tensor, familiar from the theory
of general relativity.
16
Since in this section we only consider gluons anyway, I shall here denote them by
smooth rather than wriggly lines ; this makes them somewhat easier to read, and certainly
easier to draw.
August 31, 2019 357

As before, we put a handlebar on gluon 1 :

j l j l j l
Mc1 →q1 = n + n + n (12.63)
k m k m k m

and this combination does not obviously vanish. The three-gluon vertex
is somewhat cumbersome, but we can streamline our calculations a bit by
introducing a ‘partial’ three-gluon vertex :

λ
n g
j ↔ i (q1 + q2 )λ g αβ hjkn . (12.64)
~
k q1 α
β q2

Here, the momenta are counted in the direction of the arrows. Note that
reversing the arrows leaves the vertex unchanged owing to the antisymmetry
of the h symbol. There is therefore no ambiguity. We may write

= + + . (12.65)

This makes it easier to single out particular terms in Eq.(12.63). For instance,
the terms proportional to (2 · 4 ) are given by the diagrams

j l j l j l
n + n + n
k m k m m
k

that, apart from an overall factor −i~g 2 (2 · 4 ), evaluate to

(3 · q2 + q4 ) hmkn hnlj + (3 · q1 + q2 + q4 ) hmnl hnkj + (3 · 2q2 − q3 ) hnkl hnmj
= (3 · q2 + q4 ) [mklj] − (3 · q1 + q2 + q4 ) [mlkj] + (3 · 2q2 − q3 ) [klmj]


= (3 · q2 + q4 ) −[mljk] − [mjkl]
+(3 · q1 + q2 + q4 ) [mljk] − (3 · 2q2 − q3 ) [mjlk]
= [mljk] (3 · q1 + q2 + q4 − q2 − q4 ) + [mjlk] (3 · q2 + q4 − 2q2 + q3 )


= (q1 · 3 ) [mjlk] + [mljk] . (12.66)
358 August 31, 2019

I have displayed this computation in detail in order to emphasize that we only

use momentum conservation, and not for instance the property (3 · q3 ) = 0.
In addition, in the third line the Jacobi identity comes into play. The other
terms are of course treated in the same way, so that we find

2


Mc1 →q1 = −i~g (q1 · 2 )(3 · 4 ) [jlmk] + [jmlk]


+ (q1 · 3 )(2 · 4 ) [jmkl] + [jkml]



+ (q1 · 4 )(2 · 3 ) [jklm] + [jlkm] . (12.67)

The handlebar requirement can now be satisfied by introducing a four-

gluon interaction vertex as follows:
α µ
j m g2
↔ i X(α, j; β, k; µ, m; ν, n) (12.68)
n ~
β k
ν
with

X(α, j; β, k; µ, m; ν, n) = g αβ g µν [jmnk] + [jnmk]

+ g αµ g βν [jknm] + [jnkm]

+ g αν g µβ [jkmn] + [jmkn]


. (12.69)

In addition, we have found the handlebar rule

= − − − . (12.70)

12.5 Current conservation in QCD

12.5.1 More vertices ?
After having introduced the three-gluon vertex, we have seen that we also
need a four-gluon vertex in order to save current conservation in qq → qq.
What if we now considergg → ggg ? Wil we also need a five-gluon vertex ?
And then what about gg → gggg ? Fortunately, as we shall see, no such
August 31, 2019 359

bad luck : with the three- and four-gluon vertices all amplitudes will vanish
under the handlebar. We shall prove this in the same manner as for QED,
only the proof will be obviously somewhat more involved. Before we turn to
our reliable working horse, the SDe, we first need one more result.

12.5.2 The Antkaz

Consider the four-gluon vertex in the following form :
j α
k β
l µ
m
ν

and let us now attach a fifth gluon in all possible ways, and slash the inter-
mediate propagators. This leads to the following expression :

j j j
n j
k n k
k k
n
l l
V = + l + l
+ n
. (12.71)
m
m m m

In V , no gluon momenta enter. Let us now concentrate on the term with

g αβ g µν , say. The colour part of this term reads

Vαβ;µν = [plmk] hpjn + [pmlk] hpjn + [jlmp] hpkn + [jmlp] hpkn

+ [jpmk] hpln + [jmpk] hpln + [jlpk] hpmn + [jplk] hpmn
= −[mkljn] − [lkmjn] + [jlmkn] + [jmlkn]
+[mkjln] − [jmkln] − [jlkmn] + [lkjmn] (12.72)

We now call upon the various symmetry properties and Jacobi identities :

Vαβ;µν = +[mklnj] + [lkmnj] + [jlmkn] + [jmlkn]

+[mkjln] + [lkjmn] + [jlknm] + [jmknl]
= −[mknjl] − [jlnmk] − [lknjm] − [jmnlk] = 0 . (12.73)

So we find that V vanishes completely, and we shall employ that fact in E96
what follows next.
360 August 31, 2019

12.5.3 Proof of current conservation

Let us consider an amplitude with a gluon line sticking out, and see what
the SDe has to tell us about it

= + + (12.74)

Let us now apply the handlebar rules we have developed in this chapter :

= −

+ − . (12.75)

It is important to realize that the internal lines in the SDe take on all pos-
sible identities, and therefore the third diagram stands for both graphs in
Eq.(12.39), and the fourth stands for all three diagrams in Eq.(12.70)17 . We
now iterate the SDe for the slashed propagators in the first three diagrams,
and this gives us

= −

+ + +

− = 0 , (12.76)

since the first diagram on the second line cancels against those of the first line,
and the last diagram on the second line vanishes all by itself, as we have seen
in the previous section. This establishes the proof of current conservation in
QCD. It is important to realise that the above proof relies on our use of the
axial gauge for the gluon propagator. Indeed, that choice is what makes the
identity (12.39) possible.
17
Are you worried about possible double-counting here ? Don’t worry, be happy : the
symmetry factors are there for precisely that reason.
August 31, 2019 361

12.6 Selected topics in QCD

12.6.1 White and coloured states
So far we have simply assumed quarks to have some colour label running
from 1 to N for quarks/antiquarks, and gluons to have another label running
from 1 to N 2 − 1 ; in the rest of this section we shall use the observed value
N = 3. The whole idea of quark (and gluon) colours implies that we assume
our QCD physics to be invariant if we ‘rotate’ in colour space, or rather, if
we apply to the quarks and antiquarks a unitary operation :

q a → Rba q b , q̄a − → q̄b (R† )ba , (12.77)

with the summation convention implied. Here the R are unitary 3×3 matrices
with unit determinant : they constitute the group SU(3). The quarks are
said to be triplets (3) under colour, and the antiquarks to be anti-triplets
(3̄). It pays to examine what kind of colour combined states of quarks and
antiquarks can have18 . In the first place, the combination q̄a q a transforms as

q̄a q a → q̄b (R† )ba Rca q c = q̄b δcb q c = q̄c q c : (12.78)

this combination doesn’t change colour at all, it is the ‘white’ singlet. The
other eight possible (orthogonal) q q̄ colour combinations are the colour octet
(8). In group-theory speak we say that 3 ⊗ 3̄ = 1 ⊕ 8. A typical octet state is,
for instance, q̄1 q 2 . The fact that the matrices T j are traceless combinations
of a colour and and anticolour index tells us that gluons are colour octets
as well. We can also investigate how the special quark-quark combination
abc q b q c transforms, where abc is the three-dimensional Levi-Civita symbol :

abc q b q c → abc Rm
b
Rnc q m q n . (12.79)

We start by observing that the object abc R`a Rm

b
Rnc is totally antisymmetric
in its three indices `, m, n, and must therefore itself be proportional to the
Levi-Civita symbol ; furthermore, we have
1 `mn
 abc R`a Rm
b
Rnc = det (R) = 1 , (12.80)
3!
18
We shall not delve deeply into group theory here, but only consider some of the
representations of SU(3).
 362 August 31, 2019

so that it actually is the Levi-Civita symbol. Judiciously multiplying with

R† we then find that

abc q b q c → `mn q m q n (R† )`a : (12.81)

these quarks are also in the antitriplet, 3̄ state ! The remaining orthogonal
quark-quark colour combinations are the sextet (6), so that we can say that
3 ⊗ 3 = 3̄ ⊕ 6. A typical sextet state is, for instance, q 3 q 3 . Finally we
can observe that the colour combination abc q a q b q c is also a colour singlet,
a ‘white’ state. The empirical fact that we see only mesons (q q̄), baryons
(qqq), and antibaryons (q̄ q̄ q̄) occurring as more-or-less stable, more-or-less
free hadrons can therefore be nicely summed up by the postulate that only
E97 ‘white’ (singlet) colour states can occur freely.

12.6.2 The QCD Coulomb interaction

Just like we did for QED (cf section 10.5) we can also investigate what
QCD tells us about almost-static scattering of quarks. Let us start with q q̄
scattering, using the same notation as in 10.5 and indicate the various colours
as well19 :
b a
p q
1 i~g 2 1
M(~k) = p2 k j u(q1 )γ µ u(p1 ) v(p2 )γµ v(q2 ) Cbd
= ac
,
q2 |~k|2
c d
 
ac j a j c 1 a c 1 a c
Cbd = (T )b (T )d = δ δ − δ δ . (12.82)
2 d b N b d
The attraction/repulsion discussion now centers on the colour structure of
the q q̄ state. If the quark-antiquark pair is in the singlet state, we must
compute20
2
N2 − 1
  
1 d b ac 1 a d c b 1 a b c d
√ δa δc Cbd = (δd δa )(δb δc ) − δb δc δd δa = , (12.83)
N 2N N 2N
and the interaction energy is negative. On the other hand, for the alternative
octet states we may choose a = b 6= c = d and C then evaluates to −1/2N .
19
And remember the discussion of the Fermi sign !
20
Note that, even with conservation of colour, the colour assignment 11̄ → 22̄ is al-
lowed. We must therefore sum
√ over the initial- as well as over the final-state colour singlet
combination. The factors 1/ N are the normalization factors for the singlet state.
August 31, 2019 363

We conclude that a q q̄ pair in a colour singlet state attracts, but in the octet
state repels. And indeed, colour octet states are never observed21 . We can
also consider qq scattering :
b a
p q
1 1 −i~g 2
M(~k) = p k j = u(q1 )γ µ u(p1 ) u(q2 )γµ u(p2 ) Cbd
ac
. (12.84)
2 q2 ~
|k|2
d c

Two quarks cannot form a singlet state, but they can form an antitriplet
state, an antisymmetric combination of two colours, for which we may take
1 and 3, say. The colour factor now evaluates to22
 2
1 13 13 31 31

N +1
√ C13 − C31 − C13 + C31 =− (12.85)
2 2N
and the interaction is attractive. The remaining six colour combinations are
the symmetric sextet ones, for which we may simply choose a = b = c = d =
33
3, say, and then find C33 = (N − 1)/2N . A quark pair in such a state feels a
23
repulsive interaction .

12.6.3 The process q q̄ → gg

The amplitude
For an example of a QCD calculation we may look to the process

q̄(p1 )q(p2 ) → g(q1 , λ1 , j) g(q2 , λ2 , k) ,

in the high-energy approximation where the quark mass is negligible ; we

have indicated the momenta and helicities of the quarks and gluons, as well
as the gluon colour labels. We have already studied this process in order
21
This is of course horribly handwaving since we rely on tree-level perturbation theory
here, which is known to be a very poor approximation at very small scattering momentum.
Nevertheless, as a suggestion is comes out right !
22
This expression is a little bit too pretentious : the antitriplet state only makes sense
if N = 3 (otherwise the Levi-Civita symbol in Eq.(??) would not make sense. The only √
really sensible value for the colour factor in Eq.(12.85) is therefore -2/3, and the 1/ 2
factors are again the state’s normalization.
23
Quark pairs in a triplet state have not been observed. However, inside a proton, say,
all three possible quark pairs are in an antisymmetric colour combination since the whole
state is totally antisymmetric : again only a suggestion, but again in the right direction.
364 August 31, 2019

to establish the three-gluon vertex, but what interests us here is not how
the unwanted terms vanish, but rather what are the acceptable terms that
are left. As we have seen, at the tree level the amplitude is described by 3
Feynman diagrams :
p1 q1 p1 q2 p1 q1
M= (12.86)
p2 q2 p2 q1 p q2
2

and we can write it as

 
2 1 j k 1 k j 1 j k
M(λ1 , λ2 ) = ig ~ − A1 T T − A2 T T + A3 [T , T ] ,
t u s
A1 (λ1 , λ2 ) = u+ (p1 )/1 (/q1 − p/1 ) /2 u+ (p2 ) ,
A2 (λ1 , λ2 ) = u+ (p1 )/2 (/q2 − p/1 ) /1 u+ (p2 ) ,
A3 (λ1 , λ2 ) = u+ (p1 )γα u+ (p2 ) Y (q1 , 1 ; q2 , 2 ; −q1 − q2 , α) , (12.87)
where as usual s = (p1 + p2 )2 , t = (p1 − q1 )2 , u = (p1 − q2 )2 , and s +
t + u = 0. Note that we consider only the + helicity case for the quarks.
This is quite allright, since the QCD couplings do not contain any γ 5 and
overall interchange of all positive and negative helicities can only result in an
overall phase change of the amplitude. It is at this point that in calculations
like these a handlebar check like 1 → q1 is very useful to verify e.g. the
correctness of the signs. Since we have already used the handlebar in section
12.3.3, we shall not do it now. We have to calculate three A’s, with four
helicity cases for each.

The polarisations
The first (and perhaps most crucial) step24 is to make a clever choice for the
polarisation vectors. It turns out to be smart indeed to take
u+ (qi )γ µ u+ (p2 ) u− (qi )γ µ u− (p1 )
i (+)µ = √ , i (−)µ = √ , (12.88)
2 s− (qi , p2 ) 2 s+ (qi , p1 )
just like we did in QED, since then
√ √
2 u− (qi )u− (p2 ) 2 u− (p1 )u− (qi )
ω− /i (+) = , ω− /i (−) = ; (12.89)
s− (qi , p2 ) s+ (qi , p1 )
24
Again, you must realise that the choice of polarisation vector cannot change the final
result for the amplitude : but making the intermediate steps easy or even trivial can save
a lot of sweat and tears (if not blood).
August 31, 2019 365

which, if we insert the polarisation vectors into the diagrams, leads to

/i (+)u+ (p2 ) = u+ (p1 )/i (−) = 0 . (12.90)
In addition we have
1 (±) · 2 (±) = 0 ,
s− (q2 , p2 )s+ (q1 , p1 )
1 (+) · 2 (−) = ,
s− (q1 , p2 )s+ (q2 , p1 )
s− (q1 , p2 )s+ (q2 , p1 )
1 (−) · 2 (+) = . (12.91)
s− (q2 , p2 )s+ (q1 , p1 )

Spinorial workout
It is easy to see that out of the twelve cases, eight vanish straight away25 :
An (±, ±) = A1 (−, +) = 0 = A2 (+, −) = 0 (n = 1, 2, 3) , (12.92)
The four remaining cases are readily evaluated in terms of spinor products26 :
s− (p2 , q2 )
A1 (+, −) = 2s+ (p1 , q1 )2 ,
s+ (p1 , q2 )
A3 (+, −) = −A1 (+, −) ,
s− (p2 , q1 )
A2 (−, +) = 2s+ (p1 , q2 )2 ,
s+ (p1 , q1 )
A3 (−, +) = A2 (−, +) . (12.93)
Perhaps surprisingly, the A3 ends up looking just like the A1,2 ! If we neglect E98
overall complex phases, we have
A1 (+, −) ∼ 2t t/u , A2 (−, +) ∼
p p
= 2u u/t , (12.94)
and so we find
 
p 2 1 j k 1 j k
M(+, −) ∼ 2g ~ t t/u − T T − [T , T ]
t s
2 p
2g ~
t/u u T j T k + t T k T j .


= (12.95)
s
25
Note that, since in this calculation, u(p1 )/i u(p2 ) is always zero, we here have
Y (q1 , 1 ; q2 , 2 , −q1 − q2 , α) ∼ (q1 − q2 )α (1 2 ).
26
In the A3 results, it may be useful to note that there are alternative forms : u+ (p1 )(/q1 −
q2 )u+ (p2 ) equals 2s+ (p1 , q1 )s− (q1 , p2 ) as well as −2s+ (p1 , q2 )s− (q2 , p2 ).
/
 366 August 31, 2019

Here we have used −1/t − 1/s = u/ts. In exactly the same way we find
2g 2 ~ p
u/t u T j T k + t T k T j .


M(−, +) ∼ (12.96)
s
In these expressions, replacing T j,k by unity gives us the QED result for
e+ e− → γγ if we remember that u + t = −s.

Summing and averaging, and symmetry

In order to compute h|M|2 i we must count a factor 1/4 for the spin average,
a factor 1/N 2 for the colour average, and a factor 2 for the so-far neglected
case of negative quark helicity. We find
2g 4 ~2 t
 

2
 u
|M| = + ×
N 2 s2 u t
(t2 + u2 )Tr T j T k T k T j + 2utTr T j T k T j T k . (12.97)

E99 The colour sums are most easily handled using the Fierz identities (12.16) :
(N 2 − 1)2
Tr T j T k T k T j


= ,
4N
(N 2 − 1)
Tr T j T k T j T k = −


. (12.98)
4N
The final result is
 g 4 ~2 (N 2 − 1) 2
 2
(N − 1) 2N 2

2 2

|M| = (t + u ) − 2 . (12.99)
2N 3 ut s
The differential cross section is
dσ 1
2

= |M| : (12.100)
dΩ 128π 2 s
note the symmetry factor Fsymm = 1/2 for the 2-gluon final state !

12.6.4 The Landau-Yang theorem revisited

A loophole in Landau-Yang ?
We may return to the Landau-Yang theorem discussed in the context of QED
in section 10.7.3. It states that no spin-1 particle can decay into two pho-
tons. The reason was seen to be that each possible contribution to the decay
August 31, 2019 367

amplitude (10.194) was antisymmetric in the interchange of the photons, in

violation of their Bose statistics. Now suppose we replace the photons by
gluons, that have an additional colour property. Can we build an amplitude
that is also antisymmetric in the gluon colour labels ? In that case the total
amplitude would obey Bose statistics and would hence not be forbidden. To
investigate this, we shall look into quarkonium decay.

Threshold amplitudes for q q̄ annihilation

We shall compute (at tree level) the amplitude for the annihilation process
of a (now massive ! ) quark-antiquark pair into two gluons :

q(p) q̄(p) → g(q1 ) q(q2 ) ,

where the quark and the antiquark, of mass m, have the same momentum, i.e.
they annihilate at rest. As before, we shall employ the modified polarisation
vectors
(q2 1 ) µ
η1µ = µ1 − q (12.101)
(q1 q2 ) 1
(and vice versa), and q µ = (q1µ − q2µ )/2. With the colour labels of gluon 1
and 2 denoted by j and k, respectively, the amplitude is again given by the
diagrams (12.86) but now it reads

ig 2 ~
M = − 2 2
v(p)/η1 (/q1 − p/) η/2 u(p) T j T k
(p − q1 ) − m
ig 2 ~
− 2 2
v(p)/η2 (/q2 − p/) η/1 u(p) T k T j
(p − q2 ) − m
ig 2 ~
+ v(p)γα u(p) Y (q1 , η1 ; q2 , η2 ; −2p, α) [T j , T k ]
(2p)2
ig 2 ~
= + 2 v(p)/η1 q/η/2 u(p) {T j , T k } + [T j , T k ]


4m
ig 2 ~
− 2 v(p)/η2 q/η/1 u(p) {T j , T k } − [T j , T k ]


4m
ig 2 ~
+ 2 v(p)/qu(p) (η1 η2 ) [T j , T k ] . (12.102)
2m
The first two lines corresponds to the ‘QED-like’ diagrams with quark ex-
change, and the third line derives from the s-channel diagram containing the
368 August 31, 2019

three-gluon vertex. Because (q · η1,2 ) = 0 we can write, e.g.,

η/1 q/η/2 = −/q(η1 η2 ) + iγ 5 γµ µ (q, η1 , η2 ) , (12.103)

and, since q and the η1,2 are orthogonal to p,

1
γ 5 γµ µ (q, η1 , η2 ) = 2
(p, q, η1 , η2 ) γ 5 p/ (12.104)
m
and we see two things happening : the terms with the commutator [T j , T k ]
cancel amongst themselves, and for the terms with the (symmetric) anticom-
mutator {T j , T k } we have

g2~
M=− (p, q, η1 , η2 ) {T j , Y k } v(p)γ 5 u(p) . (12.105)
m3
We now consider the total spin of the q q̄ system. If it has spin-1, the spin
vectors of the quark and antiquark must be the same, and then we find
1
v(p, s)γ 5 u(p, s) = v(p, s)γ 5 1 + γ 5 s/ u(p, s)


2
1
v(p, s) 1 − γ 5 s/ γ 5 u(p, s) = 0 .


= (12.106)
2
For opposite spin vectors, the amplitude is nonzero : |v(p, s)γ 5 u(p, −s)|2 =
4m2 . Thus, a quark-antiquark pair at relative rest in a spin-1 state cannot
decay into two gluons, and the Landau-Yang theorem appears to hold here
as well - at the tree level. We might turn QCD into QED by replacing the T
matrices by unity, and so find the proof for orthopositronium.

Landau and Yang come down from the trees

We note that the vanishing of the colour-antisymmetric term (with [T j , T k ])
in the above relied on a cancellation between three contributions. If we now
consider loop effects, we see that the s-channel graph is modified by quark
loops in the gluon propagator and in the three-gluon vertex27 ,

27
This last diagram is discussed in section 12.3.2.
August 31, 2019 369

and that the t- and u-channel diagrams do not pick up similar quark loops.
Since the number (and certainly the masses) of the other quarks are not fixed
by any principle there is, therefore, no possibility of a similar cancellation
holding at the one-loop level and beyond ; and we conclude that the Landau-
Yang theorem does not hold for gluons28 .

12.7 Exercises
Excercise 96 Prove the Antkaz !
Prove the result (12.73) by using the properties of the [jklmn] symbols.

Excercise 97 A four-colour problem ?

Suppose that the number of QCD colours were 4 rather than 3. Use what
you know about fermions versus bosons, the Pauli exclusion principle, and
general relativity to argue that in that case almost all hadronic matter in the
universe would disappear into a few huge black holes, and we wouldn’t be
here to do this exercise. . .

Excercise 98 QCD spinorial workout

Using standard-spinor techniques, prove Eq.(12.93).

Excercise 99 Colour traces for q q̄ → gg

Prove the two colour traces of Eq.(12.98).

Excercise 100 Fun with Furry’s Failure

Consider the diagram K of section 12.3.2. If one of the couplings were not of
vector type (with γ µ ) but of axial-vector type (with γ 5 γ µ ), then the integral
would change sign under the above transformations. Show that in that case
the function Y would read

−q3ρ g µν − q1µ g νρ − q2ν g ρµ

and show that this is completely transverse to any external polarisation vec-
tor.
28
Note, however, that this violating would show up only at second loop order, since the
interference of the tree-level result with the one-loop result would of course still vanish.
Experimentally, therefore, the violation is very modest.
370 August 31, 2019

Excercise 101 Gluonic MHV amplitudes

We consider the purely gluonic process

g(q1 , 1 ) g(q2 , 2 ) → g(q3 , 3 ) g(q4 , 4 ) . . . , g(qn , n )

at the tree level.

1. Show, using Eq.(6.23), that the amplitude, if correct, must be strictly
prtoportional to E 4−n , where E is the total invariant energy of the
system.

2. Show that in every contributing diagram the maximum number of prop-

agators equals n − 3.

3. Show that the denominator of every diagram goes at most as E 2n−6 .

4. Show that the numerator of each diagram consists of three types of

factors : (p·p), (p·), and (·), where the p stand for some combinations
of the momenta q1 , q2 , . . . , qn , and  for some of the polarization vectors
of the external gluouns.

5. Show that the maximum number of p’s in the numerator is n − 2, and

that therefore every diagram contains at least one ( · ).

6. Choose the gluon polarizations according to Eq.(9.53), with gauge vec-

tor rj for gluon j. Show that (a) if gluon j and gluon k have the
same helicity λ, then (j · k ) is proportional to sλ (qj , qk )s−λ (rj , rk ), (b)
if gluon j has helicity λ and gluon k has helicity −λ then (j · k ) is
proportional to sλ (qj , rk )s−λ (rj , qk ).

7. By choosing the gauge vectors cleverly, show that the amplitude van-
ishes if all n gluons have the same helicity.

8. By choosing the gauge vectors cleverly, show that the amplitude van-
ishes if n−1 gluons have the same helicity and one the opposite helicity.

9. Show that, when two gluons have helicity opposite from the other n − 2
gluons we cannot choose all gauge vectors such that the amplitude
vanishes.
These amplitudes, with 2 gluons having polarization opposite from the other
ones, are called MHV amplitudes, see also section 10.7.1.
August 31, 2019 371

Excercise 102 Other MHV amplitudes

In the process q q̄ → gg · · · g the tree-level diagrams have the generic form

We conider the case where all gluons have the same helicity.
1. Let us first concentrate on a single ‘blob’ where an off-shell gluon com-
ing from the q − q̄ line decays into n gluons:
µ

We indicate the Lorentz index. We shall denote the momenta of the

outgoing gluons generically by p, and their polarization vectors by .
Show that every term in this off-shell amplitude is of one of two possible
forms:
(a) : pµ (p · )α ( · )β (p · p)−γ , (b) : µ (p · )α ( · )β (p · p)−γ

2. Show that the energy dependence of this amplitude is p1−n .

3. Show that the maximum number of gluon propagators in each term is
equal to n − 1, so that γ ≤ n − 1
4. For case (a), show that α + 1 − 2γ = 1 − n and α + 2β = n, and that
this implies β ≥ 1.
5. Show that all terms of the form (a) vanish by a suitable gauge vector
choice.
6. For case (b), show that α − 2γ = 1 − n and α + 1 + 2β = n, and that
that implies β ≥ 0.
7. Show that all nonvanishing terms of the form (b) are therefore of the
generic form µ ( · p)n−1 (p · p)1−n .
8. Show that this implies the MHV result that the amplitude for q q̄ →
gg · · · g vanishes at the tree level if all gluons have the same helicity.
372 August 31, 2019
Chapter 13

Electroweak theory

13.1 Muon decay

13.1.1 The Fermi coupling constant
Let us return to the Fermi model of muon decay as discussed in chapter 7.
There, the (phenomenological) amplitude for this decay was proposed to be
of the form of Eq.(8.36). The resulting width was
GF 2 ~2 mµ 5
Γµ ≡ Γ(µ− → e− ν e νµ ) = . (13.1)
192π 3
The measured values of the mechanical mass Mµ and the lifetime τµ of the
muon are
Mµ ≈ 1.884 10−28 kg , τµ ≈ 2.197 10−6 sec ; (13.2)
the muon mass may be more familiar under its appellation of Mµ c2 ≈
0.106 GeV. From these we can construct the more useful quantities
Mµ c 1
mµ = ≈ 5.354 1014 m−1 , Γµ = ≈ 1.518 10−3 m−1 . (13.3)
~ cτµ
From Eq.(13.1) we then find
GF ~ ≈ 4.532 10−37 m2 , (13.4)
or
GF
2
≈ 1.164 10−5 GeV−2 . (13.5)
~c

373
374 August 31, 2019

We can therefore derive the ‘energy scale’ of the interaction responsible for
muon decay1 : s
~c2
ΛW = ≈ 292.5 GeV . (13.6)
GF

13.1.2 Failure of the Fermi model in µ− ν µ → e− ν e

If the phenomenologically motivated Fermi interaction were to have any claim
to global validity, it should also govern the process
µ− (p1 ) ν µ (p2 ) → e− (q1 ) ν e (q2 ) , (13.7)
which amounts to the previous process, only with the outgoing muon neutrino
moved to an incoming anti-muon neutrino. No matter that we cannot, at
present, build µν µ colliders ; the very, very, very early universe did provide
such processes, and their description must be correct. By the rules of the
Fermi model, the amplitude is given by
GF ~
M = i √ v(p2 )(1 + γ 5 )γ µ u(p1 ) u(q1 )(1 + γ 5 )γµ v(q2 )
2
4
= i √ GF ~ v − (p2 )γ µ u− (p1 ) u− (q1 )γµ v− (q2 )
2
8
= i √ GF ~ s− (p2 , q1 ) s+ (q2 , p1 ) (13.8)
2
Here, we have neglected both the muon and the electron mass since the
scattering takes place at high energy, and we have applied the Chisholm
identity in order to remove the contracted Lorentz index. Disregarding phases
and using momentum conservation, we then find
16 GF ~
Mli √ (p1 · q2 ) . (13.9)
2
Neutrinos2 have only one helicity state, and therefore, in the centre-of-mass
system,
|M|2 = 64 GF 2 ~2 (p1 · q2 )2 = 4 GF 2 ~2 s2 (1 + cos θ)2 ,


(13.10)
1
What precisely √
constitutes the scale is of course to some extent
√ a matter of taste. If
we include a factor 2 in GF the scale is reduced by a factor ( 2)1/2 to 246 GeV, which
is the more commonly used number.
2
We shall assume, in this section, that neutrinos are strictly massless.
August 31, 2019 375

where θ is the angle between the muon and electron momenta. By taking
also the angular average we obtain
16
|M|2 GF 2 ~2 s2 .


= (13.11)
3
The total cross section is therefore given by

− − GF 2 ~2
σ(µ ν µ → e ν e ) = s (13.12)
3π
As we have seen before, only the factor 1/3 is not immediately obvious in
this expression but has to be computed from the Feynman diagrams.

The scattering cross section rises linearly with s, and will therefore violate
the unitarity bound at sufficiently high energy. Since the the muon and its
antineutrino couple with a single γ µ , they must be in a J = 1 state. The
unitarity bound on this cross section (cf appendix 15.10) is therefore
1 16π 24π
σ(µ− ν µ → e− ν e ) ≤ (2J + 1) = , (13.13)
2 s s
which leads to a fundamental failure
√ of the Fermi model (at least, at the tree
level) at a scattering energy of s ≈ 1.5 TeV.

13.2 The W particle

13.2.1 The IVB strategy
We have to modify the Fermi model in such a way that its success in the
low-energy description of muon decay is preserved, while at high energies
unitarity remains inviolate. One possible way out might be to simply make
GF depend on the energy scale of the process so that it decreases at high
energies, making the µ− ν µ → e− ν e cross section well-behaved. For this, we
would need to replace

GF → GF (0) f (s/Λ2 ) (13.14)

with f (x) decreasing at least as fast as 1/x for large arguments, and Λ
some constant energy. Such energy-dependent couplings, called form factors,
376 August 31, 2019

are employed in for instance ‘low-energy’ hadronic physics ; in such cases,

however, this approach is generally viewed as an admission of ignorance of,
and an attempt to cope with, some underlying and simpler physics at a
smaller distance scale 1/Λ ; that is, one would basically be invoking some
‘new physics’ that would make everything work out right 3 .

The more elegant, and (as it turns out) the correct way to go is to make
the Fermi model look more ‘QED-like’: instead of using a contact interaction
between four fermions, we postulate the existence of a new particle, the so-
called W boson. This couples to fermion-antifermion pairs in a way reminis-
cent of the photon. The four-fermion interaction then resolves into two f f¯W
interactions, with the W boson mediating between the two vertices ; the cor-
responding Feynman diagram for the process µ− (p) to e− (q) νµ (k1 ) ν e (k2 )
is therefore given by

p k1

M = Q (13.15)
q
k2

At the time this model was first seriously discussed, it went under the name
of Intermediate Vector-Boson (IVB) hypothesis. We take the W to couple
to the fermion pairs eνe and µνµ , so that (as we shall check!) the W must be
electrically charged, and assume that the coupling is in both cases of equal
strength4 (for now). We therefore postulate the following Feynman rules :

3
In the present case, GF , being dimensionful, sets such a scale by itself.
4
At this point, these are of course just assumptions. Since 1983, when the W boson was
first freely produced, they have been tested with great accuracy. The alternative scenario
of the charge retention form, with an electrically neutral W , is completely ruled out by
the fact that the decay W → e+ µ− is never seen. The equality of the couplings is verified
by the fact that the branching ratios for W → eν e and W → µν µ are the same up to
computable mass effects.
August 31, 2019 377

µ k −g µν + k µ k ν /mW 2 internal W lines

ν ↔ i~
k 2 − mW 2 + i

i
gW 1 + γ 5 γ µ f f 0 W vertices


↔
µ ~

EW Feynman rules, part 13.1 (13.16)

The W propagator is the standard one for a vector particle. Note that the
occurrence of the (1 + γ 5 ) in the vertex is suggested by the form of the Fermi
interaction ; and that the two fermions meeting in the vertex must be of
different type. The values of mW and gW are to be determined. Another
attractive property of this model is that here the coupling constant, gW , has
the same dimensionality as the QED one, and does not formally contain a
length scale.
With the above Feynman rules, the muon decay amplitude can now be
written as
i~gW 2

M = u(k1 )(1 + γ 5 )γ α u(p) u(q)(1 + γ 5 )γα v(k2 )
Q2 − mW 2

1 5 5
− u(k1 )(1 + γ )/Qu(p) u(q)(1 + γ )/Qv(k2 ) , (13.17)
mW 2
where the momentum of the internal W is given by
Qµ = (p − k1 )µ = (q + k2 )µ . (13.18)
The last term in Eq.(13.17) appears to deviate significantly from the spinorial
structure of the first term, which coincides with the Fermi model. Hoewever,
notice that
u(k1 )(1 + γ 5 )/Qu(p) = u(k1 )(1 + γ 5 )(/p − k/1 )u(p)
= u(k1 ) − k/1 (1 − γ 5 ) + (1 + γ 5 )/p u(p)

= mµ u(k1 )(1 + γ 5 )u(p) (13.19)

upon application of the Dirac equation to the external spinors ; and since, in
the same way,
u(q)(1 + γ 5 )/Qv(k2 ) = me u(q)(1 − γ 5 )v(k2 ) , (13.20)
378 August 31, 2019

the second term in Eq.(13.17) is actually suppressed by a factor (me mµ )/mW 2 ,

which is small if mW is sufficiently large5 . Neglecting this term, we see that
the Fermi-model amplitude
√ is recovered with the single replacement of the
coupling constant GF / 2 by gW 2 /(Q2 − mW 2 ). Now, the maximum value
that Q2 can take in this process is mµ 2 , which is attained in the improbable
case that the muon neutrino emerges with zero momentum from the decay.
If, therefore, we assume that mW is large compared to mµ , we see that the
successes of the Fermi model in describing muon decay will be completely
reproduced provided
gW 2 GF
l √ , (13.21)
mW 2 2
which we may also write in purely dimensionless terms as
mW c2
   
gW 1
√ = 1/4 . (13.22)
c ~ 2 ΛW

13.2.2 The cross section for µ− ν µ → e− ν e revisited

We can now study the modification that the IVB hypothesis makes in the
cross section for the process µ− ν µ → e− ν e , where the Fermi model fails. In
this case the total invariant mass is (assumed to be) much larger than the
W mass, so that the modified prediction can immediately be seen to be
2
2~2 gW 4 ~2 GF 2 s mW 2

− − s
σ(µ ν µ → e ν e ) = = ,
3π (s − mW 2 )2 3π s − mW 2
(13.23)
and this cross section does decrease as 1/s for large s.
Of course, the unitarity limit (13.13) still has to be observed, which puts
an upper limit6 on the useful values of mW :

mW c2 ≤ (72π 2 )1/4 ΛW ≈ 1.5 TeV . (13.24)

However, from Eq.(13.22) we see that for such large values the dimensionless
coupling constant is so large that the tree-level approximation for the cross
section is questionable.
5 −7
In fact, for the actual values of the
√ masses the suppression factor is about 10 .
6
This value is close to the value of s at which unitairy breaks down in the unmodified
Fermi model, see Eq.(13.13). This is not a coincidence. Whatever we do to the electroweak
interactions, 1.5 TeV appears to be the energy régime where things get tricky.
August 31, 2019 379

One may wonder what happens at s = mW 2 . There, the cross section

would seem to diverge ! We must realise, however, that at that energy we
are, in fact, producing an on-shell W that decays into a fermion-antifermion
pair : that is to say, the W is an unstable particle, and has a decay width.
We ought, therefore, to include the decay width into the propagator, so that
in the neighbourhood of the resonance at s ≈ mW 2 the cross section reads

2~2 gW 4 s
σ(µ− ν µ → e− ν e ) = . (13.25)
3π (s − mW ) + mW 2 ΓW 2
2 2

This is well below the unitarity limit. The IVB hypothesis therefore indeed E103
cures the unitarity problem in this process.
E104
Because of these successes, we shall adopt the notion of an existing W
particle of spin 1 (and hence obeying the lines laid out in chapter 9), coupling
to pairs of fermions separated by one unit of charge7 .

13.2.3 The W W γ vertex

Minimal coupling
Since the W particle couples to fermion pairs of unequal charge, it must
itself also be charged8 , which means that it must couple to the photon in (at
least !) a W W γ vertex. It is our aim now to find the form of such a vertex.
Both W ’s and photons are characterised by the fact that, in addition to
their momentum, they carry also a polarisation vector, i.e. a Lorentz index:
the W W γ vertex must therefore carry no fewer than 3 Lorentz indices. As a
first attempt, we can simply view the W particles as a kind of funny scalars,
and adopt the sQED vertex dressed up with a metric tensor to take care of
the W indices. That is, the Feynman rule for the vertex is taken to be
ν
+
γ(p3 )
W (p1 ) ↔
i
QW (p1 − p2 )ρ g µν (13.26)
_ ρ ~
W (p2 )
µ
where the coupling constant (the W charge) is to be determined, and the
particles are considered to be outgoing from the vertex. To this end, let us
7
Note that this automatically rules out couplings between a W , a lepton, and a quark.
8
At pain of charge nonconservation, i.e. at pain of pain.
380 August 31, 2019

examine the process

D(q1 ) U (q2 ) → γ(k1 , ) W + (k2 , + ) .
 and W denote the polarisation vectors of the photon and the W , respec-
tively, and we have indicated the particle momenta. Here, and in the follow-
ing, we shall denote by U and D two fermions of which the U has an electric
charge one unit higher than the D: for instance, U = νe and D = e, or U = u
and D = d. Their respective charges are QU and QD . At the tree level, we
then have three Feynman diagrams :
q1 q1 q1
k1 k
k2 2

M = + +
q2 k2 q2 k1 q2 k1
(13.27)
The three diagrams correspond to the three partial matrix elements
k/1 − q/1 + mD
M1 = −i~gW QD v(q1 )/ (1 + γ 5 ) /W u(q2 ) ,
(k1 − q1 )2 − mD 2
q/2 − k/1 + mU
M2 = −i~gW QU v(q1 ) (1 + γ 5 ) /W / u(q2 ) ,
(q2 − k1 )2 − mU 2
M3 = +i~gW QW v(q1 ) (1 + γ 5 ) γα u(q2 )
g αβ − P α P β /mW 2
W β (2k2 + k1 · ) , (13.28)
s − mW 2
where s = P 2 , P = q1 + q2 = k1 + k2 .
Since this process involves a produced photon, the handlebar identity
must hold : if we replace µ by k1 µ the amplitude must vanish. We shall
investigate this in some detail, by explicit computation rather than by the
diagrammatic manipulations of the previous chapters. In the first place, we
perform some simple Dirac algebra to note that

v(q1 )/(/k1 − q/1 + mD ) →k1 = v(q1 )/k1 (/k1 − q/1 + mD )
= v(q1 ) k1 2 − 2(q1 · k1 ) + (/q1 + mD )/k1

= (k1 − q1 )2 − mD 2 v(q1 ) ,


(13.29)
where in the second line we have used anticommutation between k/1 and q/1 ,
and in the third line the Dirac equation for v(q1 ). We see that
M1 c→k1 = −i~gW QD v(q1 ) (1 + γ 5 ) /W u(q2 ) , (13.30)
August 31, 2019 381

and similarly

M2 c→k1 = +i~gW QU v(q1 ) (1 + γ 5 ) /W u(q2 ) . (13.31)

For the third diagram we find

M3 c→k1 = +i~gW QW v(q1 ) (1 + γ 5 ) /W u(q2 )

−i~gW QW (k1 · W ) v(q1 ) mU (1 + γ 5 ) − mD 1 − γ 5



u(q2 ) .
(13.32)

If we were allowed to consider only the first of the two terms of the result
(13.32), we could obtain the desired cancellation :
3
%
X
Mj = 0 ⇒ QW = QD − QU : (13.33)
j=1 →k1

but the second term in Eq.(13.32) spoils this idea by having a quite different
algebraic structure ; no tuning of coupling constants is going to ensure that
a W W γ vertex of the form (13.26) can do the job.

The W -photon coupling

Treating the W W γ vertex as a prettified sQED vertex does not work. It
means that the photon-W interactions cannot be obtained by the minimal-
substitution rule. This should not come as a surprise since the vertex (13.26)
is only designed for graceful behaviour towards longitudinal photons, not
towards longitudinal W ’s. We therefore propose to replace Eq.(13.26) by a
vertex of the form
QW
(a1 p1 + a2 p2 )ρ g µν + (a3 p2 + a4 p3 )µ g νρ + (a5 p3 + a6 p1 )ν g ρµ . (13.34)


i
~
Of course, we have encountered a similar vertex when considering the three-
gluon coupling in QCD (cf section 12.3). However, we are now dealing with a
much less symmetric situation since two out of the three bosons are massive,
and it behooves us to be careful. Note that because of momentum conser-
vation each of the three terms need contain only two of the momenta; the
constants a1,...,6 are to be determined. This we shall do by considering several
situations.
382 August 31, 2019

First, we consider the process of decay of a photon in a W + W − pair :

γ ∗ (q) → W + (k+ , + ) W − (k− , − ) .
Kinematically this is only possible if the photon is quite off-shell, and there-
fore we do not give it a polarisation vector but leave its Lorentz index µ free.
The matrix element is given by
M = i~1/2 QW Aµ ,
Aµ = (a1 k+ + a2 k− )µ (+ · − )
+((a3 k− − a4 q) · + )− µ + ((−a5 q + a6 k+ ) · − )+ µ
= (a1 k+ + a2 k− )µ (+ · − )
+(a3 − a4 )(q · + )− µ + +(a6 − a5 )(q · − )+ µ , (13.35)
where in the last line we have used q = k+ + k− and (k± · ± ) = 0. Since even
for off-shell photons the current must be strictly conserved we require that
1
Aµ qµ = q 2 (a1 + a2 )(+ · − ) + (a3 − a4 − a5 + a6 )(q · + )(q · − ) = 0, (13.36)
2
which leads to the following relations between the six constants :
a1 + a2 = 0 , a3 − a 4 = a 5 − a 6 . (13.37)
In the second place, we return to the process DU → γW + discussed in
the previous section. The third Feynman diagram now reads differently :
1
M3 = +i~gW QW v(q1 ) (1 + γ 5 ) γα u(q2 ) Zα ,


2(k1 · k2 )
Z α = δ α β − P α Pβ /mW 2

((a1 k2 − a2 P ) · )+ β + ((−a3 P + a4 k1 ) · + )β



+(a5 k1 + a6 k2 )β (+ · )

= δ α β − P α Pβ /mW 2

(a1 − a2 )(k2 · )+ β + (a4 − a3 )(k1 · + )β



+(a5 k1 + a6 k2 )β (+ · ) .

(13.38)
The replacement  → k1 now leads, after some simple algebra (and use of
momentum conservation !) to the form
Z α c→k1 = (a1 − a2 )(k1 · k2 )+ α + T α ,
T α = (−a3 + a4 + a5 − a6 )k1 α
(k1 · k2 )
− (a1 − a2 − a3 + a4 + a5 + a6 )P α . (13.39)
mW 2
August 31, 2019 383

Now a complete cancellation of all diagrams in this case is only possible if

only the first term in Z α c survives. Using the assignment9 QW = QD − QU ,
we then come to the following additional relations between the a’s :
a1 − a2 = 2 , a1 − a2 = a3 − a4 − a5 − a6 = 0 . (13.40)
A third result is obtained by considering the process U D → γW − . Be-
cause of the symmetry between this amplitude and the previous one, we can
establish that also
a1 − a2 = a5 − a6 + a3 + a4 . (13.41)
For the last necessary piece of information we must turn to the handlebar
operation for the produced W rather than the photon. We can rewrite the
three Feynman diagrams as
~gW QD 5


M1 = −i v(q 1 )/
 (/
q 2 − k
/ 2 + m D ) (1 + γ ) /+ u(q2 ) ,
(q2 − k2 )2 − mD 2
~gW QU 5


M2 = −i v(q 1 ) (1 + γ ) /+ (/k2 − q/1 + mU ) /u(q2 ) ,
(k2 − q1 )2 − mU 2
~gW QW 5
γα u(q2 ) Z α ,


M3 = +i 2
v(q 1 ) (1 + γ ) (13.42)
s − mW
with Z α as in Eq.(13.38). The handlebar operation on + now gives the
slightly more complicated result
v(q1 )/ (/q2 − k/2 + mD ) (1 + γ 5 ) /+ u(q2 ) + →k2 =



= − (q2 − k2 )2 − mD 2 v(q1 ) (1 + γ 5 ) /u(q2 )

− v(q1 ) mU (1 + γ 5 ) − mD 1 − γ 5 /k/2 u(q2 )

+ mU 2 − mD 2 v(q1 ) (1 + γ 5 ) /u(q2 ) .



(13.43)
Of these three lines, the second is suppressed with respect to the first one by
a factor (mass/energy), and the third line even by (mass/energy)2 . In the
high-energy limit, therefore, the second and third line will not contribute to
any unwanted high-energy behaviour of the amplitude : we shall call such
terms safe terms10 . We can therefore write
M1 c+ →k2 = +i~gW QD v(q1 ) (1 + γ 5 ) /u(q2 ) + · · · ,


(13.44)
9
Any common factor in the a’s is always absorbed in the value of QW so this is no loss
of generality.
10
Which is not to say that they are negligible ! The point here is that they do not
contribute to any condition on the coupling constants. At the end of a cross section
calculation it is the safe terms that we want !
384 August 31, 2019

where the ellipsis denotes safe terms. For the second diagram, we find in a
similar way :
M2 c+ →k2 = −i~gW QU v(q1 ) (1 + γ 5 ) /u(q2 ) + · · · ,


(13.45)
For the third graph we find, after some algebra,
Z α c+ →k2 = (a4 − a3 )(k1 · k2 )α − a3 mW 2 α
(k2 · )(k1 · k2 )
− (a1 − a2 − a3 + a4 + a5 + a6 )P α
mW 2
+ (k2 · )(−a1 + a2 + a5 − a6 )k1 α . (13.46)
Requiring M3 to cancel against M1 + M2 up to safe terms therefore leads
to yet more relations between the a’s :
a3 − a4 = 2 , a1 − a 2 = a 5 − a 6 . (13.47)
Combining the requirements (13.37), (13.40), (13.41) and (13.47) we find the
unique solution
a1 = a3 = a5 = 1 , a2 = a4 = a6 = −1 . (13.48)
We find that the spacetime part of the vertex has, after all, exactly the same
form as that of the three-gluon vertex, Eq.(12.33) ! We have thus established
the W W γ vertex to be

µ
+
W (p1 ) γ (p3 ) i
↔ QW Y (p1 , µ; p2 , ν; p3 , ρ)
_ ρ ~
W (p2 )
ν
All particles and momenta counted outgoing
EW Feynman rules, part 13.2 (13.49)

13.3 The Z particle

13.3.1 W pair production
Unitarization from extra fermions
In the previous section we have investigated how the possible coupling be-
tween W ’s and photons are restricted by the requirements of the handlebar.
August 31, 2019 385

We shall now pursue the same strategy for different processes. Since we
shall be interested in the high-energy behaviour of amplitudes we shall allow
ourselves to neglect particle masses wherever possible.
Let us consider the process
U (p1 ) U (p2 ) → W + (q+ , + ) W − (q− , − )
With the vertices available so far, we have the following two Feynman dia-
grams
_ _
U _ U +
W W
D
γ _
U W+ U W

which contribute to the amplitude as follows :

~gW 2 5


M1 = −2i v(p 1 ) (1 + γ ) /− (/p2 − q/+ ) /+ u(p2 ) ,
(p2 − q+ )2
~QU QW
M2 = i v(p1 )γµ u(p2 ) Y (q+ , + ; q− , − , −q+ − q− , µ)(13.50)
.
(q+ + q− )2
Here we have neglected the masses as announced. The high-energy behaviour
can be investigated by putting a handlebar on the W + , say ; we then obtain
M1 c+ →q+ = 2i~gW 2 v(p1 )(1 + γ 5 )/− u(p2 ) ,
M2 c+ →q+ = i~QU QW v(p1 )/− u(p2 ) , (13.51)
and we see that these two diagrams cannot possibly cancel one another. We
must therefore introduce an additional ingredient in the model. A possible
approach is the following. In the analogous process U U → γγ the handlebar
requirement is satisfied because there are two diagrams, with the photons
interchanged. We might do the same for the W by postulating the existence
of another fermion type U 0 , with charge one unit higher than QU , and the
existence, in addition to the U DW vertex, of a U 0 U W vertex with vector
and axial-vector couplings. We then have a third diagram at hand :
_
U
W+
U’
_
U W
386 August 31, 2019

with its own contribution

~
M3 = −i v(p1 )ω/+ (/q+ − p/1 ) ω/− u(p2 ) ,
(p1 − q+ )2
ω = g1 + g2 γ 5 . (13.52)
The mass of the U 0 is also neglected, and g1,2 are to be determined. We have
M3 c+ →q+ = −i~ v(p1 )ω 2 /− u(p2 ) , (13.53)
so that
3
%
X
2
Mj =0 ⇒ g1 + g2 γ 5 = 2gW 2 (1 + γ 5 ) + QU QW . (13.54)
j=1 + →q+

We see that it is in principle possible to attain good high-energy behaviour

in the process U U → W + W − , at the cost of introducing new fermion types ;
and the same is possible for DD → W + W − . But a very serious conundrum
immediately arises. Having postulated the existence of the U 0 , we of course
0
also have to consider high-energy behaviour in the process U 0 U → W + W − .
It is easy to see that that can only be cured by postulating also a fermion
U ”, of again one unit of charge higher . . . An infinite tower of fermions with
higher and higher charge becomes unavoidable. Not only is this extremely
unattractive11 , but as the charges grow without bound perturbation theory is
bound to break down since it is based on the assumption that the couplings
governing the interactions are not large.

The Z boson to the rescue

Since introducing additional Dirac particles does not seem a viable way to
ensure good high-energy behaviour in U U → W + W − , we shall investigate
the alternative of an additional boson. That is, we shall postulate the exis-
tence of a neutral spin-1 particle, coupling to W + W − pairs and to fermion-
antifermion pairs. This particle, denoted by Z (or Z 0 ) is supposed to cure
the high-energy behaviour in both U U → W + W − and DD → W + W − simul-
taneously12 . For the W W Z vertex it stands to reason to employ the useful
11
Even leaving aside the fact that no higher-charge fermions have been found to date.
12
This is the simplest scenario. Other possibilities could be explored, in which there is
more than one type of Z, perhaps one type for the U fermions and one type for the D
fermions. Experiment, however, has taught us that the simplest option appears, as usual,
to be the one chosen by nature.
August 31, 2019 387

Yang-Mills form Y (· · ·), with a coupling constant to be determined. Since the

diagram with the Z must cancel against a combination of the purely vecto-
rial photon diagram and the D-exchange diagram with its (1 + γ 5 ) structure,
the Z must couple to the fermions with a nontrivial mixture of vector and
axial-vector terms. We therefore arrive at the following putative Feynman
rules :
µ
+
W (p1 ) Z(p3 ) i
_ ρ → ~ gWWZ Y (p1 , µ; p2 , ν; p3 , ρ) ,
W (p2 )
ν
U i
vU + aU γ 5 γ µ ,


Z →
µ ~
U
D i
vD + aD γ 5 γ µ ,


Z →
µ ~
D

where as before in the Yang-Mills vertex every participant is counted in the

outgoing manner. With these vertices a new Feynman diagram is available
in the process U U → W + W − :
_
U W+
Z
_
U W ,

which evaluates to
~ gWWZ 5


M3 = i v(p 1 ) vU + a U γ γµ u(p2 )
(q+ + q− )2 − mZ 2
Y (q+ , + ; q− , − ; −q+ − q− , µ) . (13.55)
Note that nothing has been neglected in this expression ; the second term in
the massive-boson propagator drops out when we multiply it into the Yang-
Mills vertex. Since this diagram is so similar to M2 it is easy to perform the
handlebar operation :
M3 c+ →q+ ≈ i~ gWWZ v(p1 ) vU + aU γ 5 /− u(p2 ) ,


(13.56)
388 August 31, 2019

where we have assumed that s = (q+ + q− )2 is also much larger than mZ 2 ,

and neglected safe terms. We now see that the high-energy behaviour is
acceptable provided that the non-safe terms cancel under the relations

0 = vU gWWZ + 2gW 2 + QU QW ,
0 = aU gWWZ + 2gW 2 . (13.57)

We can perform precisely the same procedure for the process DD → W + W −

and obtain

0 = vD gWWZ − 2gW 2 + QD QW ,
0 = aD gWWZ − 2gW 2 . (13.58)

A final piece of information is obtained if we realise that, the Z being a

massive spin-1 particle, it must obey its own handlebar relations ; we can
therefore investigate the process U D → W + Z, which gives a single extra
condition
0 = vD + aD − vU − aU − gWWZ . (13.59)

13.3.2 The weak mixing angle for couplings

We can handle (if not completely solve) the system of constraints as follows.
Let us subtract Eqs.(13.57) from Eqs.(13.58). We then obtain

(vD + aD − vU − aU )gWWZ + (QD − QU )QW = 8gW 2 . (13.60)

Using Eq.(13.59) and the definition of QW , we find a relation between three

couplings :
gWWZ 2 + QW 2 = 8gW 2 . (13.61)
There must, therefore, exist an angle θW such that
√ √
QW = 8 gW sin θW , gWWZ = 8 gW cos θW . (13.62)

In the following we shall use the notation sW = sin θW and cW = cos θW . This
angle is called the weak mixing angle13 , and it parametrizes essentially all of
the minimal model of electroweak interactions we are constructing here. In
13
The subscript W was originally used to refer to S. Weinberg, one of the early propo-
nents of the electroweak model, but nowadays it appears more fair to take it to mean just
‘weak’.
August 31, 2019 389

the first place, we know that the charge of the W must be equal to the charge
of the electron (since neutrinos are neutral) and therefore we might prefer to
write
QW
gW = √ (13.63)
8 sW
which leads to a parametrization of the W mass itself14 :
πα 1
(~ c mW )2 = √ 2
GeV2 , (13.64)
2 1.16 10 sW
−5

or
37.3
~ c mW = GeV . (13.65)
sW
As we see, the assumption of the existence of a single, neutral Z boson
immediately implies that the W has a mass of at least 37.3 GeV. Notice that
no prediction for the mass of the Z is obtained, however.
The other unknowns in our treatment can now be expressed in terms of
θW . Adopting the usual convention of denoting by e the positive unit charge,
we find by straightforward algebra
cW
QW = −e , gWWZ = −e ,
sW
e
aU = −aD = ,
4sW cW
 
QU
vU = aU 1 − 4sW 2 ,
e
 
2 QD
vD = aD 1 + 4sW . (13.66)
e

We note here that θW is defined at this stage as a relation between coupling E105
constants ; later on we shall encounter it in another guise !

13.3.3 W, Z and γ four-point interactions

The 2 → 2 processes involving either four fermions or two fermions and two
bosons have led us to postulate W and Z particles and their interactions with
fermions, as well as their mutual three-point vertices. Since we have pretty
14
To arrive at this expression we have used the definition (13.5) for GF , and the result
(10.35) of α.
390 August 31, 2019

much quarried all possible information15 about this sector, we now turn to
the 2 → 2 processes involving four bosons. First we consider the process
W + (p1 , 1 ) γ(p2 , 2 ) → W + (p3 , 3 ) γ(p4 , 4 )
With the available vertices we have two Feynman diagrams for this process :
+
W+
γ W+ W
M = + γ (13.67)
γ W+ γ

with the respective contributions

~QW 2
M1 = i Y (p3 , 3 ; p2 − p3 , ν; −p2 , 2 )
(p2 − p3 )2 − mW 2
× g µν + (p1 − p4 )µ (p2 − p3 )ν /mW 2

× Y (p1 − p4 , µ; −p1 , 1 ; p4 , 4 )
~QW 2
= −i − mW 2 (2 · 3 )(1 · 4 )
2(p2 · p3 )


+Y (p3 , 3 ; p2 − p3 , µ; −p2 , 2 )Y (p1 − p4 , µ; −p1 , 1 ; p4 , 4 ) ,
~QW 2
M2 = i Y (p3 , 3 ; −p3 − p4 , ν; p4 , 4 )
(p3 + p4 )2 − mW 2
× g µν + (p1 + p2 )µ (−p3 − p4 )ν /mW 2

× Y (p1 + p2 , µ; −p1 , 1 ; −p2 , 2 )

~QW 2
= i − mW 2 (3 · 4 )(1 · 2 )
2(p3 · p4 )


+Y (p3 , 3 ; −p3 − p4 , µ; p4 , 4 )Y (p1 + p2 , µ; −p1 , 1 ; −p2 , 2 ) ,
(13.68)
where we have already used Eq.(??) in the internal W lines, as well as the
fact that (pj · j ) = 0, j = 1, 2, 3, 4. Let us now proceed to check current
conservation for the outgoing photon. The following algebra applies to M1 :
Y (p3 , 3 ; p2 − p3 , µ; −p2 , 2 )Y (p1 − p4 , µ; −p1 , 1 ; p4 , 4 )c4 →p4 =
= Y (p3 , 3 ; p2 − p3 , µ; −p2 , 2 ) (p4 · 1 )(p2 − p3 )µ + 2(p2 · p3 )1 µ

= 2(p2 · p3 )Y (p3 , 3 ; p2 − p3 , 1 ; −p2 , 2 )

+ mW 2 (p4 · 1 )(2 · 3 ) , (13.69)
15
As long as the fermion masses are neglected, see later.
August 31, 2019 391

so that

M1 c4 →p4 = −i~QW 2 Y (p3 , 3 ; p2 − p3 , 1 ; −p2 , 2 ) (13.70)

In the same manner we arrive at

M2 c4 →p4 = i~QW 2 Y (p1 + p2 , 3 ; −p1 , 1 ; −p2 , 2 ) (13.71)

Adding these last two results we obtain

2
%
X
Mj =
j=1 4 →p4
2
= i~QW (2(1 · 3 )(2 · p4 ) − (1 · 2 )(3 · p4 ) − (2 · 3 )(1 · p4 )) .
(13.72)

We might also have chosen choose to put the handlebar on 2 instead ; the
result would then have been
2
%
X
Mj =
j=1 2 →p2
2
= i~QW (2(1 · 3 )(p2 · 4 ) − (1 · p2 )(3 · 4 ) − (p2 · 3 )(1 · 4 )) .
(13.73)

Going to the limit of large energies, we can also envisage putting a handlebar
on 1 or 3 . Neglecting safe terms leads to
2
%
X
Mj =
j=1 1 →p1
2
= i~QW (2(p1 · 3 )(2 · 4 ) − (p1 · 2 )(3 · 4 ) − (2 · 3 )(p1 · 4 )) ,
(13.74)

and
2
%
X
Mj =
j=1 3 →p3
2
= i~QW (2(1 · p3 )(2 · 4 ) − (1 · 2 )(p3 · 4 ) − (2 · p3 )(1 · 4 )) .
(13.75)
392 August 31, 2019

We can repair the high-energy behaviour of the amplitude, for all these cases
at once, by introducing a four-boson vertex :
µ ν _
W+ W i
↔ − QW 2 X µναβ (13.76)
γ ~
γ
α β
where
X µναβ = 2 g µν g αβ − g µα g νβ − g µβ g να . (13.77)
The occurrence of such a four-point vertex should not surprise us, with our
experience of a similar vertex in sQED and QCD. Its precise algebraic struc-
ture can, of course, not be inferred from that example16 .

From the similarity between the W W γ and W W Z vertices we can also

immediately conclude that the analogous processes W Z → W γ and W Z →
W Z will necessitate the existence of the following four-point vertices :
µ ν _
W+ W i cW µναβ
↔ − QW 2 X
Z ~ sW
γ
α β
µ ν _
W+ W i cW 2
↔ − QW 2 2 X µναβ . (13.78)
Z ~ sW
Z
α β
fFinally, we consider the process

W + (p1 , 1 ); W − (p2 , 2 ) → W + (p3 , 3 ) W − (p4 , 4 ) ,

for which we have, so far, the four diagrams

W+ W+ W+ Z,γ W+
M = Ζ,γ + . (13.79)
_ _ _ _
W W W W
16
Except, perhaps, the idea that it contains only the metric tensor, and not any of the
momenta : but momenta would only worsen the high-energy behaviour
August 31, 2019 393

It will turn out to be useful to take the γ and Z exchanges together so that
we have two contributions :
M1 = i~QW 2 Y (p3 , 3 , −p1 , 1 , p1 − p3 , µ)
g µν cW 2 g µν − (p1 − p3 )µ (p1 − p3 )ν /mZ 2
 
+
(p1 − p3 )2 sW 2 (p1 − p3 )2 − mZ 2
Y (−p2 , 2 , p4 , 4 , p2 − p4 , ν) ,
M2 = i~QW 2 Y (−p2 , 2 , −p1 , 1 , p1 + p2 , µ)
g µν cW 2 g µν − (p1 + p2 )µ (p1 + p2 )ν /mZ 2
 
+
(p1 − p3 )2 sW 2 (p1 + p2 )2 − mZ 2
Y (p3 , 3 , p4 , 4 , p2 − p4 , ν) . (13.80)
Because the masses of the external particles are all equal, the second term in
the Z propagator can be seen to drop out exactly. We can therefore afford
to take the limit s  mZ 2 without more ado, and combine the γ and Z
propagators to arrive at the following high-energy form of the contributions :
~QW 2 1
M1 = i Y (p3 , 3 , −p1 , 1 , p1 − p3 , µ)
sW (p1 − p3 )2
2

Y (−p2 , 2 , p4 , 4 , p2 − p4 , µ) ,
~QW 2 1
M2 = i Y (−p2 , 2 , −p1 , 1 , p1 + p2 , µ)
2
sW (p1 + p2 )2
Y (p3 , 3 , p4 , 4 , p2 − p4 , µ) . (13.81)
Let us now take the outgoing W − longitudinal, i.e apply the handlebar on
4 , and drop safe terms :
Y (p3 , 3 , −p1 , 1 , p1 − p3 , µ)Y (−p2 , 2 , p4 , 4 , p2 − p4 , µ)c4 →p4
= Y (p3 , 3 , −p1 , 1 , p1 − p3 , µ)
× (p1 − p3 )µ ((p1 − p3 ) · 2 ) − ((p1 − p3 )2 − mW 2 )2 µ

≈ −(p1 − p3 )2 Y (p3 , 3 , −p1 , 1 , p1 − p3 , 2 ) (13.82)

so that
~QW 2
M1 c4 →p4 = −i Y (p3 , 3 , −p1 , 1 , p1 − p3 , 2 ) ; (13.83)
sW 2
and the exactly analogous treatment gives
~QW 2
M2 c4 →p4 = −i Y (−p2 , 2 , −p1 , 1 , p1 + p2, 3 ) . (13.84)
sW 2
394 August 31, 2019

The total result of the handlebar operation is given by

M1 + M2 c4 →p4 =
~QW 2
−i (2(p4 · 1 )(2 · 3 ) − (p4 · 2 )(1 · 3 ) − (p4 · 3 )(1 · 2 )) :
sW 2
(13.85)

we arrive at precisely the same algebraical structure as before, and we can

immediately conclude that, in addition to the W W γγ, W W Zγ and W W ZZ
couplings there must also be a W W W W coupling :
µ ν
W+ W+ i QW 2 µναβ
↔ X (13.86)
_ _ ~ sW 2
W W
α β
Note, however, a slight difference of this vertex as compared to the previous
ones. There, the term that couples the two W Lorentz indices carries the
factor 2 ; here, it is the term that couples the two W + ’s that is ‘special’.

13.4 The Higgs sector

13.4.1 The Higgs hypothesis
Fully longitudinal scattering
Having pursued the consequences of unitarity in processes where a single ex-
ternal spin-1 particle is longitudinally polarised, we must of course also face
the more taxing case in which, perhaps, all external spin-1 particles are lon-
gitudinally polarised : surely this is the most dangerous case from the point
of view of unitarity. In doing so, we must however take into account the fact
that the notion of longitudinal polarisation is not strictly a Lorentz-invariant
one since a generic Lorentz boost will mix longitudinal and transverse de-
grees of freedom. We must therefore specify in which particular Lorentz
frame the particles are assumed to be longitudinally polarised. To this end
we introduce a timelike vector cµ with c · c = 1 ; the frame in which ~c = 0 is
defines the appropriate Lorentz frame. In these notes we shall take cµ to be
proportional to the total momentum involved in the scattering process, that
August 31, 2019 395

is, the external vector particles are assumed to be purely longitudinal in the
centre-of-mass frame of the scattering17 . The longitudinal polarisation of an
on-shell vector particle with momentum pµ and mass m is then given by
m2 µ m2
 
NL
µ
L = µ
p − c , NL −2 = 1 − , (13.87)
m c·p (c · p)2
which expression is well-defined as long as p~ 6= 0. We see that, as before,
L = p/m + O (m/p0 ). In the cases studied so far, the subleading terms in
L have only led to safe terms so that they could be neglected18 ; now, this
is no longer automatically the case.

Before continuing we may note that another possible choice for longitu-
dinal polarisation is to take a lightlike vector tµ and define
1 m µ
µL = pµ − t . (13.88)
m p·t
In many applications this is useful. However, the boson is purely longitudinal
only in a frame where ~t k p~. In scattering processes, not all bosons can then
be strictly longitudinal at the same time. In this section we therefore opt for
cµ rather than tµ

W W → ZZ
The first Gedanken process19 is
W + (p1 , 1 ) W − (p2 , 2 ) → Z 0 (p3 , 3 )Z 0 (p4 , 4 )
So far, we have the following three Feynman graphs available at the tree
level :
1 W Z 3 1 W Z 4
1 W Z 4
M = W + W + , (13.89)
Z Z W
2 W 4 2 W 3 2 Z 3
17
That this is not a trivial point becomes clear when we realise that in ‘W W scattering’
at the LHC, say, the centre-of-mass frame of the scattering does not coincide with the
laboratory frame, in which the detector is at rest, and in which the polarisation analysis
of the produced bosons is presumably performed.
18
From the point of view of restoring unitarity, not that of actually getting the cross
section right!
19
As usual, with improving technology and the commissioning of machines like the LHC,
Gedanken processes are gradually being turned into actual ones.
396 August 31, 2019

and the following contributions :

Nj
Mj = −i~gWWZ 2 , j = 1, 2 ,
∆j
N1 = Y (p1 − p3 , µ; −p1 , 1 ; p3 , 3 )
 
µν 1 µ ν
× −g + (p1 − p3 ) (p1 − p3 )
mW 2
× Y (−p2 , 2 ; p2 − p4 , ν; p4 , 4 ) ,
∆1 = (p1 − p3 )2 − mW 2 = mZ 2 − 2(p1 · p3 ) ,
N2 = Y (p1 − p4 , µ; −p1 , 1 ; p4 , 4 )
 
µν 1 µ ν
× −g + (p1 − p4 ) (p1 − p4 )
mW 2
× Y (−p2 , 2 ; p2 − p3 , ν; p3 , 3 ) ,
∆2 = (p1 − p4 )2 − mW 2 = mZ 2 − 2(p1 · p4 ) ,
M3 = −i~gWWZ 2 N3 ,
N3 = X(1 , 2 , 3 , 4 ) . (13.90)

Owing to the work we have done so far, we may already anticipate some
cancellations between the diagrams when we make all bosons longitudinal
and the safe terms are therefore not the subleading ones, but rather the sub-
subleading ones. We have to proceed carefully20 . Denoting by the subscript
L the ‘fully longitudinal’ case, it appears best to write the result as
3
%
X N123
Mj = −i~gWWZ 2 ,
j=1
∆ 12
L
mZ 2
N123 = N1 ∆2 + N2 ∆1 + ∆12 N3 = −4E 6 (sin θ)2 + · · · ,
mW 4
∆12 = ∆1 ∆2 = 4E 4 (sin θ)2 + · · · , (13.91)

where E = p1 0 = p2 0 = p3 0 = p4 0 and θ = ∠(~p1 , p~3 ), all evaluated in

the centre-of-mass frame. As before, the ellipses denote contributions that
can only give rise to safe terms, and that therefore do not interest us here.
Note that we have disregarded also the normalization factors NL ; since the
polarisation vectors are overall factors in the scattering amplitude, the NL
20
This is most safely done using computer algebra, using e.g. FORM.
August 31, 2019 397

can never play a rôle in any dynamical cancellation, and their subleading
terms are therefore always safe. The non-safe contribution from our three
Feynman graphs is therefore
3
%
X mZ 2
Mj = i~ gWWZ 2 E 2 4
+ ··· , (13.92)
j=1
mW
L

and it violates unitarity at sufficiently large E. Note that each individual

Mj will go as E 4 at high energy so, as already anticipated, some cancellation
has already taken place, but not enough ; and since the vertices have already
been fixed before, we have to introduce a new ingredient into the theory.

The Minimal Higgs approach

We shall assume that, in addition to the three graphs used so far, there is a
fourth one available, mediated by a new particle type. We assume this to be
a neutral, scalar particle, denoted by H, that couples to W + W − and ZZ as
follows :
µ W
H i
↔ gWWH g µν
~
ν W
µ Ζ i
H ↔ gZZH g µν (13.93)
~
ν Ζ
A fourth Feynman diagram is now possible :
1 3
1
M4 = = −i~ gWWH gZZH (1 · 2 ) (3 · 4 ) .
H 4E 2 − mH 2
2 4
(13.94)
Its contribution to the fully longitudinal scattering reads
1
M4 cL = −i~ E 2 gWWH gZZH (13.95)
mW mZ 2
2

and good high-energy behaviour will be restored in the process W W → ZZ

provided that
mZ 4
gWWH gZZH = gWWZ 2 . (13.96)
mW 2
398 August 31, 2019

Before we proceed to the next Gedanken process, a few remarks are in order.
In the first place, the choice for a scalar Higgs particle is almost unavoidable.
It certainly cannot be a fermion ; if it were a vector particle, its propagator
would contain unwanted higher powers of the energy E, the W W H and
ZZH would presumably be of Yang-Mills type hence also E-dependent. The
vertices given above are essentially the only ones possible for the interactions
between two vectors and a scalar if we want them to be energy-independent.
Note that gWWH and gZZH may √ both be expected to contain a mass, that is,
they are of dimension L−1 / ~. The assumption that there is just one type
of neutral scalar involved is, of course, based on nothing but a prejudice in
favour of simplicity. Finally, at high energy all contributions from mH end up
in safe terms, and we do not expect to glean any information on the Higgs
mass from our considerations.

W W → W W scattering
Another four-boson scattering process of interest is
W + (p1 , 1 )W + (p2 , 2 ) → W + (p3 , 3 )W + (p4 , 4 )
for which we have five purely vector-boson diagrams :
1 1
3 4 1 3
M = γ,Ζ + γ,Ζ + (13.97)
4 2 3 4
2 2
whose contributions can be conviently written as
M1 = −i~ Y (p3 , 3 ; −p1 , 1 ; p1 − p3 , µ)
µν
+ (p1 − p3 )µ (p1 − p3 )ν /mW 2
 
2 −gµν 2 −g
× QW + gWWZ
(p1 − p3 )2 (p1 − p3 )2 − mZ 2
× Y (p4 , 4 ; −p2 , 2 ; p2 − p4 , ν) ,
M2 = M1 cp3 ,3 ↔ p4 ,4 ,
QW 2
M3 = i~ X(3 , 4 , 1 , 2 ) . (13.98)
sW 2
By the same methods as used in the previous section we arrive at
3
%
X ~ E 2 QW 2
−4mW 2 + 3mZ 2 cW 2 + · · ·


Mj =i 4 2
(13.99)
j=1
mW sW
L
August 31, 2019 399

The Higgs hypothesis now provides for two additional diagrams :

1 1
3 4
M4,5 = H , H (13.100)
4 3
2 2

with the contributions

(1 · 3 )(2 · 4 )
M4 = −i~ gWWH 2 ,
(p1 − p3 )2 − mH 2
(1 · 4 )(2 · 3 )
M5 = −i~ gWWH 2 , (13.101)
(p1 − p4 )2 − mH 2

so that %
5
X gWWH 2
Mj = i~ E 2 + ··· (13.102)
j=4
mW 4
L

In this process, then, good high-energy behaviour is obtained under the con-
dition
QW 2
gWWH 2 = 4mW 2 − 3mZ 2 cW 2 .


2
(13.103)
sW
Again, no restrictions on mH occur.

HZ → W W scattering
We have now run out of four-vector Gedanken processes. ZZ → ZZ scat-
tering has no Yang-Mills contributions21 , and any four-vector process in-
volving photons will have vanishing amplitudes under a handlebar on any
photon. However, in the same spirit by which we boldly proposed the pro-
cess U D → W Z as soon as the Z was hypothesised, we can consider the
process
H(p1 )Z 0 (p2 , 2 ) → W + (p3 , 3 )W − (p4 , 4 )
Since only three out of four particles can become longitudinal here, the uni-
tarity violations are not so bad, and the safe terms are of sub- rather than
21
Under the Higgs hypothesis ZZ → ZZ scattering is described by three diagrams each
containing Higgs exchange : no Higgs means no scattering at all ! The amplitude is safe
by itself and hence does not lead to any constraints.
400 August 31, 2019

of sub-sub-leading type. We have three diagrams,

1 3 3 3
1
M = + + (13.104)
2 2 2
4 1 4 4

that contribute as
M1 = −i~ gWWZ gWWH Y (p3 , 3 ; p2 − p3 , µ; −p2 , 2 )
−g µν + (p2 − p3 )µ (p2 − p3 )ν
× (4 )ν ,
(p2 − p3 )2 − mW 2
M2 = −i~ gWWZ gWWH Y (p2 − p4 , µ; p4 , 4 ; −p2 , 2 )
−g µν + (p2 − p4 )µ (p2 − p4 )ν
× (3 )ν ,
(p2 − p4 )2 − mW 2
M3 = −i~ gWWZ gZZH Y (p3 , 3 ; p4 , 4 ; −p3 − p4 , µ)
−g µν + (p1 + p2 )µ (p1 + p2 )ν
× (2 )ν . (13.105)
(p1 + p2 )2 − mZ 2
The kinematics of this process is a little different from that of the two previous
ones, since mH and mZ cannot be assumed to be equal. Still, at high energy
we may apply massless kinematics since we only have to cancel the leading
non-safe terms. Neglecting, therefore, mW , mZ and mH in the kinematics22
we find
3
%  
X
2 mZ 1
Mj = i~ E cos θ gWWZ gWWH − gZZH +· · · (13.106)
j=1
mW 4 mZ mW 2
L

and find the final requirement

mZ 1
gWWH 4
= gZZH (13.107)
mW mZ mW 2
if good high-energy behaviour is to emerge.

13.4.2 Predictions from the Higgs hypothesis

The Higgs hypothesis has given us the three conditions of Eqs.(13.96), (13.103)
and (13.107). If we consider gWWH and gZZH as the two unknowns, this sys-
tem is overconstrained, and we obtain additional information. The system
22
But not, of course, in the longitudinal polarisations !
August 31, 2019 401

of conditions can easily be solved and we find the two couplings

Q W mW QW mZ
gWWH = , gZZH = , (13.108)
sW s W cW

and, in addition, the interesting relation

mW = mZ cW . (13.109)

It is apposite to dwell on this last result. The weak mixing angle θW was E106
introduced to parametrize the system of coupling constants, as discussed in
section 13.3.2 : we now see it come back here as a relation between masses
instead ! From the treatment of the Electroweak Standard Model presented
in these notes, it also becomes clear that the mixing angle as a description
of coupling constants is, in a logical sense, prior to that as a description of
masses. The assumption of a single Z 0 particle determines the couplings as
described in section 13.3.2 : but it takes the supposition of a single, neutral
Higgs particle to obtain Eq.(13.109). If the Higgs sector of the Standard
Model turns out to be different, with more Higgs-like particles, say, the W
and Z mass become uncorrelated ; but the couplings of W and Z with the
fermions and each other remain unaffected. In the usual textbook deriva-
tion of the model this distinction tends to be obscured by the simultaneous
obtention of all couplings at once after symmetry breaking.
As a final comment we remark that, if unitarity is restored by any Higgs-
like phenomenon whatsoever, the weak mixing angle must always obey the
bound
4 mW 2
cW 2 < (13.110)
3 mZ 2
as can be seen from Eq.(13.103)23 .

13.4.3 W, Z and H four-point interactions

The class of bosonic four-particle scattering amplitudes is not yet completely
exhausted. We can consider the process

Z 0 (p1 , 1 )Z 0 (p2 , 2 ) → H(p3 )H(p4 )

23
For the actually observed values of W and Z mass this bound is itself somewhat larger
than unity, and therefore not so significant; but it is nice to have it even so.
402 August 31, 2019

given by two diagrams so far,

1 3 1 4
M = + (13.111)
2 4 2 3

and the following amplitude :

M1+2 = −i~ gZZH 2 (1 )µ (2 )ν
 µν
−g + (p1 − p3 )µ (p1 − p3 )ν /mZ 2
(p1 − p3 )2 − mZ 2
−g µν + (p1 − p4 )µ (p1 − p4 )ν /mZ 2

+ . (13.112)
(p1 − p4 )2 − mZ 2
In the fully longitudinal case the non-safe terms are
gZZH 2
M1+2 cL = −i~ E 2 + ··· (13.113)
mZ 4
and the remedy ought to be straightforward by now. We introduce yet an-
other vertex, involving two Z’s and two H’s :
µ Z H
i
↔ gZZHH g µν (13.114)
H ~
ν Z
upon which we have a third diagram, whose nonsafe part is trivial :
gZZHH
M3 cL = 2i~ E 2 + ··· (13.115)
mZ 2
We see that the four-point coupling constant must be given by
gZZH 2 QW 2
gZZHH = = . (13.116)
2mZ 2 2sW 2 cW 2
As in the case of sQED and YM, this four-point coupling does not contain a
length scale, in contrast to the ZZH coupling. For the case of W W → HH
scattering, exactly the same treatment holds. It suffices to replace mZ by mW
and gZZH by gWWH . We find that also a W W HH vertex is required :
µ W H
i
↔ gWWHH g µν , (13.117)
H ~
ν W
August 31, 2019 403

with
gWWH 2 QW 2
gWWHH = = . (13.118)
2mW 2 2sW 2

13.4.4 Higgs-fermion couplings

Let us return to the process
U (p1 )U (p2 ) → W + (q+ , + )W − (q− , − )
which was used in section 13.3.1 to argue the existence of the Z boson. This
time, however, we shall not neglect the fermion masses ; and we shall take
both W ’s longitudinal. It can be seen that each individual diagram will go
as E 2 when the energy E of the W ’s in their centre-of-mass frame becomes
large. This means that, in the longitudinal polarisation of Eq.(13.87), the
second term will only contribute to the safe terms, and we may simply write
(± )L = q± /mW , so that

s
Y (q+ , + ; q− , − ; −q+ − q− , µ) ≈ − 2
(q+ − q− )µ + · · · (13.119)
L 2mW
where once more the ellipsis denotes safe terms. In fact, the restriction to
nonsafe terms in our treatment means that we may neglect the boson masses
in the kinematics : every occurrence of boson masses from the kinematics is
quadratic and hence gives safe terms. For the fermions this is not the case
as we shall see.
Let us revisit the diagrams of our process. The first one now reads
q/− − p/1 + mD
M1 = −i~gW 2 v(p1 )(1 + γ 5 )/− 2 2
(1 + γ 5 )/+ u(p2 ) . (13.120)
(q− − p1 ) − mD
Note that the mD in the numerator drops out by virtue of the (1 + γ 5 )’s. We
can now perform some Diracology, using the Dirac equation and dropping
safe contributions wherever opportune :
−i~gW 2
M1 c L = v(p1 ) A u(p2 ) ,
mW 2 ((q− − p1 )2 − mD 2 )
A = (1 + γ 5 ) q/− (/q− − p/1 ) (1 + γ 5 ) q/+
→ 2(1 + γ 5 ) q/− (/q− − p/1 ) (/q+ − p/2 + mU )
= 2(1 + γ 5 ) q/− (/q− − p/1 ) (/p1 − q/− + mU )
2(1 + γ 5 ) −(q− − p1 )2 q/− + (/q− − p/1 − mU )(/q− − p/1 )mU


→
→ 2(1 + γ 5 ) (mU − q/− ) (q− − p1 )2 ; (13.121)
404 August 31, 2019

so that the fully longitudinal case gives for this diagram

M1 cL = 2i~gW 2 v(p1 )(1 + γ 5 )(/q− − mU )u(p2 ) + · · · (13.122)
For the third diagram we can perform a similar analysis :
~ gWWZ −s
M3 c L = i v(p1 ) B u(p2 ) ,
s − mZ 2 2mW 2
B = (vU + aU γ 5 )(/q+ − q/− )
→ (vU + aU γ 5 )(/q+ − q/− − p/2 + mU − p/1 ) − mU (vU − aU γ 5 )
= −2(vU + aU γ 5 )/q− + 2mU aU γ 5 ; (13.123)
and up to safe terms, we therefore have
~gWWZ 5 5


M3 c L = i v(p 1 ) (vU + a U γ )/
q − − m U a U γ u(p2 ) + · · · (13.124)
mW 2
To obtain the contribution from the second diagram, we simply put gWWZ →
QW , vU → QU , and aU → 0 in the third diagram :
~QW QU
M2 c L = i v(p1 ) q/− u(p2 ) + · · · (13.125)
mW 2
If we add the three diagrams, the contributions with v/q− u cancel precisely,
as they should since that was what we imposed in section 13.3.1. We are left
with terms proportional to mU :
~mU 2 5 5


M1+2+3 cL = i v(p 1 ) −2g W (1 + γ ) − g WWZ a U γ u(p2 ) + · · ·
mW 2
~ QW 2 mU
= −i v(p1 )u(p2 ) + · · · (13.126)
mW 2 4sW 2
so that an energy behaviour of E 1 at high energy is still uncompensated.
The Higgs boson is usefully applied here as well. We may assume the U U H
vertex
U H i
↔ gUUH 1 , (13.127)
U ~
where we must realise that the Dirac unit matrix is involved24 . For the
process U U → W W we then have a fourth available diagram :

M4 = (13.128)

24
In fact, the observation that the nonsafe part in this process is proportional to vu is
the strongest argument in favour of a scalar Higgs.
August 31, 2019 405

which contributes to the amplitude the amount E107

1
M4 = −i~ gUUH gWWH v(p1 )u(p2 ) (+ .− ) . (13.129)
s − mH 2
In the fully longitudinal case we therefore have
~ gUUH gWWH
M4 c = − i v(p1 )u(p2 ) + · · · , (13.130)
2mW 2
and the following requirement on gUUH is obtained :
QW 2 mU ~ gUUH gWWH
+ =0 , (13.131)
4sW 2 2mW 2
or
QW mU
gUUH = − . (13.132)
2sW mW
This discussion can of course be applied to any fermion type25 , and we find E108
the general Feynman rule

f H i e mf
↔ 1 (13.133)
f ~ 2sW mW

13.4.5 Higgs self-interactions

The triple H coupling
There remains the issue of possible self-interactions of the Higgs particle. To
this end we examine not a 2 → 2 but a 2 → 3 process, namely
Z(p1 , 1 ) Z(p2 , 2 ) → Z(p3 , 3 ) Z(p4 , 4 ) H(p5 ) .
At the tree level, this process is described by 21 Feynman diagrams provided
we allow for three-point couplings between H’s. These belong to one of the
three following types :

M1,2,3 = , , (13.134)

25
Note that for D-type fermions, aD has opposite sign ; but also the W + and W − are
interchanged in the first diagram.
406 August 31, 2019

where as usual the dotted lines denotes Z’s and the solid lines stand for H
particles, and we have to take into account the appropriate permutations of
the external Z particles. The amplitude is given by the three corresponding
contributions :
M1 = A1 (1, 2, 3, 4, 5) + A1 (2, 1, 3, 4, 5) + A1 (3, 4, 1, 2, 5)
+ A1 (4, 3, 1, 2, 5) + A1 (1, 3, 2, 4, 5) + A1 (3, 1, 2, 4, 5)
+ A1 (2, 4, 1, 3, 5) + A1 (4, 2, 1, 3, 5) + A1 (1, 4, 3, 2, 5)
+ A1 (4, 1, 3, 2, 5) + A1 (3, 2, 1, 4, 5) + A1 (2, 3, 1, 4, 5) ,
M2 = A2 (1, 2, 3, 4, 5) + A2 (3, 4, 1, 2, 5) + A2 (1, 3, 2, 4, 5)
+ A2 (2, 4, 1, 3, 5) + A2 (1, 4, 3, 2, 5) + A2 (3, 2, 1, 4, 5) .
M3 = A3 (1, 2, 3, 4, 5) + A3 (1, 3, 2, 4, 5) + A3 (1, 4, 2, 3, 5) ,
A1 (i1 , i2 , i3 , i4 , i5 ) = i~3/2 gZZH 3 i1 µ Πµν (pi1 + pi5 ) i2 ν (i3 · i4 )
× ∆Z (pi1 + pi3 ) ∆H (pi3 + pi4 ) ,
A2 (i1 , i2 , i3 , i4 , i5 ) = −i~3/2 gZZH gZZHH (i1 · i2 )(i3 · i4 )
× ∆H (pi3 + pi4 ) ,
A3 (i1 , i2 , i3 , i4 , i5 ) = i~3/2 gZZH 2 gHHH (i1 · i2 )(i3 · i4 )
× ∆H (pi1 + pi2 ) ∆H (pi3 + pi4 ) ,
1
Πµν (q) = −gµν + qµ qν ,
mZ 2

−1
−1
∆Z (q) = q 2 − mZ 2 , ∆H (q) = q 2 − mH 2 . (13.135)
Here we have, for once, taken all momenta outgoing, which means that the
momenta of the incoming Z’s have negative zeroth component. For this 5-
particle process the phase space is of course more complicated, and here we
demonstrate a numerical method to investigate cancellations. This can be
quite sensitive if done right26 . Although naı̈vely each diagram A1 and A2
grow quadratically with the energy in the fully longitudinal case, both M1
and M2 actualy become energy-independent at sufficiently high energy E.
But this is not safe : a 2 → 3 amplitude must go at most as E −1 , and
therefore cancellations between (M1 + M2 ) and M3 are still necessary. We
find that the required HHH coupling is given by
gHHH 3 QW mH 2
↔ i , gHHH = − (13.136)
~ 2 mW sW
26
An algorithm for sampling the relevant phase space is given in appendix 15.13
August 31, 2019 407

Figure 13.1: Monte Carlo investigation of cancellations in ZZ → ZZH

if the necessary cancellations are to arise. In figure 13.1 we have, somewhat

arbitrarily, chosen mW c2 = 80.4 GeV, mZ c2 = 91.2 GeV, mH c2 = 125 GeV.
The plot shows − M3 cL / M1+2 cL for varying total energy w. The sam-
pling is performed as described in appendix 15.13. The two contributions to
the amplitude are seen to balance one another precisely, and the combined
amplitude goes as w−1 , provided the right choice of gHHH is made. The am-
plitudes themselves are heavily dependent on the various scattering angles:
but their ratio is not. A word of caution is in order on the interpretation
of this picture. The high-energy limit is, strictly speaking, only obtained
if all products of momenta grow large with respect to all masses involved.
In a sampling over phase space it can always happen that some momentum
products are comparable to squared masses ; these cases are responsible for
the outlying dots in the plot at large values of the energy scale.

The quartic H coupling

The last Gedanken process needed is

Z(p1 , 1 ) Z(p2 , 2 ) → H(p3 ) H(p4 ) H(p5 )

408 August 31, 2019

which is described by 25 Feynman diagrams in six types:

M1,...,6 = , , , , , .

(13.137)
where we have already anticipated a quartic Higgs coupling in the last dia-
gram. The contributions to the amplitude are
M1 = B1 (1, 2, 3, 4, 5) + B1 (1, 2, 4, 5, 3) + B1 (1, 2, 5, 3, 4)
+ B1 (1, 2, 5, 4, 3) + B1 (1, 2, 3, 5, 4) + B1 (1, 2, 4, 3, 5) ,
M2 = B2 (1, 2, 3, 4, 5) + B2 (1, 2, 4, 5, 3) + B2 (1, 2, 5, 3, 4)
+ B2 (2, 1, 3, 4, 5) + B2 (2, 1, 4, 5, 3) + B2 (2, 1, 5, 3, 4) ,
M3 = B3 (1, 2, 3, 4, 5) + B3 (1, 2, 4, 5, 3) + B3 (1, 2, 5, 3, 4) ,
M4 = B4 (1, 2, 3, 4, 5) + B4 (1, 2, 4, 5, 3) + B4 (1, 2, 5, 3, 4)
+ B4 (2, 1, 3, 4, 5) + B4 (2, 1, 4, 5, 3) + B4 (2, 1, 5, 3, 4) ,
M5 = B5 (1, 2, 3, 4, 5) + B5 (1, 2, 4, 5, 3) + B5 (1, 2, 5, 3, 4) ,
M6 = −i~3/2 gZZH gHHHH (1 · 2 ) ∆H (p1 + p2 ) ,
B1 (i1 , i2 , i3 , i4 , i5 ) = i~3/2 gZZH 3 1 µ Πµ λ (pi1 + pi3 ) Πλν (pi2 + pi5 ) 2 ν
× ∆Z (pi1 + pi3 ) ∆Z (pi2 + pi5 ) ,
B2 (i1 , i2 , i3 , i4 , i5 ) = i~3/2 gZZH 2 gHHH i1 µ Πµν (pi2 + pi5 ) i2 ν
× ∆Z (pi2 + pi5 ) ∆H (pi3 + pi4 ) ,
B3 (i1 , i2 , i3 , i4 , i5 ) = i~3/2 gZZH gHHH 2 (i1 · i2 ) ∆H (p1 + p2 ) ∆H (pi4 + pi5 ) ,
B4 (i1 , i2 , i3 , i4 , i5 ) = −i~3/2 gZZHH gZZH i1 µ Πµν (pi2 + pi5 ) i2 ν
× ∆Z (pi2 + pi5 ) ,
B5 (i1 , i2 , i3 , i4 , i5 ) = −i~3/2 gZZHH gHHH (i1 · i2 ) ∆H (pi3 + pi4 ) , . (13.138)
A treatment analogous to that of the previous paragraph leads to the follow-
ing, final Feynman rule :
i 3 QW 2 mH 2
↔ gHHHH , gHHHH = − (13.139)
~ 4 mW 2 sW 2
In figure 13.2 we plot the ratio − M6 cL / M1+···+5 cL obtained in the same
manner as in the previous paragraph. Again, the choice of the factor 3/4
in gHHHH is justified by the fact that the ratio goes to 1 with great accuracy
as the total energy w increases. At high energy there are no outliers : M6
depends only on w.
August 31, 2019 409

Figure 13.2: Monte Carlo investigation of cancellations in ZZ → HHH

13.5 About anomalies

In sections 10.2.7 and 12.3.2, we have discussed triangle diagrams in which
three vector particles couple to fermions running around a loop,

V1 V3

V2

which is always accompanied by the same diagram with orientation of the

fermion lines reversed. In three or more spacetime dimensions, such dia-
grams are problematic. A fermion propagator at large momentum pµ goes as
1/p, while the integration element in D dimensions goes as pD , leading to a
divergence if d ≥ 3. In four dimensions, the diagrams are therefore linearly
divergent, but this leads to trouble : there is no unambiguous integration
prescription. If we cavalierly shift integration momenta and use dimensional
regularization, then the triangle diagram will have a logarithmic UV diver-
gence of the usual type, and might be balanced with a counterterm : but E109
this only works if the counterterm has a counterpart in the bare action. This
can mean that we are forced to change the original action, with possibly
disastrous consequences. For instance, if the three vector particles V1,2,3 are
410 August 31, 2019

V1 , V2 , V3 terms total contribution

P 2
Nc QU 2 − QD 2 + Qν 2 − QL 2


γ, γ, Z Qf af
P

γ, Z, Z Q f v f af Nc QU (1 − σQU ) + QD (1 + σQD )
+ Qν (1 − σQν ) + QL (1 + σQL )
2
P
γ, W, W P Qf2gW Nc (QU + QD ) + Qν + QL

Z, Z, Z vf af Nc (1 − σQU )2 − (1 + σQD )2
P 3 + (1 − σQν )2 − (1 + σQL )2
Z, Z, Z P af 2 trivially zero

Z, W, W vf gW Nc (1 − σQU ) − (1 + σQD )
+ (1 − σQν ) − (1 + σQL )
2
P
Z, W, W P af gW2 trivially zero
Z, g, g af gs trivially zero

Table 13.1: Contributions to the triangle anomalies

γ, Z, and Z a γZZ coupling in the bare action implies that the electrically
neutral Z’s are, actually, not neutral at all. In that case, what does electric
charge mean ?
If all three fermion-fermion-vector couplings are of vector type, Furry’s
theorem (cf section 10.2.7) shows how the two diagrams have opposite sign,
all other things being equal. Therefore, in QED, where the three vector par-
ticles are photons, (V1 , V2 , V3 ) = (γ, γ, γ), such diagram pairs will completely
cancel. In QCD, with three gluons, (V1 , V2 , V3 ) = (g, g, g), the two diagrams
form a colour-antisymmetric contribution to the one-loop correction of the
three-gluon vertex. Two mixed cases are (V1 , V2 , V3 ) = (γ, γ, g), where the
diagrams vanish individually because Tr (T j ) = 0 for gluonic colour matrices
T j , and (V1 , V2
, V3 ) = (γ, g, g)

where they again cancel against one another
since Tr T j T k = Tr T k T j . We shall now investigate how things stand
if we include the electroweak sector. In particular we have to worry about
diagrams containing one or three γ 5 ’s, since in such cases the two diagrams
with oppositely-oriented fermion loops have the same sign. Since we shall be
interested only in the possible ultraviolet behaviour of the theory, we may
assume that all fermions have the same (negligible) mass. Also, we consider
the contribution of a single fermion family, that is a charged lepton (with
charge QL ), a neutrino (with charge Qν , which we alllow to be nonzero for
the moment), and an up-type quark (with charge QU ) and a down-type quark
August 31, 2019 411

(with charge QD ). The quarks are supposed to occur in Nc colour types. We

consider the collective effect of all these fermions together in table ??, where
we have a look at all possibly problematic axial-vector contributions. We use
the shorthand σ for 4sW 2 /e, sum over all fermions in a family, and drop all
common overall factors. Inspection tells us that in fact two conditions suffice
to let the sum over fermions vanish :

Nc QU 2 − QD 2 = QL 2 − Qν 2 ,




Nc QU + QD = − QL − Qν . (13.140)

The ratio of these two result in the requirement

QU − QD = Qν − QL (= −QW ) , (13.141)

which is nothing else than the charge conservation we already encountered

in Eq.(13.33). We see that the single condition, necessary to ensure all the
cancellations, is that the quark charges conform to
Qν − QL Qν + QL QL − Qν Qν + QL
QU = − , QD = − , (13.142)
2 2Nc 2 2Nc
In the Standard Model, with Nc = 3, we have indeed
2 1
Qν = 0 , QU = − QL , QD = QL . (13.143)
3 3
This is one of those curious instances where the universe appears to arrange
itself to make the theory as well-behaved as possible. Whether there is a
deeper reason for this I do not pretend to know, but I am grateful nonetheless.

13.6 Conclusions and remarks

We have now derived all vertices of the electroweak Standard Model. That
is to say, the more usual textbook derivations arrive at precisely the set of
Feynman rules that we have also obtained. There are, however, a number of
differences between the treatment given here and the usual one.
• We have not invoked any symmetry principle, but rather the (under-
lying) SU (2) × U (1) symmetry has spontaneously emerged from our
choices for the ‘minimal’ solution, for instance by insisting on only a
single Z particle while we could have opted for more.
412 August 31, 2019

• Since we have not invoked any symmetry, there is also no need to

explain its ‘breaking’ in order to arrive at massive W ’s and Z’s. Instead,
we have simply faced the observed fact of their massiveness and come
to grips with it with the help of a (minimal) Higgs sector.
• There is, as we have already discussed, a logical distinction between
the two uses of the weak mixing angle, in which the ratio of coupling
constants is logically ‘prior’ to the ratio mW /mZ .
• We have not needed to introduce any Higgs doublet, but rather only
a single, physically observable H particle. This approach elegantly
sidesteps the question whether , and if so how the Higgs field configu-
ration is ‘spontaneously broken’. This would indicate that the Higgs
particle is also, in a sense, logically prior to a complete Higgs doublet.

13.7 A look at non-minimal models

13.7.1 Non-minimal Higgs sector
The Minimal Standard Model we have derived here is, of course, by no means
the only possibility. An alternative is presented here. Let us inspect the
process
U (p1 ) D(p2 ) → W + (p3 , + ) Z(p4 , )
for which, at the tree level, we have the following Feynman graphs:
Z W
W
W Z
Z

These diagrams are given by, respectively,

p/4 − p/2 + mD
M1 = −i~gW v(p2 )(vD + aD γ 5 )/ (1 + γ 5 )/+ u(p1 ) ,
(p4 − p2 )2 − mD 2
p/1 − p/4 + mU
M2 = −i~gW v(p2 )(1 + γ 5 )/+ 2 2
(vU + aU γ 5 )/u(p1 ) ,
(p1 − p4 ) − mU
M3 = −i~gW gWWZ v(p2 )(1 + γ 5 )γα u(p1 )
−g αβ − (P )α (−P )β /mW 2
× Y (p3 , + ; −P, β; p4 , ) , (13.144)
s − mW 2
August 31, 2019 413

with P = p1 + p2 = p3 + p4 . Using

v(p2 )(1 + γ 5 )/P u(p1 ) = v(p2 ) p/2 (1 − γ 5 ) + (1 + γ 5 )/p1 u(p1 )

= mU v(p2 )(1 + γ 5 )u(p1 ) − mD v(p2 )(1 − γ 5 )u(p1 ) (13.145)

and
Y (p3 , + , −P, −P ; p4 , ) = (mW 2 − mZ 2 )(+ · ) (13.146)
we can rewrite M3 as

~ gW gWWZ
M3 = i v(p2 )(1 + γ 5 )γα u(p1 ) Y (p3 , + ; −P, α; p4 , )
s − mW 2
mZ 2
  
5 5


+ 1− (+ · ) mU v(p2 )(1 + γ )u(p1 ) − mD v(p2 )(1 − γ )u(p1 ) .
mW 2
(13.147)

In the fully longitudinal case all three diagrams actually grow as E 2 at most,
at high energy E. From this point on, therefore, we take the bosons to be
massless since their masses can only lead to safe terms. The fully longitu-
dinal expressions, in which we systematically ignore safe terms, can then be
evaluated as follows.
~gW
M1 c L = i v(p2 ) Σ u(p1 ) ,
2(p2 · p4 )mW mZ
Σ = (vD + aD γ 5 )/p4 (−/p2 + mD )(1 + γ 5 )/q3
= −2(p2 · p4 )(vD + aD γ 5 )(1 + γ 5 )/p3
+(vD + aD γ 5 )(/p2 + mD )/p4 p/3 (1 − γ 5 )
= −2(p2 · p4 )(vD + aD )(1 + γ 5 )/p3
+2mD aD aD γ 5 p/4 p/3 (1 − γ 5 )
= −2(p2 · p4 )(vD + aD )(1 + γ 5 )/p3
−4(p2 · p4 )mD aD aD (1 − γ 5 ) , (13.148)

where we have used the fact that, up to safe terms, we can write p/4 p/3 =
p/4 (/p3 + p/4 − p/1 ) = p/4 p/2 ; so that for the first diagram we have

~gW
M1 c L = i − (vD + aD ) v(p2 )(1 + γ 5 )/p3 u(p1 )
mW mZ

5
− 2aD mD v(p2 )(1 − γ )u(p1 ) . (13.149)
414 August 31, 2019

In the same way we arrive at


~gW
M2 c L = i (vU + aU ) v(p2 )(1 + γ 5 )/p3 u(p1 )
mW mZ

5
− 2mU aU v(p1 )(1 + γ )u(p1 ) . (13.150)

Finally, we can derive27 that


~gW gWWZ
M3 c L = i v(p2 )(1 + γ 5 )/q3 u(p1 )
mW mZ
mZ 2

5 5


− mU v(p2 )(1 + γ )u(p1 ) − mD v(p2 )(1 − γ )u(p1 ) .
2mW 2
(13.151)

These diagrams each contain a term that goes as E 2 , and these contributions
cancel one another under the condition

vU + aU − vD − aD + gWWZ = 0 (13.152)

which we have already encounteered in Eq.(13.59). We are left with an

amplitude that grow as E 1
~gW
M123 cL = i v(p2 ) A u(p1 ) ,
mW mZ
mZ 2
 
A = −mU gWWZ + 4aU ω+
mW 2
mZ 2
 
+mD gWWZ − 4aD ω− . (13.153)
mW 2
Using the results obtained in section 13.3.2, we have

mZ 2 mZ 2 mZ 2 cW 2
 
e
gWWZ + 4aU = gWWZ − 4aD = 1− , (13.154)
mW 2 mW 2 cW s W mW 2
so that unitarity is respected provided that mW = cW mZ . What if the observed
W and Z masses have a different ratio? In the first place, we must then
27
In the computation of this diagram, it is important to notice that the unsafe terms
going as E 1 come from terms in the brackets, and not just from the second term.
August 31, 2019 415

conclude that there moust be more than just one single type of neutral
Higgs boson, for it was this assumption that gave us the relation 13.109, and
it would now be seen to be incorrect. In the second place, there must be
a Higgs-like object that compensates the unsafe behaviour in U D → W + Z,
and it must be charged: a charged Higgs particle! We would need to introduce
couplings of this object to fermions and vector bosons as follows:

D
H+ i
↔ gUDH (mU ω+ − mD ω− )
~
U
+
ν W
_ i
H ↔ gWZH g µν
µ Z
~

so that a fourth diagram becomes available:

with the contribution

(+ · )
M4 = −i~ gUDH gWZH v(p2 ) (mU ω+ − mD ω− ) u(p1 ) . (13.155)
s − M2
The fully longitudinal form up to safe terms is given by
~gUDH gWZH
v(p2 ) mU (1 + γ 5 ) − mD (1 − γ 5 ) u(p1 ) ,


M4 cL = −i (13.156)
4mZ mW
and we see that unitarity restoration is possible provided that
mZ 2
 
gUDH gWZH
gW gWWZ 2
+ 4aU + =0 . (13.157)
mW 4
The algebraic form of this coupling is, in fact, the same as that found in
supersymmetric versions of the Standard Model, that contain charged Higgses
in addition to at least 3 neutral ones. We see that the form of the coupling is
not dependent on which supersymmetric extension of the Standard Model is
preferred, but is dictated by the simple existence of a charged Higgs particle.
416 August 31, 2019

At this point we shall assume that there is just a single type of charged
Higgs H ± in the model, but we shall allow for more than one type of neutral
Higgs, denoted therefore by Hj0 , where j can run over several values.
Let us now find some more couplings. In the first place, since the H ± is
a charged scalar, we know immediately what the structure of its couplings
to photons must be:

q i
µ ↔ Qc (p − q)µ ,
p ~

where the momenta are counted along the arrows, and of course

µ
i
↔ (2Qc 2 ) g µν
ν ~

As discussed in section 10.4. However, we must also allow for the possibility
of charged bosons coupling to H 0 H ± pairs. Simple analogy then motivates
the choices of two additional vertices:

q i
µ ↔ gWcj (p − q)µ ,
p ~

µ
i
↔ gWcjγ g µν .
~
ν
Here the momenta are counted along the arrow as before. The coupling
constants can be restricted by considering the process

H + (p1 )Hj0 (p2 ) → W + (p3 , 3 ) γ(p4 , 4 ) ,

for which we now have three Feynman diagrams at the tree level:
August 31, 2019 417

The three contributions are given, with P = p1 + p2 = p3 + p4 , by

 
~gWcj QW αβ 1 α β
M1 = −i Y (p3 , 3 ; −P, α; p4 , 4 ) −g + P P (p1 − p2 )β ,
2(p3 · p4 ) mW 2
(p1 · 4 )(p2 · 3 )
M2 = −2i~gWcj Qc ,
(p1 · p4 )
M3 = i~gWcjγ (3 · 4 ) . (13.158)

The handlebar operation on the photon, which tests for current consaerva-
tion, gives
  
M123 c4 →p4 = i~ (p1 · 3 ) QW gWcj + gWcjγ
 
+ (p2 · ) − QW gWcj + gWcjγ − 2gWcj Qc ,(13.159)

from which we derive the relations, valid for any j:

gWcjγ = −gWcj QW , Qc = −QW . (13.160)

Note that we needed a process involving W in order to establish a relation

between Qc and QW ; and, that the four-point vertex does not contain a factor
2 as it does in sQED.
Another process of interest is

H + (p1 )H − (p2 ) → W + (p3 , 3 )W − (p4 , 4 )

in which we encounter two new vertices:

W
µ i
ν ↔ gWWcc g µν
~
W
q Z i
µ ↔ gZcc (p − q)µ
p ~

so that again three diagrams are available:

γ,Ζ
418 August 31, 2019

The contributions are

X (p1 · 3 )(p2 · 4 )
M1 = 4i~gWcj 2 2−m 2
,
j
(p 1 − p 3 ) j

M2 = i~gWWcc (3 · 4 ) ,
 
QW Qc gWWZ gZcc
M3 = i~ + Y (p3 , 3 ; p4 , 4 ; −P, p1 − p2 ) (13.161)
.
s s − mZ 2
Putting a handlebar on 3 results in
  
M123 c3 →p3 = i~ (p1 · 4 ) gWWcc + QW Qc + gWWZ gZcc
" #)
X
+(p2 · 4 ) − 2 gWcj 2 + gWWcc − QW Qc − gWWZ gZcc (.13.162)
j

13.8 Exercises
Excercise 103 The width of the W
Assume that the weak coupling constant gW is universal, i.e. independent of
the fermion’s flavour.
1. Compute (at the tree level) the decay width for the decay

W − → e− ν̄e

2. Assuming that all quarks and leptons are essentially massless compared
to the W mass mW , with of course the exception of the top quark.
Determine the total W decay with ΓW . Hint : quarks have colour !
3. Insert your result into Eq.(13.25), and verify that unitarity is not vio-
lated in this process. For the unitarity limit, see Appendix 15.10.

Excercise 104 The W width with Cabibbo mixing

In the simplest form of the standard model the W couples with universal
coupling to ud and cs quark-antiquark pairs (and, of course to tb but we
disregard that here). In fact, its couplings are more complicated, and we
have the following pattern of couplings :

W ud : gW cos(θc ) , W us : gW sin(θc ) ,
W cd : −gW sin(θc ) , W cs : gW cos(θc ) .
August 31, 2019 419

Here, θc ≈ 13o is the so-called Cabibbo angle. Show that this refinement does
not change the total W decay width.

Excercise 105 The width of the Z

Compute the total Z decay width at the tree level, with the same approxi-
mation for the masses as in exercise 103.

Excercise 106 Landau-Yang strikes again, sneakily!

From the fact that there is a ZW W vertex in the electroweak model, and
Eq.(13.109), you might hope to circumvent a ‘weak’ variation of the Landau-
Yang theorem by letting cW go to zero, so that the W mass vanishes and the
Z could actually decay into two massless spin-1 particles. Show that this is
impossible. Hint : this is simple.

Excercise 107 The width of the H

Before the Higgs was found in 2012, its mass was unknown. Let us assume
that mH is sufficiently large for decays into heavy bosons to be possible.

1. Compute the decay width of the Higgs into a fermion-antifermion pair.

Show that these widths are proportional to mH .

2. Compute the widths Γ(H → W + W − ) and Γ(H → ZZ). Show that

these widths are proportional to mH 3 .

3. Esimate the value of mH for which ΓH ≈ mH . This was considered to

be the largest realistic value of mH .

Excercise 108 Unitarity-violating fermions

Assume that there are some extremely heavy fermions around, called F1 and
F2 , both with mass M , much larger than mH .

1. Compute the total cross section for F1 F̄1 → F2 F̄2 at energies large
compared to M .

2. Compare your result with the unitarity limit and find a restriction on
M.

Excercise 109 Linear or logarithmic ?

Write out a Feynman diagram where three spin-1 particles V1,2,3 are attached
to a fermion loop (possibly with three different fermion types in the three
420 August 31, 2019

propagators). Perform the Feynman trick to combine the three propagator

denominators into one, and shift the integration momentum k µ appropriately
so that this denominator contains no term linear in k. Show that the leading
divergence, with three powers of k in the numerator, always vanishes upon
integration, and that therefore the actual divergence is logarithmic.
Chapter 14

Example computations

Relevant appedix : 15.15

In this chapter we shall go through several actual computations of a
number of tree-level processes and some loop diagrams. The theory lives
in four Minkowski dimensions ; but in the loop calculations we shall use
dimensional regularization throughout.

14.1 Neutrino production in e+e− scattering

14.1.1 The cross section
In this section we consider the process
e− (p1 , λ) e+ (p2 , λ) → νj (q1 ) ν̄j (q2 ) (j = e, µ) ,
where both the e± and the neutrinos are taken to be massless, and we have
indicated the momenta and the electronic helicities. At the tree level there
are either one or two diagrams :
p1 q1 p1 q1
M = Z
+ W (14.1)
p2 q p2 q2
2

For j = e (electron neutrinos) the process is described by the two diagrams.

For j = µ (muon or tau neutrinos) only the first diagram contributes. We
shall neglect the width of the Z boson1 . Using the looser notation for spinors
1
Since the W is exchanged with negative invariant mass and hence cannot decay at all,
its width is naturally zero in this case anyway.

421
422 August 31, 2019

allowed in the massless case we can write the two diagrams as follows, where
we indicate their dependence on the helicity:
i~ aν
M1 (λ) = uλ (p2 )(ve + ae γ 5 )γ α uλ (p1 ) u− (q1 )(1 + γ 5 )γα u− (q2 ) ,
s − mZ 2
i~ gW 2
M2 (λ) = u− (q1 )(1 + γ 5 )γ α uλ (p1 ) uλ (p2 )(1 + γ 5 )γα u− (q2 ) .
t − mW 2
(14.2)

We have introduced the Mandelstam variables

s = (p1 + p2 )2 = (q1 + q2 )2 , t = (p1 − q1 )2 = (p2 − q2 )2 ,

u = (p1 − q2 )2 = (p2 − q1 )2 = − s − t . (14.3)

The first step is to get rid of the explicit γ 5 ’s :

2i~ aν (ve − λae )

M1 (λ) = uλ (p2 )γ α uλ (p1 ) u− (q1 )γα u− (q2 ) ,
s − mZ 2
4i~ gW 2
M2 (λ) = u− (q1 )γ α uλ (p1 ) uλ (p2 )γα u− (q2 ) . (14.4)
t − mW 2
Using the Chisholm identity we can compute the explicit helicity forms :

4i~ aν (ve − ae )
M1 (+) = s+ (p2 , q2 ) s− (p1 , q1 ) ,
s − mZ 2
4i~ aν (ve + ae )
M1 (−) = s− (p2 , q1 ) s+ (q2 , p1 ) ,
s − mZ 2
M2 (+) = 0 ,
8i~ gW 2
M2 (−) = s− (q1 , p2 ) s+ (q2 , p1 ) . (14.5)
t − mW 2
Keeping in mind the antisymmetry of the spinor products, and the Fermi
minus sign, we find that up to an irrelevant overall complex phase the am-
plitudes are given by

aν (ve − ae )
M(+) = 4~ t ,
s − mZ 2
2gW 2
 
aν (ve + ae )
M(−) = 4~ + u . (14.6)
s − mZ 2 t − mW 2
August 31, 2019 423

We arrive at
 2

2
 2 aν (ve − ae )2 2 aν 2 (ve + ae )2 2
|M| = 4~ t + u
(s − mZ 2 )2 (s − mZ 2 )2
4gW 2 aν (ve + ae ) 4gW 4

2 2
+ u + u . (14.7)
(s − mZ 2 )(t − mW 2 ) (t − mW 2 )2
The centre-of-mass frame is the obvious choice to work in ; in this frame
s
t = − (1 − cos θ) , (14.8)
2
where θ is the angle between p~1 and ~q1 , and the cross section has no azimuthal-
angle dependence. We may therefore write the phase space as follows :
d cos θ dφ d cos θ 1
dV (p1 + p2 ; q1 , q2 ) = 2
→ = dt , (14.9)
32π 16π 8πs
with the integration interval being t ∈ [−s, 0]. The various integrals are
easily worked out ; we have
Z0
u2 mW 2
    
s s
=s +2−2 1+ log 1 + ,
(t − mW 2 )2 mW 2 s mW 2
−s
Z0 2 !
u2 3 mW 2 mW 2
 
s
2
= s2 + − 1+ log 1 + ,
t − mW 2 s s mW 2
−s
Z0 Z0
2 1
t dt = u2 dt = s3 . (14.10)
3
−s −s

Putting everything together2 we obtain for the total cross section the expres-
sion
σ(e+ e− → νe ν̄e ) =
( 2
~2

2 2 2


2  s 
aν v e + ae  
4πs 3 s − mZ 2 
2
 2 2
  !
s 3 m m s
+ 4gW 2 aν (ve + ae )
W W
+ − 1+ log 1 +
s − mZ 2 2 s s mW 2
mW 2
    
4 s s
+ 4gW +2−2 1+ log 1 + . (14.11)
mW 2 s mW 2
2
And not forgetting the flux factor !
424 August 31, 2019

For muon (or tau) neutrinos only the first line remains :
2
~2 2 2

+ −

2  s 
σ(e e → νµ ν̄µ ) = aν (ve + ae )   . (14.12)
6πs s − mZ 2 

14.1.2 Unitarity considerations

The above neutrino cross section lends itself to a few interesting observations.

At the Z peak
As it stands, the cross section diverges for s = mZ 2 . This is of course due to
our neglecting ΓZ . To remedy this, we may replace mZ 2 by mZ 2 − imZ ΓZ , and
neglect the second (W -exchange) diagram3 . Close to the Z pole, the cross
section for each neutrino type is then given by
~2 aν 2 (ve 2 + ae 2 ) mZ 2
σ(e+ e− → ν ν̄) = (s ≈ mZ 2 ) ,
6π (s − mZ 2 )2 + mZ 2 ΓZ 2
(14.13)
while at the very peak we have4
~2 aν 2 (ve 2 + ae 2 )
σ(e+ e− → ν ν̄) = (s = mZ 2 ) , (14.14)
6πΓZ 2
This can be cast in an instructing form, using the fact that
~ (ve 2 + ae 2 ) mZ ~ aν 2 mZ
Γ(Z → e+ e− ) = , Γ(Z → ν ν̄) = . (14.15)
12π 6π
The cross section at the peak can therefore be written as
12π Γ(Z → e+ e− )
  
+ − Γ(Z → ν ν̄)
σ(e e → ν ν̄) = (s = mZ 2 ) ,
s ΓZ ΓZ
(14.16)
which is exactly the form demanded by unitarity for an intermediate state
(the Z) with unit spin (see section 15.10).
3
This is really justified ! In the sense in which ΓZ 6= 0 comes about by interactions, ΓZ
is formally of higher order in perturbation theory, and then a factor ΓZ −1 actually lowers
the order of such diagrams. Thus, around the Z pole, the W -exchange diagram is formally
of higher order in perturbation theory.
4
Remember, this is all strictly tree level...
August 31, 2019 425

Very high and very, very high energy

We may also consider what happens at very high energies, s  mZ , mW . In
that case we may approximate the cross section by
σ(e+ e− → νe ν̄e ) =
~2 gW 4 2gW 4 + gW 2 aν (ve + ae )
  
s
− log . (14.17)
π mW 2 s mW 2
At extremely high energies the cross section becomes very simple :
~2 gW 4
σ(e+ e− → νe ν̄e ) = (s → ∞) . (14.18)
π mW 2
This is indeed a nice, simple expression — but it is a constant. Does this not
conflict with unitarity, that demands a 1/s behaviour ? The solution appears
if we realize that the unitarity behaviour is required in each individual angular
momentum channel, or alternatively in a situation where s becomes very
large together with all other momentum transfers : in this case, a fixed t/s
ratio. As we can see, at high energies the process is completely dominated
by t values close to zero, so the ratio t/s approaches zero5 . Another way
of putting this is to say that in the infinite-energy limit, all intermediate
angular momentum channels contribute. A quick inspection of Eq.(14.6)
shows that for s → ∞ at fixed t/s the amplitudes are indeed just simple,
energy-independent (but angle-dependent) quantities6 .

14.2 W pair production in e+e− scattering

14.2.1 Setting up the amplitude
In sec. 13.3.1 we have saved unitarity in the process e+ e− → W + W − by
introducing the Z particle, upon which the dangerous terms that lead to
faulty high-energy behaviour are cancelled. But of course, the well-behaved
remainder is also of interest, and its computation is a nice example of get-
your-hands-dirty theoretical work. In detail, the process is described as
e+ (p1 , λ) e− (p2 , λ) → W + (q1 , ρ11 ) W − (q2 , ρ12 )
5
Since the (t−mW 2 )−1 propagator peaks for vanishing t. The distribution proportional
to (t − mW 2 )−2 , with −s < t < 0, gives expectation value hti ≈ mW 2 − mW 2 log(s/mW 2 )
so that for very large s the ratio is typically t/s ∼ − log(s)/s.
6
Since u = −s − t, when t/s is fixed then so are u/s and u/t.
426 August 31, 2019

where we explicitly indicate the momenta p1,2 (taken to be massless) and

q1,2 , and the polarisation vectors 1,2 . The electron helicity is λ = ± and
the W polarisations are denoted by ρ1,2 that can take the values ± or 0, the
latter standing for the longitudinal polarisation. There are thus 18 different
amplitudes to be considered. We write the amplitude as
2
X
M(λ, ρ1 , ρ2 ) = i~ Aj (λ) Bj (λ; ρ1 , ρ2 ) , (14.19)
j=1

where the index 1 refers to the νe -exchange diagram, 2 to the combined γ/Z
diagrams. The A’s collect the coupling constants and propagators :
1
A1 (λ) = −4gW 2δλ,− ,
∆
e2 gWWZ (ve − λae ) e2 mZ 2 4gW 2
A2 (λ) = + = − + δλ,− (14.20)
.
s s − mZ 2 s(s − mZ 2 ) s − mZ 2

where P = p1 + p2 = q1 + q2 and s = P 2 , ∆ = (p1 − q1 )2 . We see that A2 (+)

goes asymptotically as 1/s2 at high energies, whereas A1,2 (−) decreases as
1/s. The real work of calculation is in the ‘spacetime’ objects

B1 (−; ρ1 , ρ2 ) = u− (p1 ) /ρ11 (/q1 − p/1 ) /ρ22 u− (p2 ) ,

B2 (λ; ρ1 , ρ2 ) = uλ (p1 ) γα uλ (p2 ) Y (q1 , ρ11 ; q2 , ρ22 ; −q1 − q2 , α)
= (ρ11 · ρ22 ) uλ (p1 ) (/q1 − q/2 ) uλ (p2 )
+ 2 (q2 · ρ11 ) uλ (p1 ) /ρ22 uλ (p2 )
− 2 (q1 · ρ22 ) uλ (p1 ) /ρ11 uλ (p2 ) , (14.21)

where in the last expression we have used (qj · j ) = 0.

14.2.2 Momenta and polarisations

We shall, of course, work in the centre-of-mass frame, where the momenta
are chosen as follows :

pµ1 = (E, E~ep ) , pµ2 = (E, −E~ep ) , q1µ = (E, q~e) , q2µ = (E, −q~e) .
(14.22)
The beam energy is E so that s = 4E 2 , and the W velocities are β = q/E,
where q 2 = E 2 − mW 2 . The ~ep and ~e are unit vectors, with ~ep · ~e = c, the
August 31, 2019 427

cosine of the scattering angle between e+ and W + . In addition we define two

massless vectors7
aµ = (1, ~e ) , bµ = (1, −~e ) (14.23)
For the transverse W polarisations we shall use ±
1,2 = ± and

1
(± )µ = √ u± (a) γ µ u± (b) . (14.24)
8
The two longitudinal polarisations are of different : we write

2mW 2 2mW 2
   
0 1 0 1
1 = q1 − P , 2 = q2 − P . (14.25)
mW β s mW β s

14.2.3 Working out the amplitudes

It is enlightening to see explicitly how the various elements of M are com-
puted. First, we can use /ρ /ρ = 0 to write

B1 (−; +, +) = − u− (p1 ) /+ p/1 /+ u− (p2 )

= −2(p1 · + ) u− (p1 ) /+ u− (p2 )
1
= − u+ (a) p/1 u+ (b) u− (p1 ) u+ (b)u+ (a) u− (p2 )
2
1
= − s− (p1 , b)2 s+ (a, p1 ) s+ (a, p2 ) ,
2
1
B1 (−; −, −) = − s− (p1 , a)2 s+ (b, p1 ) s+ (b, p2 ) . (14.26)
2
Similarly,
1
B1 (−; +, −) = u− (p1 ) u+ (b)u+ (a) (/q1 − p/1 ) u+ (a)u+ (b) u− (p2 )
2
= (a · q1 − p1 ) u− (p1 ) b/ u− (p2 )
∆ + m2 u− (p1 ) q/2 u− (p2 )
=
2q q
2
2(∆ + m )
= u− (p1 )/q2 u− (p2 ) = B1 (−; −, +) .(14.27)
β 2s
7
Note that the definition of a and b can be made Lorentz invariant by requiring them
to be massless linear combinations of q1 and q2 .
428 August 31, 2019

The next one is a bit more involved :

2mW 2
 
1
B1 (−; +, 0) = u− (p1 ) /+ (/q1 − p/1 ) q/2 − p/1 u− (p2 )
mW β s
2mW 2

1
= u− (p2 ) −∆/+ − (2p1 · q1 )/+
mW β s
2mW 2

+ (2p1 · + )/q1 u− (p2 )
s
2mW 2
 
= s− (p1 , b)s+ (a, p2 ) −∆ − (2p1 · q1 )
s
2mW 2

+ s+ (a, p1 )s− (p1 , b) q s− (p1 , a)s+ (a, p2 )
2
2mW 2
 
s− (p1 , b)s+ (a, p2 )

= √ −∆ − (2p1 .q1 ) − 2q(a · p1 )
mW β 2 s
1
u− (p1 ) /+ u− (p2 ) ∆ + mW 2 (1 − β) .


= − (14.28)
mW β
This is the most non-trivial of them all : notice the juggling with the spinor
products. Nevertheless the result factorizes nicely, and that does not depend
on the use of the standard form. The other such cases are
1
u− (p1 ) /+ u− (p2 ) ∆ + mW 2 (1 + β) ,


B1 (−; −, 0) = −
mW β
B1 (−; 0, ±) = −B1 (−; ±, 0) . (14.29)
The last case for B1 is
B1 (−; 0, 0) =
2mW 2 2mW 2
   
1
u− (p1 ) q/1 − p/2 (/q1 − p/1 ) q/2 − p/1 u− (p2 )
mW 2 β 2 s s
2mW 2 2mW 2 4mW 4
 
1
= u− (p1 ) ∆/q2 − ∆ p/1 + ∆ p/2 + p/2 q/1 p/1 u− (p2 )
mW 2 β 2 s s s2
4mW 4
 
1
= u− (p1 ) q/2 u− (p2 ) ∆ + . (14.30)
mW 2 β 2 s
Note that we have used momentum conservation, u(p1 )(/q1 + q/2 )u(p2 ) = 0.
Computationally, for B2 things are simpler. First of all, we have trivially
B2 (λ; +, +) = B2 (λ, −, −) = 0 , (14.31)
August 31, 2019 429

and, almost equally trivially,

B2 (λ; +, −) = B2 (λ; −, +) = 2 uλ (p1 ) q/2 uλ (p2 ) , (14.32)
owing to our choice of the polarisations. Next, by
1 sβ
q1,2 · 02,1 = (q1 q2 ) − mW 2 =



(14.33)
mW β 2mW
we have
sβ
B2 (λ, ±, 0) = − uλ (p1 ) /± uλ (p2 ) , = −B2 (λ, 0, ±) . (14.34)
mW
And, finally,
s + 2mW 2
B2 (λ; 0, 0) = uλ (p1 ) q/2 uλ (p2 ) . (14.35)
mW 2
We see that the nonzero B2 always carry the same spinor sandwich as do
the B1 . Indeed they had better if they want to show (partial) cancellation
between them. As expected, the B’s behave, for large energy, as E 2 if both
W ’s are transverse, while they go as E 3 and E 4 if one or both become lon-
gitudinally polarised, respectively.
Of the various spinorial objects we can compute the absolute values :
|s− (p1 , b)2 s+ (a, p1 )s+ (a, p2 )|2 = s2 (1 + c)3 (1 − c) ,
|s− (p1 , a)2 s+ (b, p1 )s+ (b, p2 )|2 = s2 (1 − c)3 (1 + c) ,
1
|uλ (p1 )/q2 uλ (p2 )|2 = β 2 s2 (1 − c2 ) ,
4
1
|uλ (p1 )/−λ uλ (p1 )| = s(1 + c)2 ,
2
2
1
|uλ (p1 )/λ uλ (p1 )| = s(1 − c)2 .
2
(14.36)
2
Using these results, we can easily evaluate the absolute values of the ampli-
tudes. Some observations are in order here. First, there are two vanishing
amplitudes : M(+; +, +) = M(+; −, −) = 0. This is, of course, due to our
choice of ± for both transverse polarisations. Secondly, the nonzero am-
plitudes do not look nice, what with the occurence of the two couplings gW
and e, two masses mZ and mW , the denominator ∆ here and there, and β
cropping up in various places. There is a message here : real phenomenology
is messy. It is only in the more Platonic realm of very high energies that we
may hope to see our results simplify, and this we shall now investigate.
430 August 31, 2019

14.2.4 W pair production at very high energy

As has been amply discussed in these notes, the high-energy behaviour of
amplitudes often requires diagrams to partially cancel against one another.
Indeed, that is what led us to introduce the Z boson in W W production in
the first place. However, here we are interested in what is actually left after
the cancellations have taken place. We shall consider the limit where E is
very much larger than mZ , mW , at fixed c. In that limit, we may approximate
∆ ≈ −2E 2 (1 − c). We shall neglect overall complex phases in what follows.
The amplitude for this 2 → 2 process should have no energy dimension (cf.
sec. 6.3.3), so any contribution going as E −1 or lower can be neglected. Of
course all contributions going as E +1 or higher had better cancel8 ! This
actually leaves only a few surviving amplitudes. In the first place,

~e2 (1 + c)
M(−; +, +) ≈ (1 − c2 )1/2 ,
2sW 2 (1 − c)
~e2
M(−; −, −) ≈ 2
(1 − c2 )1/2 . (14.37)
2sW
The apparent singularity at c = 1 is of course due to our approximating ∆.
Note also that e2 ~ is actually the properly dimensionless quantity 4πα, see
sec. 10.2.5. Next we have
~e2 mZ 2
M(+; 0, 0) ≈ (1 − c2 )1/2 . (14.38)
2 mW 2
The most difficult one is

M(−; 0, 0) = ~ u− (p1 )/q2 u− (p2 ) K ,

4gW 2 4mW 2
 
K = − 2 ∆+
β mW 2 ∆ s
2
e mZ 2
2
4gW 2
 
s + 2mW
+ − + . (14.39)
mW 2 s(s − mZ 2 ) s − mZ 2

Performing the cancellations inside the factor K with care we find

~e2 mZ 2
  
2 1/2 1 1
M(−; 0, 0) ≈ (1 − c ) − −1 . (14.40)
2 sW 2 mW 2 2sW 2
8
This is what we did all the hard work for, after all.
August 31, 2019 431

We have now proven that the amplitudes surviving the high-energy limit are
dimensionless and well behaved ; but they still depend on the ratio mZ /mW .
Since we have taken the fermions to be massless, the Higgs boson is absent at
the tree level here. Nevertheless (but this is, as we see, an additional input !)
we may take mW /mZ = cW from our discussion of the minimal Higgs sector
in sec. 13.4.2, and then we find the even more pleasing-looking results
1 1+c
M(−; +, +) ≈ 4πα (1 − c2 )1/2 ,
2sW 2 1 − c
1
M(−; −, −) ≈ 4πα (1 − c2 )1/2 ,
2sW 2
1
M(+; 0, 0) ≈ 4πα (1 − c2 )1/2 ,
2cW 2
1
M(−; 0, 0) ≈ 4πα (1 − c2 )1/2 . (14.41)
4sW 2 cW 2

14.3 Higgs coupling to massless vectors

14.3.1 The γγH vertex
Although the photon is massless and therefore has no direct coupling to the
Higgs, such a coupling is effectively realised by fermion loops9 . At one-loop
order, we then have two contributing diagrams :

MγγH = + , (14.42)

which differ in the orientation of the fermion line. If our theory is to be con-
sistent, this amplitude must be ultraviolet-finite (since otherwise we would
have to introduce a counterterm which would mean a direct photon-photon-
Higgs coupling after all) and it must obey current conservation. We shall
verify this point first, by putting a handlebar on one of the photons :

9
And it provided the first mechanism for observing Higgs particles at the LHC by
measurement of the invariant mass of photon pairs.
432 August 31, 2019

= − + −

(14.43)

By using the identities

= , = (14.44)

we see that the third and fourth diagram are actually equal to the first and
second one (flipped over), so that the sum vanishes and current conservation
(gauge invariance) is assured.
The process we investigate is, more explicitly, given as

γ(q1 , 1 ) + γ(q2 , 2 ) → H

where we have explicitly given the momenta and polarisations of the (on-
shell) photons. We denote the Higgs mass by mH and the fermion mass by
M . The fermion charge is denoted by Q, and its coupling to the Higgs by
gffH . Although we do not expect (or hope to see) divergences, we shall still
work in a general dimension 4 − 2 for reasons that will be come clear later
on : but we shall take  → 0 wherever we can. One of our diagrams10 is
given by

(−1)i6 Q2 ~3/2 gffH

Z
N
M1 = 4
d4 k ,
(2π) D
D = k 2 − M 2 + iη (k − q2 )2 − M 2 + iη (k + q1 )2 − M 2 + iη ,




 
N = Tr (/k + M )/2 (/k − q/2 + M )(/k + q/1 + M )/1 . (14.45)

The other diagram, M2 , is obtained by interchange of the two photons. Using

the shorthand δxy as in section 11.5, we can write
Z
1 2
= δxy
3 ; (14.46)
D k 2 + 2x(k · q1 ) − 2y(k · q2 ) − M 2 + iη

accordingly we shall shift k µ = k̂ µ − xq1 µ + yq2 µ .

10
In fact, the second one.
August 31, 2019 433

We now turn to N . Since the whole amplitude is current-conserving, we

are allowed to replace the polarisations by the explicitly current-conserving
combinations we already encountered in section 10.7.3 :
(q2 · 1 ) µ (q1 · 2 ) µ
1 µ → η1 µ = 1 µ − q1 , 2 µ → η2 µ = 2 µ − q2 ,
(q1 · q2 ) (q1 · q2 )
(14.47)
with the nice properties that ηj · qk = 0. After shifting the momenta and
dropping terms linear in k̂, we have

N → 16M (k̂ · η1 )(k̂ · η2 )

 
+ 4M (η1 · η2 ) − 4k̂ 2 + 4M 2 + mH 2 (4xy − 2) . (14.48)

Considering the first term, we notice that in our integral we can again, just
like in section 11.5, replace
 
α β 1 2 αβ 1 
k̂ k̂ → k̂ g ∼ + k̂ 2 g αβ . (14.49)
4 − 2 4 8
Thus effective form of N is
 
2 2 2
N → M (η1 · η2 ) 4k̂ + 4M + mH (4xy − 2) , (14.50)

and the diagram is therefore

M1 = Q2 gffH M (η1 · η2 ) A ,
2µ2 2 2 2
4−2 4k̂ + 4M + mH (4xy − 2)
Z Z
A = δxy d k̂ 3 (14.51)
(2π)4−2

k̂ 2 − M 2 + xy mH 2 + iη

We see that in fact M2 = M1 , owing to our use of the η’s rather than the
’s11 . Proceeding with the one-loop cookbook of section 6.612 we can write
Z∞
4t − 4M 2 − mH 2 (4xy − 2)
Z
i
A= 2 δxy dt t1−
3 (14.52)
8π t + M 2 − xy mH 2 − iη
0

11
You might start to worry at this point. Weren’t the two diagrams necessary for the
cancellations proving current conservation ? How, then can they be equal ? The answer
is that by going from  to η we have exchanged contributions between the diagrams, and
indeed in the η formulation they are separately current-conserving.
12
And taking  → 0 wherever that is justified by the event.
434 August 31, 2019

Since
Z∞ Z∞
t2− 1 t1− 1
dt 3
≈ + O (1) , dt 3
≈ + O () , (14.53)
(t + a)  (t + a) 2a
0 0

we find
1 − 4xy
Z
i
A = 2 δxy (14.54)
8π (M /mH 2 ) − xy − iη
2

We can also prettify the product η1 · η2 by introducing the field strength

tensor 13
Fiµν ≡ µi qiν − qiµ νi (i = 1, 2) (14.55)
so that
1
F1 µν F2µν .
η1 · η2 = (14.56)
mH 2
The whole amplitude then takes the form
iQ2 ~3/2 √ 1/2 µν M2
 
MγγH = (G F 2) F1 F2µν F ,
12π 2 mH 2
3τ (1 − 4xy)
Z
F(τ ) = δxy . (14.57)
τ − xy − iη
where we have used Eq.(13.133)) to rewrite gffH /M . The limit of large fermion
mass is interesting. For very large τ , F(τ ) approaches unity. The amplitude
thus becomes independent of M for very large fermion mass ! That this
happens is due to the fact that in Eq.(14.54) the M 2 terms in the numerator
cancel. But for that to happen, the first term, proportional to , has to be
there : were we to set d = 4 from the outset, it would be absent. Of course,
the fact that the numerator would contain a k̂ 2 term and the loop integral
would be divergent if d would differ ever so slightly from 4 should give us
pause. This is the reason why we have to stick to variable dimension in this
calculation : if we don’t, then the amplitude will be proportional to M 2 ! For
finite values of τ we can straightforwardly find



F(τ ) = 3τ 2 + (1 − 4τ ) Li2 (1/x+ ) + Li2 (1/x− ) ,
1  p 
x± = 1 ± 1 − 4τ + iη , (14.58)
2
and we plot it in figure 14.1.
13
Perhaps more familiar in its Fourier-transformed form ∼ ∂ µ Aν − ∂ ν Aµ .
August 31, 2019 435

Figure 14.1: The absolute value of F(M 2 /mH 2 ) defined in Eq.(14.58)

14.3.2 The ggH amplitude and the Next Generation

We can easily extend the computation of the previous section to the case of
gluon fusion, that is the process
g(q1 , 1 , j) g(q2 , 2 , `) → H
where we have indicated the momenta, polarizations, and colours of the glu-
ons. It is in fact this process that enabled the famous 2012 Higgs discovery
at the LHC14 . It suffices to replace Q by the strong coupling constant g, and
to include the colour factor

1
Tr T j T ` = δj` . (14.59)
2
The gluon-gluon-Higgs amplitude is
iαs ~1/2 √ 1/2 µν
 2
M
MggH = (GF 2) F1 F2µν δj` F , (14.60)
6π mH 2
where we used g 2 = 4παs /~.

14
Since it involves both a loop-induced production of the Higgs by gluon fusion and a
loop-induced decay of the Higgs into photon pairs, this process may also be considered as
a ‘proof’ of the actual reality of loop effects ; and as a true ‘quantum’ process since at the
tree level it is impossible.
436 August 31, 2019

An interesting observation can be made here. The Standard Model15

contains three families of quarks and leptons. Since these are so similar
(except for their masses), it is natural to ask whether a fourth, fifth,. . .
generation may exist, with particles that are simply too heavy to have been
produced so far16 . The Higgs is very helpful here, since we can see that each
extra doublet of super-heavy U, D quarks would enhance Higgs production
by gluon fusion by an amount that is about twice that delivered by the top
quark, independently of how heavy these newcomers are ! From the observed
rate of Higgs production we can be quite sure that the three generations we
know are all there are17 .

15
At least, its version of August 31, 2019.
16
Indeed, if you present the Standard Model to a lay audience, you should always try
to elicit this question.
17
You might construct theories in which the extra generations are somehow masked by
some cancellation mechanism - but then they are no longer copies of the known ones.
Chapter 15

Appendices

15.1 Perturbative (non)convergence issues

15.1.1 Punishment at the singular point
Let us reinspect the Green’s functions G2n for zero-dimensional ϕ4 theory, in
perturbation theory : G2n = H2n /H0 , with
s  n
2π~ ~ X  λ4 ~ k (n)
H2n = H2n , H2n = − 2
ak ,
µ µ k≥0
24µ
√
(n) (4k + 2n)! k n−1 2 2
ak = 2k+n ≈ k! 16 (4k) , (15.1)
2 (2k + n)! k! π
where in the last line we have used the Stirling approximation (cf appendix
15.15.2). Although we have treated the expressions for the H’s as if they were
well-defined objects, in fact these series do not converge ! For large k and
(n)
fixed n the coefficient ak increases superexponentially with k, so that the se-
ries has a radius of convergence equal to zero. This should not come as a sur-
prise. For, in the discussion of the perturbation expansion we have assumed
the coupling constant λ4 to be small, but positive. If, on the other hand,
it was small but negative, perturbation theory would look very different : in
fact it would look like nothing at all since for negative λ4 the path integral is
completely undefined. Therefore, the perturbative expansion is not regular
around λ4 = 0, and in the set of all ϕ4 theories the point λ4 = 0 constitutes
an essential singularity. And this is what we are expanding around ! Small
wonder that things go haywire.

437
438 August 31, 2019

15.1.2 Borel summation

All may not be lost, however. The method of Borel summation sometimes1
enables us to assign a value to a sum with vanishing radius of convergence.
Suppose that a function of a positive variable x is given by the sum
X
f (x) = ck x k , (15.2)
k≥0

where the coefficients ck grow superexponentially. Clearly it is difficult to

make sense of such a sum ; but it may be possible to make sense of a related
sum : X ck
g(x) = xk , (15.3)
k≥0
k!

simply because the coefficients do not grow as rapidly. Let us suppose that
this is indeed the case. We then may employ the formula
Z∞
dy exp(−y) (xy)n = n! xn , n = 0, 1, 2, . . . (15.4)
0

to arrive at the rule

Z∞
f (x) = dy e−y g(xy) . (15.5)
0

Notice that here, we have again interchanged summation and integration,

thus in a sense repairing the damage done when we arrived at the perturba-
tion expansion in the first place. This approach is called Borel summation.
We can illustrate this in a simple example, using ck = 1 :
Z∞
1 1
f (x) = → g(x) = ex → dy e−y exy = . (15.6)
1−x 1−x
0

However, an important observation is to be made here. The sum for f (x)

converges (conditionally) for the region |x| ≤ 1, whereas the sum for g(x)
converges everywhere, and the Borel integral converges in this case as long
1
In zero dimensions this will work. In four-dimensional Minkowski space things are not
nearly as simple. . .
August 31, 2019 439

as <(x) < 1, thus immeasurably enlarging the region of x values where the
Borel-summed version makes sense. A more challenging example :
X
f (x) = k! (−x)k , x > 0 . (15.7)
k≥0

for which Borel-summed form is therefore

Z∞ Z∞  
−y
X
k exp(−y) 1 1
f (x) = dy e (−xy) = dy = e1/x E1 (15.8)
k≥0
1 + xy x x
0 0

where the function E1 , the exponential integral, is discussed in appendix

15.15.8. F (x) is a function that starts (obviously) at F (0) = 1 and then
gently decreases. Borel summation works ! But how do we actually compute
the series F (x) ? The theory of asymptotic functions provides an answer. Let
us consider not the infinite sum F (x) as given in Eq.(15.7) but its truncated
version
K−1
X
FK (x) = k! (−x)k (15.9)
k=0

It can be shown that the difference

 between f (x) and fK (x) isK of

the or-
2
der of the first neglected term : F (x) − FK (x) = O K! (−x) . Taking
 
‘order’ to mean ‘roughly equal in magnitude, barring accidents’3 we might
therefore conclude that the optimal value of K is that for which the error
term is minimal, that is, we truncate around K ≈ 1/x. In that case k!xK
behaves roughly as exp(−1/x), so that the numerical error can be very small
indeed for small x. small indeed for small x. As an illustration, figure 15.1
shows the actual and asymptotically-inspired-truncated result for the func-
tion (15.7).The exact and truncated results for the function F (x) of (15.7).
The smooth curve is the exact, the zigzagging one the truncated result. The
approximate value oscillates around the true one, but for small x the differ-
ence is negligible. This shows that, even if a sum is divergent, it may still
2
Also this statement needs interpretation. In the theory of asymptotic series it means
that the difference will go to zero at least as fast as the first neglected term goes to zero,
not that these two numbers must be necessarily comparable in magnitude. As an example,
the object 1012 /x2 is formally of the order of 1/x as x → ∞, but x has to be really large
for them to be of equal size. Fortunately, it often happens that the difference and the
neglected term are of similar magnitude.
3
Only to be justified by its success.
440 August 31, 2019

Figure 15.1: Example of Borel summation

be possible to make sense out of it by Borel summation. Note that in our

example we have required x to be positive, so that (−x)n oscillates in sign.
That this is essential becomes clear when we try to Borel-sum
X
F (x) = n! (x)n , x > 0 : (15.10)
n≥0

the Borel integral reads

Z∞
1
F (x) = dy e−y , (15.11)
1 − xy
0

and this integral runs into problems around y = 1/x. One may of course
extend the integral to complex y values, and then skirt around the singularity.
But it is not clear whether we should pass the point y = 1/x above, or below,
the real axis. The ambiguity, that is, the difference between the results from
the alternative contours, is of course given by the number
e−y e−1/x ,
I
dy = 2πi
1 − xy x
y∼1/x
August 31, 2019 441

and, since during the integration we might decide to circle around the sin-
gularity any number of times, arbitrary multiples of the ambiguity may be
added. We see that the Borel integral becomes ambiguous : it may be some
consolation that the ambiguity is nonperturbative in nature, i.e. it has no
series expansion for infinitesimal but real and positive x. We conclude that
the function F (x) is given by
Z∞
e−y e−1/x
F (x) = P dy − (2n + 1)iπ ,
1 − xy x
0
(15.12)

where n is an undetermined integer. The real part of F is a principal-value

integral (cf appendix 15.15.6) that evaluates to the finite result

Z∞ Z1/x
 
−y −1/x z −z
P dy
e
=
e −E1 (1/x) + dz e − e  . (15.13)
1 − xy x z
0 0

15.2 More on symmetry factors

15.2.1 The origin of symmetry factors
In this section we shall return to the ‘simple’ world of zero dimensions, since
symmetry factors do not depend on the dimensionality of the theory. Let us
consider again the ϕ3/4 theory, with a path integrand
 

µ 2 λ3 3 λ4 4
exp −S(ϕ) + Jϕ = exp − ϕ − ϕ − ϕ + Jϕ
2! 3! 4!
3 n3
n
1 −λ4 ϕ4 4 1
  
−µϕ2 /2!
X 1 −λ3 ϕ
n1
= e Jϕ
n
n3 ! 3! n4 ! 4! n1 !
1,3,4≥0

(15.14)

We see that a diagram with n3 three-vertices, n4 four-vertices and n1 source

vertices carries an a priori factor of
1
.
n1 ! n3 ! n4 ! (3!)n3 (4!)n4
442 August 31, 2019

We then have to consider the number of ways in which a particular diagram

can be formed by connecting external lines and the vertices in appropriate
ways. The best way to learn this is, of course, to see how it is done.

15.2.2 Explicit computation of symmetry factors

Let us start with the very simple diagram

It is built up from 2 external lines (distinguishable !) and one four-vertex :

the a-priori factor is therefor 1/24. We can lay out the ingredients as a tool
kit4 : We can denote this as follows :
1
4!

To build the diagram, we first connect one of the external lines to the vertex :
there are 4 possible ways to do so since all legs of the vertex are indistin-
guishable. We then have

4
4!

To attach the other external leg, there are now 3 choices :

4×3
4!

For the remaining operation of linking the two other legs of the vertex there is
of course only one possibility. The resulting symmetry factor is 4×3/24 = 1/2
as advertised.

We next turn to a slightly more complicated diagram :

4
If you have ever played with K’NEX this may look familiar
August 31, 2019 443

The tool kit is now

1
2! × 4! × 4!
The first external leg can now be attached in 8 ways, since the vertices and
all their legs are indistinguishable :
8
2! × 4! × 4!
For the other external line there are now 4 possibilities :
8×4
2! × 4! × 4!
Now we arbitrarily consider the upper vertex leg on the left. It has to be
connected to one on the right, which can be done in 3 ways :
8×4×3
2! × 4! × 4!
The next leg on the left now has 2 ways to go :
8×4×3×2
2! × 4! × 4!
and then the diagram is closed uniquely, leading to a symmetry factor of
8.4.3.2/2!4!4! = 1/6.
The symmetry factors of vacuum diagrams are no exception, as we can
see from the construction of

The tool kit is now

1
4! × (3!)4
We arbitrarily pick one of the vertices and attach to an arbitrarily chosen leg
one of the 9 remaining vertex legs :

9
4! × (3!)4
444 August 31, 2019

The other two vertices are attached in 6 and 3 ways, respectively :

9×6×3
4! × (3!)4

Now take one of the outside legs. It must be attached to another outside leg
that comes from another vertex, hence 4 possibilities :

9×6×3×4
4! × (3!)4

In the same way, for the next outside leg there are possibilities :

9×6×3×4×2
4! × (3!)4

and then the diagrams is closed uniquely. The symmetry factor is therefore
1/24, corresponding to the symmetry of permuting the vertices.

We finish with a truly nontrivial one :

Note that the two crossing lines do not touch but form an ‘overpass’ : this
diagram is nonplanar in the sense that an overpass remains no matter how
you try to draw it. The toolkit is now even more impressive :
1
5! × (3!)5

We can attach the external lines to different 3-vertices in 15, 12, and 9 ways,
respectively :

15 × 12 × 9
5! × (3!)5
August 31, 2019 445

Now the leftmost object can be attached to the two remaining 3-vertices in
6 and then in 3 ways :

15 × 12 × 9 × 6 × 3
5! × (3!)5

The topmost leg of the object on the left has to be connected to both the
upper and the lower objects on the right, for which there are thus 4 and 2
options, respectively :

15 × 12 × 9 × 6 × 3 × 4 × 2
5! × (3!)5

By now, it ought to be obvious that there are 2 ways to finish off the dia-
gram, leading to a symmetry factor of (15.12.9.6.3.4.2.2)/(5!(3!)5 ), in simpler
terms 1/2. The actual symmetry is, once again, the permutation of the two
innermost vertices.

15.3 Derivation of the diagrammatic sum rules

In this section we shall derive the diagrammatic sum rules
For an arbitrary given diagram let us define the characteristics

E = number of external lines,

I = number of internal lines,
Vq = number of vertices of q-point type,
L = number of closed loops,
P = number of disjunct connected pieces.

An example is

E = 2 , I = 6 , V1 = 1 , V3 = 3 ,
V4 = 1 , P = 1 , L = 2 .

We now look for linear combinations T of these numbers that are the same
for all diagrams. That is, whatever we do to a diagram, the value of T must
remain unchanged. It is easy to see that any diagram can be transformed into
446 August 31, 2019

any other diagram by application of the following four basic transformations,

or their inverse :
(i) coalescing a q-vertex and a 3-vertex :

(ii) adding an external line onto any other line :

1111111
0000000
0000000
1111111 1111111
0000000
0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
↔ 0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111

(iii) cutting through a line such that the graph falls apart :
0000000
1111111
1111111
0000000 1111111
0000000 0000000
1111111
1111111
0000000 1111111
0000000
0000000
1111111 0000000
1111111
0000000
1111111 0000000
1111111 0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111 0000000
1111111 0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
↔ 0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111 0000000
1111111 0000000
1111111
0000000
1111111

(iv) cutting through a line which is part of a loop :

0000000000000000
1111111111111111
1111111111111111
0000000000000000 0000000000000000
1111111111111111
0000000000000000
1111111111111111 1111111111111111
0000000000000000
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111 0000000000000000
1111111111111111
0000000000000000
1111111111111111 0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
↔ 0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111 0000000000000000
1111111111111111
0000000000000000
1111111111111111

These four operations modify the characteristics as follows :

(i) : V3 → V3 − 1 , Vq → Vq − 1 , Vq+1 → Vq+1 + 1 , I → I −1 ;
(ii) : E → E+1 , I → I +1 , V3 → V3 + 1 ;
(iii) : I → I −1 , E → E+2 , P → P +1 ;
(iv) : I → I −1 , E → E+2 , L → L−1 .
If the combination
X
T = αE E + αI I + αq Vq + αL L + αP P (15.15)
q

is to be invariant under the four basic transformations, then the coefficients

α must obey
(i) : −αq + αq+1 − α3 − αI = 0 ,
(ii) : αE + αI + α3 = 0 ,
(iii) : αI − 2αE − αP = 0 ,
(iv) : αI − 2αE + αL = 0 . (15.16)
August 31, 2019 447

Adding (i) and (ii) we find

−αq + αq+1 + αE = 0 , (15.17)

with the general solution

αq = β − qαE . (15.18)
(ii) then gives us
αI = 2αE − β (15.19)
and (iii) and (iv) yield αP = −αL = −β. The invariant T can therefore be
written as
! !
X X
T = αE − qVq + E + 2I − β − Vq + I + P − L , (15.20)
q q

where αE and β are undetermined. We see that we have precisely two di-
agrammatic sum rules. By inspection of an arbitrary5 diagram we see that
T = 0, so that the sum rules are
X
qVq = 2I + E , (15.21)
q
X
Vq = I + P − L . (15.22)
q

These two sum rules each have their own ‘physical’ interpretation. In the
first place, the sum rule 15.21 simply tells us that external lines have one
endpoint to be attached to a vertex, and internal lines have two endpoints
that must be accomodated. The second rule, 15.22, comes into its own when
we assign momenta to flow along the lines. The momenta in the external lines
are assumed to be given In every vertex the total momentum is conserved,
and each connected piece must contain an overall momentum conservation.
We see that in that picture L is simply the number of internal lines whose
momentum is not fixed by conservation. This derivation of the sum rules
given above is to convince you that there is not a third type of sum rule,
based on yet again different arguments, lurking about.
5
Arbitrary, except that it must contain at least one vertex. There are two connected
diagrams without vertices: the first one, , conforms to the sum rules by choosing
I = −1, and the second one, , fits in if we choose I = 0. But these choices are
obviously somewhat forced.
448 August 31, 2019

15.4 Alternative solutions to the Schwinger-

Dyson equation
15.4.1 Alternative contours for general theories
In section 1.2.5, it was mentioned that ϕ3 theory is not well-defined for real
fields since the action will go to infinity whenever ϕ → +∞ or ϕ → −∞. It
is instructive to lift the requirement that ϕ be real. In that case, we see that
different integration contours become available for which the path integral is
well-defined (albeit not necessarily real). Let us consider a zero-diensional
theory with general action
m
X λp p
S(ϕ) = ϕ . (15.23)
p=1
p!

The requirement for the path integral to be defined is that at both endpoints
(still assumed to be at infinity in some complex direction) the real part of
the action goes to positive infinity. That is,
<(ϕm ) → +∞ ⇒
π 2π π 2π
− + k < arg(ϕ) < + k , k = 1, 2, . . . , m .(15.24)
2m m 2m m
The argument of the endpoints are restricted to certain intervals. Inside
each interval the precise value of the argument is irrelevant since the path
integral will be precisely the same: we may therefore say that for a theory
(m)
with highest interaction term ϕm the admissible endpoints are ∞n , with
n = 0, 1, 2, . . . , m − 1, where
 
(m) iφn 2π 1 2π 1
∞n = lim re , φn ∈ (n − ), (n + ) . (15.25)
r→∞ m 4 m 4
Since the path integrand is analytic, the theory is completely determined by
the endpoints. We see that for a theory with highest interaction of the form
ϕm there are precisely m − 1 independent solutions to the SDe, as necessary
since the SDe is a linear differential equation of order m − 1. We may take
(m)
these as given by a contour running between ∞0 and any of the m−1 other
(m)
∞n . By suitably combining several integrals we can of course also obtain
(m)
a theory based on a contour running between any two distinct ∞n .
August 31, 2019 449

An interesting observation can be made on the limit of vanishing cou-

pling. Consider an action in which the highest coupling is λm ϕm /m!, and
the next highest is λk ϕk /k!. We can immediately see that the theory will
remain well-defined in the limit λm → 0, provided that its endpoints ∞(m)
are chosen such as to overlap with two distinct endpoints of the subleading
coupling, ∞(k) . If this is not the case the path integral will not be defined in
the limit of vanishing leading coupling constant.

15.4.2 Alternative contours for ϕ3 theory

As an example, let us look again at ϕ3 theory. There are three endpoints
(3)
∞0,1,2 . Since the point −∞ is not inside one of the admissible edpoints,
the real axis is not a valid contour as we have remarked. An interesting
(3) (3)
well-defined choice is the contour between ∞1 and ∞2 : by symmetry
we see that, as long as the action’s parameters and the source are real, the
path integral and φ(J) are well-defined and real. On the other hand, both
(2)
endpoints overlap with the same endpoint ∞1 , which means that in the limit
λ3 → 0 the theory must become ill-defined. A quick look at the tree-level
form of the theory bears this out : for the action
λ 3 µ 2
S(ϕ) = ϕ + ϕ (15.26)
6 2
the classical solution is given by
s !
µ 2λJ
S 0 (φc (J)) = J ⇒ φc (J) = −1 ± 1+ 2 . (15.27)
λ µ

Choosing the − sign we obtain a clasical tadpole φc (0) = −2µ/λ, which

corresponds to the contour discussed above6 ; and indeed it becomes ill-
defined as λ → 0. The choice of the + sign gives a classical solution that
has a Taylor series expansion around λ = 0. It corresponds to the contours
(3)
running from ∞0 to either ∞(1) or ∞(2) ; it is not possible to tell which of
the two contours is intended. In fact the situation appears to be even worse.
6
It should be observed that the classical SDe allows us to construct the full classical
solution from the tadpole φc (0), and that from the classical solution we can construct the
full quantum solution — all perturbatively, and some care has to be taken if the tadpole
is nonzero.
450 August 31, 2019

If λ, µ and J are all real, the SDe can be iteratively solved starting from the
classical solution, and the perturbation series is completeley fixed as well as
real ; whereas the fact that the two integration contours are really distinct
from the real axis tells us that the path integral (and hence φ(J)) ought
to be complex, with the results from the two contours related by complex
conjugation. We conclude that the difference between the two alternative
path integrals must be non-perturbative in nature.

15.4.3 Alternative contours for ϕ4 theory

For ϕ4 theory, with action
1 4 1 2
S(ϕ) = λϕ + µϕ , (15.28)
4! 2
(4)
there are three independent contours. Since ∞1,3 do not overlap with any
∞(2) , we see that only the real axis gives a theory in which the limit λ4 → 0
(4)
is well-defined. Another interesting contour is that running between ∞3
(4)
and ∞1 : we may take this contour to be the imaginary axis. By the simple
variable transformation ϕ → iϕ0 we see that the theory we are actually
investigating here is that with real field ϕ0 but action
1 4 1 2
S(ϕ0 ) = λ4 ϕ0 − µϕ0 , (15.29)
4! 2
that is, a theory with the ‘wrong’ sign for the quadratic term. Such mod-
els are regularly studied in connection with the phenomenon of spontaneous
symmetry breaking7 . As we see, this theory does not have a standard per-
turbative expansion around λ4 = 0 even though the tadpole vanishes.

15.5 Diagram counting

15.5.1 Tree graphs and asymptotics
Direct counting
An interesting and useful application of zero-dimensional field theory lies
in the topic of counting diagrams. To the extent that we may consider
7
In zero dimensions, spontaneous symmetry breaking does not occur.
August 31, 2019 451

every diagram as being of the same ‘order of magnitude’ this gives an idea,
however crude, of the amplitude to be expected. Of particular interest is
the behaviour of the number of graphs under extreme circumstances such
as when the number of external lines becomes very large. In this section
we shall consider the simplest case, that of tree-level Green’s functions of a
single self-interacting field.
In order to count diagrams, we can simply consider the zero-dimensional
theory so that we are not bothered by summing diagrams over internal de-
grees of freedom. Secondly, we replace every vertex, and every propagator by
unity. This reduces every Feynman diagram to just its symmetry factor. For
tree diagrams, the symmetry factor is unity; for loop graphs, the symmetry
factors are nontrivial and getting rid of them is quite cumbersome8 . The
appropriate action reads
1 X k
S(ϕ) = ϕ2 − F (ϕ) , F (ϕ) = ϕk , (15.30)
2 k≥3
k!

where k is unity for every k-point interaction proposed in the theory, oth-
erwise zero. Since we only consider counting graphs, the fact that S may
become negative infinity for infinite ϕ does not bother us. The number-of-
diagrams generating function
X Nn
Φ(J) = Jn , (15.31)
n≥0
n!

where Nn is the number of tree graphs with n + 1 external lines, is given by

the classical version of the SDe :
Φ = J + F 0 (Φ) . (15.32)
There are several ways of solving for Φ. We may directly solve Eq.(15.32) as
an algebraic equation and then expand in powers of J, but this is practical
only in the simplest cases such as the ϕ3/4 theory. Alternatively, we can
approach the root of Φ = J + F 0 (Φ) by Lagrange expansion9 :
X 1  ∂ n−1  n
0
Φ=J+ F (J) . (15.33)
n≥1
n! ∂J
8
In the literature ‘counting diagrams’ is usually understood to mean ‘counting diagrams
with symmetry factors’.
9
This is proven in detail in section 15.15.12.
452 August 31, 2019

This is useful in theories with only a single coupling, such as pure ϕ4 the-
ory. In more complicated theories, the best approach for n not too large is
simply to iterate Eq.(15.32) by computer algebra. For pure ϕp theories we
can explicitly work out the result of the Lagrange expansion. The counting
equation is
1 m
φ=J+ φ , m=p−1 , (15.34)
m!
so that Lagrange’s formula gives
 n−1
X 1 ∂
φ = J+ J mn
n>0
n! (m!)n ∂J
X (mn)!
= J mn−n+1 . (15.35)
n≥0
n!(m!)n (mn − n + 1)!

The nonvanishing N ’s are therefore

(mn)!
Nn(m−1)+1 = , n = 0, 1, 2 . . . (15.36)
n!(m!)n

As expected, for m > 2 some connected Green’s functions vanish identically

at the tree level since no diagrams contribute.

Asymptotic methods
For asymptotically large n, we can estimate the form of Nn by realizing that
these must be given by the behaviour of Φ(J) near that of its singularities
that lies closest to the origin in the complex-J plane. Now, if Φ(J) is singular,
then Φ0 (J) is divergent10 , so that dJ/dΦ must vanish. We therefore solve the
equation
∂
J = 1 − F 00 (Φ) = 0 (15.37)
∂Φ
for Φ. If the highest power of interaction in the theory is ϕm , this equation
has m − 2 complex roots Φ1 , Φ2 , . . . , Φm−2 , and

Jp = Φp − F 0 (Φp ) , p = 1, 2, . . . , m − 2 . (15.38)
10
The divergence might also show up in higher derivatives only, but in every actual case
that I have studied the divergence shows up in Φ0 .
August 31, 2019 453

Now, single out that Jp that has the smallest absolute value11 , which we shall
call J0 , and its corresponding Φp will be writtten Φ0 . For J and Φ very close
to the values J0 and Φ0 , respectively, we may use Taylor expansion to write
1
J ≈ J0 − F 000 (Φ0 )(Φ0 − Φ)2 , (15.39)
2
since the linear term vanishes by definition. Hence
 1/2 s
J 2J0
Φ ≈ Φ0 − 1 − 000
(15.40)
J0 F (Φ0 )

close to the singularity. From the standard Taylor expansion12

√ X (2n)!
1− 1−x= xn+1 (15.41)
n≥0
(n + 1)!n!22n+1

we then recover the asymptotic form for Nn :

s
(2n − 2)! 1 8J0
Nn ≈ n 000
. (15.42)
(n − 1)! (4J0 ) F (Φ0 )

This estimate grows roughly as n!, as ought to have been immediately obvious
from the fact that Φ(J) has a finite radius of convergence ; the above, more
careful, treatment gives an estimate that is quite good even for non-huge n.
As an application, we may consider purely gluonic QCD. In this theory, the
only interactions are between 3 or 4 gluons, and the theory is equivalent, as
far as counting is concerned, to the ϕ3/4 theory, with
1 3 1 4
F (ϕ) = ϕ + ϕ . (15.43)
3! 4!
The solutions of Eq.(15.37) and the corresponding J values are
√ 4 √ √ 4 √
Φ1 = −1 + 3 , J1 = − + 3 ; Φ2 = −1 − 3 , J2 = − − 3 ,
3 3
(15.44)
11
The case that there are several such values is discussed in the next paragraph.
12
This can be proven by applying the Lagrange expansion to the object u = y + u2 /2 =
√
1 − 1 − 2y, and putting y = x/2.
454 August 31, 2019

n Nn (exact) Nn (asymptotic)
1 1 0.85
2 1 1.07
3 4 4.01
4 25 25.17
5 220 220.94
6 2485 2493.60
7 34300 34397.35
8 559405 560754.85
9 10525900 10547973.57

Table 15.1: Exact and asymptotic number of tree diagrams in ϕ3/4 theory

√ √ √
so that J0 = 3 − 4/3, Φ0 = 3 − 1, and F 000 (Φ0 ) = 3. In table 15.5.1 we
give the exact number Nn , and its asymptotic estimate. The approximation
is better than one per cent for n ≥ 3. The non-polynomial (that is, n!) growth
of the number of diagrams with n can be seen as an immediate indication
of the failure of perturbation theory as a convergent series, as discussed in
Appendix 1.

Coarse-graining effects
In the above we have assumed that there is only a single J0 . This is indeed
usually the case ; for pure ϕp theories, however, Eq.(15.37) reads
1 q
ϕ =1 , q =p−2 , (15.45)
q!
and this has solutions
 
1/q n
φn = (q!) exp 2iπ , n = 1, 2, . . . , q ; (15.46)
q
the corresponding values for J are
1 q
Jn = 1 − φn q+1 = φn , n = 1, 2, . . . , q , (15.47)
(q + 1)! q+1
and these have all the same absolute value. The thing to do is therefore
to take the asymptotic contributions from all these q singular points into
August 31, 2019 455

account, and sum them. We then obtain

s
q
X (2k − 2)! 8(q − 1)!Jn
Nk ≈ (4Jn )−k
(k − 1)!
n=1
φn q−1
 k r X q
(2k − 2)! q + 1 8
= φn −(k−1) . (15.48)
(k − 1)! 4q q n=1

The sum over the n values of φ will vanish completely, except when k − 1 is a
multiple of q, and then it evaluates to q/(q!)k−1 ; this is exactly the behaviour
we found using Lagrange expansion.
We might have proceeded otherwise, by simply taking the single real
solution φq = (q!)1/q as the only singular point. The number of diagrams
Nk will then be nonvanishing for every k value, while in the asymptotic
expression (15.48) the sum over n φ’s is replaced by φq −(k−1) , that is precisely
q times smaller than the nonvanishing sums of Eq.(15.48). We see that the
taking into account of only the single, real solution causes the asymptotic
values of Nk to be ‘smeared out’ ; Nk is then never zero anymore, but its
average value13 is still correct.

15.5.2 Counting one-loop diagrams

The SDe approach to counting diagrams has a number of interesting or useful
applications, one of which we discuss here. We can extend the treatment of
the previous section as follows. For the case of purely gluonic QCD the num-
ber of one-loop diagrams including their symmetry factors can be counted
by iterating the appropriate Schwinger-Dyson equation :
1 1 ~
Φ(J) = J + Φ2 + Φ3 + (1 + Φ) Φ0 (15.49)
2 6 2
and taking care to discard terms of order ~2 or higher. As an example, the
gluonic 20-point function is given by

N (19) = N0 (19) + ~ N1 (19) + O ~2 ,

N0 (19) = 11081983532721088487500 ,
N1 (19) = 2900013601350201168582750 . (15.50)
13
For the correct definition of ‘average’.
456 August 31, 2019

The number N0 (19) is the actual number of diagrams since tree diagrams
always have unit symmetry factor ; but the number N1 (19) underestimates
the actual number of diagrams since the symmetry factors are not trivial. We
can see, however, that the only possible nonntrivial symmetry factor at the
one-loop level is 1/2, as evidenced by the factor ~/2 in Eq.(15.49). Inspection
tells us that in this theory the only elementary Feynman diagrams that have
symmetry factor 1/2 are

E1 = , E2 = , E3 = ,

E4 = , E5 = . (15.51)

All diagrams that contain one of these elementaries as a subgraph will have
a symmetry factor 1/2, and it will suffice to determine their number and
multiply it by two14 . Alternatively, we may get rid of all such diagrams,
and work with the difference. This is the more useful approach ; and it
illustrates how we may go about using counterterms to impose constraints
on the structure of Feynman diagrams. The procedure is best explained by
going through it step by step. In the first place, it will become necessary to
again distinguish betwee three- and four-point vertices. We therefore modify
Eq.(15.49) be reinserting labels for these couplings:

g3 2 g4 3 ~
Φ(J) = J + Φ + Φ + (g3 + g4 Φ) Φ0 (15.52)
2 6 2
Iterating this gives for the first N :

~
N (0) = g3 ,
2  
1 2
N (1) = 1 + ~ g4 + g3 ,
 
2 7
N (2) = g3 + ~ 4g3 + g4 g3 ,
2
 
2 7 2 4 59 2
N (3) = g4 + 3g3 + ~ g4 + 24g3 + g4 g3 . (15.53)
2 2
14
This relies, of course, on the fact that there can be no diagrams containing two (or
more) of the elementaries, since that would be a two-loop diagram (or even higher).
August 31, 2019 457

We can now start to remove graphs. We shall get rid of all diagrams with a
tadpole by introducing a tadpole counterterm ~T in the SDe:
g3 2 g4 3 ~
Φ(J) = J + Φ + Φ + (g3 + g4 Φ) Φ0 − ~T (15.54)
2 6 2
We see that this amounts to replacing J by J − ~T , and the N ’s become
 
1
N (0) = ~ g3 − T ,
2
 
1 2
N (1) = 1 + ~ g4 + g3 − g3 T ,
 
2 7 2
N (2) = g3 + ~ 4g3 + g4 g3 − g4 T − 3g3 T ,
2
 
2 7 2 4 59 2 3
N (3) = g4 + 3g3 + ~ g4 + 24g3 + g4 g3 − 10g4 g3 T − 15g3 T .
2 2
(15.55)

The tadpole N (0) is removed by choosing T = g3 /2 ; and by the recursive

structure of the SDe all diagrams containing the elementariy E1 are removed
as well. The remaining low-order N s are now
 
1 1 2
N (1) = 1 + ~ g4 + g3 ,
2 2
 
5 3
N (2) = g3 + ~ 3g4 g3 + g3 ,
2
 
2 7 2 49 2 33 4
N (3) = g4 + 3g3 + ~ g4 + g4 g3 + g3 . (15.56)
2 2 2
Next, we want to get rid of the two self-energy bubbles E2 and E3 . To this
end, we again modify the SDe:
g3 2 g4 3 ~ ~B
Φ(J) = J + Φ + Φ + (g3 + g4 Φ) Φ0 − ~T + Φ , (15.57)
2 6 2 1 + ~B
where the strange-looking form of the counterterm is justified by the fact
that we can rewrite Eq.(15.57) into
  
g3 2 g4 3 ~ 0
Φ(J) = J + Φ + Φ + (g3 + g4 Φ) Φ − ~T 1 + ~B , (15.58)
2 6 2
458 August 31, 2019

which lends itself better to the purpose of iteration. We then obtain

 
1 1 2
N (1) = 1 + ~ g4 + g3 + B ,
2 2
 
5 3
N (2) = g3 + ~ 3g4 g3 + g3 + 3Bg3 ,
2
 
2 7 2 49 2 33 4 2
N (3) = g4 + 3g3 + ~ g4 + g4 g3 + g3 + 4Bg4 + 15Bg3 .
2 2 2
(15.59)
Requiring N (1) = 1 leads to B = −(g4 + g3 2 )/2, and we are left with
 
3 3
N (2) = g3 + ~ g4 g3 + g3 ,
2
 
2 3 2 2 4
N (3) = g4 + 3g3 + ~ g + 15g4 g3 + 9g3 . (15.60)
2 4
Now, the one-loop contribution to the three-point function N (2) must not
be completely cancelled, since it contains the diagram

which has symmetry factor 1 and must be retained. We therefore add a

counterterm to the three-point coupling in the SDe:
  
(g3 − ~δ3 ) 2 g4 3 ~ 0
Φ(J) = J + Φ + Φ + (g3 + g4 Φ) Φ − ~T 1 + ~B ,
2 6 2
(15.61)
where the counterterm is needed only at one place since we are working to
one-loop accuracy. The result of the iteration is
 
3 3
N (2) = g3 + ~ g4 g3 + g3 − δ3 ,
2
 
2 3 2 2 4
N (3) = g4 + 3g3 + ~ g + 15g4 g3 + 9g3 − 6δ3 g3 . (15.62)
2 4
The condition now is that
N (2) = g3 + ~g3 3 , (15.63)
August 31, 2019 459

which requires δ3 = 3g4 g3 /2 to remove all elementaries E4 , and leads to

 
2 3 2 2 4
N (3) = g4 + 3g3 + ~ g4 + 6g4 g3 + 9g3 . (15.64)
2
The same trick can be applied to the four-point coupling: the SDe is then
  
(g3 − ~δ3 ) 2 (g4 − ~δ4 ) 3 ~ 0
Φ(J) = J + Φ + Φ + (g3 + g4 Φ) Φ − ~T 1 + ~B ,
2 6 2
(15.65)
which gives
 
2 3 2 2 4
N (3) = g4 + 3g3 + ~ g4 + 6g4 g3 + 9g3 − δ4 . (15.66)
2
For the four point coupling, we only want to retain the diagrams

, , and ,

which occur respectively 3,6, and 3 times. Therefore, δ4 = 3g4 2 /2 removes

all occurrences of E5 . With these choices, the SDe Eq.(15.65) can be iterated
(and truncated to one-loop order!) to give all diagrams that do not contain
any of the elementaries E1,...,5 as subdiagrams15 .
For the 20-point gluonic amplitude we find that the number of diagrams
with symmetry factor unity is given by

N̂ (19) = N0 (19) + ~M1 (19) , M1 (19) = 2013070318716871853439000 .

(15.67)
The total number of one-loop diagrams is therefore given by


N̂1 (19) = M1 (19) + 2 N1 (19) − M1 (19) = 3786956883983530483726500 .
(15.68)
In table 15.5.2 we give the results for the amplitudes from two to twenty
external lines. It is seen that the ratio of one-loop to tree diagrams increases
with the number of external legs ; while the average symmetry factor per
one-loop diagram seems to slowly approaches unity. It can indeed be proven
that asymptotically it does do so.
15
The actual implementation of the approach described here in computer algebra may-
have to be somewhat modified in the interest of speed : simply iterating Eq.(15.65) as it
stands may lead to unwieldily large expressions.
460 August 31, 2019

n+1 N0 (n) N̂1 (n)/N0 (n) avg.symm.

2 1. 3. 0.5000
3 1. 14. 0.5357
4 4. 24.75 0.5758
5 25. 37.88 0.6066
6 220. 52.09 0.6309
7 2485. 67.47 0.6506
8 34300. 83.86 0.6672
9 5.594 105 101.2 0.6813
10 1.053 107 119.4 0.6936
11 2.244 108 138.5 0.7044
12 5.349 109 158.3 0.7140
13 1.409 1010 178.9 0.7226
14 4.064 1012 200.2 0.7305
15 1.274 1014 222.2 0.7376
16 4.315 1015 244.9 0.7441
17 1.569 1017 268.2 0.7502
18 6.101 1018 292.1 0.7558
19 2.525 1020 316.6 0.7609
20 1.108 1022 341.7 0.7658

Table 15.2: Computing the average symmetry factors

August 31, 2019 461

The above strategy can of course be applied to other problems as well.

For instance, we may remove all one-loop three- and four point elementaries
instead of just those with symmetry factor one-half : in that case we are
essentially renormalising the theory. It should also be clear that in that case,
in which we just want to remove subdiagrams rather than count them, it is
easy to go to more loops in an order-by-order approach.

15.6 Concavity of the effective action

In the zero-dimensional case of a single field variable, the effective action
is concave. Let us now investigate whether this persists in case of more
fields. Let the collection of all fields be denoted by {ϕ} as before, and the
collection of all sources, one for each field, by {J}. We shall denote the
combined probability density of all fields, including the effects of the sources,
by P ({ϕ}, {J}). The effective action is now that function of the collection
of all field functions {φ} that has the correct classical equation :
∂
Γ({φ}) = Jn . (15.69)
∂φn
Concavity of the effective action in the many-field case means that the matrix
∂ ∂ ∂
Γnm ≡ Γ({φ}) = Jn (15.70)
∂φn ∂φm ∂φm
has only positive eigenvalues. If this is the case, then also its inverse, the
matrix
∂
Hmn = φn (15.71)
∂Jm
must have only positive eigenvalues16 . That is, for any eigenvector a of H
the eigenvalue λ must be positive :
X
Hmn an = λ am , λ > 0 . (15.72)
n

In turn, this is guaranteed if

X
Hmn am an > 0 (15.73)
m,n

16
Since Γmn is symmetric, so is Hmn although this is not obvious from the form it is
written here.
462 August 31, 2019

for any vector a. Now, we have

R Q
( n dϕn ) P ({ϕ}, {J}) ϕm
φm = R Q , (15.74)
( n dϕn ) P ({ϕ}, {J})
and therefore
R Q
1 ( n dϕn ) P ({ϕ}, {J}) ϕm ϕn
Hmn = R Q
~ ( n dϕn ) P ({ϕ}, {J})
R Q R Q
( n dϕn ) P ({ϕ}, {J}) ϕm ( n dϕn ) P ({ϕ}, {J}) ϕn
− R Q R Q (.15.75)
( n dϕn ) P ({ϕ}, {J}) ( n dϕn ) P ({ϕ}, {J})
We now employ the following trick : duplicate the set of fields {ϕ} by the
addition of another set of fields, {ϕ̂}, with the combined probability density
P ({ϕ}, {ϕ̂}, {J}) = P ({ϕ}, {J})P ({ϕ̂}, {J}) . (15.76)
By this construction, the random variables ϕ and ϕ̂ are statistically indepen-
dent. We can then write the matrix H as
1
Hmn = hϕm ϕn − ϕm ϕ̂n i , (15.77)
~
with the average taken with respect to the new probability density. Using
the fact that this density is symmetric in ϕ ↔ ϕ̂, we can write this as
1 1
Hmn = hϕm ϕn − ϕm ϕ̂n − ϕ̂m ϕn + ϕ̂m ϕ̂n i
~ 2
1
= h(ϕm − ϕ̂m )(ϕn − ϕ̂n )i , (15.78)
2
and we arrive at
* !2 +
X ~ X
Hmn am an = (ϕn − ϕ̂n )an , (15.79)
m,n
2 n

which is necessarily positive. The matrix H has, therefore, only positive

eigenvalues, and the effective action is always concave. It is of course possible
(and even likely in the case of continuum theories that have a noncountable
infinity of field values) that the eigenvalue is actually infinite. In that case
the effective action contains flat directions. So perhaps the more careful
statement is that the effective action cannot be convex anywhere.
A final point to note is that our proof relies only on the fact that the ϕ
values are randomly distributed over some nonvanishing region, no matter
how small. Of course, by restricting the values that the ϕ are allowed to take
we will change the effective action ; but it will never be convex.
August 31, 2019 463

15.7 Functional derivatives

To understand how functional derivatives work, let us assume that the action
of a theory, as in section 4.3.3, can be written as
Z  
0
S[ϕ] = F ϕ(x); ϕ (x) dx . (15.80)

Upon ‘discretisation’ using the Weyl ordering, this becomes

 
X 1 1
S = ∆F (ϕk+1 + ϕk ); (ϕk+1 − ϕk )
k
2 ∆
 
1 1
= ∆F (ϕn+1 + ϕn ); (ϕn+1 − ϕn ) +
2 ∆
 
1 1
∆F (ϕn + ϕn−1 ); (ϕn − ϕn−1 ) +
2 ∆
lots of terms not containing ϕn . (15.81)
The classical equation then reads ∂S/∂ϕn = 0, with
 
1 ∂ 1 1 1
S = F1 (ϕn+1 + ϕn ); (ϕn+1 − ϕn )
∆ ∂ϕn 2 2 ∆
 
1 1 1
+ F1 (ϕn + ϕn−1 ); (ϕn − ϕn−1 )
2 2 ∆
 
1 1 1
− F2 (ϕn+1 + ϕn ); (ϕn+1 − ϕn )
∆ 2 ∆
 
1 1 1
+ F2 (ϕn + ϕn−1 ); (ϕn − ϕn−1 ) , (15.82)
∆ 2 ∆
where Fj denotes the partial derivative of F with respect to its j-th argument.
Re-inserting the Weyl ordering, we can write this equation as
   
1 0 1 0
0 = F1 ϕ(x); ϕ (x) + F1 ϕ(x − ∆); ϕ (x − ∆)
2 2
   
1 0 1 0
− F2 ϕ(x); ϕ (x) + F2 ϕ(x − ∆); ϕ (x − ∆) (15.83)
.
∆ ∆
By Taylor expansion we get, for arbitrary f :
   
0 0 0 00
f ϕ(x − ∆); ϕ (x − ∆) ≈ f ϕ(x) − ∆ϕ (x); ϕ (x) − ∆ϕ (x)
464 August 31, 2019
      
0 0 0 00 0
≈ f ϕ(x); ϕ (x) − ∆ ϕ (x)f1 ϕ(x); ϕ (x) + ϕ (x)f2 ϕ(x); ϕ (x)
   
0 d 0
= f ϕ(x); ϕ (x) − ∆ f ϕ(x); ϕ (x) . (15.84)
dx
The classical equation thus takes the form
   
0 d 0
F1 ϕ(x); ϕ (x) − F2 ϕ(x); ϕ (x) = 0 . (15.85)
dx
we can cast this in the formal language of functional derivatives : we define,
using the Dirac delta function,
δϕ(y) δϕ0 (y)
= δ(x − y) , =0 ,
δϕ(x) δϕ(x)
δϕ0 (y) δϕ(y)
0
= δ(x − y) , =0 , (15.86)
δϕ (x) δϕ0 (x)
where, as we see, ϕ(x) and ϕ0 (x) are treated as independent variables. Apply-
ing these rules to the continuum form of the action, we find that the formal
form of the classical field equation is therefore that of the Euler-Lagrange
equation. The language of functional derivatives is, in these notes, treated
as an effective method, valid in the continuum limit, of writing the more
fundamental discrete classical field equation. In the functional formalism,
the Euler-Lagrange equation reads
 
δ d δ
S[ϕ, J] − S[ϕ, J] = 0 . (15.87)
δϕ(x) dx δϕ0 (x)
For ϕ4 theory, the Euler-Lagrange equation takes precisely the form (4.35).

15.8 Frustrated and unusual actions

15.8.1 Frustrating your neighbours
The one-dimensional action we have studied was based on ‘nearest-neighbour’
interactions. We can, of course, extend this treatment to include ‘next-to-
nearest-neighbour’ interactions as well. Let us take
X 1 
2
S({ϕ}) = ∆ µϕn − γ1 ϕn ϕn+1 − γ2 ϕn ϕn+2
n
2
August 31, 2019 465

X 1 1
= ∆ (µ − 2γ1 − 2γ2 )ϕn 2 − (γ1 + 4γ2 )(ϕn+1 − ϕn )2
n
2 2

1 2
− γ2 (ϕn+2 − 2ϕn+1 + ϕn ) , (15.88)
2

with the continuum behaviour of µ, γ1 and γ2 to be determined. We disregard

any other interactions since we shall only be interested in the propagator.
Setting up the SDe for the discrete propagator is trivial: we have
 
~
Π(n) = δn,0 + γ1 Π(n + 1) + Π(n − 1)
µ
 
+γ2 Π(n + 2) + Π(n − 2) , (15.89)

so that Fourier transformation gives us

un−1
I
~
Π(n) = du ,
2iπ f (u)
|u|=1
   
1 2 1
f (u) = µ − γ1 u + − γ2 u + 2 . (15.90)
u u

In the continuum limit, the only relevant poles of the integrand are those at
values of u such that |u| = 1 − O (∆). Let uj (j = 1, 2, . . .) be these poles:
then
X uj |x|/∆
Π(x) = ~ . (15.91)
j
f 0 (uj )
Writing u = 1 − v∆, we can approximate

f (u) = (µ − 2γ1 − 2γ2 ) − (γ1 + 4γ2 )(v 2 ∆2 + v 3 ∆3 )

−(γ1 + 5γ2 )v 4 ∆4 + O ∆2 .


(15.92)

There are now two possible continuum limits. In the first case, we can assume
that γ1 + 4γ2 does not vanish. In that case, we can take γ1 + 4γ2 ∼ 1/∆, and
the resulting continuum limit is indistinguishable from the nearest-neighbour
case. For later reference we shall denote this propagator by
~
P1 (x) = exp(−m|x|) . (15.93)
2m
466 August 31, 2019

The more curious solution is provided by the special choice γ1 = −4γ2 .

The only sensible continuum limit in that case is to take

1 4 1
γ1 + 5γ2 ∼ → γ1 ∼ , γ2 ∼ − 3 , (15.94)
∆3 ∆3 ∆

and
6
µ = m4 ∆ + 2γ1 + 2γ2 ∼ m2 ∆ + . (15.95)
∆3
The poles of the integrand are therefore approximately given by

f (u) = ∆ m4 + v 4 + O ∆2 = 0 ,



(15.96)

so that the solutions are

 2k−3
1+i
uk ≈ 1 − ∆m √ , k = 1, 2, 3, 4 . (15.97)
2

Only u1 and u2 are inside the unit circle, and we obtain the propagator
     
~ −m|x| m|x| m|x|
Π(x) = √ exp √ cos √ + sin √ (15.98)
,
m3 8 2 2 2

which we shall denote by P2 (x): it has the interesting property that Π2 (x)
is negative for mx between 3π/4 and 7π/4, modulo 2π. An discrete action
such as the one belonging to this continuum limit, in which nearest-neighbour
and next-to-nearest-neighbour couplings have opposite sign, are called frus-
trated17 . The continuum limit of the propagator can also be written as
Z
~ exp(ikx)
P2 (x) = dk , (15.99)
2π k 4 + m4

and that of the action reads

Z  
1 4 1
S[ϕ] = m ϕ(x)2 + ϕ00 (x)2 . (15.100)
2 2
17
Frustrated in the sense that ‘not all couplings can have it their own way’.
August 31, 2019 467

15.8.2 Increasing frustration

It is quite possible to construct even more frustrated actions, as follows. Let
us suppose that the action is given by
p
" #
X 1 X
S({ϕ}) = µϕn 2 − γj ϕn ϕn+j . (15.101)
n
2 j=1

The propagator is given by Eq.(15.91), where now

p  
X
j 1
f (u) = µ − γj u + j . (15.102)
j=1
u

We shall now arrange for the only the highest possible power of 1 − u to
survive in this expression. We first put u = exp(ik∆), so that the function
f (u) becomes
p
X X
f (u) = µ − 2γj cos(jk∆) = µ − (k∆)2r Br ,
j=1 r≥0
p
X 2(−)r
Br ≡ j 2r γj . (15.103)
j=1
(2r)!

We now seek to find the γ’s such that

1
B1 = B2 = · · · = Bp−1 = 0 , Bp = − . (15.104)
∆2p−1
In that case, we can take arbitrary constants cr , with cp = 1, and always
have p p
X X
cr Br = γj Q(j) = Bp , (15.105)
r=1 j=1

with p
X 2(−)r
Q(j) = cr j 2r . (15.106)
r=1
(2r)!
The polynomial Q(j) is even and of degree 2p in j, and Q(0) = 0. We can
now, for any preassigned q with 1 ≤ q ≤ p, choose the numbers cr such that
Q(0) = · · · = Q(q − 1) = Q(q + 1) = · · · = Q(p) = 0 , Q(q) 6= 0 ,
(15.107)
468 August 31, 2019

upon which
γq = Bp /Q(q) . (15.108)
Obviously, the polynomial Q(j) is given by

2(−)p Y  
2 2
Q(j) = j −n , (15.109)
(2p)!
0≤n≤p
n 6= q

from which we derive

(−)q−1 (2p)!
γq = , 1≤q≤p . (15.110)
∆2p−1 (p − q)!(p + q)!
The continuum limit of the propagator is, then
Z
~ exp(ikx)
Πp (x) = dk 2p (15.111)
2π k + m2p
The poles of the integrand are located at k = mωj , where
 
2j + 1
ωj = exp iπ , j = 0, 1, 2, . . . , 2p , (15.112)
2p

so that Cauchy integration gives

p
−i~ X
Πp (x) = ωj exp(iωj m|x|) . (15.113)
2pm2p−1 j=0

We may even investigate the limit p → ∞: in that case we may approximate

 −2p
1 m if −m < k < m
≈ (15.114)
2p
k +m 2p 0 elsewhere

so that the propagator takes the form

Zm
~ 1 sin(mx)
Πp (x) ≈ dk exp(ikx) = . (15.115)
2πm2p m2p−1 π mx
−m

Figure 15.2 shows the propagators Πp (x) for ~ = m = 1, as a function of

August 31, 2019 469

Figure 15.2: Propagators for frustrated actions

x. The values of p are 1,3,10, and also the asymptotic form of Eq.(15.115)
is plotted. For large p the asymptotic form is approximated smoothly. The
higher the value of p, the more frustrated the lattice is, and the more diffi-
cult it becomes for momentum modes with high wave number to propagate
through the lattice, as is evident from Eq.(15.111). For the totally frustrated
lattice, all wave numbers smaller than m propagate equally, and all wave
nubers larger than m do not propagate at all.

15.9 Newton’s First Law revisited

15.9.1 Introduction : the matter of sources
In our discussion of Newton’s first law in section 5.2.5 we have used a par-
ticular expression for the shape of the time-dependent part of the source,
motivated by mathematical convenience. Here we shall redo the analysis of
5.2.5 but with several different time dependences of the source. The response
of the field function to the source is
0 0 ~
e−ik x +ik·~x
Z
1
0
φ(x , ~x) = 4
0 3
dk d k 0 2 Jt (k 0 ) Js (~k) , (15.116)
(2π) 2
k − ω + i
470 August 31, 2019
p
with ω = ~k 2 + m2 as usual. The space part of the source will be Gaussian :
in position language it reads
~x2
 
2 −3/4 i
Js (~x) = (2πσ ) exp − 2 + p~ · ~x (15.117)
4σ ~
which corresponds to unit strength,
Z
d3 x |Js (~x)|2 = 1 . (15.118)

In momentum language, we have

Z  2 !
~ 1
Js (~k) = d x Js (~x) e
3 −i k·~
x 2
= (8πσ )3/4
exp −σ ~k − p~
2
. (15.119)
~

For the time dependence of the source we examine three alternatives, which
may be called slow, fast, and abrupt, respectively : in position language,
|x0 |
 
(1) 0 1 i 0 0
Jt (x ) = exp − − px ,
(a1 )1/2 a1 ~
!
2
(2) 0 1 x0 i 0 0
Jt (x ) = exp − 2 − p x ,
(2πa22 )1/4 4a2 ~
 
(3) 0 1 0

i 0 0
Jt (x ) = θ −a3 < x < a3 exp − p x . (15.120)
(2a3 )1/2 ~
These three sources are all normalised to unit strength :
Z
(j)
dx0 |Jt (x0 )|2 = 1 , j = 1, 2, 3 . (15.121)

To compare the spread of the sources in the time domain we can use
D E Z
2 (j) a1 2 a3 2
x0 = dx0 (x0 )2 |Jt (x0 )|2 = = a2 2 = . (15.122)
2 3
In momentum language we have
(1) 2a1 1/2
Jt (k 0 ) = ,
∆2 + 1/a1 2
(2) 2 2
Jt (k 0 ) = (8πa22 )1/4 e−a2 ∆ ,
(3) −i −ia3 ∆
Jt (k 0 ) = ia3 ∆


e − e , (15.123)
(2a3 )1/2 ∆
August 31, 2019 471

with ∆ = k 0 − p0 /~. The response of the field to the timelike part of the
source is 0 0
e−ik x
Z
0 (j)
ψj ≡ dk 0 2 Jt (k 0 ) , (15.124)
k − ω 2 + i
and this is what we now investigate for positive times : x0 > 0.

15.9.2 Slow, fast and abrupt

Slow source: exponential behaviour
The case j = 1 is what we have already investigated. Recalling the discussion
of how to close the contour in the k 0 plane, we have
exp(−iωx0 )

√
ψ1 = −2iπ a1 ω
(ω − p0 /~)2 + 1/a1 2
ia1 ω exp(−x0 /a1 − ip0 x0 /~)

+ (15.125)
(p0 /~ − i/a1 )2 − ω 2
As we have remarked before, the second term dies out with the source, so
that for large enough times x0  a1 we can disregard it.

Fast source : Gaussian behaviour

The second, ‘moderate’ time dependence has to be treated more carefully.
2
This is due to the fact that the exponential exp(−k 0 ) diverges when k 0 → ∞
if the argument of k 0 lies in (π/4, 3π/4) or (5π/4, 7π/4), so we cannot close
the contour simply as in the previous case. Instead, we write
!
02
x
ψ2 = a2 (8πa2 )1/4 exp − 2 − ix0 p0 /~ A2 ,
4a2
Z∞
exp (−(y + iτ )2 )
A2 = dy ,
(y − b+ + i)(y − b− − i)
−∞

b± = a2 (±ω − p0 /~) , τ = x0 /(2a2 ) , (15.126)

and we have chosen y = a2 (k 0 −p0 /~). The y integral runs over the real axis ;
we may shift it downwards by an amount τ provided we include a contour
integral around the point b+ − i, as indicated in the figure 15.9.2. The poles
at b± ∓ i are indicated.
472 August 31, 2019

Figure 15.3: Contour shift for a Gaussian source

The result of this operation on A2 is

exp (−(y + iτ )2 )
I
A2 = − dy
y∼b+ −i (y − b+ + i)(y − b− − i)
∞
exp(−z 2 )
Z
+ dz
(z − b− iτ )(z − b− − iτ )
−∞
2π
exp −(b+ + iτ )2


= −i
b+ − b−
Z∞
exp(−z 2 )
 
1 1
+ dz − . (15.127)
b+ − b− z − b− − iτ z − b+ − iτ
−∞

For large times (τ → ∞) the two integral terms in the second both behave
as ∼ 1/τ so, as in the previous case, they give rise to a contribution that dies
out with the source. We therefore have

lim ψ2 ∼ a2 (8πa2 )1/4 exp −a22 (ω − p0 /~)2 − iωx0 .

(15.128)
τ →∞

As before, the on-shell condition implied by ω ≈ p0 /~ is enforced.

Abrupt source : Heavyside behaviour

(3)
This case is the easiest one to analyse once we realise that Jt is perfectly
regular at ∆ = 0 notwithstanding the denominator. So for x0 > 0 we may
August 31, 2019 473

simply close the contour in the lower half complex k 0 plane to find
0 0 /~)
−π eia3 (ω−p /~) − e−ia3 (ω−p
ψ3 = √ exp(−iωx0 ) √ . (15.129)
ω 2 a3 (ω − p0 /~)
The numerator in the last factor remains bounded in absolute value. There-
√
fore, as long as ω 6= p0 /~, ψ3 goes to zero as 1/ a3 for large a3 , while at
√
ω = p0 /~ it approaches infinity as a3 : yet another situation in which the
on-shell condition is enforced.

15.9.3 Conclusion : general effect of the sources

We have seen that all three type of sources exhibit a large-time behaviour
0
ψj ∼ e−iωx dj (aj ; ω − p0 /~) , (15.130)

characterised by a plane wave exp(−iωx0 ), with the contributing values of

ω governed by a distribution dj such that for aj becoming large (hence the
source becoming broader in time) the energies ω are closely clustered around
the central value p0 /~. The reasoning leading to the mass-shell condition
E 2 ≈ |~p|2 c2 + M 2 c4 , as well as the motion along straight trajectories, ~x ≈
p~ t/p0 , follow in each case as discussed in section 5.2.5.

15.10 Unitarity bounds

15.10.1 Resonances
In this appendix we shall establish bounds on total cross sections as implied
by the unitarity of the theory. We are interested in upper bounds on cross
sections, that is we want to investigate the most efficient way to get rid
of the initial state in favour of some final state. Now, as is known from
the elementary theory of coupled oscillators, the most efficient way to pump
energy (i.e. the energy content of the initial-state particles) into another
state is by resonance. In our language, this means that we shall consider two
initial-state particles colliding and coupling to another particle with just the
right energy to put that particle on its mass shell. Unavoidably, if the new
particle can be made in such a way it can also decay, and it therefore must
have a nonzero decay width which protects its propagator from exploding.
We shall investigate this process in some detail.
474 August 31, 2019

15.10.2 Preliminaries : decay widths

We shall investigate the unitarity bound on the cross section for a given
initial two-particle state 1 to evolve into a given n-particle state 2 by way of
a resonant particle X of rest mass M and total decay width Γ. This means
that particle X must couple both to 1 and to 2. There is therefore a possible
decay X→1, described by
00000
11111
11111
00000
MX→1 = X
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
1 = i Ak · uj . (15.131)
00000
11111
00000
11111
00000
11111

In this admittedly abstract expression, u stands for the external-line factor18

for the incoming X particle that has, in addition to energy and momentum, a
discrete quantum number j denoting its angular momentum (for brevity we
shall use the smaller word ‘spin’ throughout this section). We shall assume
that j runs from 1 to N , so that there are in total N spin states : for a spin-J
particle, therefore, N = 2J + 1. Similarly the final state is characterized by a
discrete quantum number k alongside the continuous energy and momentum
variables, and k is assumed to run from 1 to K. For instance, if 1 stands
for an electron-positron state, K = 4 since there are two spin states for
the electron and two for the positron. Thus, Ak denotes the total of the
connected diagrams (the blob) and any external-line factors for a final state
with discrete quantum number k. The total decay width Γ1 for X to go into
the two-particle state 1 is given by

1 dΩ λ(M 2 , m2 , m02 )1/2

Z
1 1 X
Γ1 = Ak ·uj uj ·Ak S1 , (15.132)
2M N j,k (2π)2 8 M2

where m and m0 are the masses of the two particles in 1. The symmetry
factor S1 equals 1 if the particles are distinguishable, and 1/2 if they are not.
Ω is of course the solid angle of one of the particles in the rest frame of X.
The angle- and spin-averaged transition rate is therefore

16πM Γ1 N M 2
Z
1 X dΩ
Ak · uj uj · Ak = , (15.133)
K j,k 4π S1 K λ1/2

with λ1/2 = λ(M 2 , m2 , m02 )1/2 .

18
This might be just a number, or a spinor, or a polarisation vector,. . . take your pick.
August 31, 2019 475

The process X→2 is described by

11111
00000
00000
11111
00000
11111
MX→2 = X 00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
2 = i B l · uj , (15.134)
00000
11111
00000
11111
00000
11111
00000
11111

where l denotes the discrete quantum numbers in the state 2. The width for
the process is given by
Z
1 1 X
Γ2 = B l · uj uj · Bl dVn S2 , (15.135)
2M N j,l

where dVn is the n-particle phase space factor going with the state 2, and S2
is the appropriate symmetry factor.

15.10.3 The rôle of angular momentum conservation

Let us consider the process X→2 in some greater detail. It is easy to conceive
of a final state 2 that couples only to a particle of spin J and to no other
spin. Now, our important supposition : if the initial particle is at rest, and
if space is isotropic so that there is no preferred direction, this does not only
mean that angular momentum is conserved but also that the various 2J + 1
spin states of the X particle are to be treated on the same footing, so that
each spin state must have the same decay width. This in its turn implies that
the integrated-over final state must form a projection onto the pure spin-J
state : X X
Bl B l = B(M 2 ) un un (15.136)
l n

where n runs, of course, from 1 to N. Obviously, under the isotropy assump-

tion B can only depend on M 2 . We find that
Z X X
Bl · uj uj 0 · Bl = B(M 2 ) un · uj uj 0 · un ∝ δj,j 0 , (15.137)
l n

or, in other words,

XZ
B l · uj uj 0 · Bl = 2 M Γ2 δj,j 0 . (15.138)
l
476 August 31, 2019

15.10.4 The unitarity bound

We now consider
√ the process 1→2 by X exchange. For total scattering in-
variant mass s, it is given by the diagram
11111
00000
00000
11111 11111
00000
00000
11111 −i
M = 1
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
X 00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
2 = B l · Π · Ak . (15.139)
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
s − M 2 + iM Γ
Here, Π is the numerator of the X propagator : on the mass shell, therefore,
we must have X
Πcs=M 2 = uj uj . (15.140)
n

For the total cross section we therefore have

1 1 1
σ =
2λ(s, m2 , m02 )1/2 (s − M 2 )2 + m2 Γ2 K
X Z
× (B l · Π · Ak ) (Ak · Π · Bl ) dVn S2 . (15.141)
k,l

On the X mass shell, we can write, with the help of Eq.(15.138),

Z
1 1 1 X
σ = (B l · uj uj · Ak ) (Ak · uj 0 uj 0 · Bl ) dVn S2
2λ1/2 M 2 Γ2 K k,l,j,j 0
Z
1 1 1 X
= (B l · uj uj 0 · Bl ) (Ak · uj 0 uj · Ak ) dVn S2
2λ1/2 M 2 Γ2 K k,l,j,j 0
1 2M2Γ X
Z
1 dΩ
= 1/2 2 2
Ak · uj 0 uj · Ak δj,j 0 , (15.142)
2λ M Γ K k,j,j 0 4π

where it must be realized that we have rewritten the integral over B-cum-A
by the integral over B times the average over A. Due to angular-momentum
conservation we can now write, using Eq.(15.133),
  
Γ1 Γ2 N 16π s
σcs=M 2 = . (15.143)
Γ Γ S1 K λ(s, m2 , m02 )

Now, the factor Γ2 /Γ is understandable since the X particle has only a frac-
tional probability to decay into state 2 (there may be other decay channels
available, in fact at least the decay X→1), and then symmetry between the
reactions 1→2 and 2→1 requires also the presence of the factor Γ1 /Γ. We
August 31, 2019 477

conclude that the cross section for the initial state 1 to go into any final state
with spin J is bounded by the unitarity limit

2J + 1 16π s
σUL = , (15.144)
S1 K λ(s, m2 , m02 )

where as mentioned before S1 is 1/2 for indistinguishable particles and 1 for

distinguishable ones, and K is the total number of possible discrete quantum
numbers for the initial state19 .

15.11 The fundamental theorem for Dirac

matrices
15.11.1 Proof of the fundamental theorem
In this appendix we prove the following statement : if we have two sets of
four matrices, γ µ and γ̂ µ (µ = 0, 1, 2, 3) , satisfying Dirac’s anticommutation
relation

γ µ γ ν + γ ν γ µ = 2 g µν , γ̂ µ γ̂ ν + γ̂ ν γ̂ µ = 2 g µν , (15.145)

then there is a matrix S such that

γ̂ µ = S γ µ S −1 . (15.146)

To this end, we first set up a basis of the Clifford algebra as follows :

Γ0 = 1 , Γ1 = γ 0 , Γ2 = iγ 1 , Γ3 = iγ 2 , Γ4 = iγ 3 ,
Γ5 = γ 0 γ 1 , Γ6 = γ 0 γ 2 , Γ7 = γ 0 γ 3 , Γ8 = iγ 1 γ 2 ,
Γ9 = iγ 1 γ 3 , Γ10 = iγ 2 γ 3 , Γ11 = iγ 0 γ 1 γ 2 , Γ12 = iγ 0 γ 1 γ 3 ,
Γ13 = iγ 0 γ 2 γ 3 , Γ14 = γ 1 γ 2 γ 3 , Γ15 = iγ 0 γ 1 γ 2 γ 3 , (15.147)

which we denote by Γk , k = 0, 1, 2, . . . , 15 ; and using the γ̂ µ we construct an

analogous set Γ̂k in the same way. These have a few interesting properties.
19
For example, for an initial e+ e− state we have S1 = 1, K = 4 : for an initial state of
two photons S1 = 1/2, K = 4, and for an initial state of two gluons S1 = 1/2, K = 256
since gluons come with 2 possible spin states and 8 different colour states.
478 August 31, 2019

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1 1 0 5 6 7 2 3 4 11 12 13 8 9 10 15 14
2 2 5 0 8 9 1 11 12 3 4 14 6 7 15 10 13
3 3 6 8 0 10 11 1 13 2 14 4 5 15 7 9 12
4 4 7 9 10 0 12 13 1 14 2 3 15 5 6 8 11
5 5 2 1 11 12 0 8 9 6 7 15 3 4 14 13 10
6 6 3 11 1 13 8 0 10 5 15 7 2 14 4 12 9
7 7 4 12 13 1 9 10 0 15 5 6 14 2 3 11 8
8 8 11 3 2 14 6 5 15 0 10 9 1 13 12 4 7
9 9 12 4 14 2 7 15 5 10 0 8 13 1 11 3 6
10 10 13 14 4 3 15 7 6 9 8 0 12 11 1 2 5
11 11 8 6 5 15 3 2 14 1 13 12 0 10 9 7 4
12 12 9 7 15 5 4 14 2 13 1 11 10 0 8 6 3
13 13 10 15 7 6 14 4 3 12 11 1 9 8 0 5 2
14 14 15 10 9 8 13 12 11 4 3 2 7 6 5 0 1
15 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0

Table 15.3: Multiplication table for the Clifford algebra

In the first place, Γk 2 = 1 for all k. Secondly, for every pair j and k there
are numbers n and cn such that
Γj Γk = cn Γn , cn = 1, −1, i or − i. (15.148)
From these properties it follows that simultaneously
1
Γk Γj = Γn (15.149)
cn
We can thus construct the multiplication table 15.11.120 , where the possible
values of j define the rows, and those for k the columns: the corresponding
entry is then the value of n. For instance,
Γ6 Γ4 = Γ13
20
Kids! Don’t do this at home, since constructing this multiplication table is extremely
tedious. The numbers cn are not given: they are anyhow only defined up to a sign, since
we can always replace Γj by −Γj (using γ 2 γ 0 instead of γ 0 γ 2 , say) without changing the
Dirac anticommutation relation.
August 31, 2019 479

(in this case c13 happens to be 1). Note that in this table every row and every
column contains each of the numbers from 0 to 15 precisely once. Hence, if
we keep j fixed and let k run from 0 to 15, the value of n will also take on
all values from 0 to 15 (although generally in a different order). Obviously,
for the set Γ̂ exactly the same multiplication table holds.
We are now ready to prove the theorem. Let A be an arbitrary matrix,
and define S by
X15
S≡ Γ̂k A Γk . (15.150)
k=0

This has the desired property since

15
X
Γ̂j S Γj = Γ̂j Γ̂k A Γk Γj
k=0
15
X 1
= cn Γ̂n A Γn = S , (15.151)
n=0
cn

in other words,
Γ̂j S = S Γj . (15.152)
It remains to ensure that the matrix S actually has an inverse. Since A can
be chosen at will (except A = 0) this is not a problem. Let us pick another
matrix B and construct
X 15
T = Γk B Γ̂k . (15.153)
k=0

For this matrix we obviously have

Γj T = T Γ̂j . (15.154)

Combining Eq.(15.152) and (15.154) we see that the product T S commutes

with Γj (and the product ST commutes with Γ̂j ). Therefore T S is propor-
tional to the unit matrix and we can adjust the elements of B such that
T = S −1 .
It is an interesting observation that the dimensionality of the γ µ and that
of the γ̂ µ does not have to be the same. In that case the matrices A and
B are simply not square matrices but have different numbers of rows and
columns.
480 August 31, 2019

15.11.2 The charge conjugation matrix

An application of the fundamental theorem is the following. The anticom-
mutation relation, if satisfied by the Dirac matrices γ µ , is automatically also
satisfied by the matrices −(γ µ )T where T stands for the transpose. There
exists, therefore, a matrix C such that
γ̂ µ = C γ µ C −1 = −(γ µ )T . (15.155)
This is called the charge conjugation matrix. In the representation given in
section ??, we have
γ̂ 0 = −γ 0 , γ̂ 1 = γ 1 , γ̂ 2 = −γ 2 , γ̂ 3 = γ 3 ; (15.156)
and we see that a good choice is
C = C −1 = γ 0 γ 2 . (15.157)
Since this form is not proof against change of representation, the use of the
charge conjugation matrix in arguments and derivations lacks somewhat in
elegance.

15.12 States of higher spin

15.12.1 The spin algebra for integer spins
In this Appendix we shall consider how to represent particles with higher
spins. We shall start with integer spin. Such particles states can be repre-
sented, in the Feynman rules, as tensors of some rank r :
|s, miµ1 µ2 µ3 ···µr
where s stands for the total spin of the particle, and m denotes the spin along
some quantization axis, for which we shal take the z direction here. That is,
once we have found the correct operators of the spin algebra
(Sx,y,z )µ1 µ2 ···µr ν1 ν2 ···νr and (S 2 )µ1 µ2 ···µr ν1 ν2 ···νr
Then we have, by definition,
(S 2 )µ1 µ2 ···µr ν1 ν2 ···νr |s, miν1 ν2 ···νr = ~2 s(s + 1) |s, miµ1 µ2 ···µr ,
(Sz )µ1 µ2 ···µr ν1 ν2 ···νr |s, miν1 ν2 ···νr = ~ m |s, miµ1 µ2 ···µr . (15.158)
August 31, 2019 481

It is easy to see that the spin algebra is correctly constructed once we have
raising and lowering operators

(S± )µ1 µ2 ···µr ν1 ν2 ···νr , S− = (S+ )† ,

with
[[S+ , S− ], S+ ] = 2~2 S+ . (15.159)
We can then find the other algebra elements via
1
1
1
Sx = S+ + S− , S y = S+ − S− , Sz = [S+ , S− ] ,
2 2i 2~
1
S2 = S+ , S− + (Sz )2 .

(15.160)
2
We will start with particles in their rest frame21 . The spin representations
are built using four unit vectors, with obvious notation, as tµ , xµ , y µ and z µ ,
which obey

t·t = 1 , x·x = y·y = z·z = −1 , t·x = t·y = t·z = x·y = x·z = y·z = 0 .
(15.161)
Things will become easier if we also define
1
x± µ = √ xµ ± i y µ


(15.162)
2
so that

x± · x± = 0 , x+ · x− = −1 , x± · z = x± · t = 0 . (15.163)

Since the spin of a particle informs us about its behaviour under rotations in
the three-dimensional spacelike part of Minkowski space, we always require,
for particles in their rest frame,

|s, miµ1 µ2 ···µr tµj = 0 , j = 1, 2, . . . , r . (15.164)

This means that the appropriate tensors in fact contain only the three vec-
tors x+ , x− , and z ; for instance the rank-4 tensor |s, miµ1 µ2 µ3 µ4 may contain
a term x+ µ1 x− µ2 z µ3 x+ µ4 . In general, the particle’s tensor is a linear combi-
nation of such terms : which precise linear combination it is depends on s
and m, and this is what we want to look into.
21
This implies that the particles are massive. For massless partices, see later on.
482 August 31, 2019

15.12.2 Rank one for spin one

The simplest nontrivial case is that of a rank-1 tensor, that is, a vector. We
have already considered these in Chapter 9. We can define

|1, 1iµ = x+ µ , |1, 0iµ = z µ , |1, −1iµ = −x− µ , (15.165)

so that

h1, 1|µ = x− µ , h1, 0|µ = z µ , h1, −1|µ = −x+ µ . (15.166)

For brevity, we shall use the easily interpretable notation

|1, 1i = |+i , |1, 0i = |0i , |1, −1i = − |−i . (15.167)

These states are properly normalized, since

h1, m1 |1, m2 i = h1, m1 |µ |1, m2 iµ = − δm1 ,m2 . (15.168)

In addition, the states are complete in the sense that

X
|1, λiµ h1, λ|ν = tµ tν − δ µ ν ≡ ∆µ ν . (15.169)
λ=+,−,0

Note that
∆µα ∆αν = −∆µ ν , ∆µ µ = −3 . (15.170)
We now proceed to set up the spin algebra. A general raising operator can
always be written in the form
√
 
S+ = 2~ a |+i h0| + b |0i h−| , (15.171)

where a and b are some complex numbers ; and so

√
 
∗ ∗
S− = 2~ a |0i h+| + b |−i h0| . (15.172)

From
 
2
2 2
S+ S− = −2~ |a| |+i h+| + |b| |0i h0| ,
 
2 2 2
S− S+ = −2~ |a| |0i h0| + |b| |−i h−| , (15.173)
August 31, 2019 483

we find that to get the correct form of Sz we have to take |a| = |b| = 1, since
only then22  
Sz = −~ |+i h+| − |−i h−| ; (15.174)

furthermore, we find automatically

 
2 2
S = −2~ |+i h+| + |0i h0| + |−i h−| , (15.175)

which shows that we have here indeed a spin-one system. For reasons that
will become clear later on we shall choose a = −1 and b = 1. Thus,
√ √
S+ |+i = 0 , S+ |0i = 2~ |+i , S+ |−i = − 2~ |0i . (15.176)

In more explicit tensorial language, we have the following matrix forms :

√
 
µ µ µ
S+ ν = 2~ −x+ zν + z x+ ν ,
√
 
µ µ µ
S− ν = 2~ −z x− ν + x− zν ,
 
µ µ µ
Sz ν = ~ −x+ x− ν + x− x+ ν ,
 
2µ 2 µ µ µ
S ν = −2~ x+ x− ν + x− x+ ν + z zν . (15.177)

15.12.3 Rank-2 tensors

By taking tensor products of vectors we can build more complicated systems.
Let us attempt rank-2 tensors. We can easily construct the spin algebra for
this system as follows :

Σj µν αβ = Sj µ α δ ν β + δ µ α Sj ν β , j = +, −, z , (15.178)

and it is easily checked that these also obey the correct commutation relations

[Σ+ , Σ− ] = 2~ Σz , [Σz , Σ+ ] = ~ Σ+ ; (15.179)

22
Do not be confused with the overall minus signs emerging here ! Remember that the
states are normalized to minus unity. This is a consequence of our dealing with spacelike
objects in an essentially Minkowski space.
484 August 31, 2019

the operator for the total spin is of course

µν µ 2ν µ ν µ ν µ ν
Σ2 αβ = S2 ν µ
α δ β +δ α S β +S+ α S− β +S− α S+ β +2Sz α Sz β . (15.180)

There is precisely one rank-2 tensor with a spin 2~ along the z axis : it is
the tensor product

|2, 2iµν = |1, 1iµ |1, 1iν = x+ µ x+ ν ≡ |++i , (15.181)

with obvious notation. It is straightforward to check that the total spin of

this object is, indeed, equal to 2~. By applying the lowering operator as
given in Eq.(15.178), and normalizing, we can immediately recover the other
states in the spin-2 sector :

|2, 2i = |++i ,
√
 
|2, 1i = |+0i + |0+i / 2 ,
√
 
|2, 0i = − |+−i − |−+i + 2 |00i / 6 ,
√
 
|2, −1i = − |−0i − |0−i / 2 ,

|2, −2i = |−−i . (15.182)

These five objects are totally symmetric. They are also traceless in the sense
that |2, miµν gµν = 0 ; this is due to our choice for the constants a and b made
above. The one object made up from |+0i and |0+i that is orthonormal to
|2, 1i is |+0i − |0+i, which forms the basis of a spin-1 sector :
√
 
|1, 1i = |+0i − |0+i / 2 ,
√
 
|1, 0i = |−+i − |+−i / 2 ,
√
 
|1, −1i = |−0i − |0−i / 2 . (15.183)

Finally, one single state is left :

√
 
|0, 0i = |+−i + |−+i + |00i / 3 , (15.184)
August 31, 2019 485

which upon inspection is seen to have zero spin. The orthonormality of these
nine states is easily checked. Some simple algebra also tells us that
2
X 1 µ 1 1
|2, miµν h2, m|αβ = ∆ α ∆ν β + ∆µ β ∆ν α − ∆µν ∆αβ ,
m=−2
2 2 3
1
X 1 µ 1
|1, miµν h1, m|αβ = ∆ α ∆ν β − ∆µ β ∆ν α ,
m=−1
2 2
1 µν
|0, 0iµν h0, 0|αβ = ∆ ∆αβ , (15.185)
3
so that there is a completeness relation of the form
2
X s
X
|s, miµν hs, m|αβ = ∆µ α ∆ν β . (15.186)
s=0 m=−s

This confirms that no states have been overlooked.

15.12.4 Rank-3 tensors

For the sake of illustration we also give the complete set of rank-3 tenso-
rial states. These fall apart in one spin-3, two spin-2, three spin-1 and
a single spin-0 sector, giving the correct total of 27 possible orthonormal
states, listed below. For reasons of typography I have left out the normaliz-
ing denominators ; these can of course be trivially recovered.

spin-3 :
|3, 3i = |+ + +i
|3, 2i = |+ + 0i + |+0+i + |0 + +i
|3, 1i = 2 |+00i + 2 |0 + 0i + 2 |00+i
− |+ + −i − |+ − +i − |− + +i
|3, 0i = 2 |000i − |+0−i − |0 − +i − |− + 0i
− |−0+i − |+ − 0i − |0 + −i
|3, −1i = |+ − −i + |− + −i + |− − +i
−2 |−00i − 2 |0 − 0i − 2 |00−i
|3, −2i = |− − 0i + |−0−i + |0 − −i
486 August 31, 2019

|3, −3i = − |− − −i
spin-2(1) :
|2, 2i = |+0+i + |0 + +i − 2 |+ + 0i
|2, 1i = 2 |00+i − |+ − +i − |− + +i
− |+00i − |0 + 0i + 2 |+ + −i
|2, 0i = |+0−i + |0 + −i − |−0+i − |0 − +i
|2, −1i = 2 |00−i − |+ − −i − |− + −i
− |0 − 0i − |−00i + 2 |− − +i
|2, −2i = 2 |− − 0i − |0 − −i − |−0−i
spin-2(2) :
|2, 2i = |+0+i − |0 + +i
|2, 1i = |+00i − |0 + 0i − |+ − +i + |− + +i
|2, 0i = − |0 − +i + |−0+i − |+0−i
+ |0 + −i − 2 |+ − 0i + 2 |− + 0i
|2, −1i = |+ − −i − |− + −i + |−00i − |0 − 0i
|2, −2i = |0 − −i − |−0−i
spin-1(1) :
|1, 1i = 6 |+ + −i + 3 |0 + 0i + 3 |+00i
+ |+ − +i + |− + +i − 2 |00+i
|1, 0i = 3 |0 + −i + 3 |+0−i + 3 |−0+i
+3 |0 − +i − 2 |+ − 0i − 2 |− + 0i + 4 |000i
|1, −1i = 2 |00−i − 3 |−00i − 3 |0 − 0i
− |+ − −i − |− + −i − 6 |− − +i
spin-1(2) :
|1, 1i = |+00i − |0 + 0i + |+ − +i − |− + +i
|1, 0i = |0 + −i − |+0−i + |0 − +i − |−0+i
|1, −1i = |+ − −i − |− + −i + |0 − 0i − |−00i
spin-1(3) :
|1, 1i = |+ − +i + |− + +i + |00+i
|1, 0i = |+ − 0i + |− + 0i + |000i
|1, −1i = |+ − −i + |− + −i + |00−i
August 31, 2019 487

spin-0 :
|0, 0i = |+ − 0i + |−0+i + |0 + −i
− |0 − +i − |+0−i − |− + 0i (15.187)

Note that the spin-0 state is totally antisymmetric : obviously, this is the
only possible such state in three space dimensions. We can also compute the
‘partial’ completeness relations pertaining to each spin sector. Some algebra
teaches us that these are the following set of mutually orthogonal projection
operators :

3
X
spin-3 : |3, miµνρ h3, m|αβγ =
m=−3

1
∆µ α ∆ν β ∆ρ γ + ∆µ β ∆ν γ ∆ρ α + ∆µ γ ∆ν α ∆ρ β
6

µ ν ρ µ ν ρ µ ν ρ
+ ∆ β ∆ α∆ γ + ∆ α∆ γ ∆ β + ∆ γ ∆ β ∆ α

1
∆µν ∆ρ α ∆βγ + ∆ρ β ∆γα + ∆ρ γ ∆αβ


−
15
+ ∆νρ ∆µ α ∆βγ + ∆µ β ∆γα + ∆µ γ ∆αβ



ρµ ν ν ν


+ ∆ ∆ α ∆βγ + ∆ β ∆γα + ∆ γ ∆αβ
2
X
spin-2(1) : |2, miµνρ h2, m|αβγ =
m=−2
 
1 µ ν
∆ α ∆ β + ∆ α ∆ β ∆ρ γ ν µ
3
 
1 µ ν ν µ

ρ µ ν ν µ

ρ
− ∆ β ∆ γ + ∆ β ∆ γ ∆ α + ∆ α∆ γ + ∆ α∆ γ ∆ β
6
 
1 µν ρ ρ
+ ∆ ∆ α ∆βγ + ∆ β ∆αγ
6
 
1 µρ ν νρ µ
+ ∆ ∆ γ + ∆ ∆ γ ∆αβ
6
 
1 µρ ν µρ ν νρ µ νρ µ
− ∆ ∆ α ∆βγ + ∆ ∆ β ∆αγ + ∆ ∆ α ∆βγ + ∆ ∆ β ∆αγ
12
488 August 31, 2019

1
− ∆µν ∆ρ γ ∆αβ
3
2
X
spin-2(2) : |2, miµνρ h2, m|αβγ =
m=−2
 
1
∆ α ∆ β − ∆ α ∆ β ∆ρ γ
µ ν ν µ
3
 
1 µ ν ρ µ ν ρ ν µ ρ ν µ ρ
+ ∆ γ ∆ β ∆ α − ∆ γ ∆ α∆ β − ∆ γ ∆ β ∆ α + ∆ γ ∆ α∆ β
6
 
1 µρ ν µρ ν νρ µ νρ µ
+ ∆ ∆ α ∆βγ − ∆ ∆ β ∆αγ − ∆ ∆ α ∆βγ + ∆ ∆ β ∆αγ
4
X 1
spin-1(1) : |1, miµνρ h1, m|αβγ =
m=−1
1 µν ρ
∆ ∆ γ ∆αβ
15  
1 µν ρ ρ
− ∆ ∆ α ∆βγ + ∆ β ∆αγ
10
 
1 µρ ν νρ µ
− ∆ ∆ γ + ∆ ∆ γ ∆αβ
10
 
3 µρ ν µρ ν νρ µ νρ µ
+ ∆ ∆ α ∆βγ + ∆ ∆ β ∆αγ + ∆ ∆ α ∆βγ + ∆ ∆ β ∆αγ
20
X1
spin-1(2) : |1, miµνρ h1, m|αβγ =
m=−1
 
1 µρ ν µρ ν νρ µ νρ µ
− ∆ ∆ α ∆βγ − ∆ ∆ β ∆αγ − ∆ ∆ α ∆βγ + ∆ ∆ β ∆αγ
4
1
X
spin-1(3) : |1, miµνρ h1, m|αβγ =
m=−1
1 µν ρ
∆ ∆ γ ∆αβ
3
spin-0 : |0, 0iµνρ h0, 0|αβγ =

1
∆µ α ∆ν β ∆ρ γ + ∆µ β ∆ν γ ∆ρ α + ∆µ γ ∆ν α ∆ρ β
6
August 31, 2019 489

ν µ ρ ν µ ρ ν µ ρ
−∆ α ∆ β ∆ γ − ∆ β∆ γ∆ α − ∆ γ ∆ α∆ β . (15.188)

The total completeness relations is also valid :

3
X s
X
|s, miµνρ hs, m|αβγ = ∆µ α ∆ν β ∆ρ γ , (15.189)
s=0 m=−s

provided we sum over all sectors with the same s.

15.12.5 Massless particles : surviving states

So far, we have taken our particles to be at rest, with a momentum p for
which
pµ = m tµ .
For moving particles, we can obtain the correct states by simply performing
the appropriate Lorentz boost. As already indicated, we shall take the motion
of the particles to be along the z axis ; our states have been prepared for this
by taking z as the spin quantization axis. The momentum of the particle
will then be
pµ = m tµ → pµ = p0 tµ + |~p| z µ , (15.190)
and the vector z µ becomes, under the same boost

p0
   
µ |~p| µ
z → t + zµ . (15.191)
m m

The vectors x± are not affected by the boost. It is therefore sufficient to

replace, in Eqns.(15.165),(15.182), (15.183),(15.184), and (15.187), z by its
boosted form.

Let us now consider the extreme case : that of a massless particle. We

can view this as the limit p0 /m → ∞ of a massive particle. In that limit,
z µ diverges badly, and we must again adopt the point of view presented in
chapter 9 : the theory wil only be viable if those tensors that diverge
in the massless limit decouple completely. That is, the only observable
states must be those that do not diverge, i.e. those that contain x+ ’s and
490 August 31, 2019

x− ’s but not any trace of a z. A quick inspection in our inventory of states

reveals that only a handful of states are left :

rank-1, spin-1 : |1, 1i = |+i , |1, −1i = − |−i

rank-2, spin-2 : |2, 2i = |++i , |2, −2i = |−−i
√
rank-2, spin-1 : |1, 0i = (|+−i − |−+i)/ 2
rank-3, spin-3 : |3, 3i = |+ + +i , |3, −3i = − |− − −i(15.192)

15.12.6 The Bermuda triangle

With the exception of the rank-2, spin-1 state, the so-called Kalb-Ramond
state, all the surviving states have m = ±s and are totally symmetric. Is this
general ? In other words, how do we know that there is no rank-31, spin-17
state that is built up from only x+ ’s and x− ’s ? We can answer this question
by the following pleasing argument. Since the ladder operators Σ± transform
physical states into one another, any physical state must be an eigenstate√of
Σ+ Σ− or Σ− Σ+ 23 . Disregarding, for simplicity, minus signs and factors 2,
the effect of Σ+ is 0 → +, − → 0, and that of Σ− is + → 0, 0 → −. We can
therefore write

Σ+ Σ− |+−i → Σ+ |0−i → |+−i + |00i . (15.193)

Let us now consider a hypothetical massless-particle candidate state. It will

be a linear combination of kets with lots of +’s and −’s. Among these we
concentrate on three kets in particular :

T1 = |· · · + + − · · ·i , T2 = |· · · + − + · · ·i , T3 = |· · · − + + · · ·i .
(15.194)
The rest of the content of the kets (indicated by the ellipses, and consisting of
some sequences of +’s and −’s) is identical for the three kets. The candidate
state contains these T ’s in some linear combination :

C1 T1 + C2 T2 + C3 T3 + lots of other terms

Let us now consider what happens if we let Σ+ Σ− work on these kets. T1

will turn into a lot of terms, among which we can recognize two important
23
It is of course possible that Σ+ acting on our state, say, will give zero, and then it is
an eigenstate of Σ− Σ+ with eigenvalue zero. We may avoid this trivial case by choosing,
instead, Σ+ Σ− , under which our state will have a nonzero eigenvalue.
August 31, 2019 491

ones :
T1 → |· · · + 00 · · ·i + |· · · 0 + 0 · · ·i + · · · . (15.195)
Similarly, we find for T2 and T3 :

T2 → |· · · 00 + · · ·i + |· · · + 00 · · ·i + · · · ,
T3 → |· · · 00 + · · ·i + |· · · 0 + 0 · · ·i + · · · . (15.196)

We now note a few things. In the first place, a resulting ket like |· · · 0 + 0 · · ·i
can only come from the T ’s (in this case, from T1 and T3 ). In the second
place, our candidate state cannot contain this ket by itself, since it must be
free of 0’s. In the third place, such unwanted kets must drop out because our
state is an eigenstate of Σ2 . We must therefore rely on cancellations between
the T ’s. In fact, we need simultaneously

C1 = −C2 , C2 = −C3 , C3 = −C1 . (15.197)

Obviously, C1,2,3 = 0 : our three T ’s do not occur at all24 ! But of course

we can repeat the same argument for any other such three kets. We see that
the only possibilities to have admissible massless-particle states are twofold:

• Only +’s, or only −’s, occur. These are precisely the rank-s, spin-s
states such as we have found, and this persists also for s > 3. Note that
these states are totally symmetric — not for some deep field-theoretical
reason, but because they can’t help it.

• Precisely one + and one − occur. This is the Kalb-Ramond state,

which now stands revealed as a lone exception.

15.12.7 Massless propagators

For massless states, the spin sums cannot be built up from objects like ∆µ α
since these diverge. An often-used recipe is the following. For a massless
particle of momentum pµ , define

pµ = (p0 , p~) , p̄µ = (p0 , −~p) . (15.198)

24
A three-cornered argument such as this, in which all T ’s disappear, deserves to be
called a Bermuda triangle.
492 August 31, 2019

Obviously, this is not a Lorentz-invariant definition, but as we shall see that

is not a problem. The point is that a p̄ can be found whatever the Lorentz
frame is. We can now write
|+iµ h+|ν + |−iµ h−|ν = x+ µ x− ν + x− µ x+ ν
 
1
= p p̄ν + p̄ pν − δ µ ν ≡ ∇µ ν .
µ µ
(15.199)
p · p̄
In analogy to Eq.(15.170) we now have
∇µα ∇αν = −∇µ ν , ∇µ µ = −2 . (15.200)
If, as we must promise ourselves, massless states only couple to conserved
sources (on which the handlebar operation gives zero), the terms containing
p̄ will always drop out. We can now write the spin sums for the surviving
massless states as follows :
rank-1, spin-1 : ∇µ α ,
 
1 1
rank-2, spin-2 : ∇ α ∇ β + ∇ β ∇ α − ∇µν ∇αβ ,
µ ν µ ν
2 2
 
1
rank-2, spin-1 : ∇µ α ∇ν β − ∇µ β ∇ν α ,
2

1
rank-3, spin-3 : ∇µ α ∇ν β ∇ρ γ + ∇µ β ∇ν γ ∇ρ α + ∇µ γ ∇ν α ∇ρ β
6

µ ν ρ µ ν ρ µ ν ρ
+ ∇ β ∇ α∇ γ + ∇ α∇ γ ∇ β + ∇ γ ∇ β ∇ α

1
∇µν ∇ρ α ∇βγ + ∇ρ β ∇γα + ∇ρ γ ∇αβ


−
12
+ ∇νρ ∇µ α ∇βγ + ∇µ β ∇γα + ∇µ γ ∇αβ



ρµ ν ν ν


+ ∇ ∇ α ∇βγ + ∇ β ∇γα + ∇ γ ∇αβ (15.201)

Compared to the massive case, some coefficients are different : -1/2 rather
than -1/3 in the spin-2 case, and -1/12 instead of -1/15 for spin-3. This is
due, of course, to the different traces of ∆ and ∇. The spin sum for the
massless vector particle (rank-1, spin-1) is in fact that of the axial gauge
discussed in Chapter 9, with the gauge vector r chosen to be p̄. Note that,
whatever rµ , we can always move to the centre-of-mass frame of pµ and rµ ,
and in that frame we have precisely rµ = p̄µ .
August 31, 2019 493

15.12.8 Spin of the Kalb-Ramond state

Concerning the Kalb-Ramond (KR) state, there may be some controversy.
For a massless particle in this state, the spin along the axis of motion must,
under measurement, always come out zero. It is not easy to see how such a
particle can be distinguished from a scalar one. Indeed, in string theory where
the KR state comes up naturally, it is considered to describe a (pseudo)scalar
particle called the axion. In order to talk sensibly about the spin of the KR
state it is useful to consider how it may be measured, for instance using
fermions. We therefore consider the coupling of a rank-2, spin-1 state to
fermions. The interaction vertex must have the properties that (a) it is an
antisymmetric rank-2 tensor, and (b) it is current-conserving, in order to
make sense in the massless limit. Denoting the two fermions by ψ and ψ the
simplest choice appears to be

ψ µνρσ pρ (A + Bγ 5 ) γσ ψ

where p is the momentum of the antisymmetric tensor state, and A and B

are constants. This interaction vertex vanishes trivially under the handlebar
operation. For the process

f¯(p1 ) f (p2 ) → f (p3 ) f¯(p4 )

by the exchange of a KR state of mass M , we then have the amplitude

M = i~ v(p1 ) µνρσ pρ (A + Bγ 5 ) γσ u(p2 )

∆µα ∆νβ − ∆µβ ∆να
×
2(s − M 2 )
× u(p3 ) αβκλ pκ (A0 + B 0 γ 5 ) γλ v(p4 ) ,
s = p · p , p = p1 + p2 = p 3 + p4 . (15.202)

Because of the current conservation and the antisymmetry of the vertices,

we may replace ∆µα ∆νβ − ∆µβ ∆να by 2gµα gνβ . Furthermore, since

µνρσ pρ µν κλ pκ = 2 pσ pλ − s g σλ


(15.203)

we have
1
M = −2i~
s − M2
494 August 31, 2019

v(p1 ) A(m2 − m1 ) + B(m1 + m2 )γ 5 u(p2 )


0 0 5


× u(p3 ) A (m3 − m4 ) − B (m3 + m4 )γ v(p4 )

−s v(p1 ) A + Bγ 5 γ µ u(p2 )


0 0 5


× u(p3 ) A + B γ γµ v(p4 ) . (15.204)

Here mj is the mass of momentum pj . Note that, in contrast to e.g. the

case of QED, m1 = m2 or m3 = m4 is not necessary for current conservation.
We can now investigate several situations. In the first place, if M 6= 0 the
amplitude has a pole for some nonzero s value, which we may take as the
signal of a particle. The second term in brackets in Eq.(15.204) then tells us
that, indeed, a spin-1 particle has been exchanged25 . The occurrence of the
first term is, then, not surprising : a similar contribution is found in e.g. the
W exchange in muon decay. Secondly, we may take M = 0. In that case, the
second term no longer has a pole. It can therefore not survive a truncation
argument, and must not be counted as coming from any particle propagation.
The first term does survive ; if we also assume flavour conservation so that
m1 = m2 and m3 = m4 , the only degree of freedom that propagates is,
indeed, that of a pseudoscalar.

15.12.9 Spin 1 from Dirac particles

We can, of course, also use Dirac spinors to build spinning states. As before,
m is the particle mass, and pµ the particle momentum ; there is no particular
advantage to take it at rest here. The spin vector is sµ and p2 = m2 , s2 = −1,
(ps) = 0. We shall use the standard representation as in sec.8.1.5, and use
1
k0µ = pµ − sµ (15.205)
m
so that (pk0 ) = m. Thus,
1 1
u± = u(p, ±s) = √ (/p+m)u∓ (k0 ) , v± = v(p, ±s) = √ (/p−m)u± (k0 ) .
2m 2m
(15.206)
25
We can measure this, for instance by looking at the angular distribution of the pro-
duced fermion-antifermion pair ; see also Appendix ??.
August 31, 2019 495

We can easily check that γ 5 u+ = v− and γ 5 u− = −v+ , which is handy since

it allows us to write everything using only the u’s if we want. We choose to
work in a Lorentz frame where p~ k ~s : then these spinors are pure helicity
states. Let us now consider the object
1
eµ+ = √ v + γ µ u+ . (15.207)
8m2
We see immediately that
e+ · p ∝ v + p/u+ ∝ v + u+ = 0 , e+ · s ∝ v + s/u+ = v + γ 5 u+ = v + u− = 0 .
(15.208)
The Chisholm and reversal identities for e+ are
e/+ = v + γ µ u+ gµ = 2 (u+ v + − v− u− ) , eµ+ = −u− γ µ v− .


(15.209)
From the first of these we find immediately
e+ · e+ = 0 , e+ · e∗+ = −1 . (15.210)
Therefore, eµ+ is an object with spin 1 in the direction of p~. We can now
lower the spin, using the lowering operator u+ → u− , v + → v − . This gives
us the spin-0 part if the triplet :
1
eµ0 = (v − γ µ u+ + v + γ µ u− )
4m
1
−u+ γ 5 γ µ u+ + u− γ 5 γ µ u− = sµ ,


= (15.211)
4m
exactly what you would expect for the spin-0 part. A final lowering gives us
1
eµ− = √ v − γ µ u− = − (eµ+ )∗ , (15.212)
8m2
which yields the part with spin minus one. The spin sum is therefore
X µ 1
ej (eνj )∗ = −g µν + 2 pµ pν . (15.213)
j=+,0,−
m

So now we have constructed three spin-1 states using two spin-1/2 states.
There ought to be another, spin-0 state as well. Indeed there is, somewhat
trivially : it is the state
−1 1
v − γ 5 γ µ u+ = u+ γ µ u+ = pµ , (15.214)
2m 2m
496 August 31, 2019

the only object with one Lorentz index orthogonal to the eµ ’s. That it has
zero spin can be seen by lowering26 which turns v − γ 5 γ µ u+ into v − γ 5 γ µ u− =
−u+ γ µ u− = 0.

15.12.10 Spin 3/2 particles

Adding one more Dirac spinor we can describe spin-3/2 particles. These
states have one Dirac index and one Lorentz index. The highest spin resides
in the state
1
uµ3/2 = √ u+ v + γ µ u+ = u+ eµ+ . (15.215)
8m 2

That this state has the properties of a spin-1/2 as well as of a spin-1 particle
can be seen from
(/p − m) uµ3/2 = 0 , pµ uµ3/2 = 0 . (15.216)

There is one more property : again using Eq.(15.209) we see that also
γµ uµ3/2 = 0 . (15.217)

The properties (15.216) and (15.217) are the so-called Rarita-Schwinger con-
ditions. Now, repeated application of the lowering operator (and normaliza-
tion) give us, successively :
1
uµ1/2 = √ u− v + γ µ u+ + u+ v − γ µ u+ + u+ v + γ µ u−


24m2
1  √ 
= √ u− eµ+ + 2u+ eµ0 ,
3
1
uµ−1/2 = u− v − γ µ u+ u− v + γ µ u− + u+ v − γ µ u−


√
24m2
1  √ 
= √ u+ eµ− + 2u− eµ0 ,
3
1
uµ−3/2 = √ u− v − γ µ u− = u− eµ− . (15.218)
8m 2

These three states also obey the Rarita-Schwinger conditions. Getting the
spin sum is now less trivial than before. It helps to introduce the shorthand
[αρβστ ] = uα uα γ 5 γ ρ uβ uβ γ 5 γ σ uτ uτ , α, β, τ = ±
26
And, conversely, also by raising.
August 31, 2019 497

We can then write

3/2
X 1
uµj uνj =


[+ν − µ+] + [−ν + µ−]
8m2
j=−3/2
1
+ [−µ + ν−] + [+µ − ν+]
24m2
− 2[+µ + ν−] − 2[−µ − ν+]
− 2[+µ − ν−] − 2[−µ + ν+]


+4[+µ + ν+] + 4[−µ − ν−] . (15.219)

With some careful algebra we can turn this into

3/2  
X 1 µ ν
uµj uνj µν
= (/p + m) −g + 2 p p
m
j=−3/2
1
− 2
(/p + m)γ 5 γ µ (/p + m)γ 5 γ ν (/p + m) . (15.220)
12m
This is an interesting object ! Each of the two terms obeys the first two
Rarita-Schwinger conditions separately, but for the third one (with the γµ )
we need a special combination of the two. In fact, Eq.(15.220) is the only
combination with up to three p’s that obeys the Rarita-Schwinger conditions
‘from the left’ and ‘from the right’. The first term is what you would probably
write down as your first guess for ‘a product of a spin-1/2 and a spin-1 state’,
but that has 2 × 3 degrees of freedom : the second term kills off the two
unwanted degrees of freedom. And yes, our three-Dirac world contains of
course also a spin-1/2 sector : the two states

µ 1 √ µ µ
 1 √
µ µ µ

w1/2 =√ 2 u− e+ − u+ e0 , =√ 2 u+ e− − u− e0 ,
w−1/2
3 3
(15.221)
µ
are orthogonal to the four u ’s. Now, note that
√ 1
2 u− eµ+ − u+ eµ0 = (u− v + γ µ u+ − u+ v − γ µ u+ )
2m
1
u− u− γ 5 γ µ u+ + u+ u+ γ 5 γ µ u+


=
2m
1
= (/p + m)γ 5 γ µ u+ , (15.222)
2m
498 August 31, 2019

so, indeed
X 1
wjµ wνj = (/p + m)γ 5 γ µ (/p + m)γ 5 γ ν (/p + m) . (15.223)
12m2
j=±1/2

The total spin sum over both sectors is therefore

3/2 1/2  
X X 1 µ ν
uµj uνj + wjµ wνj µν
= (/p + m) −g + 2 p p . (15.224)
m
j=−3/2 j=−1/2

Like before, we should be careful in the massless limit, in which sµ ∼ pµ

blows up. Any of the Dirac-built states that contains e0 must decouple in
that limit ; on the other hand, the spinors u± and v± themselves remain
perfectly finite, as discussed in sec.7.4.6. We again conclude that massless
spin-3/2 particles can only be seen in the two states of maximal helicity, uµ3/2
and uµ−3/2 .

15.12.11 The spin algebra for Dirac particles

In the above we have simply stated that the spinors u+ and v + correspond
to positive helicity, and u+ andv − to begative helicity ; but we ought to be a
bit more formal, by construviting the correct algebra. The lowering operator
in this algebra, M− , must be such that

M− u+ = ~ u− , v + M− = ~ v − . (15.225)

This implies that, here, the lowering and raising operators must be
~ ~
M− = (u− u+ − v+ v − ) , M+ = M − = (u+ u− − v− v + ) .
2m 2m
(15.226)
The spin operator in the s direction is then
1 ~
Ms = (M+ M− − M− M+ ) = p/ γ 5 s/ . (15.227)
2~ 2m
Since Ms M+ = −M+ Ms = −~M+ /2 we also have [Ms , M+ ] = ~M+ as we
ought to. And the operator for total spin, finally, is
1 3~2
M2 = (M+ M− + M− M+ ) + Ms 2 = . (15.228)
2 4
August 31, 2019 499

So everything appears to be fine. However, looking at the spin of the various

spinors, we find
~ ~
Ms u± = ± u± , Ms v± = ∓ v± , (15.229)
2 2
so the spinor v± would appear to have the ‘wrong’ spin component. However,
we employ v rather than v itself, and the transition v → v involves time
reversal27 . Under time reversal, axial vectors like sµ change sign. Therefore
the apparent problem disappears if you realise what the spin means for these
particles. At any rate, the fact that eµ+ , say, is an honest spin-1 object justifies
our spin assignments.

15.13 Generating three-particle kinematics

Here we describe how to do Monte Carlo surveys of three-particle phase space,
as used in section 13.4.5. The process is that of two incoming particles with
momenta pµ1,2 and masses M1,2 , and three outgoing particles with momenta
µ
q1,2,3 and masses m1,2,3 , respectively. This is more general than the discussion
of three-body phase space in muon decay (section 8.3.2). The total energy is
w, where w2 = (p1 + p2 )2 . We shall consistently work in the centre-of-mass
frame, where p~1 + p~2 = ~q1 + ~q2 + ~q3 = 0. In terms of two of the energies, q10
and q20 , say, the phase space integration element is still constant : rather it is
the boundaries that become very nontrivial. From (p1 + p2 − q3 )2 = (q1 + q2 )2
we find

w2 − 2wq30 + m23 = 2w(q10 + q20 ) − w2 + m23

= m21 + m22 − 2q10 q20 + 2|~q1 ||~q2 | cos θ , (15.230)

where θ is the angle between ~q1 and ~q2 . The allowed phase space points
therefore obey
 
2w(q10 + q20 ) − w2 + 2q10 q20 + m23 − m21 − m22  ≤ 2|~q1 ||~q2 | . (15.231)

The actual boundary, at which the inequality (15.231) becomes an equality, is

0
of multinomial type in q1,2 with powers up to (q10 q20 )2 , hence quite horrible28 .
27
Since v describes an outgoing antifermion, and v an incoming antifermion.
28
The actual volume of phase space is only known for some special cases.
500 August 31, 2019

What we do know are the lower and upper limits on the energies :
1
mj ≤ qj0 ≤ w2 + m2j − (mk + m` )2 = qjmax , {j, k, `} = {1, 2, 3} .


2w
(15.232)
Armed with this knowledge we can use (pseudo-)random numbers ρ, uni-
formly and independently distributed in the interval [0, 1], to
sample the
final-state momenta29 . First, we repeat qj0 ← mj + ρ qjmax − mj for j = 1, 2
until Eq.(15.231) is satisfied : typically, one has to try only a few times. This
gives us |~q1 | and |~q2 |, and cos(θ) from Eq.(15.230). We then put

~q1 ← (0, 0, |~q1 |) , ~q2 ← (0, |~q2 | sin(θ), |~q2 | cos(θ)) , ~q3 ← −~q1 − ~q2 .

This gives us the final-state momenta, sampled uniformly. For the initial-
state momenta we generate a uniformly distributed solid angle by

cos(η) ← −1 + 2ρ , φ ← 2π ρ

and the two energies by

1
p01 = w2 + M12 − M22 p02 = w − p01


,
2w
and this allows us to construct the incoming momenta :

p~1 ← (|~p1 | sin(η) sin(φ), |~p1 | sin(η) cos(φ), |~p1 | cos(η)) , p~2 ← −~p1

The fact that ~q1 always points in the positive z direction is not a problem in
any theory that respects angular momentum conservation, since the incoming
momenta will take on all directions. You might want to have the direction
of p~1 fixed instead : in that case a global rotation on all the momenta will
do the trick.

15.14 The CPT theorem

In this appendix, we shall discuss the very fundamental CPT theorem30 for
theories with interacting particles. This theorem deals with what happens (or
29
Everytime we make a reference to ρ it is to be understood that a new random number
is picked. Therefore, a = ρ and b = ρ does not imply a = b !
30
Also known as the CTP theorem, the TCP theorem, the TPC theorem, the PTC
theorem, or the PCT theorem.
August 31, 2019 501

ought to happen) to scattering amplitudes when we relate various physical

scattering processes31 . As usual, we shall start by looking at Dirac particles.

15.14.1 Transforming spinors

In Chapter 7, we defined the standard form for the various spinors corre-
sponding to an on-shell (anti)-particle with mass m and momentum pµ . We
recapitulate them here :

u± (p) = N (p) (/p + m) u∓ (k0 ) ,

v± (p) = N (p) (/p − m) u∓ (k0 ) ,
u+ (k0 ) = k/1 u− (k0 ) , u− (k0 )u− (k0 ) = ω− k/0 ,
p
N (p) = 1/ 2(pk0 ) , k0 2 = (k0 k1 ) = 0 , k1 2 = −1 . (15.233)

This is, of course, only a phase convention, where the phase choice is not
explicit but implied by the choice of k0 , k1 and the complex phase of u− (k0 ).
Now, let us apply γ 5 to these states. It is easy to see that

γ 5 u+ (p) = v+ (p) , γ 5 u− (p) = − v− (p) ,

u+ (p) γ 5 = −v + (p) , u− (p) γ 5 = v − (p) . (15.234)

In words, what this transformation does is to change an incoming, right(left)-

handed fermion into an outgoing, left(right)-handed antifermion (and vice
versa). Thus we have (a) the interchange of particle and anti-particle (charge
conjugation, C), (b) the interchange of right- and left-handedness32 (parity
inversion, P), and (c) the interchange of initial and final state (time reversal
T), which goes by the name of CPT transformation33 . Applied to Feyn-
man diagrams, we can depict this as follows (where we have indicated the
31
Recall that, in these notes, we concentrate on the (perturbative)processes that are
going on, that is, scattering described by diagrams and amplitudes.
32
Recall that for a particle + means right-handed, but for an antiparticle it means
left-handed (cf section 8.1.5)
33
There is a slight subtlety here. An ingoing particle with three-momentum p~ is trans-
formed into an outgoing antiparticle with the same momentum p~. Under P, momenta are
inverted so that p~ becomes −~ p : but under T the velocities are again inverted. The same
holds, of course, for spin vectors. It is only the fact that ”+” means right-handed for
particles and left-handed for antiparticles that ensures that the net result is just a change
of handedness.
502 August 31, 2019

helicity) :

+ → + , − → − −

+ → − + , − → − . (15.235)

As we can see, the effect of CPT on any diagram is not only to interchange
initial and final states, but also to reverse the arrows on intermediate fermion
lines.

15.14.2 CPT transformation on sandwiches

Let us consider the scalar current for two on-shell momenta p1,2 , with respec-
tive masses m1,2 :
Jλ1 λ2 = uλ1 (p1 )uλ2 (p2 ) . (15.236)
Under CPT, this scalar current behaves as follows :

J++ → Jˆ++ = −v + (p1 )v+ (p2 ) ,

J+− → Jˆ+− = v + (p1 )v− (p2 ) . (15.237)

At first sight, these CPT transforms look nothing like the original. Note,
however, that using the standard form we can write them as traces :

J++ = N (p1 )N (p2 ) Tr (ω− k/0 (/p1 + m1 )(/p2 + m2 )) ,

J+− = N (p1 )N (p2 ) Tr (ω− k/0 (/p1 + m1 )(/p2 + m2 )/k1 ) , (15.238)

whereas

Jˆ++ = − N (p1 )N (p2 ) Tr (ω− k/0 (/p1 − m1 )(/p2 − m2 )) ,

Jˆ+− = N (p1 )N (p2 ) Tr (ω− k/0 (/p1 − m1 )(/p2 − m2 )/k1 ) , (15.239)

Keeping track of which terms in these traces actually survive34 , we see that,
appearances notwithstanding,

Jˆλ1 λ2 = Jλ1 λ2 . (15.240)

34
For J±± , these are the terms that contain an odd number of masses, for J±∓ those
with even numbers of masses survive.
August 31, 2019 503

Similar (almost trivial) trace arguments show that, under CPT,

Jλ1 λ2 µ = uλ1 (p1 ) γ µ uλ2 (p2 ) → − Jλ1 λ2 µ ,

Jλ1 λ2 µν = uλ1 (p1 ) γ µ γ ν uλ2 (p2 ) → + Jλ1 λ2 µν ,
Jλ1 λ2 µνα = uλ1 (p1 ) γ µ γ ν γ α uλ2 (p2 ) → − Jλ1 λ2 µνα , (15.241)

and so on.

15.14.3 CPT transformation on diagrams

Consider a nontrivial but very simple diagram, for simplicity taken from the
electroweak process e− (p1 )γ(q1 ) → e− (p2 )Z 0 (q2 ) :

D= (15.242)

Leaving out overall constants and denominators, this can be written as

D = Aµν ¯µλZ (q2 ) λγ ν (q1 ) ,

Aµν = uλ2 (p2 ) ωγµ (/q + m) γν uλ1 (p1 ) ,
q = p1 + q1 = p2 + q2 , ω = gv + ga γ 5 , (15.243)

where we have indicated the handedness (helicity) of the external particles.

For the polarisation vectors we take the representation given in Eq.(9.32),
and for q/ we may, if we wish, use Eq.(??) to write
1
q/ = γα u+ (q)γ α u+ (q) . (15.244)
2
Let us now see what happens if we apply CPT. In the first place,

q/ → − q/ , (15.245)

following immediately from Eq.(15.244)35 Therefore, Aµν transforms as

Aµν → −λ1 λ2 v λ1 (p2 ) ωγµ (−/q + m) γν vλ2 (p1 ) . (15.246)

35
Antoher approach might be to find a set of timelike, positive-energy P momentaα
k1,2,3,... with masses m 1,2,3,... , and a set of constants c1,2,3,... such that j cj kj =
α
P
q and Pj cj mj = m. Obviously, this is always possible. We can then write
qP+ m = j cj (u+ (kj )u+ (kj ) + u− (kj )u− (kj )), which under CPT are transformed into
/
j cj (−v+ (kj )v + (kj ) − v− (kj )v − (kj )) = −/q + m.
504 August 31, 2019

The arguments given in the previous section show that this evaluates again
to Aµν itself. Finally, for the polarisation vectors we have, for instance,

λγ ν → − λγ ν = − ¯ν−λγ , (15.247)

so that the CPT transform of an incoming, left(right)-handed photon can be

interpreted as that of an outgoing, right(left)-handed photon with the same
momentum, up to an overall minus sign. The same goes of course for λZ µ .
We see that, under CPT, the amplitude M remains unchanged36 : but the
interpretation is now that of the process e+ (p2 )Z 0 (q2 ) → e+ (p1 )γ(q1 ), with
the understanding that left(right)-handed particles have been replaced by
right(left)-handed ones. The corresponding Feynman diagram is now

DCPT = , (15.248)

which may help you to understand the replacing of q/ by −/q : in diagrammatic

terms, it comes from the fact that now q runs against the propagator’s arrow.
It is now easy to see that we can perform similar operations on every
conceivable Feynman diagram in our theory37 , and we shall always find that
it transforms into itself. We say that our theory is CPT-invariant : if we
(a) replace every external particle by its antiparticle (and vice versa), (b)
interchange the initial and final states, and (c) interchange right- and left-
handed, then all amplitudes remain the same. This is the CPT theorem.

15.14.4 How to kill CPT, and what it costs

Like all such theorems, the CPT theorem can only be valid under a number
of circumstances. Here, we mention the most important of these.
In the first place, comparing the diagrams (15.242) and (15.248) we see
that we have implicitly assumed that the vertices of the theory are insensitive
to what is the ‘incoming’, and what the ‘outgoing’ particle : for instance, the
36
You might think that the fact that the two minus signs coming from the polarisation
vector cancel so nicely is suspicious : but you should realize that if three external bosons
were involved there would be two internal fermion propagators instead of one.
37
If push comes to shove, we can always write every vector quantity in the diagram
with spinors : we then end up with a massively complicated object containing loads of
(anti)spinors and their conjugates, but for the rest only fixed numbers or matrices ; for
such structures, we have already proven everything that is needed.
August 31, 2019 505

two vertices µ µ

_ and
e e+
µ
are both assigned the value iQγ /~. More poignantly, in the electroweak
sector we use the same vertex for
+ _
W W

and D
U D U

It is, of course, possible to let the vertex depend on the ‘orientation’ of

the (sub)process : such theories, which as we see are not easily expressed di-
agrammatically38 , are called non-Hermitian. A non-Hermitian action would
ruin CPT.
In the second place, and more subtly, we have assumed that there is,
at least, the very possibility of a vacuum state through which particles can
move ; in the literature, this means that there is a state with lowest energy.
If the spectrum of the theory is not bounded from below, CPT is ruined :
but, again, it is not easy to see how any ordinary particle physics could be
alive under such circumstances39 , whether CPT invariant or not.
In the last place, there is the issue of Lorentz invariance. We have as-
sumed that every vector hµ will, under CPT, turn into −hµ , and this is very
important for proving the CPT invariance of amplitudes. Suppose, now, that
we introduce into our theory a fixed vector40 f µ , simply a set of four uni-
versally defined41 numbers which enter nontrivially into the Feynman rules.
Such a vector would, under CPT, not turn into its opposite ; but neither
would it change under Lorentz transformations, it would simply remain f µ .
CPT would be ruined together with Lorentz invariance. A theory violating
CPT will therefore manifest itself in being Lorentz-noninvariant. You might
hope to avoid this by having, built into the fabric of the universe, some
physically meaningful vector quantity f µ , that does change with Lorentz
transformations42 . Still, CPT would be ruined, but we must also conclude
38
At least in the way we have formulated things.
39
In these notes, we take the existence of particles with a perturbative description for
granted.
40
We speak of a ‘vector’ here in the sense that it has four components, not in the sense of
its behavior under coordinate transformations : indeed, the whole point is that it doesn’t
transform at all.
41
Think of having some inspiration, or a voice from heaven engraving these numbers on
stone tablets.
42
Such a thing would be, for instance, the ‘momentum of the æther’.
506 August 31, 2019

that the ‘vacuum’ state is itself simply not Lorentz invariant since there is a
‘preferred momentum’.

Note that it is, in principle, possible to violate Lorentz invariance without

destroying CPT. For instance we can use a fixed ‘tensor’ f µν rather than a
vector f µ . Such a tensor does not change sign under CPT, exactly as it
should. We can then construct theories where Lorentz invariance is violated
but CPT invariance is not43 .
We see that the conditions under which CPT symmetry holds are very
plausible and general, but they are not unavoidable. CPT may be ruined,
but we can see that by the concomitant violation of Lorentz invariance, either
in the interactions of the theory or in the structure of the vacuum itself !

15.15 Mathematical Miscellanies

In this section some collected mathematical issues are discussed which are
usueful in the main text, or maybe just of some interest.

15.15.1 The Gaussian doubling trick

The Gaussian integral
Z∞
dx exp −x2


G= (15.249)
−∞

is not easily computed in the standard manner. However, there is a ‘once

seen, never forgotten’ way of doing it, by doubling the integral and going
over to polar variables :

Z∞ Z2π Z∞
2 2 2
dr r exp −r2




G = dx dy exp − x + y = dφ
−∞ 0 0
Z∞ Z∞
dr r exp −r2


= 2π = π ds exp(−s) = π . (15.250)
0 0

43
As an example, we can use, for the kinetic part of a Lagrangian, the object f µν ∂µ ϕ ∂ν ϕ
rather than the usual g µν ∂µ ϕ ∂ν ϕ.
August 31, 2019 507

Thus we find
x2 √ x2 √
Z   Z  
dx exp − 2 = σ 2π , dx exp ±i 2 = σ π(1 ± i) ,
2σ 2σ
(15.251)
44
where the last result is obtained by analytic continuation .

15.15.2 Stirling’s approximation for n!

The following discussion will bring us right back to the start of these notes.
The factorial function is defined by the integral
Z∞
n! ≡ dz z n e−z (n = 0, 1, 2, 3, . . .) (15.252)
0

We can write the integrand as

z n e−z = exp (n log(n) − n + n log (1 − ϕ) + nϕ) , (15.253)
where z = n(1 − ϕ). We recognise a path-integral-like expression :
Z1
n! = nn+1 e−n dϕ e−S(ϕ) ,
−∞
n 2 n 3 n 4 n 5
S(ϕ) = ϕ + ϕ + ϕ + ϕ + ··· , (15.254)
2 3 4 5
that is, a zero-dimensional theory with µ = n and λk = −n (k − 1)!. We find
the following asymptotic result :
!
√ X D`+1
n! ≈ 2π nn+1/2 e−n 1 + , (15.255)
`≥1
n`

where D`+1 /n` is the sum of the (` + 1)-loop vacuum diagrams for the above
action, including the disconnected ones. We can make life simpler by concen-
trating only on the sums of the connected vacuum diagrams, C`+1 /n` , and
then we write
1 X C`+1
log(n!) ≈ (n + 1/2) log(n) − n + log(2π) + `
. (15.256)
2 `≥1
n
44
And inspection of cos(x2 ) and sin(x2 ) which shows that these both integrate to positive
numbers.
508 August 31, 2019

Curiously, Ck vanishes for all odd values of k ! This is nontrivial since,

for instance, the C3 terms arise from 15 three-loop diagrams. The first few
nonzero terms are
X C`+1 1 1 1 1 1
`
≈ − 3
+ 5
− 7
+
`≥1
n 12 n 360 n 1260 n 1680 n 1188 n9
691 1 3617
− 11
+ 13
− + · · · (15.257)
360360 n 156 n 122400 n15
The terms alternate in sign, which is quite usual for a summable asymptotic
expansion of this type. It is worth noting that the action (15.254) is actually
nothing but the ‘wonderful action’ of exercise 13. There we prove that the
Green’s functions are completely free of loop corrections beyond the first
order. In this section, however, we deal with the vacuum diagrams ; there is
therefore no contradiction. Of course no-one has actually computed C16 , say,
by evaluating all 16-loop diagrams. Rather we can take the more primitive
perturbative idea, by expanding the exponent of the interactions terms in
powers of ϕ, dropping the odd powers of ϕ, and substituting for ϕp the
Gaussian result (p − 1)!!/np/2 . The above discussion is rather intended to
show how the asymptotic part of the Stirling approximation is, in fact, closely
related to Feynman diagrammatics.

15.15.3 Manipulating asymptotic series

Comparing asymptotic series
In spite of the problematic convergence behaviour of asymptotic sums, we can
still obtain some surprising insights. This is the field of resurgence theory,
to which we do not full justice here ; but a relatively crude and simple
approach is enough for our
P purpose. Let us consider asymptotic sums, f (x) =
n n
P
n≥0 x f n and g(x) = n≥0 x gn , and so on. We shall take f0 = g0 = 1. The
coefficients fn and gn are growing superexponentially with n : in particular,
we shall assume that fn behaves, for very large n, as
fn ≈ af cnf Γ(n + βf ) (n → ∞) , (15.258)
and similar for g(x) and so on. Let us define some comparison between the
sums : we shall say that
fn fn
f (x) ≺ g(x) if lim =0 , f (x) ∼ g(x) if lim =1 , (15.259)
n→∞ gn n→∞ gn
August 31, 2019 509

and f (x)  g(x) if g(x) ≺ f (x). Obviously, f (x)+g(x) ∼ f (x) if f (x)  g(x).
Let us now consider the product f (x)g(x) :
X X n
f (x)g(x) = xn fk gn−k . (15.260)
n≥0 k=0

Because the coefficients grow so fast, the sum over k is (for n → ∞) com-
pletely dominated by gn + fn , and we conclude that
f (x)g(x) ∼ g(x) + f (x) . (15.261)
For instance, we find 1/f (x) ∼ −f (x). Also, for finite p, f (x)p ∼ pf (x).

Asymptotic pullbacks
Consider two asymptotic sums f (x) and g(x) as above, and assume that
f (x)  g(x). We want to study the asymptotic form of the pullback f (x g(x)),
that is, an asymptotic sum of powers of an asymptotic sum ; at first sight
this would seem hopelessly tremendous. However, we can write45
X
f (x(g(x)) = fn xn g(x)n
n≥0
  2 
X
n n 2
= x fn + fn−1 (ng1 ) + fn−2 g1 + ng2
n≥0
2!
 3  
n 3 2
+fn−3 g1 + n g1 g2 + ng3 + · · · (15.262)
3!
Discarding the 1/n corrections, the leading behaviour of this sum is
n
X X nk
f (xg(x)) ∼ xn fn−k g1 k , (15.263)
n k=0
k!

and, for very large n, we have fn−k nk ≈ af cfn−k Γ(n + βf ), so that

n  k
X
n
X 1 g1
f (x(g(x)) ∼ fn x ∼ f (x) exp(g1 /cf ) . (15.264)
n k=0
k! c f

Here, it is important that f  g to ensure that only the powers of g1 con-

tribute in the asymptotic regime, and that the sum over k is dominated by
only the lower k values.
45
Here we introduce the falling powers : ns ≡ n(n − 1) · · · (n − s + 1) = n!/(n − s)!. For
asymptotically large n, we can approximate ns ≈ ns up to 1/n corrections.
510 August 31, 2019

Asymptotic inversion
The equation xf (x) = y can (at least formally) be inverted to give x as a
function of y. We can write
X fn+1
xf (x) = y ⇒ x = y+x(1−f (x)) = y−f1 x2 h(x) , h(x) = xn .
n≥0
f1
(15.265)
As explained in appendix 15.15.12, the Lagrange expansion is basically just
the iteration of this equation, and we can (formally) write
X 1  d n−1
n
x = y+ −f1 y 2 h(y)
n≥1
n! dy
 n−1 X
X 1 n d fm+1 m+2n
∼ (−f1 ) y
n≥1
(n − 1)! dy m≥0
f 1
XX 1
∼ − (−f1 )n−1 fm+1 mn−1 y m+n+1
n≥1 m≥0
(n − 1)!
k  n−1
XX
k 1 f1
∼ −y fk y −
k≥0 n=1
(n − 1)! cf
∼ −yf (y) exp(−f1 /cf ) . (15.266)

That is, y = xf (x) = y(x) and x = x(y) are asymptotically equal up to an

overall factor. Here it is important to note that for large k = m + n the sum
over n is dominated by the smaller values of n (as we can see), so that the
approximation h(x)n ∼ nh(x) is justified.

A case study: ϕ4 theory in zero dimensions

As an example we will study the ‘renormalization improvement’ discussed in
section 2.4. The zero-dimensional path integral reads, after the perturbation
expansion (with λ = u, and putting m = ~ = 1 for simplicity),
Z   X
2n 1 2 u 4 u k (4k + 2n)!
H2n = dϕ ϕ exp − ϕ − ϕ = −
2 24 k≥0
24 k!(2k + n)!22k+n
X  2u k 4n
∼ − Γ(k + n) √ . (15.267)
k
3 π 2
August 31, 2019 511

It follows that H2m  H2n if m > n. That in its turn implies G2n =
H2n /H0 ∼ H2n and C2n ∼ G2n , leading to the counter-intuitive result that
at very high loop order, almost all diagrams are connected ! Moreover, in low
order, we have C2 = 1 − u/2 + O (u2 ) and C4 = −u + 7u2 /2 + O (u3 ) so that
û = u(1 − 5u/2 + O (u2 )), and û(u) ∼ −H4 (u) up to an overall factor. This
gives us u = û(1+5û/2+O (û2 )), giving g1 = 5/2 in Eq.(15.264). In addition,
Eq.(15.267) tells us that cf = −2/3, where cf is defined in Eq.(15.258). Since
Hp (u)  û(u) for p = 6, 8, 10, . . . we then find

Hk (u(û)) ∼ Hk (û) exp(g1 /cf ) = Hk (û) e−15/4 . (15.268)

The ‘improvement factor’ therefore has the asymptotic value exp(15/4).

15.15.4 Gamma, Digamma and Bernoulli

The Gamma function Γ(z) has a well-known integral representation :

Z∞
Γ(z) = dt tz−1 exp(−t) , (15.269)
0

but that works only if <(z) > 0. A slightly less well-known but very useful
representation, that works for almost all z, is a special limit46 :
n
Y
Γ(z) = lim n! nz (k + z)−1 . (15.270)
n→∞
k=0

This shows immediately that

n nz
Γ(1) = lim = 1 , Γ(z + 1)/Γ(z) = lim =z , (15.271)
n→∞ n + 1 n→∞ n + 1 + z

thus proving the correctness of Eq.(15.270). Also we find Γ(m) = (m − 1)!

for natural m. For z close to −m (m = 0, 1, 2, . . .) we also find very simply

(−)m
Γ(z) ≈ . (15.272)
m! (z + m)
46
Found by — who else ? — Euler.
512 August 31, 2019

Finally, we can work out how Γ(z) behaves for small z :

n
!

X
log zΓ(z) = lim log(n!) + z log(n) − log(k + z)
n→∞
k=1
n
! n
!
X 1 X (−z)p X 1
= lim z log(n) − +
n→∞
k=1
k p≥2
p k=1 k p
X (−z)p
= −zγE + ζ(p) , (15.273)
p≥2
p
Pn 47
where γP
E = lim(− log(z)+ 1 (1/k)) is the Euler-Mascheroni constant , and
∞
ζ(p) = 1 1/k p is the Riemann zeta function48 . Using section 15.15.11, we
can then write, for small z :
 2
π2

1 γE
Γ(z) = − γE + z + + ··· (15.274)
z 2 12
The digamma function is defined as
n
!
X 1
ψ(z) = Γ0 (z)/Γ(z) = lim log(n) − . (15.275)
n→∞
k=0
k+z

We can immediately derive

ψ(1) = − γE , ψ(1/2) = − γE − log(4) . (15.276)

We can also find an asymptotic expansion for ψ(z) as follows, by taking an

extra derivative and letting n → ∞ :
Z∞
0
X 1 X
ψ (z) = = dt t exp(−(k + z)t)
k≥0
(k + z)2 k≥0
0
Z∞ ! Z∞  
−zt −zt
X
−kt t
= dt te + te e = dt te−zt + e−zt t
.
k≥1
e −1
0 0
(15.277)
47
In short : γE = .57721566490153286060651209008240243104215933593992359881 · · ·.
48
We have interchanged the sums over k and p in Eq.(15.273). Since the result is
convergent this is actually allowed.
August 31, 2019 513

The Bernoulli numbers Bm are defined as

t X tm 1 1
t
= Bm , B0 = 1 , B1 = − , B2 = . . . (15.278)
e − 1 m≥0 m! 2 6

Performing the integral over t and then integrating with respect to z yields
1 X Bm
ψ(z) = log(z) − − z −m , (15.279)
2z m≥2 m

where the constant of integration vanishes as we can check from the Stirling
approximation of section 15.15.2. Finally, from the fact that t/(et − 1) + t/2
is actually symmetric in t we can see that Bm vanishes for odd m ≥ 3.

15.15.5 The Dirac delta distribution

The Kronecker delta, as introduced in chapter 0, is defined for integer argu-
ments and reads
δm,n = θ(m = n) , (15.280)
so that ∞
X
δn,m = 0 for m 6= n , δm,n = 1 . (15.281)
n=−∞

The Dirac delta distribution49 is the continuum variant of this. Being a

distribution, it is really defined in the context of integration with a test
function50 . The Dirac delta is commonly denoted by (surprise ! ) δ(x) and
its definition is
Z∞
dx δ(x − a) f (x) = f (a) (15.282)
−∞

for all test functions f (x). Viewed as some kind of function it therefore has
properties analogous to those in Eq.(15.281) :
Z∞
δ(x) = 0 for x 6= 0 , dx δ(x) = 1 . (15.283)
−∞

49
Colloquially, the Dirac delta function, but it is really a distribution in the sense of
Schwartz.
50
A test function has compact support and is inifinitely many times differentiable :
simplistically, it is a nice function.
514 August 31, 2019

Applying partial integration (and assuming cavalierly that this is allowed ! )

we also find properties of its derivatives :
Z∞ Z∞
dx δ 0 (x − a) f (x) = −f 0 (a) , dx δ 00 (x − a) f (x) = +f 00 (a) ,
−∞ −∞
(15.284)
and so on. The Dirac delta can be viewed as the limit of a set of nonnegative
functions with unit integral, that are increasingly narrow and more and more
peaked at zero.

An important result that we use extensively in the text is

Z∞
dx exp(ixz) = 2π δ(z) , (15.285)
−∞

the proof of which forms exercise 32.

15.15.6 The principal-value distribution

Like the Dirac delta, this distribution is best defined in the context of an
integral. For a test function f (x) it acts as follows :

Zb>0 Z− Zb>0

1 f (x) f (x)
dx f (x) P ≡ lim ↓0 dx + lim ↓0 dx . (15.286)
x x x
a<0 a<0 

The tiny region where the singularity resides is cut away. If we view distri-
butions as limiting cases of functions, we can write

1 1
lim ↓0 = P + iπ δ(x) . (15.287)
x − i x

15.15.7 Generating the Bell numbers

In order to arrive at the generating function for the Bell number B(n), we
start with a more basic concept. By Bn (k) we denote the number of ways
to divide n distinct objects into k non-empty groups: we shall then have
August 31, 2019 515
P
B(n) = k≥0 Bn (k). For zero objects, there is obviously only one way to
divide them, namely in zero groups:

B0 (k) = δk,0 . (15.288)

If we have n − 1 objects distributed into k groups, we can let the nth object
form its own group, or add to one of the existing groups in k different ways.
This gives us the recursion

Bn (k) = Bn−1 (k − 1) + k Bn−1 (k) , n ≥ 1 . (15.289)

Let now form the set of generating functions

X zn
φk (z) = Bn (k) , k = 0, 1, 2, . . . (15.290)
n≥0
n!

From Eq.(15.288) we have that φ0 (z) = 1, and from Eq.(15.289)

φ0k (z) = k φk (z) + φk−1 (z) , φ0 (z) = 1 (k ≥ 1) . (15.291)

It is easily checked that the unique solution to these inhomogeneous first-

order differential equations is

1 z
k
φk (z) = e −1 , (15.292)
k!
so that
X zn X z −1)
B(n) = φk (z) = e(e . (15.293)
n≥0
n! k≥0

15.15.8 The exponential integral E1 and the Bessel K

functions
The exponential integral is defined by

Z∞
exp(−y)
E1 (z) = dy . (15.294)
y
z
516 August 31, 2019

We can find its series expansion as follows51 :

Z∞ Z∞ n−1
!
dy 1 X nk
E1 (z) = lim = lim dy −
n→∞ y(1 + y/n)n n→∞ y(n + y) k=1 (n + y)k+1
z z
"   n−1
#
n+z 1 X
= lim log −
n→∞
k=1
k(1 + z/n)k
z
" n−1  #
 n  X  z ` X 1 k−1+`
= lim log − −
n→∞ z `≥0
n k=1 k k−1
"   n−1 n−1
#
1 X 1 X  z ` 1 X `−1
= lim log + log(n) − − − k
n→∞ z k=1
k `≥1
n `! k=1
`
 
1 X (−z)
= log − γE − . (15.295)
z `≥1
`! `

The modified Bessel function of the second kind, of order ν, can be defined
by its integral representation :
Z∞   
1 ν−1 z 1
Kν (z) = dx x exp − x+ , (15.296)
2 2 x
0

We can see (by x → 1/x) that Kν (z) = K−ν (z). For small z and ν > 0, the
integral will be dominated by the large-x region, and we can approximate
Z∞  ν
1 ν−1
 zx  1 2
Kν (z) ∼ dx x exp − = Γ(ν) . (15.297)
2 2 2 z
0

For ν = 0 the regions 0 ≤ x ≤ 1 and 1 ≤ x ≤ ∞ contribute equally, but for

small z we can afford to simplify the exponent in each one:
Z1 Z∞
1 dx  z  1 dx  zx 
K0 (z) ∼ exp − + exp − = E1 (z) . (15.298)
2 x 2x 2 x 2
0 1

51
In the penultimate line of Eq.(15.295) we use the fact that, for n → ∞, the last
sum over k is dominated by its highest terms, and for large k the binomial coefficient is
approximately k ` /`!.
August 31, 2019 517

Figure 15.4: The functions K0 , K1 , K2 and E1

For large z, on the other

√ hand, the integral
√ is dominated by the narrow x
region from 1 − O (1/ z) to 1 + O (1/ z), we can approximate
Z∞  rπ
1  z
2
Kν (z) ∼ dx exp − (1 + x ) = exp(−z) . (15.299)
2 2 2z
−∞

By methods analogous to those of sec.15.15.2 the higher-order terms can

in principle be obtained, but the following method is easier. The modified
Bessel function satisfies the differential equation52
z 2 Kν00 (z) + zKν0 (z) − (z 2 + ν 2 )Kν (z) = 0 . (15.300)
Under the Ansatz (inspired by Eq.(15.299))
r
π
Kν (z) = exp(−z) k(z) , (15.301)
2z
this implies for the function k(z) the differential equation
4z 2 k 00 (z) − 8z 2 k 0 (z) − (4ν 2 − 1)k(z) = 0 . (15.302)
52
This can be checked by inserting the form (15.296) and partial integration.
518 August 31, 2019

Im z
Γ

Re z
x

Figure 15.5: Integration contour for the ramers-Kronig relation

If we now insert the formal power series

X cn
k(z) = 1 + (15.303)
n≥1
zn

and collect the various powers of 1/z, we find

1 (4ν 2 − (2n − 1)2 )
c1 = (4ν 2 − 1) , cn = cn−1 (n > 1) , (15.304)
8 8n
and we therefore have the following asymptotic expansion :
r n
!
π X 1 Y
4ν 2 − (2r + 1)2


Kν (z) ∼ exp(−z) 1 + n
.
2z n≥0
n!(8z) r=1
(15.305)

15.15.9 The Kramers-Kronig relation

We consider a function f (z) that is analytic for =(z) > 0, and goes to zero
sufficiently fast as |z| → ∞, =(z) > 0. We may then construct a contour Γ
as indicated in figure 15.15.9. The contour runs along the real axis from −∞
to +∞. At the point x it circles around it, and a big half-circle then leads
back from +∞ to −∞. By Cauchy’s theorem, we have
I
f (z)
dz = 0 ,
z−x
Γ
August 31, 2019 519

since also f (z)/(z − x) is analytic on, and inside, Γ. Splitting the integral
into its various contributions, we therefore have
Zx− Z+∞ I
f (z) f (z) 1 f (z)
0= dz + dz − dz , (15.306)
z−x z−x 2 z−x
−∞ x+ z∼x

where we have assumed that the big half-circle does not contribute since
f (z)/(z − x) vanishes fast enough. The number  is infinitesimal, and the
sum of the first two terms is called the principal value integral :
 x− 
Z+∞ Z Z+∞
f (z)  f (z) f (z) 
P dz ≡ lim dz + dz . (15.307)
z−x →0  z−x z−x 
−∞ −∞ x+

We therefore have the following equality:

Z+∞
f (z)
P dz = iπ f (x) ; (15.308)
z−x
−∞

and by inspecting the real and imaginary parts separately we arrive at the
Kramers-Kronig relations
Z+∞ Z+∞
1 =f (z) −1 <f (z)
<f (x) = P dz , =f (x) = P dz . (15.309)
π z−x π z−x
−∞ −∞

15.15.10 The dilogarithm function

The dilogarithm function Li2 (z) is defined by the following integral :
Zz
1
Li2 (z) = − du log(1 − u) , (15.310)
u
0

where the integration contour should not cross the cut in the logarithm (this
is usually chosen to be the real axis at z values larger than 1). Figure 15.6
shows the function Li2 (z + iη) for real values of z. Along to the cut we have

= (Li2 (z ± iη)) = ± π log(z) (z > 1) . (15.311)

520 August 31, 2019

Figure 15.6: The dilogarithm function

By expanding the logarithm we find its series expansion :

X zn
Li2 (z) = , |z| < 1 , (15.312)
n≥1
n2

which is handy for evaluating the dilogarithm for small arguments. We also
have Li2 (1) = π 2 /6 (cf section 15.15.11). There are a number of useful
identities for the dilogarithm53 , of which we give several below. You can
prove them by differentiating the left-hand and right-hand sides with respect
to z, and additionally checking them at z = 0 or z = 1 or so.

1
Li2 (z) + Li2 (−z) = Li2 (z 2 ) ,
2
π2
Li2 (z) + Li2 (1 − z) = − log(z) log(1 − z) ,
6
π2 1
2
Li2 (z) + Li2 (1/z) = − − log(−z) ,
6 2
53
A word of caution ! Sometimes the definition of the dilogarithm differs from the one
given here. In particular, in the computer-algebra system MAPLE the function dilog(z)
corresponds to our Li2 (1 − z).
August 31, 2019 521

1 π2
Li2 (−z) − Li2 (1 − z) + Li2 1 − z 2 = − − log(z) log(1 + z) ,


2 12
1
2
Li2 (1 − z) + Li2 (1 − 1/z) = − log(z) . (15.313)
2
With all these relations, small wonder that expressions involving dilogarithms
can be made to look bewilderingly different ! Some other special values can
also be derived using the above identitites :

π2 1
2
Li2 (0) = 0 , Li2 (1/2) = − log(2) ,
12 2
π2 π2
Li2 (−1) = − , Li2 (2) = − iπ log(2) . (15.314)
12 4
As for the numerical evaluation of the dilogarithm function : by applying
Eqs.(15.313) we can always arrange for the argument z to have |z| ≤ 1 and
0 ≤ <(z) ≤ (1/2) ; and then we can judiciously change the integration
variable from x to y = − log(1 − x), to see that

− log(1−z)
Z
y X Bn
n+1
Li2 (z) = dy = − log(1 − z) , (15.315)
ey − 1 n≥0 (n + 1)!
0

where Bn are the Bernoulli numbers

√ of section 15.15.4. Since | log(1 − z)| is
now always at most log((1 + i 3)/2) = π/3 this sum converges very rapidly.

15.15.11 Some values of the ζ function

The value of ζ(2) given in the previous section, and more, can be found
using a beautiful almost-rigorous method due to Euler, which I cannot resist
including here. Consider the function sin(x)/x. This is an analytic function
that equals 1 for x = 0 and has zeroes whenever x = nπ, n = ±1, ±2, ±3, . . ..
We therefore have
∞ 
x2

sin(x) Y
= 1− 2 2
x n=1
nπ
∞ ∞
x2 X 1 x4 X θ(n < m)
= 1− + − ··· (15.316)
π 2 n=1 n2 π 4 n,m=1 n2 m2
522 August 31, 2019

Also, by Taylor expanding we have

sin(x) x2 x4
=1− + − ··· (15.317)
x 6 120
By comparing the various powers of x in both expressions, we can immedi-
ately see that ζ(2) = π 2 /6. Also ζ(4) can be inferred :
∞ ∞
!2 ∞
X 1 X 1 X θ(n 6= m)
ζ(4) = 4
= 2
−
n=1
n n=1
n n,m=1
n2 m2
 2 2  4
π π π4
= −2 = ; (15.318)
6 120 90
and you can go much higher by hand — if you are Euler.

15.15.12 The Lagrange expansion

Here we deal with the solution of the equation
ξ = x + f (ξ) . (15.319)
We shall assume that f (x) = O (x2 ) so that x and ξ ‘tend to be close’,
especially for small x. The task is to express ξ as a function of x and f (x)
only. First we note that ξ is given as the root of the equation
φ(y) = 0 , φ(y) = y − x − f (y) , (15.320)
which by our assumption of smallishness has only a single simple root in a
sufficiently small neighbourhood of y = 0. The function φ0 (y)/φ(y) therefore
only has a simple pole at y = ξ, and we may write
y 1 − f 0 (y)


φ0 (y)
I I
1 1
ξ= dy y = dy , (15.321)
2πi φ(y) 2πi y − x − f (y)
where the contour is taken inside the region that contains only the single
pole. Since f (y) is small if x and y are in the neighbourhood of zero, we may
expand
I
1 X 1 0
f (y)n


ξ = dy n+1
y 1 − f (y)
2πi n≥0
(y − x)
X 1  n
∂
x f (x)n − x f 0 (x) f (x)n


= (15.322)
n≥0
n! ∂x
August 31, 2019 523

The first term in the second line reads

X 1  ∂ n
x f (x)n


x+
n≥1
n! ∂x

which we can rewrite as

X 1  ∂ n−1 X 1
 n−1
∂
n
x f 0 (x) f (x)n−1 :


x+ f (x) +
n≥1
n! ∂x n≥1
(n − 1)! ∂x

and the last term of this expression cancels against the last term in Eq.(15.322).
We are left with the following :

X 1  ∂ n−1
ξ = x+ f (x)n . (15.323)
n≥1
n! ∂x

A note is in order here about the computational properties of this relation.

One might simply iterate Eq.(15.319) to arrive at the result, in the following
manner :

ξ = x ,
ξ = x + f (x) ,
ξ = x + f (x + f (x)) ≈ x + f (x) + f 0 (x)f (x) ,
ξ = x + f (x + f (x) + f (x)f 0 (x))
1
≈ x + f (x) + f (x)f 0 (x) + f (x)f 0 (x)2 + f (x)2 f 00 (x) (15.324)
2
and so on, assuming f (x) and its derivatives to be small enough to warrant
Taylor expansion. This reproduces the Lagrange expansion.

As an application, let us consider the equation

1
ξ = x + ξ2 . (15.325)
2
The solution that vanishes with x is
√
ξ =1− 1 − 2x . (15.326)
524 August 31, 2019

Lagrange expansion, on the other hand, leads to

X 1  ∂ n−1  x2 n X 1 (2n)! xn+1
ξ =x+ = n
; (15.327)
n≥1
n! ∂x 2 n≥0
n! (n + 1)! 2

and by differentiating to x, then replacing x by 2x, we find

1 X (2n)!
√ = xn . (15.328)
1 − 4x n≥0 (n!)2

15.15.13 Determinants from traces

The determinant of a n × n matrix A with elements Ajk (j, k = 1, . . . , n) can
be defined using the antisymmetrizer :
 
1 k1 k2 · · · kn
det (A) = Ajk11 Ajk22 · · · Ajknn . (15.329)
n! j1 j2 · · · jn
Since the antisymmetrizer can be expressed in terms of Kronecker δ’s that
tie indices together, the determinant can be computed in terms of traces of
powers of A. To arrive at explicit results we can do the following. If an
n × n matrix A has an eigenvalue λ, then matrix 1 + zA has, for the same
eigenvector, eigenvalue 1 + zλ. Let λj , j = 1, . . . , n be the eigenvalues of A.
We can then employ the following truism :
n n
!
Y
X

1 + zλj = exp log 1 + zλj . (15.330)
j=1 j=1

For the left-hand side we can write

Yn n
X
z p Rp ,


1 + zλj = 1 + (15.331)
j=1 p=1

where X
Rp = λj1 λj2 · · · λjp , (15.332)
1≤j1 <j2 <···<jp ≤n

so that R1 is the trace, and Rn is the determinant, of A. For the right-hand

side we have
n
X X zn
(−)n−1 Tr (An ) ; (15.333)


log 1 + zλj = Tr (log(1 + zA)) =
j=1 n≥1
n
August 31, 2019 525

and then we equate the powers of z on both sides. The coefficient of z n is

the sought-for determinant expression. Denoting Tr (Ap ) by Tp , we find for
the first few powers of z :
1
T12 − T2 ,


R2 =
2!
1
T13 − 3 T1 T2 + 2 T3 ,


R3 =
3!
1
T14 − 6 T2 T12 + 3 T22 + 8 T1 T3 − 6T4 ,


R4 =
4! 
1
R5 = T15 − 10 T13 T2 + 15 T1 T22 + 20 T12 T3 − 30 T1 T4
5!

−20 T2 T3 + 24 T5 ,

1
R6 = T16 − 15 T14 T2 + 40 T13 T3 + 45 T12 T22 − 15 T23
6!
−90 T12 T4 − 120 T1 T2 T3 + 144 T1 T5 + 90 T2 T4

2
+40 T3 − 120 T6 . (15.334)

Since the left-hand side of Eq.(15.330) is a polynomial of degree n in z, we

see that Rk = 0 for k > n.
For Dirac matrices the traces are representation-independent, and so
therefore is their determinant : Tp = 0 for odd p and |T2 | = T4 = 4, so

det γ µ = det γ 5 = 1 , det (ω± ) = 0 .

(15.335)
Index

R , 177 Casimir trick, 208

causal dynamical triangulations, 125
action, 30 Cayley-Hamilton theorem, 201
for ϕ4 theory charge retention form, 237
Euclidean, 104 chirality spinors, 199
Minkowskian, 124 Chisholm identity, 195
zero-dimensional, 33 closed loops, 48
free, 32 colour conservation, 342
interacting, 33 colour democracy, 342
interaction terms, 33 Compton scattering, 283
kinetic part, 33 conservation of probability, 160
active-passive distinction, 214 Coulomb interaction
annihilation, 142 in QCD, 362
antiparticles, 141 in QED, 303
antisymmetrizer, 24 coupling quantisation, 355
auxiliary fields, 145 Cutkosky rules, 165

Bell number, 45 digestive processes, 141

Bernoulli numbers, 513 dimensional regularization, 174
Bessel function, 116 dimensionalities, 157
Bhabha scattering, 288 Dirac conjugation, 195
biker’s privilege, 258 Dirac equation, 224
bootstrap, 152 Dirac matrices, 189
Borel summation, 438 discontinuous limits, 332
boustrophedonic diagrams, 148 Dispersion relations, 179
Bremsstrahlung, 290 subtracted, 323
collinear, 295 dispersion-absorption, 161
hard, 293 divergence
soft, 292 IR, 166
UV, 166
career decision, 30 doublethink, 174

526
August 31, 2019 527

dyad, 186 Gauge

Dyson summation, 47 unitary, 248
gluon fusion, 435
Electron-positron annihilation , 286 Gordon decomposition, 276
Euler’s integral, 173 convection term, 277
Euler-Lagrange equation, 105 spin term, 277
Euler-Mascheroni constant, 512 Green’s functions, 30
experimentalist, 30 connected, 31
External fields, 304 generating function, 31
gyromagnetic ratio of the electron, 307
Fermi interaction, 236
failure, 374 half-renormalisability, 79
Fermi-Dirac condensate, 222 Hall of Fame, 26
Feynman diagrams, 37 handlebar, 258
complexity, 48 helicity methods, 229
connected, 37 helicity suppression, 243
disconnected, 37 hemicranogenic arguments, 328
empty, 38
multiplicity, 41 infrared divergence, 166
one-particle irreducible, 55 integrating out, 146
semiconnected, 44 IVB hypothesis, 376
symmetry factor, 39 Källén function, 157
tree, 38 Kalb-Ramond state, 490
Feynman rules, 37 Klein-Gordon equation, 132
Feynman rules for Dirac particles, 221 charged, 304
field function, 32 Kronecker symbol, 23
field strength tensor, 434
Fierz identity Landau-Yang theorem
for T matrices, 347 for gluons
for Dirac matrices, 193 higher order, 369
Fierzing, 194 tree order, 368
fine structure constant, 22 for photons, 316
foliation, 125 Levi-Civita symbol, 24
form factors, 375 lines, 37
functional derivatives, 105 external, 37
Furry’s theorem internal, 37
failure in QCD, 350 longitudinal part, 251
in QED, 278 loose terminology, 26
528 August 31, 2019

Lorentz transformation, 213 radiation cone, 297

in Dirac space, 213 random number, 29
in Minkowski space, 250 real distance, 115
infinitesimal, 213 relativistic Pauli equation, 305
minimal, 213 resurgence, 508
reverse of Dirac string, 195
mass gap, 333 rotation group, 213
mental flip, 197
metric, 23 satellite diagram, 150
MHV amplitudes, 309 scalar part, 251
muon pair production, 279 Schouten identity, 231
Mœller scattering, 290 Schwinger-Dyson equation, 35
for diagrams, 44
Newton’s constant, 21 for the field function, 36
nonpertubative, 53 for the path integral, 35
normalization factor, 30 sea-gull vertex, 299
notion of particle, 132 slash notation, 190
soft-photon approximation, 292
optical theorem, 161 source, 31
speed of light, 20
parenthetophiliacs, 25
spin-statistics theorem
parity violation, 242
difficult, 256
partial-wave unitarity, 161
easy, 256
path integral, 31
spinor sandwich, 196
Pauli exclusion principle, 222
spinor techniques, 231
Pauli identity, 195
static scattering, 171
periodophobia, 25
step function
perturbation theory, 33
logical, 23
phenomenology, 236
numerical, 23
pimped propagator, 187
Stern-Gerlach term, 305
pion decay constant, 242
super-renormalizable, 83
Planck’s constant, 20
superexponential, 437
polyparenthetophobia, 25
positronium test functions, 513
ortho-, 315 theorist, 30
para-, 315 transition amplitudes, 125
propagator, 98 typographies, 32

quantum field, 29 ultraviolet divergence, 166

August 31, 2019 529

unitary gauge, 247

units, 20
fundamental HEP, 20
Planck, 21

vacuum bubbles
absence, 46
definition, 42
vacuum polarization
hadronic, 326
leptonic, 324
vertices, 37

Ward identity, 274

Ward-Takahashi identity, 272
Weyl ordering, 104
Weyl representation, 202
Weyl-van der Waerden spinors, 202
Wick rotation, 174

Yang-Mills four-boson vertex, 358

Yang-Mills three-boson vertex, 351

zero-dimensional space, 29