Renormalization: An Introduction to Renormalization, the Renormalization Group and the Operator-Product Expansion 0521242614, 9780521242615, 9780511867392, 0511867395

Most of the numerical predictions of experimental phenomena in particle physics over the last decade have been made poss

531 105 4MB

English Pages 390 Year 1984

Report DMCA / Copyright

DOWNLOAD FILE

Polecaj historie

Renormalization: An Introduction to Renormalization, the Renormalization Group and the Operator-Product Expansion
 0521242614, 9780521242615, 9780511867392, 0511867395

Table of contents :
cover
title
contents
1 Introduction
2 Quantum field theory
2.1 Scalar field theory
2.2 Functional-integral solution
2.3 Renormalization
2.4 Ultra-violet regulators
2.5 Equations of motion for Green's functions
2.6 Symmetries
2.7 Ward identities
2.8 Perturbation theory
2.9 Spontaneously broken symmetry
2.10 Fermions
2.11 Gauge theories
2.12 Quantizing gauge theories
2.13 BRS invariance and Slavnov-Taylor identities
2.14 Feynman rules for gauge theories
2.15 Other symmetries of (2.11.7)
2.16 Model field theories
3 Basic examples
3.1 One-loop self-energy in phi3 theory
3.1.1 Wick rotation
3.1.2 Lattice
3.1.3 Interpretation of divergence
3.1.4 Computation
3.2 Higher order
3.3 Degree of divergence
3.3.1 phi3 at d = 6
3.3.2 Why may Z be zero and yet contain divergences?
3.3.3 Renormalizability and non-renormalizability
3.4 Renormalization group
3.4.1 Arbitrariness in a renormalized graph
3.4.2 Renormalization prescriptions
3.5 Dimensional regularization
3.6 Minimal subtraction
3.6.1 Definition
3.6.2 d = 6
3.6.3 Renormalization group and minimal subtraction
3.6.4 Massless theories
3.7 Coordinate space
4 Dimensional regularization
4.1 Definition and axioms
4.2 Continuation to small d
4.3 Properties
4.4 Formulae for Minkowski space
4.4.1 Schwinger parameters
4.4.2 Feynman parameters
4.5 Dirac matrices
4.6 gamma_5and epsilon
5 Renormalization
5.1 Divergences and subdivergences
5.2 Two-loop self-energy in (phi3)6
5.2.1 Fig. 5.1.2
5.2.2 Differentiation with respect to external momenta
5.2.3 Fig. 5.1.3
5.3 Renormalization of Feynman graphs
5.3.1 One-particle-irreducible graph with no subdivergences
5.3.2 General case
5.3.3 Application of general formulae
5.3.4 Summary
5.4 Three-loop example
5.5 Forest formula
5.5.1 Formula
5.5.2 Proof
5.6 Relation to L
5.7 Renormalizability
5.7.1 Renormalizability and non-renormalizability
5.7.2 Cosmological term
5.7.3 Degrees of renormalizability
5.7.4 Non-renormalizability
5.7.5 Relation of renormalizability to dimension of coupling
5.7.6 Non-renormalizable theories of physics
5.8 Proof of locality of counterterms; Weinberg's theorem
5.8.1 Degree of counterterms equals degree of divergence
5.8.2 \bar{R}(G) is finite if delta{G) is negative
5.8.3 Asymptotic behavior
5.9 Oversubtractions
5.9.1 Mass-shell renormalization and oversubtraction
5.9.2 Remarks
5.9.3 Oversubtraction on IPR graphs
5.10 Renormalization without regulators: the BPHZ scheme
5.11 Minimal subtraction
5.11.1 Definition
5.11.2 \bar{MS} renormalization
6 Composite operators
6.1 Operator-product expansion
6.2 Renormalization of composite operators: examples
6.2.1 Renormalization of phi2
6.2.2 Renormalization of (phi)2(x)(phi)2(y)
6.3 Definitions
6.4 Operator mixing
6.5 Tensors and minimal subtraction
6.6 Properties
Property 1. Linearity
Property 2. Differentiation is distributive
Property 3. Simple equation of motion
Property 4. Equation of motion times operator
Property 5. Ward identities
Property 6. Non-renormalization of current
6.7 Differentiation with respect to parameters in L
6.8 Relation of renormalizations of phi2 and m2
7 Renormalization group
7.1 Change of renormalization prescription
7.1.1 Change of parametrization
7.1.2 Renormalization-prescription dependence
7.1.3 Low-order examples
7.2 Proof of RG invariance
7.3 Renormalization-group equation
7.3.1 Renormalization-group coefficients
7.3.2 RG equation
7.3.3 Solution
7.4 Large-momentum behavior of Green's functions
7.4.1 Generalizations
7.5 Varieties of high- and low-energy behavior
7.5.1 Asymptotic freedom
7.5.2 Maximum accuracy in an asymptotically free theory
7.5.3 Fixed point theories
7.5.4 Low-energy behavior ofmassless theory
7.6 Leading logarithms, etc.
7.6.1 Renormalization-group logarithms
7.6.2 Non-renormalization-group logarithms
7.6.3 Landau ghost
7.7 Other theories
7.8 Other renormalization prescriptions
7.9 Dimensional transmutation
7.10 Choice of cut-off procedure
7.10.1 Example: phi4 self-energy
7.10.2 RG coefficients
7.10.3 Computation of g0 and Z; asymptotically free case
7.10.4 Accuracy needed for g0
7.10.5 m20
7.10.6 Non-asymptotically free case
7.11 Computing renormalization factors using dimensional regularization
7.12 Renormalization group for composite operators
8 Large-mass expansion
8.1 A model
8.2 Power-counting
8.2.1 Tree graphs
8.2.2 Finite graphs with heavy loops
8.2.3 Divergent one-loop graphs
8.2.4 More than one loop
8.3 General ideas
8.3.1 Renormalization prescriptions with manifest decoupling
8.3.2 Dominant regions
8.4 Proof of decoupling
8.4.1 Renormalization prescription R* with manifest decoupling
8.4.2 Definition of R*
8.4.3 IR finiteness of C*(\Gamma)
8.4.4 Manifest decoupling for R*
8.4.5 Decoupling theorem
8.5 Renormalization-group analysis
8.5.1 Sample calculation
8.5.2 Accuracy
9 Global symmetries
9.1 Unbroken symmetry
9.2 Spontaneously broken symmetry
9.2.1 Proof of invariance of counterterms
9.2.2 Renormalization of the current
9.2.3 Infrared divergences
9.3 Renormalization methods
9.3.1 Generation of renormalization conditions by Ward identities
10 Operator-product expansion
10.1 Examples
10.1.1 Cases with no divergences
10.1.2 Divergent example
10.1.3 Momentum space
10.1.4 Fig. 10.1.3 inside bigger graph
10.2 Strategy of proof
10.3 Proof
10.3.1 Construction of remainder
10.3.2 Absence of infra-red and ultra-violet divergences in r(\Gamma).
10.3.3 R(T) — r(T) is the Wilson expansion
10.3.4 Formula for \bar{C}(U)
10.4 General case
10.5 Renormalization group
11 Coordinate space
11.1 Short-distance singularities of free propagator
11.1.1 Zero temperature
11.1.2 Non-zero temperature
11.2 Construction of counterterms in low-order graphs
11.3 Flat-space renormalization
11.4 Externa] fields
12 Renormalization of gauge theories
12.1 Statement of results
12.2 Proof of renormalizability
12.2.1 Preliminaries
12.2.2 Choice of counterterms
12.2.3 Graphs with external derivatives
12.2.4 Graphs finite by equations of motion
12.2.5 Gluon selfenergy
12.2.6 BRS transformation of A.
12.2.7 Gluon self-interaction
12.2.9 Quark-gluon interaction; introduction of B
12.3 More general theories
12.3.1 Bigger gauge group
12.3.2 Scalar matter
12.3.3 Spontaneous symmetry breaking
12.4 Gauge dependence of counterterms
12.4.1 Change of Xi
12.4.2 Change of Fa
12.5 R-gauge
12.6 Renormalization of gauge-invariant operators
12.6.1 Caveat
12.7 Renormalization-group equation
12.8 Operator-product expansion
12.9 Abelian theories: with and without photon mass
12.9.1 BRS treatment of massive photon
12.9.2 Elementary treatment of abelian theory with photon mass
12.9.3 Ward identities
12.9.4 Counterterms proportional to A2 and d-A2
12.9.5 Relation between e0 and Z3
12.9.6 Gauge dependence
12.9.7 Renormalization-group equation
12.10 Unitary gauge for massive photon
13 Anomalies
13.1 Chiral transformations
13.2 Definition of gamma5
13.3 Properties of axial currents
13.3.1 Non-anomalous currents
13.3.2 Anomalous currents
13.3.3 Chiral gauge theories
13.3.4 Supersymmetric theories
13.4 Ward identity for bare axial current
13.4.1 Renormalization of operators in Ward identities
13.5 One-loop calculations
13.6 Non-singlet axial current has no anomaly
13.6.1 Reduction of anomaly
13.6.2 Renormalized current has no anomalous dimension
13.7 Three-current Ward identity; the triangle anomaly
13.7.1 General form of anomaly
13.7.2 One-loop value
13.7.3 Higher orders
14 Deep-inelastic scattering
14.1 Kinematics, etc.
14.2 Parton model
14.3 Dispersion relations and moments
14.4 Expansion for scalar current
14.5 Calculation of Wilson coefficients
14.5.1 Lowest-order Wilson coefficients
14.5.2 Anomalous dimensions
14.5.3 Solution of RG equation - non-singlet
14.6 OPE for vector and axial currents
14.6.1 Wilson coefficients - electromagnetic case
14.7 Parton interpretation of Wilson expansion
14.8 W4 and W5
References
Index

Citation preview



TH



S H

SICS

RENORMALIZATION An introduction to

renormalization, the renormalization group, and the operator-product expansion JOHN

C.COLLINS

Illinois Institute of Technology

The right of the University of Cambridge to print and sell all manner of books was granted by Henry VIII in 1534. The University has printed and published continuously since 1584.

CAMBRIDGE UNIVERSITY PRESS CAMBRIDGE LONDON

NEW YORK

MELBOURNE

NEW ROCHELLE SYDNEY

Published by the Press Syndicate of the University of Cambridge The Pitt Building, Trumpington Street, Cambridge CB2 1RP 32 East 57th Street, New York, NY 10022, USA 296 Beaconsfield Parade, Middle Park, Melbourne 3206, Australia ? Cambridge University Press 1984 First published 1984 Printed in Great Britain at the University Press, Cambridge Library of Congress catalogue card number: 83-23927

British Library cataloguing in publication data Collins, John C. Renormalization.—(Cambridge monographs on mathematical physics) 1. Renormalization (Physics) 2. Quantum field theory I. Title 530.T43

QC174.45

ISBN 0 521 24261 4

This work was supported in part by the US Department of Energy under contract number DE-AC02-80ER10712.

Contents



Introduction



Quantum field theory

2.1

Scalar field theory

1 4 4 6 9

2.2

Functional-integral

2.3

Renormalization

2.4

Ultra-violet regulators Equations of motion for Green's functions

12

2.5 2.6

Symmetries

15

2.7

Ward identities

19

2.8

Perturbation theory

21

solution

13

2.9

Spontaneously broken symmetry

25

2.10

Fermions

27

2.11

Gauge theories

28

2.12

Quantizing gauge theories BRS invariance and Slavnov-Taylor identities

29

34

2.15

Feynman rules for gauge theories Other symmetries of (2.11.7)

2.16

Model field theories

36

2.13 2.14

32

35

38



Basic examples

3.1

One-loop self-energy in

3.1.1

Wick rotation

39

3.1.2

Lattice

40

3.1.3

Interpretation of divergence

41

3.1.4

Computation Higher order

43 45

3.3.1

Degree of divergence c/>3 aid 6

3.3.2

Why may Z be zero and yet contain divergences?

48

3.3.3

49

3.4

Renormalizability and non-renormalizability Renormalization group

3.4.1

Arbitrariness in a renormalized

3.4.2

Renormalization prescriptions Dimensional regularization Minimal subtraction

3.2 3.3

3.5 3.6

</>3 theory

38

44 47



graph

50 50 52 53 56



Contents

VI

3.6.1

Definition

3.6.2



3.6.3

Renormalization group and minimal subtraction



56



57 58

3.6.4

Massless theories

59

3.7

Coordinate space

59



Dimensional regularization

62

4.1

Definition and axioms

64

4.2

Continuation to small d

68

4.3

Properties

73

4.4

Formulae for Minkowski space

81

4.4.1

83

4.5

Schwinger parameters Feynman parameters Dirac matrices

4.6

7 s and eK^v

86



Renormalization

88

5.1

Divergences and subdivergences Two-loop self-energy in (c/>3)6 Fig. 5.1.2 Differentiation with respect to external momenta Fig. 5.1.3 Renormalization of Feynman graphs One-particle-irreducible graph with no subdivergences General case

4.4.2

5.2 5.2.1 5.2.2 5.2.3 5.3 5.3.1 5.3.2

81

83

89 93 94 97 99 101 102 103 105

5.3.3

Application

5.3.4

Summary

106

5.4

Three-loop example

106

5.5

Forest formula

109

5.5.1

Formula

109

5.5.2

Proof

110

5.6

Relation to if

112

5.7

Renormalizability

116

5.7.1

Renormalizability and non-renormalizability

116

5.7.2

Cosmological term

118

5.7.3

Degrees of renormalizability

118

5.7.4

Non-renormalizability

120

5.7.5

Relation of renormalizability to dimension of coupling

121

5.7.6

Non-renormalizable theories of physics

122

5.8

Proof of locality of counterterms; Weinberg's theorem Degree of counterterms equals degree of divergence

125

5.8.1 5.8.2

R(G) is finite if d(G) is negative

127

5.8.3

Asymptotic behavior Oversubtractions Mass-shell renormalization

130

5.9 5.9.1

of general formulae

126 129

and oversubtraction

130

Contents 5.9.2

Remarks

5.9.3

Oversubtraction

vn

131 on 1 PR

graphs

131

5.10

Renormalization without regulators: the BPHZ scheme

133

5.11

Minimal subtraction

135

5.11.1

Definition

135

5.11.2

MS renormalization

137

5.11.3

Minimal subtraction with other regulators

137



Composite operators

138

6.1

Operator-product expansion

139

6.2

Renormalization of composite operators: examples

142

6.2.1

Renormalization of

6.2.2

c/>2 c/>2(x)(/>2(>!)

142

Renormalization of

6.3

Definitions

146

6.4

Operator mixing Tensors and minimal subtraction

150

6.6

Properties

152

6.7

Differentiation with respect to parameters 'm<f Relation of renormalizations of <f>2 and m2

166

6.5

6.8

7 7.1 7.1.1 7.1.2

145

149

Renormalization group Change of renormalization prescription Change of parametrization

163

168 169 169

7.1.3

Renormalization-prescription dependence Low-order examples

7.2

Proof of RG invariance

176

7.3

Renormalization-group equation Renormalization-group coefficients

180

183

7.3.3

RG equation Solution

184

7.4

Large-momentum behavior of Green's functions

185

7.4.1

Generalizations

187

7.5

Varieties of high- and low-energy behavior Asymptotic freedom

187 189

7.5.3

Maximum accuracy in Fixed point theories

7.5.4

Low-energy behavior of massless theory

190 191

7.6

Leading logarithms, etc.

193

7.6.1

193

7.6.2

Renormalization-group logarithms Non-renormalization-group logarithms

7.6.3

Landau ghost

196

7.7

Other theories

197

7.8

200

7.9

Other renormalization prescriptions Dimensibnal transmutation

7.10

Choice of cut-off procedure

206

7.3.1

7.3.2

7.5.1 7.5.2

an

asymptotically

171 172

180

free theory

187

196

203

Contents

Vlll

c/>4 self-energy

7.10.1

Example:

7.10.2

RG coefficients

7.10.3

Computation of g0 and Z; asymptotically free case

7.10.4

Accuracy needed for g0

7.10.5 7.10.6

ml Non-asymptotically

7.11

Computing

free

case

208 208 209 213 216 216

renormalization factors using dimensional

regularization

7.12

Renormalization group for composite operators



Large-mass expansion

8.1

A model

217 219

8.2

Power-counting

8.2.1

Tree

8.2.2

Finite graphs with heavy loops

222 224 225 225 226

8.2.3

Divergent one-loop graphs

226

8.2.4

More than

229

8.4.2

loop General ideas Renormalization prescriptions with manifest decoupling Dominant regions Proof of decoupling Renormalization prescription R* with manifest decoupling Definition of R*

8.4.3

IR finiteness of C*(T)

237

8.4.4

Manifest decoupling for R*

239

8.4.5

Decouplmg theorem

239

8.5

Renormalization-group analysis

8.5.1

Sample calculation

8.5.2

Accuracy



Global symmetries

244

9.1

Unbroken symmetry

245

9.2

247

9.2.1

Spontaneously broken symmetry Proof of invariance of counterterms

9.2.2

Renormalization of the current

251

9.2.3

divergences Renormalization methods Generation of renormalization

252

8.3 8.3.1 8.3.2 8.4 8.4.1

9.3 9.3.1

graphs

one

230 231 232 233 233 236

240 242 243

Infra-red

249

253

conditions by Ward identities

254

10

Operator-product expansion

257

10.1

Examples

258

10.1.1

Cases with

10.1.2 10.1.3 10.1.4

divergences Divergent example

258

Momentum space Fig. 10.1.3 inside bigger

262

no

graph

260 262

Contents

IX

269

10.4

Strategy of proof Proof Construction of remainder Absence of infra-red and ultra-violet divergences in r(T) R(F) r(T) is the Wilson expansion Formula for C(U) General case

10.5

Renormalization group

274

11

Coordinate space

277

11.1

Short-distance singularities of free propagator

278

11.1.1

Zero temperature

278

11.1.2

Non-zero temperature Construction of counterterms in low-order graphs

280 286

11.4

Flat-space renormalization External fields

12

Renormalization of gauge theories

293

12.1

Statement of results

296

12.2

298

12.2.1

Proof of renormalizability Preliminaries

12.2.2

Choice of counterterms

300

12.2.3

Graphs with external derivatives

301

12.2.4

302

12.2.5

Graphs finite by equations of motion Gluon self-energy

12.2.6

BRS transformation of

303

12.2.7

Gluon self-interaction

12.2.8

SRc

12.2.9

Quark-gluon interaction,

12.3

More general theories

307

12.3.1

Bigger

307

12.3.2

Scalar matter

12.3.3

Spontaneous symmetry

12.4

Gauge dependence

12.4.1

Change of c,

310

12.4.2

Change of Fa

313

Krgauge

314

10.2 10.3 10.3.1 10.3.2 10.3.3 10.3.4

11.2 11.3

12.5



263 266 266 270 272 273

281 291

298

302

A^

304 304

Sip, Sip; introduction of PMail

gauge group

305

307 308

breaking

309

of counterterms

12.6

Renormalization of gauge-invariant operators

316

12.6.1

Caveat

319

12.7

Renormalization-group equation

319

12.8 12.9

Operator-product expansion Abelian theories: with and without photon

12.9.1

BRS treatment of massive

12.9.2

Elementary

12.9.3

Ward identities

321

photon theory with photon

treatment of abelian

322

mass

323 mass

324 325

Contents



326

12.9.5

Counterterms proportional to A2 and d-A2 Relation between e0 and Z3

12.9.6

Gauge dependence

327

12.9.7

Renormalization-group equation Unitary gauge for massive photon

329

12.10 13

Anomalies

331

13.1

Chiral transformations

333

13.2 13.3

Definition of y5 Properties of axial currents

12.9.4

327

330

334 338

13.3.1

Non-anomalous currents

338

13.3.2

Anomalous currents

338

13.3.3

Chiral gauge theories Supersymmetric theories

339

Ward identity for bare axial current Renormalization of operators in Ward identities

339

One-loop calculations Non-singlet axial current has

342

13.3.4 13.4 13.4.1 13.5 13.6

339

no

anomaly-

340

346

13.6.2

Reduction of anomaly Renormalized current has

13.7

Three-current Ward

13.7.!

General form of anomaly

349

13.7.2

One-loop value

351

13.7.3

Higher orders

352

14

Deep-inelastic scattering

354

14.1

Kinematics, etc.

355

14.2

Parton model

358

14.3

359

14.4

Dispersion relations and moments Expansion for scalar current

14.5

Calculation of Wilson coefficients

364

14.5.1

Lowest-order Wilson coefficients

364

14.5.2

Anomalous dimensions

365

14.5.3

Solution of RG

14.6

OPE for

14.6.1

Wilson coefficients

13.6.1

14.7 14.8

346 no anomalous dimension

identity:

equation



the triangle anomaly

non-singlet

vector and axial currents

347 349

361

367 367

electromagnetic case Parton interpretation of Wilson expansion W4 and W5 References

368

Index

377



369 371 372

1 Introduction

The structure of a quantum field theory often simplifies when

one

considers

involving large momenta or short distances. These important in improving one's ability to calculate predictions from the theory, and in essence form the subject of this book. The first simplification to be considered involves the very existence of the theory. The problem is that there are usually ultra-violet divergences caused by large fluctuations of the field(s) on short distance scales. These manifest themselves in Feynman graphs as divergences when loop momenta go to infinity with the external momenta fixed. The simplification is that the divergences can be cancelled by renormalizations of the parameters of the action. Consequently our first task will be to treat the ultra-violet processes

simplifications are

renormalizations. Renormalization is essential, for otherwise most field theories do not exist. We will then expose the methods needed to handle high-energy/short- distance problems. The aim is to be able to make testable predictions from a

strong interaction theory,

improve the rate of convergence of the weakly coupled theory. The simplifications

or to

perturbation expansion generally take the form of a factorization of a cross-section or of amplitude, each factor containing the dependence of the process phenomena that happen on one particular distance scale. Such a in



an on

perturbation expansion for a process are large when the process involves widely different distance scales.

factorization is useful, because the coefficients of the

The industry called 'perturbative QCD' consists of deriving such factorization theorems for strong interactions (Mueller (1981)) and exploring their phenomenological consequences. We will only study the earliest of

these factorizations, the operator product expansion of Wilson (1969). We will also discuss the theorems that describe the behavior of a theory when of its fields get large (Appelquist & Carazzone (1975) and Witten These large-mass theorems have their main uses in weak (1976)). the

masses

interaction theories. The presence of ultra-violet

divergences,

by renormalization counterterms,

means



even

though they

are

cancelled

that in any process there

are



Introduction

contributions from quantum fluctuations on every distance scale. This is both a complication and an opportunity to find interesting physics. The

complication is that the derivation of factorization theorems is made difficult. The opportunity is given by the observable phenomena that directly result from the existence of the divergences. A standard example is given by the scaling violations in deep-inelastic scattering. It is the renormalization group (Stueckelberg & Petermann (1953) and Gell-Mann & Low (1954)) that is the key technique in disentangling the complications. The infinite parts of the counterterms are determined by the requirement that they cancel the divergences, but the finite parts are not so determined. In fact, the partition of a bare coupling g0 into the sum of a finite renormalized coupling gR and a singular counterterm Ag is arbitrary.

One can reparametrize the theory by transferring a finite amount from gR to Ag without changing the physics: the theory is invariant.

renormalization-group

This trivial-sounding observation is in practice very useful, and far from trivial. Suppose one has some graph whose renormalized value is large (so that it is

inadequate

to use a few low orders of the

perturbation theory

to

compute the corresponding quantity). Then in appropriate circumstances it is possible to adjust the partition of g0 (viz., g0 gR + Ag) so that the =

counterterm Ag cancels not only the

large piece

but also the

excessively divergence graph's finite part. The large piece is now in the lowest higher orders. Construction of factorization theorems of

of the

order instead of

the sort reviewed by Mueller (1981) provides many circumstances where this trick is applicable. Without it the factorization theorems would be

almost

powerless.

that the subjects of renormalization, the renormalization group, and the operator product expansion are intimately linked, and we will treat We

see

explain the general methods that examples we will examine but in many other situations. We will not aim at complete rigor. However there are many pitfalls and traps ready to ensnare an unwary physicist. Thus a precise set of concepts and notations is necessary, for many of the dangers are essentially combinatorial. The appropriate basis is then that of Zimmermann (1969, them all in this book. The aim will be to are

applicable not only

to the

1970, 1973a, 1973b). One other problem is that of choice of an ultra-violet cut off. From a fundamental point of view, the lattice cut-off seems best as it appears in non-

perturbative treatments using the functional integral (e.g., Glimm & Jaffe (1981)). In perturbation theory one can arrange to use no regulator whatsoever (e.g., Piguet & Rouet (1981)). In practice, dimensional reg-



Introduction

deservedly become very popular. This consists of replacing physical space-time dimensionality 4 by an arbitrary complex number d. The main attraction of this method is that virtually no violence is done to the structure of a Feynman graph; a second attraction is that it also regulates infra-red divergences. The disadvantage is that the method has not been formulated outside of perturbation theory (at least not yet). Much of the treatment in this book, especially the examples, will be based on the ularization has

the

use

of dimensional

strongly

that

none

regularization.

However it cannot be

of the fundamental results depend

on

emphasized this choice.

too



Quantum field theory Most of the work in this book will be

strictly perturbative. However it is important not to consider perturbation theory as the be-all and end-all of field theory. Rather, it must be looked on only as a systematic method of approximating a complete quantum field theory, with the errors under control. So in this chapter we will review the foundations of quantum field theory starting from the functional integral. The purpose of this review is partly to set out the results on which the rest of the book is based. It will also introduce our notation. We will also list a

number of standard field theories which will be used throughout the book. Some examples are physical theories of the real world; others are simpler theories whose only purpose will be to illustrate methods in the absence of

complications. integration is not absolutely essential. Its use is to provide a systematic basis for the rest of our work: the functional integral gives an explicit solution of any given field theory. Our task will be to investigate a certain class of properties of the solution. The

use

For

more

of functional

details the reader should consult



standard textbook

on

field

theory. Of these, probably the most complete and up-to-date is by Itzykson & Zuber (1980); this includes a treatment of the functional integral method. Other useful references include: Bjorken & Drell (1966), Bogoliubov & Shirkov (1980), Lurie (1968), and Ramond (1981).

2.1 Scalar field theory The simplest quantum field theory is that of a single real scalar field t/>(xM). The theory is defined by canonically quantizing a classical field theory. This

classical theory is

specified by <?





Lagrangian density:

(dct>)2/2-P((t>\

(2.1.1)

from which follows the equation of motion D</> + P'(0) 4



O.

(2.1.2)

2.1 Scalar



field theory

P(0) is a function of </>(x), which we generally take to be a polynomial like P(4>) m24>2/2 + #4/4!, and P'((/>) dP/dt/>. (Note that we use units with h c 1.) In the Hamiltonian formulation of the same theory, we define a canonical Here

--=







momentum field: n(x)



d^/dcj)



<j)



(2.1.3)

dtfr/dt,

and the Hamiltonian

-J-

d3x(;r2/2 + V02/2 + P(0)).

(2.1.4)

Physically, we require that a theory have a lowest energy state.

If it does not

unstable against decay into a lower energy state plus a collection of particles. If the function P(t/>) has no minimum, then the then all states

are

catastrophic situation exists (Baym (I960)). require P((j>) to be bounded below. in the Quantization proceeds Heisenberg picture by reinterpreting (j>(x) as a hermitian operator on a Hilbert space satisfying the canonical equal- formula (2.1.4) implies that just such Thus



the function

we

time commutation relations, i.e.,

[4>(xU(y) ̄\



WMy)]



y'

oj

l"'

given by (2.1.4) so the equation of motion (2.1.2) Heisenberg equation of motion

The Hamiltonian is still

follows from the

i34>/dt



(2.1.6)

\_4>9ir].

theory is specified by stating what the space of states is by giving the manner in which <f> acts on the states. We will construct a solution by use of the functional integral. It should be noted that <j)(x) is in general not a well-behaved operator, but rather it is an operator-valued distribution. Physically that means that one cannot measure 0(x) at a single point, but only averages of (/>(x) over a space-time region. That is, A solution to the

and

4>f



0(x)/(x)d4x,

for any complex-valued function /(x), is

an

(2.1.7)

operator. Now, products of

always make sense (e.g., <5(x)2). In particular, the Hamiltonian H involves products of fields at the same point. Some care is needed to define these products properly; this is, in fact, the subject of

distributions do

not

renormalization, to be treated shortly. The following properties of the theory

are

standard:



Quantum field theory

(1) The theory has



|0>,

Poincare-invariant ground state

called the

vacuum.

(2) The

states and the action of

(j>

on

them

can

be reconstructed from the

time-ordered Green's functions

GN(x1,...,xN)



<0|T^(x1)...^(xN)|0>.

(2.1.8)

The T-ordering symbol means that the fields are written in order of increasing time from right to left. The Green's functions have appropriate causality properties, etc., so (3) that they are the Green's functions of a physically sensible theory. Mathematically, these properties are summarized by the Wightman axioms (Streater &

Wightman (1978)).

Bose symmetry of the

(/>-field

symmetric under interchange of motion of

<f>

means

that the Green's functions

and from the commutation relations

of motion for the Green's functions. The

For



general (N



l)-point Green's function,



we

of

be derived equations example is

can

simplest

DyG2(y9x) + <0| TP'(d>(y))ct>(x)\Oy

are

equations

any of the x's. From the



i<5<4>(*



y).

(2.1.9)

have N <5-functions

on

the

right:

□,GN+1(y,x1,...,xiV) + <O|TP'(^))0(x1)...^(xN)|O> =

-i£ S?{y-xJ)GN_l(xl9...9xJ_l9xJ+i9...9xN).

(2.1.10)

j=i

summarizes both the equations of motion and the commutation relations. Solving the theory for the Green's functions means in

This

equation

essence

solving this

set of

functions that are the easiest

theory

can

be calculated

2.2

coupled equations. It is in fact the Green's objects to compute. All other properties of the

once

the Green's functions

Functional-integral

are

known.

solution

quantum field theory is a non-trivial problem in two cases are elementary: free field theory (P m2(/>2/2), consistency. Only 1 is a rather \. The case d and the case of one space-time dimension, d The solution of









just the quantum mechanics of a particle with Heisenberg position operator <f>(t) in a potential P((/>). (In Section 2.1, we explained the case d 4. It is easy to go back and change the formulae to be valid for a general value of d) trivial field theory, for it is =

2.2 For the

case

of

Functional-integral solution

(/>4 theory, P((/>)



with

(2.2.1)

m2</>2/2 + #74!,



are rigorously known to exist if d 2 or 3 (Glimm & Jaffa (1981)). If d > 4 then no non-trivial solution exists (Aizenman (1981)). The case d 4 is difficult; the difficulty is to perform renormalization of the ultra-violet

solutions





divergences beyond perturbation theory. As we will see the theory at d 4 is 'exactly renormalizable' in perturbation theory; this is the most interesting case. For the most part we will ignore the difficulties in going beyond perturbation theory. We will return to this problem in Section 7.10 =

when

we

discuss the application of the renormalization group outside of

perturbation theory. If we ignore, temporarily, the renormalization problem, then a solution for the theory can be found in terms of a functional integral. The formula for the Green's functions is written

GN(xl9...,xN)



as

jr

[d^]eis^M(x1)...^l(xJV).

(2.2.2)

(See Chapter 9 of Itzykson &Zuber (1980), or see Glimm &Jaffe(1981).)On the

right-hand

integration is

side of this equation ^(x) represents a classical field, and the the value of ^(x) at every space-time point. The result of

over

the integral in (2.2.2) is the N-point Green's function for the corresponding quantum field, (/>. In the integrand appears the classical action, which is

S|>4]= d4*^- The normalization factor Jf is to



Equivalent

to

(2.2.2) is



the

(2-2-3)

give <0|0>

{[[^]eiSM1j integral



1,

so

that



for the

(2.2.4)

generating functional

of

Green's functions:

Z[J]



Jf

I[(L4]exp jiS[v4]



\d*xJ(x)A(x) 1,

(2.2.5)

where J(x) is an arbitrary function. Functionally differentiating with respect to J(x)

gives

the Green's functions, e.g.,

<0|T^(,)|0>



^^^ZW|^o)

It is somewhat delicate to make over A.

The

principal steps

are:

precise

the definition of the

(2.2.6) integration



theory

Quantum field

(1) 'Wick-rotate' time

to

imaginary

values: t

Euclidean. The exponent in the

"SEudM= With

our

metric,

we



integral





i-r,

[drd3x[-dA2/2

have 3 A2





so

that space-time is

is then:

(dA/dr)2



(2.2.7)

P(A)l



VA2. We may subtract

P(A); this subtraction gives an overall factor in the functional integral, and it cancels between the integral and the normalization factor (2.2.4). Therefore the Euclidean action 5Eucl is positive definite. The factor exp(-SEucl) gives much out from ¥£ the minimum value of

better convergence for large A and for

rapidly varying A than does in Minkowski exp(iS) space. (2) Replace space-time by a finite lattice. We may choose a cubic lattice with spacing a. Its points are then x" where the n^s are

box



n'a.

integers. They are bounded to keep x inside a spatial T/2 to + 7/2. The keep t within a range

of volume V and to



integral

t[dA}A(xl)...A(xN)exp(-SEucl[A]) is of

now an

absolutely convergent ordinary integral

variables.

The

action

approximation. (3) Take the continuum limit

5Hucl

a ->

infinite time T. (4) Analytically continue back

is

given

over a

(2.2.8) finite number

obvious

its

discrete

0, and the limits of infinite volume V and

to Minkowski

space-time.

The difficulties occur at step 3. Taking the limits of infinite T and V gives divergences of exactly the sort associated with taking the thermodynamic limit of a partition function see below. Further divergences occur when -

the continuum limit

(2.2.2) gives or

as

an

a ->

overall normalization factor which goes to

the number of

absorbed

by

0 is taken. In addition, the canonical derivation of

space-time points

goes to

infinity as a ->0 infinity; this factor is

the normalization Jf.

The limits of infinite volume and time

are

under

good

control. They

are

literally thermodynamic limits of a classical statistical mechanical system in four spatial dimensions. Recall, for example, that in write

S[A]=g ̄l

cj)4 theory

\d*x(dA2/2-m2A2/2-A4/4\)

one can

(2.2.9)

2.3 Renormalization =

gulA.

Thus the integral

J<

where A

f [cL4] exp(





S[/l]) is

proportional to

f[(L3]exp{-(l/fif)S[yl-^ff-l]}.

(2.2.10)

This is the partition function of a classical system at temperature l/g, when the phase space is spanned by the field A, and when the energy of a given configuration is

^EuclMJ



d4x(



3A2/2 + m2A2/2 + A*/4\).

identity between Euclidean field theory and certain classical statistical mechanics systems has been fruitful both in working out the rigorous The

mathematical treatment of quantum field theory (Glimm & Jaffe (1981)) and in finding new ways to treat thermodynamic problems (Wilson & Kogut (1974)). As is particularly emphasized in Wilson's work, there is a lot of cross-fertilization between field

theory

and the theory of

transitions. The methods of the renormalization group fields, and the continuum limit in field theory can be

phase

are common to

both

usefully regarded as a particular type of second-order phase transition. The thermodynamic limit gives a factor exp( pTV), where p is the ground state energy-density. This factor is clearly cancelled by Jf. All the remaining divergences are associated with the continuum limit a -> 0. These are the divergences that form the subject of renormalization. They are —

called the ultra-violet (UV) divergences. One notational change needs to be made now. In more complicated theories, there will be several fields, and the functional-integral solution of

theory involves an integral over the values of a classical field for each quantum field. It is convenient to have a symbol for each classical field that is clearly related to the corresponding quantum field. The standard notation is to use the same symbol. Thus we change the integration variable such



in (2.2.2) from A(x) to 0(x), with the result that

<O|T(/>(xJ...0(x^)|O> This is somewhat of whether

is

one

or to mean

an

using (/>



^|1[d0]eiS^(xO..>(%)-

(2.2.11)

abuse of notation. However, it is usually obvious quantum field, as on the left-hand side,

to mean the

the corresponding classical field,

as on

the

right-hand

side.

2.3 Renormalization

divergences in this limit; this has been known from the earliest days of quantum elec-

The difficult limit is the continuum limit a

-?

0. There

are

10

Quantum field theory

trodynamics (e.g., Oppenheimer (1930)). It is possible to say that the UV divergences mean that the theory makes no physical sense, and that the subject of interacting quantum field theories is full of nonsense (Dirac (1981)). Luckily we can do better, for our ultimate aim need not be to construct a field theory literally satisfying (2.1.2)—(2.1.5). Rather, our aim is to construct a relativistic quantum theory with a local field as its basic observable. These requirements are satisfied if we construct a collection of Green's functions satisfying sensible physical properties (for example, as formulated in the Osterwalder-Schrader axioms (1981)). We may further ask that

we

find





see

Glimm & Jaffe

theory that is close in some sense

satisfying the defining equations (2.1.2)—(2.1.5). Combining the integral with suitable renormalizations of the parameters of theory satisfies these requirements. to

functional

The basic idea of renormalization

one-loop

comes

the

from the observation that in

graphs divergences example, they change the mass of the particles described by cj)(x) the

amount to shifts in the parameters of the

action. For

from the value

m to some

other effective value, which is infinite if m is finite.

Renormalization is then the procedure of cancelling the divergences by the action. To be precise, let us consider the c/>4

adjusting the parameters in

theory

with <£



(3A0)2/2



m20A2/2



gA*/41



A0.

(2.3.1)

subscript zero is here used to indicate so-called bare quantities, i.e., those that appear in the Lagrangian when the (3A0)2/2 term has unit coefficient. (We also introduce a constant term. It will be used to cancel a UV divergence in the energy density of the vacuum.) Then we rescale the

The

field by writing A0 so

that, in





ZdA2/2 ZdA2/2

dropped A0 from





(2.3.2)

Zil2A,

terms of the 'renormalized field'

<£

We have



A, the Lagrangian is

m2ZA2/2 g0Z2A*/4! m2A2/2 gBA4/4\. -



SC since it has

no

effect

The Green's functions of the quantum field <f> (2.3.3) as the Lagrangian in the functional integral

on

(2.3.3)

the Green's functions.

obtained by using (2.2.2). We let Z, m0, and

are now

g0 be functions of the lattice spacing a, and we choose these functions (if possible) so that the Green's functions of </> are finite as a-> 0. If this can be

constructing a continuum field theory, and it is termed'renormalizable'. The theory may be considered close to solving

done, then

we

have succeeded in

11

2.3 Renormalization

(2.1.2)—(2.1.5). This is because the theory is obtained by taking a discrete (i.e., lattice) version of the equations and then taking a somewhat odd continuum limit. We will call m0 the bare mass, and g0 the bare coupling, and we will call Z the wave-function, or field-strength, renormalization. It is also common to call mB and gB the bare mass and we will not do this in this book.

viewing the

Another way of c£



coupling;

but for the sake of consistency

renormalization is to write (2.3.3)

as

dA2/2 m2A2/2 gA*/4! + 6Z3A2/2 6m2A2/2 SgA4/4!. -





(2.3.4)



We will call the first three terms the basic

Lagrangian and the last three the

mass m and the renormalized Lagrangian. coupling g are finite quantities held fixed as a->0. The counterterms Z SZ 1, Sm2 m^ m2, and Sg gB g are adjusted to cancel the divergences as a-^0. This form of the Lagrangian is useful in doing perturbation theory; we treat 3A2/2 m2A2/2 as the free Lagrangian and

The renormalized

counterterm















interaction. The expansion is in powers of the renormalized coupling g. The counterterms are expanded in infinite series, the remainder

as

each term cancelling the divergences of one specific graph. The form (2.3.4) for S£ also exhibits the fact that the theory has

independent parameters, and of



and g. The counterterms

are

two

functions of m, g,

a.

We will discuss these issues in much greater depth in the succeeding chapters. For the moment it is important to grasp the basic ideas:

(1) The self-interactions of the field create, among other things, dynamical contributions to the mass of the particle, to the potential between

particles, and to the coupling of the field to the single particle state. Thus the measured values of these parameters are renormalized relative to the values

(2)

appearing

These contributions,

in the or

Lagrangian.

renormalizations,

are

infinite, in many

cases.

The most important theorem of renormalization theory is that they the only infinities, in the class of theories called 'renormalizable'. (3) The infinities

counterterms,

are

cancelled by wave-function, mass, and coupling that the net effect of the interactions is finite.

are so

quantitative the sizes of the infinities, the theory is constructed a lattice theory. The infinities appear as the lattice when divergences spacing goes to zero.

(4) To make as

the continuum limit of

12

Quantum field theory 2.4 Ultra-violet regulators

In the last sections

we

showed how to construct field theories

the functional integral

Ultraviolet zero,

as

the continuum limit of

by defining

lattice theory.



divergences appear as divergences when the lattice spacing, a, goes to and are removed by renormalization counterterms. The lattice

therefore is a

To be able

regulator,

or

cut-off, for the UV divergences.

divergences quantitively and to construct a theory involving infinite renormalizations, it is necessary to use some kind of UV cut-off. Then the theory is obtained as an appropriate limit when the cut-offis removed. There are many possible ways of introducing a cut-off, of which going to a lattice is only one example. The lattice appears to be very natural when

to discuss the

working with the functional integral.

But it is cumbersome to

within perturbation theory, especially because of the loss of Poincare invariance. There are two other very standard methods of making an use

ultraviolet cut-off: the Pauli-Villars method, and dimensional regularization. The Pauli-Villars (1949) method is very traditional. In its simplest version it consists of replacing the free propagator field theory by

SF(p,m;M)





i/(p2



m2)

in



scalar

p2-M2

(m2 M2) (p2-m2)(p2-M2)' -

(2A1)

approaches the original propagator. The behavior for large been clearly improved. Thus the degree of divergence of the Feynman in the graphs theory has been reduced. All graphs in the t/>4 theory, except for the one-loop self-energy are in fact made finite. In the </>4 theory it is necessary to use a more general form in order to make all graphs finite: As M

->

oo, this

p has

SJp,m;MuM7) FKy i9 V

=—

\- (p2-M*) (M\-M¥)

—-^

(p2-m2)

(m2









M2)

 ̄(p2-m22)(m2-mJ) i  ̄ ̄

(p2 It is

usually convenient

Now the

and

p2





(Ml-m2)(M22-m2) m2) (7-M2) (p2 M22)' -

to set Mx M2. regulated propagator has extra poles at p2

M2. Since

to the pole at

p2



one

(2.4.2)



of the extra poles has





M2, or at p2



M\

residue of the opposite sign

m2, the regulated theory cannot

be

completely physical.

2.5 Equations It is normally true unphysical features.

that a

of motion for Green's functions

theory with

an

13

ultra-violet cut-off has

some

Perhaps the most convenient regulator for practical calculations is dimensional regularization. There it is observed that the UV divergences are removed by going to a low enough space-time dimension d, so d is treated

as a

continuous variable. In perturbation theory this

can

be done

consistently (Wilson (1973)), as we will see when we give a full treatment of dimensional regularization in Chapter 4. However it has not been possible to make it work non-perturbatively, so it cannot at present be regarded as a fundamental method. Since it is only the renormalized theory with no cut-off that is of true interest, the precise method of cut-off is irrelevant. In fact, all methods of ultra-violet cut-off differences

are

equivalent,

are

mainly

a matter

of

at least in perturbation theory. The practical convenience (or of personal

taste). Thus dimensional regularization is very useful for perturbation theory. But the lattice method is maybe most powerful when working

beyond perturbation theory; it is possible, for example, to compute the functional integral numerically by Monte-Carlo methods (Creutz (1980, 1983), and Creutz & Moriarty (1982)). Within

perturbation

theory

one

need

not

even

use



cut-off.

Zimmermann (1970, 1973a) has shown how to

apply the renormalization procedure to the integrands rather than to the integrals for Feynman graphs. The lack of fundamental dependence on the procedure of cut-off is thereby made manifest. The application of this procedure to gauge theories, especially, is regarded by most people as cumbersome. 2.5 Equations of motion for Green's functions We have defined



collection of Green's functions by the functional integral the definition are a certain number of limiting below (2.2.6).) This definition we will take as the basis

(Implicit in procedures, as listed

(2.2.2).

for the rest of our work. First

we must

check that it in fact gives a solution of that we are to derive the equations of

the theory. This means, in particular, motion (2.1.10) for the Green's functions, thus ensuring that both the

operator equation of motion (2.1.2) and the commutation relations (2.1.5) hold. (For the remainder of this chapter we will not specify the details of how renormalization affects these results.) It is convenient to work with the generating

functional (2.2.5). We make where e is a small number, and A(x) A(x) ef(x), of x". an Since the is function integration measure is invariant f(x) arbitrary the change of variable

->



14

Quantum field theory

under this shift, the value of the integral is unchanged:

|[d^]expjiS|>4+e/]

M+e/)j[= |[d^]expji5[^]+



IA

A. (2.5.1)

Picking

out the terms of order e

dVW where,

as

usual,

gives

[cL4]exp^iS|X|

we





8S =

SA(y)

0,

define the functional derivative SS

dP

(2.5.3)

SA(y) ̄ ̄nA ̄dA' _

Since f(y) is arbitrary,

we

get

[di4]exp|iS[^]+ J

(2.5.2)

\aj\

SS





8A(y)

0.

(2.5.4)

Functionally differentiating JV times with respect to J, followed by setting 0, gives the equation of motion (2.1.10). For example,



(5





,/Ttttt [left-hand side of (2.5.4) ]J=0 SJ(x)









jr

JT



i[dA]eiSlA]

A(x)i

SS

SA(y)



Sw(x-y)

[dA]eiSW{iA(x)[- UA(y)- P'(A(y))] + 3w(x-y)}

J^iDy\[dA']e[SiA]A(x)A(y)

-in/O|70(x)</>(y)|O>-i<O^^ (2.5.5)

which is equivalent to (2.1.9).

Note that in the fourth line we have exchanged

the order of integration and of differentiation for the □ y term. We have also used the normalization condition (2.2.4). It is important that the derivative of the quantum field (next-to-last line) is outside the time ordering, and SS/S(j)(y) in the last line is defined to be a shorthand for the combination of

operators in the previous line.

15

2.6 Symmetries This is somewhat

paradoxical since

we

have the operator equation of

motion:

from

<0|

which

it

is

tempting to deduce that the Green's function Tcj)(x)SS/d(j)(y)\Oy should be zero. However, in view of the work above

it is convenient

by the functional-integral

to define this Green's function

formula

<0lT^|0>=^j[d^eiS[^(x)^ Then,

have seen, the Dy is implicitly outside the time-ordering. inside the time-ordering gives a commutator, so that we get the

as we

Bringing it

(5-function term in (2.1.9)

or (2.5.5). The momentum-space version of the equation of motion (2.1.10) is often useful. We define the momentum-space Green's functions

GN(pt,



? ?,

PN)



GN(Pi,..., pN){2n)*8?{Pl



The momenta pj

d4xt.. .d% exp {i(pl-xl

be regarded





















pN-xN)}GN(xu..., xN) (2.5.6)

pN).

flowing theory implies the (5-function for momentum conservation that is explicitly factored out in the last line of (2.5.6). A convenient notation (which we will use often) is to write are to

out of the Green's functions.

as

Translation invariance of the

GN(p1,...,pJV)



<O|r0(p1)...^(pN)|O>.

(2.5.7)

Implicit in this formula is the definition that the integrals over x defining the momentum-space field </>(p) are all taken outside the time-ordering, as stated in (2.5.6). We will use a tilde over the symbol for any function to indicate the Fourier-transformed function. Fourier transformation of the

equation

of motion

(2.1.10) gives

-q\0\U(q)¥(Pl)...4>(pN)\0y + <0\TP\¥)(q)¥(pd...¥(pN)\0> =

-i

I <O|T<£(pJ...0(p._J^

7=1

2.6

We

Symmetries

the consequences of symmetries. As we will see, interesting problems in renormalization theory that stem

now turn to

many

(2.5.8)

there

are

from the

16

Quantum

field theory

following question: If

a classical field theory has certain symmetries, does the symmetry survive after quantization? Generally, it is the need for renormalization of the theory that makes this a non-trivial question. The symmetry properties are expressed in terms of Green's functions by

the Ward identities. (Historically the earliest example was found by Ward (1950) in QED.) If the symmetry is not preserved by quantization there are extra terms called anomalies. In many cases there are no anomalies, so we will derive the Ward identities in this section ignoring the subtleties that in some cases

Chapter

lead to anomalies. Discussion of anomalous

cases

is given in

13.

theory of N fields which we collectively denote by a vector 0 (0,,..., </>N). Our discussion is general enough to include the case of fields with spin. We consider a symmetry group of the action S[0]. This is a Consider





group of transformations

on

the classical fields

0-+F[>;ctf]



(2.6.1)

0',

which leaves the action invariant:

S[*'] a

(2.6.2)

S|>].

w (w<z) is a set of parameters of the group, which we assume here to 0 Lie group, i.e., the a/s take on a continuous set of values. We let w

Here

be



be the





identity: F[0;O]



0.

It is easiest to work with infinitesimal

transformations:

dF- =



(A summation convention

on a

(2.6.3)

aA5a0,

is understood.)

In the quantum theory the symmetry is implemented as a unitary representation U(a)) of the group on the Hilbert space of states such that

U(a>)jU{a>)-! Since the representation is unitary, U(o) where the generators Qa algebra of the group:

The normalizations

are

are



we



(2.6.4)

F[0;o>].

may

parametrize

exp(i(D"Q,\

the group

so

that

(2.6.5)

hermitian operators which represent the Lie

such that the structure constants cafiy are totally are then given by:

antisymmetric. The infinitesimal transformations <5^iW



i[a^,W]-

(2-6.7)

2.6 Symmetries There

(1)

are a

17

number of special cases, each with its

special features:

own

Global internal symmetry: A finite-dimensional Lie group acts on the point of space-time, with the same transformation at each

fields at each

point. Thus

yf=-iWi% where the ta form



(2-6.8)

hermitian matrix representation of the Lie algebra:

Single-particle states carrying this representation Lagrangian is invariant.

are

annihilated by </>,..

The

(2) Global space-time symmetry: The group effectively is on

space-time; the Poincare



transformation

group and its extensions are the usual

cases. For a Poincare transformation xM->

corresponding transformation

A£xv



aM,

have the

we

of the fields:

(2.6.10)

(/>;(*)-+(/>.(Ax +a)jR(A)V

Here R is a finite-dimensional matrix representation of the Lorentz group (never unitary if non-trivial), acting on the spin indices of 0. The Lagrangian is not invariant. It transforms

as

J^[0,x]-+J^|>,Ax + a], so

that the action S



Jd4xif is invariant. Infinitesimal transformations

of (j> involve the derivative of (/>. (3) Global chiral symmetry: This looks like

global internal symmetry but differently right-handed parts of Dirac fields (which we have yet to discuss). Anomalies are often present see Chapter 13. (4) Supersymmetry: This is a generalized type of symmetry where Bose and Fermi fields are related (Fayet & Ferrara (1977)). The only case that we acts

on



the left- and



will discuss is the BRS-invariance (Becchi, Rouet & Stora (1975)) of gauge theory. (5) Gauge, or local, symmetry: Any of the above symmetries

symmetry whose parameters depend



may be

o(x). In quantum theories, these are not really implemented by unitary transformations. Their treatment is rather special. The elementary examples are general coordinate in variance in General Relativity, and extended to



gauge invariance in The basic tool for

on x.w

electromagnetism.

discussing symmetries

is Noether's theorem, which

relates them to conservation laws. This theorem in its most

form

applies only

to



symmetries

of the first three types.

straightforward

18

Quantum For

Noether's theorem asserts that

global symmetry,



field theory a

conserved

current^ exists for each generator of a symmetry. Let the Lagrangian have

the infinitesimal transformation S£-+S£ + ofb^ so



&+

(2.6.11)

co*d^ Y£,

that the action is invariant. Define

Then the

equations

of motion

imply

dj£/dx? The generators of the symmetry

conservation of y£, i.e., =

group

(2.6.13)

0.

are

Q^Wxjl

(2.6.14)

The canonical commutation relations imply that

tiMx)#3\x [ft?^)]=-i^M

D,°W, </?,(>-)]

as







y),

(if *°



A (2-6.15)

required by (2.6.7). only transformations that are symmetries of the 'broken symmetries'. There are several cases (not

We need to consider not

quantum theory, but also mutually exclusive). Let us define them, since there is



certain amount of

confusion in the literature about the terminology: (1) Explicit breaking: The classical action has a non-invariant term. If 6aJ? d^Y£ + Aa then the Noether currents are not conserved: =

Aa. An important case is where this term is small, so that it can be as a perturbation. (2) Anomalous breaking: Even though the classical action is invariant, the quantum theory is not, and there is no conserved current. The classical

dj*



treated

action is important for the quantum theory, since it appears in the functional integral defining the theory. The cause of anomalous breaking is generally an ultra-violet problem: dj% ^ 0 in the UV cut-off

theory, and the non-conservation does not disappear when the cut-off is removed. (Cases

are

conformal transformations and

some

chiral

theories.) (3) Spontaneous

breaking:

The action is invariant and the currents

conserved (in the quantum

theory),

but the

vacuum

are

is not invariant

under the transformations. Whether

or not a

symmetry is broken either spontaneously

anomalously is a dynamical question. That is,

one must

or

solve the theory, at least

19

2.7 Ward identities

partially, to find the answer. Frequently, perturbation theory is adequate to do this and

lowest

order or next-to-lowest order calculations suffice.

integral part of treating anomalous breaking (see renormalization-group methods are sometimes

Renormalization is an

Chapter 13),

while

necessary in treating spontaneously broken symmetry (Coleman & Weinberg (1973)). The case of spontaneous symmetry breaking that is not visible in perturbation theory is often termed dynamical (Jackiw & Johnson (1973), Cornwall & Norton (1973), and Gross (1976)). Anomalous breaking is sometimes called spontaneous, but this is a bad terminology, because it

gives

two very different

phenomena the

same name.

2.7 Ward identities Ward identities express in terms of Green's functions the consequences of a

it is broken). One derivation applies the equation of motion (2.1.10) to the divergence of a Green's function of the current jj. There are two terms: one in which the current is differentiated, and one in

symmetry (whether

or not

which the ^-functions

defining the time-ordered product are differentiated. identity expresses not only conservation of its current but also commutation relation (2.6.15), which is equivalent to the

Thus the

Ward



transformation

The

law.

Ward

renormalization of a

identities

are central

to

a discussion

of the

with symmetries, expecially if spontaneously

theory

broken. Our derivation of Ward identities begins by making the following change of variable:

Ai(x)-+Ai(x)+f(x)8aAi(x) in the functional integral for the generating functional Z[J]. before, the variation of the field A{ under fa(x) is x -?

oo.

a set

we

variables

0=



HereyJ,

Here

SaAt is, as

symmetry transformation, and



of arbitrary complex-valued functions that vanish rapidly

as

We get

Z[J] (Here

(2.7.1)



j[dA]

exp

{iS[A +/*<5aA]

assumed that the

(2.7.1).) The

measure

terms in



J^. +f'8gcAi)}.

is invariant under the

(2.7.2) that

are



exp

(iS



jJ-



W/^OO



change

of

linear in f" give

f[dA]expMS fj-AV-5S[A+/^aA]/5/^)

f[dA]

(2.7.2)

i^My)}-

is the Noether current in the classical theory.



iJ'M.-()')} <2-7-3)

20

Quantum

field theory

The Ward identities follow by functionally differentiating with respect to J(x). Thus one differentiation gives

the sources

^7<0| r£0#,(x)|0>





oy

while a double differentiation

i<5<4>(*



yK0\SMy)\0>,

(2.7'.4)

gives

^<O|77^)0f(w)^(x)|O> -iS^w-yK0\T5My)<t>jM\°>





i<5<4>(*



><)<0| T0,(w)^(/>.(y)|O>.

(2.7.5)

Note that, just as in our derivation of the equation of motion for Green's functions in Section 2.5, the derivative d/dy" is outside the time-ordering. The general case is:

A<oiTOn*j*,)io> °y











X <5(%





*/)<0| T^GO J] ^,.W|0>.

(2.7.6)

Important consequences of these Ward identities are obtained by integrating over all y (with y° fixed). The spatial derivatives give a surface term, which vanishes, so that we have, for example,

jd3>'^o <0| 7£0#,(x)|0> The spatial



i<5(*°



/)<0|^f(x)|0>.

integral of;'0 is just the charge Q°. The time derivative acts either

the charge that on



or on

the ^-functions defining the time-ordering;

<0|rdea/dt^(x)|0> + <0|[ea,^(x)]|0>?5(x0-y0) -i5(x°-y°K0\SMy)\0y. =

In this

equation

find

(2-7.7)

(2.7.6), we may choose the y°. Therefore an arbitrary operator dga/dt is zero. The

and its generalizations from

times of the fields

</>f(xf)

not to coincide with

Green's function of dga/dr is zero, remaining part of (2.7.7) therefore

so

that the

gives:

<O|[ea,0,-M]|O>= -i<0|^.(y)|0>. From

so we

(2.7.8) and its generalizations with

more

fields,

we

(2.7.8) find that the

ga's

have the correct commutation relations with the elementary fields 0f to be the generators of the symmetry group. Finally, another specialization of (2.7.6) is to integrate it

over

all y* and to

21

2.8 Perturbation theory

drop

the

resulting surface

term. The result is that

I<o|r5A/*;)n6,((*.)|o>

o=

<5a<0|7n^,(xi)|0>.



(2.7.9)

i= 1

All the above

equations

are true

for the

case

of



completely

unbroken

symmetry. The derivation breaks down at the first step if we have anomalous breaking. (In Chapter 13 we will discuss the anomaly terms that

must then be inserted in the Ward identities to make them correct.) For

explicitly

broken symmetry, where d-j



Aj=Q,

we must

add

an

a term

<0\TAx(y)Y\4>n,(XiP>

(2-7.10)

* = 1

to the

side of (2.7.6).

right-hand

In the case of a spontaneously broken theory the basic Ward identities (2.7.4)-(2.7.6) remain true-we still have an exact symmetry. But the integrated Ward identities (2.7.7) and (2.7.9) are no longer true. Equation (2.7.9) must be false if the vacuum is not invariant, and the derivation fails because the surface term is not zero. This is caused by the existence of zero-

particles. These Weinberg (1962)) are

mass

Nambu-Goldstone bosons (Goldstone, Salam & characteristic of theories with a spontaneously

broken symmetry.

2.8 Perturbation theory As

an

example, consider again (/>4 theory, with classical Lagrangian S£

We will



Z(dA)2/2



m2BA2/2



(2.8.1)

gBA*/4L

the Green's functions in powers of the renormalized

expand coupling g, for small g. powers of g,

we

To

expand the functional-integral formula (2.2.2) in

write i?



J5?0



(2-8-2)

^i>

where if0 is the free Lagrangian: (2.8.3)

&0=(dA)2/2-m2A2/2, and <£x is the interaction Lagrangian: Se,





We will

gA*/4!

expand



(Z



l)(dA)2/2



(m2



m2)A2/2



the renormalization counterterms, Z

(gB —



1,

g)A*/41 (2.8.4)

m|



m2,

and

22

Quantum

gB



field theory

0,in powers of #,so that all of the terms in if, have at least one power of expansion of the Green's functions is then obtained from (2.2.2)

g. The series as:

GN(x1,x2,... ,xN)



(i"/n!)



[[d^x,)- ̄A(xJ fd4y^iW

exp(iS0M)

(2.8.5)

f[d^]|" fdV^wTexp^o^])

(i-/n!)

Here

[d4y^o= [d4y(^2/2-m2/(2/2)

SoW= is the free action.

Each of the terms in the series is (aside from



Green's function in the free-field theory

normalization), (1951) formula:

a common

Mann-Low

so

(2.8.5) is equivalent

to the Gell-

uN(xl9... x;V) ,

_

£ (iWfl U*yJ)<0\T<t>F(x1)...ct>F(xN) f[ J^(yi)|0> \j=l J

n^O



_

jfj



tr/ni)(f\ Uyj]<o\Tf\ ^(yj)\oy (2.8.6) Here t/>F is a free quantum field of mass m. It is the field generated from the free Lagrangian if0 (d(j)F)2/2 m2t/>F/2. Then i^ is the quantum interaction Lagrangian, if j£f0, which is a function of the free field 0F. =





To compute the integrals in (2.8.5) it suffices to compute the functional of free-field Green's functions:

Z0[ J]



J-

|[<U]expl iS0IXI+ yA J- ^ ^ -r

generating

(2.8.7)

[d^]exp(iS0[^]) This is done by of variable:

completing the

A(x)

Here, GF(x), is



square, i.e.,

A(x)



by making the following change

fd*yGF(x



y)J(y).

(2.8.8)

the Feynman propagator satisfying (□ +

m2)GF(x)





i<5(4)(x),

(2.8.9)

23

2.8 Perturbation theory and

x°

boundary condition that,







it, GF(x)

-?



as x -?

oo.

after rotating to Euclidean space by

Thus

The result is that

Z0[J] Green's functions



exp

j± fd4xd V(*)GF(*-y)J(y)\.

of free fields

(2.8.11)

obtained by differentiating with respect

are

to J; for example

(0\T<t>(x)<t>(y)\0}=- SJ(x)SJ(y)\J





(2.8.12)

GF(x-y).



derive the well-known Feynman rules for the interacting from (2.8.6). These can be given either in momentum or coordinate theory space. In either case the Green's function GN is written as a sum over all We

can now

possible

topologically distinct Feynman graphs. Each graph T consists of a by lines. It has N 'external vertices', one for each

number of vertices joined

0(xf),

with

one

line attached, and

number,

some

n, of interaction vertices.

corresponding to the terms in Lagrangian (2.8.4). The vertex for the A* interaction has four

The interaction vertices are of several types, the interaction

lines attached and the vertices for the dA2 and A2 interactions have two lines attached. The value of the

position yj of the



graph, denoted

interaction vertices. The

integral over the integrand is a product of

/(T), is the

factors: (1)

GF(w

z) for each line, where





and

z are

the positions of the vertices at

its end. (2) A combinatorial factor 1/S(r). (3) i#B for each A4 interaction. —

(4)

(5)





i(m| i(Z





m2)

for each A2 interaction.

l)d2/dw"dwM

respect to

for each

w act on one

(dA)2

interaction: the derivatives with

of the propagators attached to the vertex.

For each Feynman graph a number of equal expanding (2.8.5). If T has no symmetries and if

contributions arise in it has no counterterm

vertices, then this number is n !(4!)" so that the explicit n! in (2.8.5) and the 4! in each interaction

are

cancelled.

Graphs with symmetries have a number of

factor of the symmetry number 5(r). (For by 6.) The combinatorial example, the self-energy graph Fig. 2.8.1 has 5

contributions smaller





factor is then the inverse of S(T).

24

Quantum field

theory

CO Fig. 2.8.1. A graph with symmetry fac- tor 5

Fig. 2.8.2. A graph with

6.





vacuum

bubble.

is the sum of all graphs with no external lines. graphs in the numerator that have disconnected

The denominator of (2.8.5)

The result is to cancel all vacuum bubbles

(like Fig. 2.8.2.).

In momentum space each line is

Feynman rules

(1)

A factor

(2) A factor

i/[(27r)4(/c2 (2n)4'

or

m2



times

vertex (external

(3) An integral

assigned



(directed)

momentum k. The

are:



ia)]

for



a momentum

line with momentum k.

conservation ^-function for each

interaction).

the momentum of every line. (4) A combinatorial factor 1/S(T). (5) i#B f°r eacri ^4 interaction. over



(6) (7)

\{ml m2) for each A2 interaction. i(Z \)p2 for each (dA)2 interaction, where p is the momentum flowing —





of the propagators attached to the vertex.

on one

The perturbation series in (2.8.5) need not be convergent, but only asymptotic. Let GN be the sum up to order gn of the perturbation series for ?

the Green's function GN. Then it is asymptotic to GN if for any satisfies

\GN ̄GNJ as g





the

error

(2.8.13)

0(gn+')

general, perturbation theory is asymptotic but not convergent. rigorously known (Glimm & Jaffe (1981)) for the </>4 theory in the

->0. In

This is cases

that the space-time dimension is d



0,1, 2, 3. (d



0 is the

case

of the

ordinary integral dx exp ( while d





m2x2/2



gx*/4!),

1 is the quantum mechanics of the anharmonic

Physically, the reason for non-convergence is that when g unbounded below and

so

oscillator.)

the energy is

the vacuum-state continued from g>0 is phenomenon in quantum

unstable. (Dyson (1952) first observed this

electrodynamics.)

< 0

Spontaneously broken symmetry

2.9 In later

chapters

we

will

assume

(2.8.13).

momentum behavior, it will be important

When

25

compute large-

we

to understand the maximum

possible validity and accuracy of the calculations if the perturbation theory is asymptotic but not convergent. 2.9

Spontaneously

broken symmetry

(/>4 interaction. If m2 is positive and g is small, we have a theory

Consider the

of particles of mass

slightly perturbed by the interaction. This interaction is basically a repulsive (5-function potential, as can be seen by examining the Hamiltonian in the non-relativistic approximation. There is a symmetry m

But if m2 is negative this interpretation is incorrect. The true situation can

by noticing that the functional integral (in Euclidean space) is dominated by classical fields with the lowest Euclidean action, which is be discovered

SEuciM (Remember



(dA)2

that

[d4x[ =





are constant

A_



(2.9.1)

(dA)2/2 + m2A2/2 + gA*/4i].

is negative.) If m2 > 0, then 0. But, if m2 < 0, then there are two minima;

-(dA/dz)2-VA2

the minimum action field is A

these





fields with P'(A)



0, i.e., A=A+

-J(-6m2/g).

=y/(



6m2jg)

and

in the impose the boundary condition A(x) ->A integral. (The condition A^A_ gives equivalent physics, A symmetry of the action.) Then field configurations because of the A with A close to A will dominate. We may understand this by observing A that field configurations with large regions where A is not close to A will give small contributions to the Euclidean functional integral (2.2.8) because their action SEucl is so big. Indeed, a constant field with A not equal to A A and its A has infinitely more action than one with A A contribution to the integral is zero. One's first inclination then is that the only configurations that contribute have A^A+ or A^A_ as x->oo. We choose to



as x ->oc

functional

->









or



_



However, other configurations contribute, because there

or

are

or

_

_,

many of

integrate over all possible fluctuations. However, one can that in general A will be argue even rigorously (Glimm & JafTe (1981) A typical configuration of the classical field ^(x) will to A close to A be close to one of these values over almost all of space-time. them

one



has to







_.

choice of boundary condition A^>A + , even more is true: a configuration is close to A A + almost everywhere, rather than to

Given

typical

or

either A

our



_

or A +.

The reason is that if it had



large region with A(x) close to

26

Quantum field theory A ±A

















d=\

d>\

Fig. 2.9.1. Illustrating transitions between regions with fields close minima of the potential

A_ (Fig. 2.9.1), then there would be

to different

contribution to the action



proportional to the size of the boundary between the regions and of negative A(x). Only if the space-time dimension is d

of positive A(x) —

1 will

have

we

finite contribution from the boundary. This special case is quantum mechanics of a particle in a potential with two wells. The particle can tunnel a

between the two wells. In the

have discussed, of a discrete rather than of a continuous the argument that A is close to A + almost everywhere for the symmetry, case we

important configurations is than

one.

correct in all

The quantum field therefore has

space-time dimensions greater expectation value close

a vacuum

to A+:

<0|<£(x)|0>





\&A ̄]A{x)e-S{A]^A





In the case of a continuous symmetry, there is a continuous series of minima of the potential. A field configuration can interpolate between different minima without going over a big hump in the potential. The only

from the gradient terms in the action. This suppresses that do not stay close to one minimum, but only in more configurations than two space-time dimensions. In one space-time dimension there is no penalty

comes

spontaneous breaking of

a continuous symmetry (Mermin & Wagner (1966), Hohenberg (1967), and Coleman (1973)). Perturbation theory can be considered as a saddle point expansion about

the minimum of the action. We write A(x) A\x) + v where v we treat A'{x) as the independent variable. We have =

if Here C

have











(3Af/2

V/2

theory

of





M2A'2/2

gv*/4!,

particles



and M2 of

mass

gvA'3/3\- gA'*/4\ =

gv2/2 + m2





with both



an







Now

(2.9.2)

+ C.

2m2

+.

0. We

A'3 and

an

now

A'4

interaction. The symmetry is hidden; its only obvious manifestation is in the relation between the A'3 coupling and the mass and Af4r terms: gv



M(3gyt2.

(2.9.3)

27

2.10 Fermions

We will show later that renormalization counterterms are correctly given by continuing in m2 from positive m2. The vacuum expectation value of cf> has corrections which can be computed in perturbation theory

<0|</>|0>



? +

<0|</>'|0>



v +

0(g1/2).

(2.9 A)

Exactly similar methods can be applied if there is a continuous symmetry. Then the Goldstone theorem tells us that there will be a massless scalar particle for each broken generator. 2.10 Fermions The field theories obtained by functional integration as in Section 2.2 all theories of bosons. This follows from the symmetry of the Green's

are

functions

under

exchange

<0|7\/>(y)<£(x)|0>).

of

fields

<O|!T0(x)0()O|O>

(e.g.,



In turn, this symmetry property follows from the

functional-integral formula (2.2.2)

because the integration variables (the values of the classical field A(x)) commute with each other. To get a theory with quantized fields, it is necessary to define something like an integral over anticommuting variables. A rather small number of

properties of integration are needed to derive the equations of motion for Green's functions. Requiring these properties determines the integration operation uniquely (Itzykson & Zuber (1980)). As an example, consider the following Lagrangian £e Here





^ is

7*d^



generating

The fields and the a



free Dirac field:

(2.10.1)

¥(\¥-M)\l,.

¥ a row vector, while functional of Green's functions is written as:

four-dimensional column vector and

The

Z[rj,rj]=jr sector of



for

[[d^d^expji

J2?+

#+

UA

(2.10.2)

rj(x) and fj(x) take their values in the fermionic algebra. In the lattice approximation the definition (2.10.2) is really algebraic (Itzykson & Zuber (1980)).

sources

Grassmann

of the integration in

Green's functions

defined by differentiating with respect to the sources. One important difference between ordinary integration and Grassmann integration will be important in treating gauge theories. The simplest case are

{?

of this difference is in the integral over two variables x and where a is a real number. For ordinary real variables the dxdxeiiflx



27t

dx<5(xa)



2n/a.



of exp (ixax),

integral

is

(2.10.3)

28

Quantum field theory

For Grassmann variables

get

we

\dxdx&*ax

ia.



The overall normalization is irrelevant for

our

(2.10.4)

applications, for it is always

cancelled by the overall normalization factor in the functional integral. What matters is that the a-dependence of (2.10.4) is inverse to that in (2.10.3). 2.11 Gauge theories A gauge symmetry is an invariance under a group G where the group transformation is different at each space-time point. The earliest examples

General Relativity (where G is GL(4), the group of linear transformations of the coordinate system), and electrodynamics (where G is the were

rotations). Yang & Mills (1954) and Shaw (1955) generalized general group. Beg & Sirlin (1982) and Buras (1981) explain some of the uses of gauge theories as theories of physics. Let G be a simple group and let a matter field \\/ transform as group of phase

the idea to



Mx)-exp(



i^(x)t>(x)



(2.11.1)

l/(ai(x))"V(4

The field ^ is a column vector of components, and the hermitian matrices ta form a representation of the group, with structure constants caPy defined by

[W/J=i<Wr The matrices U(a)) form



(2-1L2)

representation of the group.

In order that the action be gauge invariant, we need



covariant

derivative:

(2.11.3)

D^^id^igA^

we have introduced the gauge potential A^. It is a vector under Lorentz transformation. As far as its gauge symmetry properties are or in terms of components concerned, it can be written as a matrix A^:

Here

AM

(2.11.4)

A,= I^r.. 3

It transforms under the gauge group

as:

\(x)-> UHxr'Wixi-ig-'d^UHx)). To build

an

action,

we

need the

which transforms as F-*U ̄lFU.

field-strength

tensor

(2.11.5)

2.12 A

gauge-invariant Lagrangian "



where

we

assumed In

tjp <5a/J/2. fields. If there are tr

Quantizing



with



spin \

nP



29

matter fields is

(2.11.7)

MW,

the conventional normalization

of the

ta's, viz.,

exactly similar way, an action can be set up using scalar matter fields in several irreducible representations a term

an

for each is needed in

coupling

F?VF"72

tr

gauge theories

<£. The transformation (2.11.5)

ensures

that the

g is the same for all matter fields if the group is non-abelian.

The form of the infinitesimal transformations is needed:

Ji 8,0*1 S

^F^ If the group is

not

e.g., SU(2)?SU(2).







simple,

igco^t^ ^co' + gc^caPAl,

dcapyFlv.

(2.11.8)

then it is the product of several simple groups, a gauge field and an independent

For each there is

coupling. 2.12 A gauge

theory

such

as

solved by the functional fermion propagator.

<o|rwx)0oo|o>

Quantizing

gauge theories

defined by the Lagrangian (2.11.7) can be integral. Thus, as an example, we can write for the

the



one

^^^

(2.12.1)

In fact,

a lattice approximation to the functional integral forms the basis of Monte-Carlo calculations (Creutz (1983) and Creutz & Moriarty (1982)). The only trouble with (2.12.1) is that it is exactly zero. To see this we observe

that, given any field configuration, we can make a gauge transformation on it, as in (2.11.1) and (2.11.5). The new field configuration has the same action as the old field. Thus the only dependence on the gauge transformation is in

explicit \^{x)^{y). Now the gauge transformation is independent at each space-time point. So (2.12.1) contains a factor

the



joints z



\dU(z))u-\x)?U(y)





(number))

\<iU(x)U(xyl V

dI7(y)l%)\

(2.12.2)

30

Quantum

which is

field theory

(Note that the propagator is

zero.



matrix in the representation

space of the gauge group.) The vanishing of (2.12.1) is not a fundamental problem, for we may choose only to work with Green's functions of gauge-invariant operators (e.g., tfty, F?vFa"v, the Wilson loop (Wilson (1974) and Kogut (1983))). But the

vanishing is



disaster for

formulating perturbation theory; for

the basic objects needed to write the Feynman rules the elementary fields. An elegant solution to this

Faddeev

are

among

the propagators for

problem

was

given by

& Popov (1967). The integral over all gauge fields is written as the of the integral over fields satisfying some given gauge condition

product (such as d'Aa 0) and of the integral over all gauge transformations. Any field configuration can be obtained by gauge transforming some configuration that satisfies the gauge condition. For a gauge-invariant Green's =

function, the integral over gauge

over gauge

transformations amounts to

an

overall

inverse factor in the normalization. So the integral transformations can be consistently omitted.

factor which cancels

an

The new integral over fields with the gauge-fixing condition imposed also provides a solution to the theory. But the gauge-variant Green's functions like (2.12.1) no longer vanish. It is necessary, moreover, to find the correct measure for the integral; this was the key point of the work of Faddeev and Popov. These authors also constructed



slightly

formulation that is most often used, and that

different formulation; it is this we

will review

treatment and further references are to be found in

Itzykson

A detailed

& Zuber (1980).

summarize the argument and derive the Ward identities in the form that we will use them. Here

we

will

now.

merely

We will consider gauge conditions of the form Fa[A,x] =/a(x). There is condition at each point of space-time and for each generator of the

one

group. The functional Fa might be d Aa, for any real valued functions of x. Faddeev and

<0|TX\0}

Popov write =

JfGl

an

example. The functions/a(x) are

arbitrary Green's function

as

f[cU][#][diAJXexp(iSlnv)A[^]\\S(Fa-/J.



x'"

(2.12.3) is the gauge invariant action, and X is any product of fields. The 5inv factor A|\4] is a Jacobian that arises in transforming variables to the set of fields that satisfy the gauge condition plus the set of gauge transformations. Here

The key result is that

A[A ̄\ A=

is



determinant,

so

that it

f[dcj[dcjexp(ii?gc).

can

be written

as

(2.12.4)

2.12 Here ca and ca are

ghosts. The

Quantizing

anticommuting scalar fields, called the Faddeev-Popov gauge-compensating Lagrangian is

so-called

^gc where

ScFa

For the

31

gauge theories



(2.12.5)

caScFa[A,xl

is the infinitesimal transformation of

case

Fa



Fa

with

a)

replaced by

c.

dA*

Se%c



d?c(d?ca



(2.12.6)

gcafi7cfiAl) + divergence.

independent fields. They are not genuine physical fields, as they do not obey the usual spin-statistics theorem. A convenient form of solution to the theory is obtained by averaging over  ̄1 ally's, with weight exp (— £ \fl/2). This leaves gauge-invariant Green's functions unaltered, and gives the following formula: We treat



and

c as

<0| TX\0} with



Jf

different normalization. The



The action

integral

is

over



ifgc).

(2.12.7)

all fields (A, \j/9



|d4x(i?inv



i?gf

The

gauge-fixing



c).

an

chosen parameter.

arbitrarily Fa.)

and

term is

J?gf=-iF,2/(2£), where £ is

c,

(2.12.8)

already defined the gauge-invariant Lagrangian by (2.11.7)

JSfgc by (2.12.5).

into

\J/,

S contains three terms: S

We have

I[d fields]Xeis

(If desired,

(2.12.9) it may be absorbed

redefinition of

The advantage of the form (2.12.8) is that Green's functions of the elementary fields

are

defined

invariant observable X the

objects

as

<0|T*|0> If

as in a simple non-gauge theory. For a gauge- equations (2.12.3) and (2.12.7) define the same



JTGI [d^d^di?]A:exp(i5inv).

(2.12.10)

variant, then all the definitions give different results, and (2.12.7) depends on £. Quantities that depend on the choice of a gauge fixing X is gauge

are

called gauge dependent, of course. We see that gauge in variance of the a Green's function implies gauge independence.

operators in

important to distinguish the concepts of gauge in variance and gauge independence. Gauge in variance is a property of a classical quantity and is in variance under gauge transformations. Gauge independence is a property It is

of a quantum quantity when quantization is done by fixing the gauge. It is

32

Quantum field theory

independence of the method of gauge fixing. Gauge invariance implies gauge independence, but only if the gauge fixing is done properly. Gauge theories such as (2.11.7) have a dimensionless coupling if space- time is four dimensional. General results, which we will treat in later chapters, imply that the theories need renormalization. However these same

results

available by

imply that many more renormalizing (2.11.7).

extra couplings

are

counterterms may

Chapter

In

absent. The tools needed

are

12

we

be needed than

are

will prove that the

the Ward identities for the

will prove in the next section. It is also necessary to prove that the unphysical degrees of freedom represented, for example, by the ghost fields ca and ca do not enter unitarity relations. This proof also gauge symmetry. These

needs the gauge Zuber (1980)).

we

properties exhibited in

the Ward identities (see Itzykson &

2.13 BRS invariance and Slavnov-Taylor identities

fixing, the gauge invariance of a gauge theory is no longer manifest in the functional-integral solution. Slavnov (1972) and Taylor (1971) were the first to derive the generalized Ward identities that carry the After gauge

consequences of gauge invariance. Their derivation

simplified by Becchi,

Rouet & Stora

was

very much

(1975) through the discovery of what is

called the BRS symmetry of the action (2.12.8). BRS symmetry is in fact a supersymmetry, that is, its transformations

now

involve parameters that take their values in a Grassmann algebra. Let 8X be a

fermionic Grassmann variable. Then the BRS transformation of a

or a

gauge field is defined to be

^brs^A

<Ws*A









gauge transformation with of

ig(c"SX)tail/





matter

caSL Thus

igt^ij/SX,

ig¥tac*SA9

Observe that 8k is fermionic, so it anticommutes with fermion fields (c, c, ^, \ji). The ghost fields transform as

<W'a

W.





-igcafiycpcy5l9 (2-13-2)

W?-

(Note that ca and ca are not related by hermitian conjugation, contrary to appearances.) Since ^inv is invariant under gauge transformations, it is BRS invariant. Hence

^BRS^







<W^gf + ^gc) (l/£)FaSMFalA;x] -

0.



(l/S)Fa5McFa



*Ars(W (2.13.3)

2.13 BRS invariance and

In the second line

we

used

anticommuted 8k and

we

addition



33

identities

to prove the last line zero in the first two terms of the second line. In

J^gc c



cx8cFx, while



nilpotence of

used the

we

Slavnov-Taylor

the BRS transformation:

\2

^fY*.



0,

(2.13.4)

which follows from anticommutativity of the c's. By applying the Noether theorem we find a conserved current: ih*



gfrtjc

F*">Dv<r





(l/w-A'iyd' (2.13.5)

-ig(d"ca)cl'cycalh.

the BRS transformations involve Grassmann-valued

Although

given in Section 2.7 goes through need the integrated Ward identity only

parameters, the derivation of Ward identities

unchanged. (2.7.9). A case

For

our

we

purposes, of (2.7.9) applied to BRS invariance is called a Slavnov-

Taylor identity. A simple example is 0 &H*s<0\TAl(x)c,(y)\0y/8A. =

-<0\T(dflcx + gc,Syc*A?)cl!(y)\Oy + {VZ)<0\TAHx)d-A'(y)\0>.



We have defined the notation

<5BRS (quantity)/<5A to

(2.13.6) mean that the 8). in the

BRS variation is commuted or anticommuted to the

The most used o





Here

X is a

right and then deleted.

of the Slavnov-Taylor identities

cases

are:

WO|rxc.M|o>/<5;. -<0\T(SBRSX/81)ca\0> + (l/ZK0\TXd-A°\0y.

product of fields with total ghost number equations of motion. Let:

(2.13.7)

zero.

We will also need

#*=-Zr-d?JT7r=W-MW, ^

Se^

<ei(t



?A(-i^-M),









DVF">"



gh'%^



i\/Wd-A' +

0C??(3"c%

^d"c",

Se',= -Uc.- gc^id^U". Then each of these is

theory

we

zero.

Furthermore, for <j> equal

(2.13.8) to any field in the

have; <o|rjs^(x)x|o>



i<o|Ta*/^(x)|o>,

(2.13.9)

34

Quantum field

theory

where time derivatives in ^ are taken outside the time-ordering (as see

usual



Section 2.5).

2.14 Feynman rules for gauge theories Feynman rules are given in Fig. 2.14.1 for the Lagrangian of (2.11.7), with the gauge-fixing term Fa d-Aa. Note that these agree with the figures but =

not the

equations of Marciano & Pagels (1978). They are the rules for quantum chromodynamics (QCD), the theory of strong interactions, if the

gauge group is SU(3). The fermions

are

the quarks and consist of several

own mass term. (In the conventional the field is called the gluon field and the gauge terminology, gauge

triplets of SU(3), each with its

symmetry is called the color symmetry of strong interactions. Each

representation in the quark field is called

irreducible



flavor, and has



label: u, d, s, c, b, etc.) The same Lagrangian

also describes quantum electrodynamics (QED) if the change gauge group to C/(l). In that case there is but one gauge field (the photon) and, since the group is abelian, the three- and four-point self- we



ij/





M + \E

■wX 'a, n 4" p

\J^

ApX

fAi Af



ca

->-r> ̄ J









n-^v

S P'

9caPy[(qK





rK)gXu



(rx



px)g,K



(/>?- q,)gK J

ig2[caPecyd {g^gkv gKVg?J -



ca8tcfiyt(gKkg^



gKftgXv)]

=-gcafiyP'>

At 2.14.1. Fig. Feynman rules for

the gauge theory defined by the

Lagrangian (2.11.7).

2.15 Other symmetries

35

of (2.11.7)

interactions of the gauge field vanish. Moreover, with the gauge fixing term F\A\ 8-A there is no coupling to the ghost fields, so we may drop them =

from consideration. The transformations

phase

rotations:

\j/-+e ̄lq<°\l/ where

the matter fields

are simple field of the (negative for charge

on

q is the

the electron field). The transformation of the gauge field is

A^A^-V d^co.

2.15 Other symmetries of (2.11.7) The

Lagrangian (2.11.7)

is gauge invariant. After gauge

fixing

we

get

invariant, but which has BRS symmetry. The action (2.12.8) also is invariant under global gauge transformations those

(2.12.8), which is

not gauge



with constant

a>



because

chose the gauge fixing not to break this

we

symmetry. There are also what in strong interactions are called flavor symmetries. These are transformations that act identically on every member of an irreducible representation of the gauge group. In this case they give

conservation

of the number of each of the different flavors of quark. Other symmetries of parity and time-

flavor symmetries include the discrete reversal invariance.

Charge-conjugation is also an invariance of (2.12.8), and its action on the ghost fields is rather interesting. Let us define ea by the parity of the representation matrices under transposition: ta In this and the



Ejl(%).

(2.15.1)

following equations, the symbol (J) means that the on repeated indices is suspended. The fermion and

summation convention gauge fields transform

as

usual:

(2.15.2)

^->?c(iyV)(<A)T. where rjc is



real matrix such that

tl The

ghosts transform



(2.15.3)

nXnc.

as: ca-+



eaca,

c*->



eaca.

(2.15.4)

AI Green's functions. They so only two couplings of the gauge indices are possible: c^y which is antisymmetric, and a symmetric coupling which we can call dxPy. Charge-conjugation invariance prohibits Consider the ca



cp



A\ and the A\



A*



are

invariant under global gauge transformations

the

symmetric,coupling.

36

field theory

Quantum

2.16 Model

field theories

Although the concepts of quantum field theory are very general, we have reviewed them by examining mainly two specific models. The first was the theory of a real scalar field (2.1.1), mostly with the (/>4 interaction (2.2.1). The second was a gauge theory (2.11.7) with matter described by some Dirac fields. It should be clear that the general principles apply to any Lagrangian ££. A field theory is specified by listing its elementary fields and giving a formula for ¥£\ It is solved by a functional-integral representation of its Green's functions.

is to describe the real world. To the extent that a field- is description the correct one, the fundamental problem in physics

The aim of physics theoretic

is to find

theory. In fact the Lagrangian (2.11.7) appears to electromagnetic interactions if the gauge group and matter fields are correctly chosen. Weak interactions can be included by the Weinberg-Salam theory, and many speculations have been made about extensions (see the proceedings of most recent conferences on high-energy physics). Our aim in this book would be badly served by only treating real theories. One reason is that we wish to develop techniques and concepts applicable to any field theory, for example not only to the many Grand Unified Theories currently under discussion (see Langacker (1981) and the correct field

do this for strong and

Ross (1981) for reviews), but also to the theories to be invented in the future. As is usual in the subject, we will make use of field theories that are more properly called models. The (/>4 theory is an obvious case. Another

important reason for using models is to be able to discuss particular aspects of the methods without having other complications to clutter up the presentation. Particular models will be introduced

as needed. Some will recur often, such as the t/>4 theory and the simple gauge theory (2.11.7). Another frequently used model is the (/>3 interaction of a real scalar field:

<£ This is much



{d(t>)2ll





003/3!.

(2.16.1)

than the other models. It is not

unphysical completely consistent. Because of more

m2(j)2/2

even

the (/>3 interaction, the energy is not bounded below. This is manifest in the classical theory and true in the quantum theory (Baym (I960)). Hence any state must catastrophically

decay. But the perturbation theory is well-defined, and somewhat simpler than for the t/>4 theory. So it proves very convenient to use the cj)3 model in treating the elements of the theory of renormalization within perturbation theory.

2.16 Model

field theories

37

constructing models is to change the dimension, d, of physical value 4. One motivation for this is that the renormalization problem becomes easier as d is reduced; the degree of divergence of a Feynman graph decreases. As we will see in Chapter 4, it is Another way of

space-time from

its

both useful and possible to treat d

as a

continuous variable, for the purpose

of computing the values of terms in the

perturbation expansion.

3 Basic

examples

chapters we will develop methods to treat large-momentum behavior. The complete treatment becomes rather intricate at times, so this In the later

chapter is devoted to

exposing in their simplest form the issues we will be discussing. We will do this by examining the self-energy graph in t/>3 theory. This will exhibit the basic phenomena which we will later be treating in detail. We will

see

that (in four-dimensional

space-time)

the

graph is

re-

normalized by a mass counterterm. Then the concept of 'degree of divergence' will be introduced by varying d, the dimensionality of space- time. This device will enable us to see how simple power-counting methods

determine what

counterterms are needed. It will also introduce us to the

regularization. The renormalization group will be introduced by examining the behavior

method of dimensional

of the graph

as its external momentum, p, is made large. By exploiting the arbitrariness in the renormalization procedure, we can reduce the size of

higher-order contributions when p" is large. 3.1

One-loop self-energy

in

<£3 theory

Consider the graph shown in Fig. 3.1.1 in the (j)3 theory of (2.16.1). We define its contribution to the self-energy to be i times the value of the graph with the external propagators removed:

^i{p2)



li^rk[k^-m2

The overall factor \ is





u-\i{p + kf-m'+^

symmetry factor. k

-O-

Fig. 3.1.1. One-loop self-energy graph in <f>3 theory. 38

(3U)

3.1

One-loop self-energy

in

<f>3 theory

39

When all components of k* get

large, this integral diverges of an ultra-violet divergence. As we will see, logarithmically. simplest example the divergence can be cancelled by a mass counterterm. But to explain the renormalization properly, we must discuss a number of other issues as well: It is the

(1) The fact that if | k° |



| k | the divergence as k goes to infinity appears to

be much worse.

(2) A precise

way of

cancelled

formulating

the statement that the divergence is

by (3) The arbitrariness inherent in the renormalization. (4) The interpretation in coordinate space. a mass counterterm.

3.1.1 The first of these

Wick rotation

problems is handled by recalling

the Wick rotation into

Euclidean space that was used to define the functional integral. This rotation determined the sign of the ie in the free propagator. The Wick

it, then performing the integral, and finally analytically continuing back to real time. In momentum space, this forces us to work with k° + ica, the opposite sign rotation involved starting with

imaginary

time t







appearing

so

that in the Fourier transformation &kx is always

In the Euclidean formulation let us

perform

the



phase.

/c°-integral first. The pole

/c°-plane is shown in Fig. 3.1.2, when p° is imaginary. In this and (p + k)2 are both negative, so that the integrand is positive

structure in the situation k2

definite. Observe that the factor i coming from the Wick rotation combines with the overall factor of i in (3.1.1) to make Z, real. (We have d4/c =

idaxl3fc.) Now rotate

shown in

Fig.

p°

back to

real value. If



3.1.3: there is

no

| p° | < m then we have the situation

obstruction to

rotating the k° contour to run a problem from the

along the real axis. It is only at this last step that there is

L p°  ̄EP+k





^p

+ fc

-X- E*

Fig. 3.1.2. The

fc°'-plane

when

imaginary.

p°

is

Fig. 3.1.3. The fc°-plane when p° is real, but \p°\ is less than m.

Basic examples

40

IN

m-

Fig. 3.1.4. The /c°-plane when p° is real, but

region of |fc°| |fc|. integral by rotating  ̄

we

pick

pole

up a

run

m.

we merely have to define along the imaginary axis.

the

region

to the

illustrated in Fig. 3.1.4. Again

(p + k)2

problem

the /c°-contour to

is greater than

|p°|>m. The case of negative p° is Wick-rotate to imaginary k°, but this time term. Now the pole term occurs only when (p0)2 > m2 + p°

Now continue

To avoid the

\p°\

we

E2+k.

Thus it contributes only in a finite region of k; the UV still from the integration over imaginary k°. comes divergence The moral of all this is that the UV divergence is essentially Euclidean,

i.e.,



we may

regard k°

as

imaginary 3.1.2

and k2 < 0, (p +

k)2

< 0.

Lattice

quantify the divergence. The divergence comes from the asymptotic large-/c behavior of the integrand which is l/(k2)2. Let us add We next need to

and subtract =

:{{'

z1

a term

Iff 32n

d4k

with this behavior: 1



l(k2

-m2



ie)[(p + k)2

-m2 +

ie]

(k2



fi2





d4fc- (k2-n2+ie)2



is)2 (3.1.2)

integral is manifestly finite, for we have subtracted off the leading asymptotic behavior of the integrand. To avoid introducing an extra

The first

divergence Since

we

at

k2



0 we have subtracted

l/(/c2



pi2)2

rather than

l/(/c2)2.

of/i is irrelevant; 2 i is unchanged. is independent of p. This is the fact that

add this term back on, the value

The second term, while

divergent, divergence by a counterterm. Of course we are manipulating divergent integrals, so that (3.1.2), as it stands, makes no sense. We will remedy this defect by using the fact that the theory is defined initially on a lattice. The propagators will then be different will enable

us to

cancel the

functions of momentum. However, the structure of (3.1.2) will be unchanged after

imposing



cut-off,

as we

will

now

show.

One-loop self-energy

3.1

To define the functional integral also had to put the

theory, limit qa

theory

on a

let the free propagator ->

0, this is just i/(q2



in

<f>3

41

theory

only had to Wick rotate time, but space-time lattice, of spacing a. In the lattice of a particle of mass M be SF(q;M, a). In the we not

M2 + ie). But it is

zero

if qa



1, since high-

is, of course, that when one makes a Fourier transformation on a discrete space, one only uses

(The

momentum states do not exist on the lattice.

modes with

momentum

self-energy

wave-lengths longer

the lattice is finite, and (3.1.2)

on

Zx(p, m\a)

-i2 10

than now

lattice



spacing.)

The

reads:

(r

\d*k[SF(k;m,a)SF(p + k;m,a)-SF(k;ti,a)2]



32n







reason

\d*kSF(k;ix,a)2\ (3.1.3)

Zlfin(p,m,/i,a) + Zldiv(m,/i,a).

integrals are now convergent, so (3.1.3) is a correct version of (3.1.2). As the lattice spacing goes to zero, the first integral approaches the first convergent integral in (3.1.2). The second integral diverges, but is All the

independent of p. Thus (3.1.2) is not nonsensical, provided that the propagators are implicitly replaced by lattice propagators wherever necessary.

3.1.3 Interpretation of divergence No matter how it is limit. The From

use

manipulated, the self-energy diverges in the continuum us to quantify the divergence.

of a lattice cut-off now enables

(3.1.3)

12_ Zi=T&jd4^F(^;/i,a)2 32n4

327U4 2

16tc2



-Q2

Jfl)2+fc2<1



Zlf

dcod3/c—-—=r; k2 (w2



r-r



n2)2

+ finite

fl/a





(/c2 + /i2)2

—^lnl/a +finite

+ finite

asa-+0.

(3.1.4)

interpret the divergence as follows: Let £ be (i times) the sum of self-energy graphs. (As usual, the self-energy graphs are graphs for the Thus we can

propagator that have the external lines amputated and that cannot be split

42

Basic

G*

Fig.





-#- -#-#- +

3.1.5. Summation of

into disconnected parts

examples

self-energy graphs



into propagator.

by cutting a single line.) Then

G2(p2)



i/(p2



m2



I +



the full propagator is

ie).

(3.1.5)

This equation is illustrated in Fig. 3.1.5; the propagator is the

sum

of



geometric series involving the self-energy. The actual mass mph of the particle is determined by the pole position, p2 m2h. Evidently m2h is not m2 but m2+ E(p2 m2h,m2). In other words, the self-energy represents =



the dynamical contribution to the mass coming from the interactions. The divergence (3.1.4) is independent of p2, so it is precisely a contribution to the mass. (We ignore higher orders for now.) Traditionally, one observes that it is convenient to parametrize the theory, not by the mass parameter m that cannot be observed directly but by the physical mass mph. One writes the mass term in <£ as -

m2ct>2/2

=--



m2h</>2/2



The first term is left in the free

i/(p2



that the free propagator is Lagrangian, second term called the mass counterterm is

m2h + ie). But the





put into the interaction propagator has a pole at

dynamical contribution

(3.1.6)

?5m2</>2/2.

so

Lagrangian,

p2

m2h.



to the

and

adjusted

so

that the full

The counterterm exactly cancels the

particle's

mass.

This is the basic idea of

renormalization. It is physically irrelevant that dm2 happens to diverge. In perturbation theory, dm2 is determined as a power series in g. To 0(g2) we have, in addition to Fig. 3.1.1, the graph of Fig. 3.1.6 corresponding to the mass counterterm in

is the

parametrization

(3.1.6). The self-energy sum

of Figs. 3.1.1

to

0(g2)

with the

new

and 3.1.6. We call it the

renormalized self-energy: ^iR





{^i(p2;mph^) + ^2}|fl-o Ilfin(p2;mp2h;/i,0)-f[<5m2 + Ildiv].

(3.1.7)

X Fig.

3.1.6. Counterterm graph to the self-energy.

adjust dm2 first of all to cancel the divergence in £ ldiv, so that the term in square brackets is finite as a->0. Then we adjust the finite part so that We

<5m2 + Zldiv=

-XUi>p2h,mp2h,/z).

(3.1.8)

One-loop self-energy

3.1

For the renormalized

*..-£K

self-energy

(*



value at



we



43

find

mp2h + ie)[(p + k)2 p2

<fi3 theory

in



m2h + ifi]

(k2



\i2

m2h.



ie>

(3.1.9)

^-dependence cancels. Equation (3.1.9) gives the value of the self-energy in the continuum theory

Note that the

correct to

0(g2). 3.1 A Computation

One way to calculate course a Lorentz

integration

Z1R

is to differentiate with respect to

scalar.) Integrating the result gives Z1R; the

is fixed by the renormalization condition that

p2. (It

is of

constant of

TlR(m2h) is zero.

We have:

asu

_

<>





'J'

 ̄'e'

U,?(k2 -m2,,.,*? ??, is) [(p k)2 32;r4p2 J This is identical to what

we







(3.1.1(0

..,?■



m2



i£]2

would have obtained from the unrenormalized

expression (3.1.1) without regard to the fact that it is divergent. We could have written it down directly without going through the long explanation that we used. But then there would have been no defense to the argument

that

we

are

manipulating meaningless quantities

and that therefore

quantum field theory makes no sense. Since (3.1.10) is finite it can be easily calculated by using



Feynman

parameter representation and then by shifting the /c-integral: <5I1R

ig2

dp2

16a V

f J

ig2



16kVJ 32*2

g2.

J0 o

32n2dp:

d**

dxx

dx

p-(p+k)

[Wph

,4,.



P2x



2p-kx -k2- ie]3

P-jk + pjl-x)]

a\m2fh-p2x(l-x)-k2-ieY x(l-x)

[m2h-p2x(\-x)-\

|rdxln[mp2h-p2x(l-x)]|.

(3.1.11)

44

Basic

Using the condition Z1R



0 at

*lR ̄32n2 This integral

can

p2

examples =

dxln

be worked out

3.2

m2h

now

mp2h(l



gives x +

x2)

(3.1.12) .

analytically.

Higher

order

graph of Fig. 3.1.1 is not the only divergent graph in the theory. In Chapter 5 we will discuss the general theory of renormalization and we will see how to extend the removal of divergences to all orders. In this section we will only consider a class of graphs which have divergences generated because Fig. 3.1.1 occurs as a subgraph. Examples are given in Fig. 3.2.1.

The

One property should be clear. This is that the divergences come from subgraphs all of whose lines are part of a loop. A general way of

characterizing these subgraphs is to define the concept of a one-particle- irreducible graph or subgraph. A one-particle-irreducible (1PI) graph is one which is connected and cannot be made disconnected by cutting a single line. A graph which is not 1PI is called one-particle-reducible (1PR). The in Fig. 3.2.1 are all one-particle-reducible, since they all have one or lines that when cut leave the graph in two disconnected pieces. The

graphs more

self-energy subgraph of Fig. 3.1.1 consisting of the two lines in the loop is 1 PI. This identical subgraph occurs several times in the graphs of Fig. 3.2.1. We introduced

a mass



counterterm

into the interaction,

so

that the

o^>



(c) Fig. 3.2.1. Graphs containing the one-loop self-energy

as a

subgraph.

3.3 Degree

of divergence

(a)

45

(b)

Fig. 3.2.2. Counterterm graphs to Fig. 3.2.1.

counterterm

graph Fig.

the counterterm vertex

3.1.6 cancels the

can

be used

as an

in Fig. 3.1.1. Clearly interaction anywhere in a graph.

divergence

In fact, all graphs containing it can be found as follows: (a) take a graph with the loop of Fig. 3.1.1 occurring as subgraph one or more times, but with no mass counterterm

vertices; (b) replace

one

(or more) of the

occurrences of

the loop by the counterterm. The terms generated from Figs. 3.2.1 (a)-(c) are shown in Fig. 3.2.2. sum of the original graph and its counterterm graph(s) is the just original graph with every occurrence of the loop replaced by its renormalized value iZ1R. It is sensible to keep the counterterm associated with the loop, thereby considering the loop plus the counterterm as a single

Evidently, the



entity. In the

case

of the graphs of Fig. 3.2.1 the result of this

graphs finite. The generalization out in Chapter 5.

make the worked

3.3 We

Degree

of

to an

procedure is arbitrary graph will

to be

divergence

how the Wick rotation ensured that the UV divergence of the one- self-energy is a purely Euclidean problem. The divergence is from the

saw

loop

region

| fc^ |

-^

oo, without any

regard

to direction.

Thus, simple power-

a logarithmic divergence. The power- counting involves merely counting the powers of k in the integral for large k.

counting determined that there is The divergence is logarithmic

as

Power-counting in this fashion

the lattice spacing, a, goes to zero. works for a general graph to determine

46

Basic examples

degree of divergence' or the 'superficial Chapter 5. There we will also see how the value of the degree of divergence determines the particular counterterm vertices needed for a theory. This will enable us to determine whether or not a theory is renormalizable by invoking arguments revolving what is called either the 'overall

degree of divergence'. This

we

will discover in

around the dimension of the

coupling. dimensionality, d, of space-time in the self-energy graph, Fig. 3.1.1. We can then explore the the degree of divergence, as determined by power-

In this section we will vary the

calculation of the

relation between

counting, and the momentum dependence of the counterterm. The integral for Fig. 3.1.1 is now ,2

^(p2,m2,d) The

space-time

1&



has

ddk(k2 -m2

2{2iif one



ie)[(p + k)2 -m2

time dimension and d





ie]'

(3'3J)

1 space dimensions. In the

Feynman rules the factors (2n)A get replaced by {2nf, since they arise from the result



ddxQikx



(3.3.2)

(2n)dS^(k).

The number of powers of k in the

divergence of the convergent. But whenever degree of

integral is now d 4; we call this the If graph. d is less than four, then the graph is d is greater than or equal to four, the graph —

diverges. Our discussion in the previous sections tells

by adding

renormalize it

Now, differentiating

try

to

a counterterm.

once

with respect to p" gives convergence if d

^1 Z2t U =

(27r)d

dp*

us we must

^±^

J V-m2)[(p + fc)2-m2]2'



4:

l(3 3 3)} *

degree of divergence of the integral being reduced by one to d 5. integral diverges logarithmically. However one might surmise that the divergence comes from the piece of the integrand proportional to kfl k would kill the divergence. This is in fact and that symmetrizing by k-+

with the

If d





5 the



true, but let us be more simple-minded. Differentiating again with respect to divergence d

9% dp?dpv ^dpv To recover





result with degree of

6:

in'

[^AHp + kUp + kl-gJip + kf -m']}

(2n)d (2n)d]J

Z1R

p gives

(k2

at d= 5 we

integration that appear



m2)[(p + k)2

integrate twice.

as an



There

m2]3

(3J4)

are more constants

additive contribution of the form: A +

of

B^p*.

3.3 Degree

of divergence

47

require Lorentz invariance of £, so the Bfi term is eliminated, and we are left with a mass term as the only counterterm. This is However,

we

must

the first and simplest example of the

use

of



symmetry argument

to

determine the form that we will allow for a counterterm.

(j)3

3.3.1

atd





Differentiating Zj three times gives a finite result with degree of divergence d 7. Integrating to obtain Z1R gives arbitrary integration constants of the form A Bp2, where again we have used Lorentz invariance. If we went to the lattice we would find divergences proportional to I/a2, m2ln(a) and p2ln(a). The fact that these terms have dimension 2 corresponds to the fact that the integral for Ex has degree of divergence 2. To make it finite we must not only use a mass counterterm but also a counterterm proportional to p2; the total we will call dm2 dZp2. Let

us now

go to d



6.







This is generated by

a counterterm



(3.3.5)

6m2(j)2l2 + 6Z{d(f))2l2

in the Lagrangian. Evidently the value of the

degree

maximum number of derivatives

or

of divergence is reflected as the powers of p in the counterterms.

Equally, it is reflected in the integration constants that appear when we recover Z1R from the differentiated T,l. These two phenomena happen for a general graph, as we will see later. The method of proof will in fact be to differentiate each graph enough times with respect to its external momenta until it is finite.

example of the wave- can interpret it the bare field for the The physically by examining propagator propagator. can be expressed in terms of the propagator of the renormalized field: The

(d(f))2

counterterm in

(3.3.5) is of

course an

function renormalization introduced in Section 2.3. We

e2(O)

<o|r#o(p)0o(O)|o>



Z<0|7>(p)4>(0)|0>



iZ/(p2





m2h



£1(0)



dm2



SZp2



ie)

iZ/(p2-mp2h-Z1R).



(3.3.6)

Note the distinction between bare and renormalized fields. The residue of

particle pole in unity: the

G2(0)



the propagator of

iR(0)/(P2



"*ph



ie)

an



interacting field is

finite>

as

P2

->

in

general

(3-3-7)

mph-

Examination of its spectral representation demonstrates that 0

not



R(0)



1,

48

Basic examples

(j>0 has canonical equal-time commutation relations. (The proof is given, for example, in Section 16.4 of Bjorken & Drell (1966).) Now we can always change the definition of (/> Z" 1/20o by multiplying Z by a finite factor. So it is possible to adjust Z so that the renormalized propagator has a pole of unit residue at p2 mp2h: because the field





<52 In this

case we



i/(p2

integrating



ie)



Wph

identify R{0) Z,R

When



finite,



with Z. The renormalized

5Z1R/3p2





p2

at

(3.3.8)

p2->m2h.

as



self-energy

satisfies

(3.3.9)

m2h.

Z1R from the finite derivative of Zl9 this integration constants to be determined. It is called the

to obtain

condition enables the

mass-shell renormalization condition.

3.3.2

Why

may Z be

zero

and yet contain divergences!

a property of the exact theory we know that 0 < Z < 1, if we adopt the renormalization condition (3.3.9); Z is definitely finite. However its perturbation expansion starts at 1 + g2[C In (a) + finite] + \ (Here C is a constant.) This seems to be infinite as a -? 0, rather than finite. We resolve the contradiction by realizing that we should not expect higher-order terms to be small if the one-loop correction is large. For example, we could have

As





the series

Ll-02[(C/D)ln(a) + const.]J where D is



loop (otherwise if

we

fix

It would appear



through the one- positive a and let #->0 there is a region with Z > 1). impossible to derive a formula like (3.3.10) since it

positive number. infinity. In

term goes to



Then Z -? 0 any event

as a -? 0 even

we see

that C must be

involves summing all orders of perturbation theory. Moreover it involves an analytic continuation from within its radius of convergence | In (a) |



D/g2C

perturbation

to

series

| In (a) | are



oc.

This

in general

seems to

make

no sense at

asymptotic series rather

all since

than convergent

series. However, we will see in Chapter 7, on the renormalization group, that we can find a systematic method of calculating Z and the other

renormalizations in the limit a-?0.

It relies

theory. The behavior (3.3.10) asymptotically free theories.

freedom of the

on

the so-called

asymptotic

will turn out to be essentially

correct, in No matter what the truth is, it should be clear that the divergences in perturbatioh theory as a -? 0 need not be reflected as divergences in the exact

theory, but only

as

singularities.

3.3 Degree 3.3.3

Suppose the

Renormalizability

we now go to a

and non-renormalizability

dimension d

one-loop self-energy has



49

of divergence

> 6.

For example, let

quartic divergence. The

us set





8. Then

necessary counter-

terms contain up to four derivatives of the field: -

The quartic term

Sm2(j)2/2



SZ(d(j))2/2 + E( □ c/>)2/2.

is not of the form of any term in the

(3.3.11)

original Lagrangian,

it cannot be obtained by renormalization of the Lagrangian. When this situation occurs the theory is called non-renormalizable. Non- so

renormalizability

is



consideration. There

good reason possible ways to

priori are



for

dropping

the

avoid this, but

theory

we

from

will leave

discussion of this until later.

simple argument that helps in the determination of whether or theory is renormalizable. It links dimensional analysis and power counting. Consider a one-particle-irreducible graph T. Let its degree of divergence be S(T), and let its mass dimension be d(T). Now T is the product of a numerical factor, a set of couplings, and an integral. As we see from (3.3.1) the degree of divergence is the dimension of the integral. Therefore, if There is



not a

we

let A(r) be the dimension of the couplings in T,

d(r) Now consider the

3(r)





we

have

(3.3.12)

A(r).

polynomial of degree S(T) in the external momenta. For each term C in the polynomial, let S(C) be the number of derivatives and let A(C) be the dimension of the coefficient, so counterterms to T. These form a

that its dimension is the

same as T:

<5(C) + A(C)



(3.3.13)

d(r).

Now, the maximum number of derivatives in the

counterterms is

d(T)=d(T)-A(n If the couplings

have

large by going

to

negative dimension, then 8(T) can graph of high enough order.



miraculous cancellations this tells

couplings have

If the

as

arbitrarily

In the absence of

expect non-renormalizability. If the

positive dimension, we have a finite number of since A(r) is bounded above by zero and d(T)

zero or

counterterm vertices, decreases

us to

be made



the number of external lines increases.

couplings

never

have negative dimension,

coefficients of the counterterms

satisfy

A(C)



d(T)-d(C)

>d(T)-S(r) =

A(r),

we

observe that the

50

Basic examples

so that these coefficients also have non-negative dimension. This, in the simplest cases, is sufficient to ensure renormalizability. For example, in the (/>3 theory cj> has dimension d/2 1. (Recall that ¥£ has the dimension of an energy-density.) Then A(g) 3 d/2. So if d > 6 the < 6, there are only theory is non-renormalizable, as we saw by example. If d -





a finite number of possible counterterms, since these coefficients of non-negative dimension.

are

restricted to have

3.4 Renormalization group 3.4.1 Arbitrariness in The infinities of

renormalized graph

renormalizable theory amount to divergent dynamical that renormalize the parameters in the Lagrangian.

contributions Traditionally





one

thinks of renormalization

as

the

procedure

of

working

with measured

quantities instead of the corresponding bare quantities. The most obvious case is that of the mass mph of the particle corresponding to an elementary field.

However to take the traditional view is much too

restrictive. This issue can be understood

by looking at strong interactions. There we theory, QCD, particles corresponding to the do fields not to exist appear elementary (pace LaRue, Phillips & Fairbank (1981)). So arises the hypothesis of quark confinement not proved from in which free



have



so far according to which quarks are never isolated particles. Even has the so, theory quark masses, which can be measured (up to considerable uncertainties for the light quarks). But one cannot identify these masses

QCD,



with the directly measurable

masses

of free

quarks.

One must

only speak

parameters, measured, in this case, rather indirectly. There is no problem in taking this point of view. For example,

of

mass

the

(f>3 Lagrangian &



d(j)2l2



as a

basic

m2(/>2/2



Lagrangian plus

#3/3!



we

write

a counterterm

Lagrangian:

dm24>2/2

Sg(f)3/3!. (3.4.1)

6Zd(j)2l2





avoid identifying the renormalized mass with the mass mph of a particle. Similarly we do not identify the renormalized coupling, g, with any specific measured quantity, and we do not define Z by requiring that the But

we

residue of the

propagator's pole

be

unity.

Consider the calculation of the one-loop choose SZ



0 for this case; the

self-energy (3.1.7)

self-energy at d 4. We can only divergence is in dm2. The renormalized =

is

SiR



^ifin



(<5m2



Zldiv).

3.4 Renormalization group We must choose dm2 to have a

divergent part

51

to cancel the

divergence

at

0 of £ldiv. But the finite part of dm2 is not determined; the arbitrariness is the same as that of the integration constant when obtaining I1R by





integrating dyLl/dpfl. At first sight it might appear that the arbitrariness ruins the theory unless one pins down m to be the physical mass mph. This is in fact not so; the arbitrariness is

more like the arbitrariness in choosing a coordinate system. first Suppose computes the propagator with the mass-shell condition £ir(P2 0 0. Then one





G2







i/[p2







m2h

^idiv





Ilfin(p2) + Ilfin(p2 Slfin(mp2h) + 0(g*).

One could also compute with

G2

m20





i/[>2 m2





m2

Ildiv









m2h) + 0(a4)], (3.4.2a) (3.4.26)

different finite part to

5m2,

Ilfin(p2, m2) + g2C + 0(a4)], g2C + 0(a4),

where C is any chosen number. The

X¥(p\m2)



self-energy i:Uin



is

with



result

(3.4.3a) (3.4.36)

now

g2C.

Evidently the two ways of renormalizing the theory give the same results if we require that the bare mass ml is the same in both of (3.4.2) and (3.4.3). In the

ml

complete solution of the theory, say by the functional integral, it is only that matters, not the partition into a renormalized mass squared m2 and

a counterterm



Zldiv m2

with the



g2C. Clearly we have m2h + g2C Ilfin(mp2h) + 0(a4),





(3.4.4)

0(a4) terms depending on the renormalization of higher-order self-

energy graphs. We come then to the central idea of the renormalization group. The

physically irrelevant, for a change in the arbitrary constant C can be exactly compensated by a change in m2. A change in C merely gives a different parametrization of the set of theories arbitrariness in the definition of Z1R is

be obtained by varying the mass parameter m. The renormalization group is the set of transformations on the parametrizations of the

that

can

theory. The transformations are accomplished by moving parts of the terms in ¥£ from the basic Lagrangian to the counterterm Lagrangian. In the case of



it is

a move

from the free

Lagrangian



to the interaction

Lagrangian.

rearrangement of the perturbation series, which is the

gives key to the many practical applications It might be objected that

This of course

of the renormalization group.

p2-m2-Ilfin(p2,m2) + a2C

52

Basic examples

is not

equal

to

p2 since the

m2h





Ilfin(p2,m2h) + Zlfin(mp2h,mp2h),

parameters in

mass

Zlfin(p2)

parametrized in either way is the

are

same.

different, whereas the theory

But since

m2h



m2 is

0(g2)

the

difference in the two expressions is in fact 0(#4). Thus the rearrangement of the perturbation series does not leave the p2-dependence of the coefficients invariant. The

0(g*)

terms will cancel the difference

(up

to even

higher-

order terms), etc.

utility of the renormalization group is precisely in its ability to reorganize the perturbation series. Since one effect of the interaction is to The

induce dynamical contributions to the

good idea

mass

and couplings, it is evidently



to arrange that these contributions are small. The result is to

higher-order corrections and thus improve the perturbative calculation. Now the effective size of the dynamical mass or coupling must be treated as dependent on the situation under consideration. This can be seen by examining Z1R given in (3.1.12) at large p2: reduce

the

reliability of

values

of



L1R



(02/327t2)[ln(



p2/m2h)

+ constant

+???].

(3A5)

If | p21 is large enough this can be large. Since the graph occurs as a subgraph of higher-order graphs, it is likely (and often is true) that higher-order graphs are as important as low-order graphs at large enough p2. This situation is undesirable and can be remedied by a renormalization-group transformation. We absorb the large part of Z1R into a redefinition of the renormalized mass

m2. We must examine higher-order graphs

that there are no further sources of

systematically

in

Chapter

large p2 to demonstrate large coefficients. We will do this at

7.

3.4.2 Renormalization prescriptions

resolving the ambiguity in constructing given theory, each of these ways corresponding to a particular parametrization. It is essential that, whenever a particular divergent graph occurs as a subgraph of a bigger graph, the ambiguity is There are

infinitely

many ways of

the counterterms for

resolved in the



same way at

each occurrence, since the corresponding

counterterm vertex is generated by

perform

concrete

calculations



one

ambiguity., Such a rule is called renormalization scheme.

single term in the Lagrangian. So to adopts some rule to resolve the a

renormalization prescription

or

3.5 Dimensional

53

regularization

infinitely many possible renormalization prescriptions, a few have become standard, because they are especially convenient either for practical use or for theoretical considerations. In this section we will explain two of the standard ones with the aid of the example of the one-loop graph Fig. 3.1.1. We have already encountered the mass-shell, or physical, scheme. The Of the

is defined to be the physical mass, i.e., the position of the Wave-function renormalization is fixed by requiring the propagator pole. residue of the pole to be unity (see (3.3.9)). Couplings can be defined by renormalized

specifying

mass

the value of



suitable S-matrix element.

A possibility that is much used in discussions of renormalization theory is the BPH or BPHZ scheme (Bogoliubov-Parasiuk-Hepp-Zimmermann), as zero-momentum subtraction. Let Y be a one-particle- irreducible (1P1) graph that is divergent, i.e., it has (5(r)>0. The prescription is that at zero external momentum its renormalized value R(T) and its first S(T) derivatives with respect to external momentum are zero.

otherwise known

this by subtracting off the first S(T) integrand (Zimmermann (1970)), that is, the renormalization is performed before the integration over loop momenta. No explicit UV cut-off is needed. In this scheme the self-energy already discussed is, at d 4, The BPHZ scheme is to

implement

terms in the Taylor expansion of the



£(1RPHZ) while



at d



y(BPHZ)

Jd4/C{[(fc2

32^4 6,

we

^2



_

_

m2}]

{k2

^2

_

j>

(3A6)

have:

fJ6,f

*M\(k2



m2){{p + k)2

128tu6 J





1 -

m2)[(p + k)2



m2]

(k2



m2 + i£)2

[Vfc2-(/c2-mV]|

Irk

(/c2-m2)3

3.5 Dimensional

(fc2-m2)4

J*

[X)

regularization

we used the lattice as a cut-off, or we in is the renormalized what are interested really regulator. However, theory with no cut-off. We could equally well use some other kind of

In

our

initial treatment of UV divergences

regulator. For example,

higher derivative S£





Pauli-Villars type of cut-off is achieved by

term in the

dA2/2

Lagrangian.

m2^2/2





qA 3/3!



[(Q



For



example, from

m2)^]2/2(M2

+ counterterms



m2) (3.5.1)

54

Basic

examples

obtain the free propagator (2.4.1). When the cut-off M goes to infinity, the propagator is the ordinary one, but when p2 is much bigger than M2, it

we

is smaller by a factor p2/M2. in six

or

Thus UV divergences

fewer space-time dimensions; the

take the limit M

-?

are

cut-off for the theory

divergences

reappear when we

oc.

If we defined the theory by a functional integral the lattice would appear intermediate step, but the a -? 0 limit would give no divergences, if M is

as an

Although the Euclidean Green's functions for the cut-off theory (3.5.1) exist, the Minkowski space field theory is not physical. A symptom of this is that the pole of the free propagator at p2 M2 has the wrong sign of residue; it implies a particle with negative metric. A theory with no cut-off can be obtained by adding counterterms with

finite.



appropriate M-dependences to cancel the divergences and then taking the M limit. As an example we showed that counterterms cancelled the divergences of the one-loop self-energy graph. Although we assumed a lattice regulator, we used no properties of the lattice propagator that are not true for the Pauli-Villars case. We assumed only that: (1) If the cut-off is taken away (i.e., M 0) with p and m fixed, then the propagator goes to i/(p2 m2). -?

oo

-?

oo or a -?



(2)

If

p2

-> oo

with fixed cut-off, then the propagator is sufficiently much 1/p2 that the graph is not UV divergent.

smaller than

(3)

In the Euclidean

region there

are no

propagator poles.

In principle, any method of imposing a UV cut-off is good enough, but in

practice

methods

some

are

more

convenient than others. For most

purposes dimensional regularization is the most convenient. The method starts from the observation that U V divergences are eliminated by going to a small enough space-time dimension d. We can use the space-time

dimension

as a

that the cut-off

regulator provided can

we treat



as a

continuous variable (so

be removed by taking the limit d->4). This idea has



long history, but its popularity roughly started after the papers by Wilson (1973) in statistical mechanics and by't Hooft & Veltman (1972a), Bollini & Giambiagi (1972), Ashmore (1972), and Cicuta & Montaldi (1972) in field theory (especially non-abelian gauge theories). Since vector spaces of non-integer dimension do not exist as such, it is not obvious that the

concept has any consistency, let alone validity,

even

in



purely formal sense. This we will remedy in the next chapter. For the present we

will

uniqueness integral

assume

manipulations to the

and existence, and

apply

standard

3.5 Dimensional

55

regularization

until it is of a form where we can sensibly assign a value. We will (following Wilson (1973)) express (3.5.2) in terms of a standard Gaussian integral:

|ddfcexp(/c2)



i \dco

give this the value ind/2, which is correct if d is

It is sensible to

(3.5.3)

\dd-lkexp(-w2-k2). an

integer.

In

following calculation of the value of (3.5.2) with non-integer d, the assumed properties of the integration are italicized. All the manipulations are valid for any integer value of d for which the integral converges. the

We use the

Schwinger representation for each propagator:

rdaexp[-a(m2-fc2-ie)].

l/(m2-k2-ie)=

Observe that because of the Wick rotation

exchange

we treat

k2

as

(3.5.4)

negative. Then

we

the order of integration to obtain

2(27r)" Jo

da

db

Jo

\d"kexp[-(a + b)m2 + bp2 + 2bpk + (a + b)k2l



(3.5.5) We shift k" by x



a/z

an amount

p"b/(a + b)

change variables

and

to z



a +

b,

to get:

Zi=^M dxrdzz[ddfcexp{-z[m2-p2x(l-x)]+z/c2}. z(Zn> Jo Jo J After scaling k by



factor





we

find that

E1=J|—I dx|=°dzz1-d/2exp{-z[m2-p2x(l-x)]}

d"kexp(k2). (3.5.6)

brings in the dimensionality d. We have now reduced the ^-dimensional integral to the form (3.5.3), which we defined to be i7rd/2.

It is this stage which

The z-integral in (3.5.6) gives



T-function,

so we

finally obtain:

^i=^F2T(2-d/2)^dx[m2-p2x(l-x)r12-2.

This result is unique (except possibly for

an

(3.5.7)

overall normalization, which is

for every ^-dimensional integral). The divergences The now reside in the T-function which has simple poles at d 4, 6, 8, residue of each pole is a polynomial in p of degree equal to the degree of universal



the

same





divergence. One of the main advantages of dimensional regularization is immediately apparent. Not only was the integral unchanged from its form in a theory in

56

Basic examples

integer dimensional space-time with no cut-off, but the method of calculation was unchanged. Use of the representation (3.5.4) is an efficient way of obtaining a parametric representation like (3.5.7) for a Feynman graph. A second advantage that we will see later is that it preserves not only Poincare invariance in the regulated theory, but also gauge symmetries. This was a main motivation for its use by 't Hooft & Veltman (1972a) and many others. Most methods of introducing a cut-off fail in this respect. (For example, gauge invariance is preserved invariance is lost.)

on

the lattice but full Poincare

importance in practice is that continuous space-time dimension is also a gauge-invariant cut-off for A third

advantage



also of great





infrared divergences in theories with massless fields (Gastmans & Meuldermans (1973), Gastmans, Verwaest & Meuldermans (1976), and Marciano & Sirlin (1975)). A trivial example, but without any gauge invariance, is by the (/>3 self-energy with m 0 at d 2. It is =

8tt2J which is divergent at k

increasing

d. Care is

present in the

same



(fc2

given



i£)[(p + k)2 + i£]'



0 and at p



—k. The divergence is regulated

required in using this method if UV divergences graph, for they are regulated by reducing d.

by are

3.6 Minimal subtraction 3.6.1

Definition

self-energy (3.5.7) we compute the renormalized 4 by adding a mass counterterm Sm2(g,m2,d) and d Suppose we choose m to be the physical mass. Then

From the unrenormalized

self-energy Z1R

at

then letting d->4.



=^r(2-j/2)|^dx{K-rti-x)f-2

-[mp2h(l-x + x2)r2"2}. (3.6.1)

Now

T(z)

has



pole

at z



0:

r(z)

where

yE

0.5772





...

^h'

in agreement with



l/z-yF



is Euler's constant. So at d =

our

(3.6.2)

0(z), =



#,|"dxlnH-p2x(1-2Xn earlier calculation.

(3.6.3)

57

3.6 Minimal subtraction

divergence in Z, amounts to a simple pole at d 4, a rather obvious way of renormalizing it is to define dm2 to cancel just the singularity, i.e., the pole ('t Hooft (1973)). This, of course, means that we are Since the

changing see



renormalization

our

prescription. By referring back

to (3.5.7),

we

that dm2 in this scheme is: Sm2

from which

we



(g2/32n2)/(2

get I1R by expanding Zj in '



ilR ̄327t2





d/2),

(3.6.4)

power series in d

\ ̄m2-p2x(l-x) ̄\



4. We find



Unfortunately

this contains the logarithm of a dimensional quantity. The is that in the expansion in powers of d 4 we did not allow for the fact that g has a dimension dependent on d. Therefore we implicitly introduced a mass scale. reason



To make this scale

explicit,

we

rewrite the coupling

(3.6.6)

g^tx2-d/2g,

we have introduced a parameter fi with the dimensions of mass, called the unit of mass ft Hooft (1973)). The redefined coupling g now has

where

fixed dimension

equal

to



y(MS)_ ^1R

1, and the renormalized self-energy becomes



dxiln

"32tt2J0 jln

We derived this

by observing

 ̄m2 -p2x(l -x)

47i/r

+ yE

(3.6.7)



that

?2-d/2 =e(2-d/2)lnM =

1 + (2



d/2)\n fi +i(2



d/2)2 In V +

''' ?

This renormalization prescription, where counterterms are pure poles at physical value of d, is called minimal subtraction (MS). The unit of mass fi is entirely arbitrary. Thus the self-energy (3.6.7) now

the

depends

on

amounts to a

three parameters instead of two. However a change of fi change of renormalization prescription, so the change can be

compensated, in this case, by a change in m. a

In effect minimal subtraction is

one-parameter family of renormalization prescriptions. 3.6.2 d

We can also

There

we





minimal subtraction to the six-dimensional

apply

theory.

define Sm2



Sz



¥0



poles

at d



6,

poles

at d



6,

jx3-d'2(polesati







6)J

(3.6.8)

58

Basic examples

Note that the renormalized

coupling

is

dimensionless. Since

now

r(z-l)=-l/z + yE-l as z ->

0,

we



0(z)

(3.6.9)

find that i2

_

y(MS). 128tt

(m2-p/6)(yE-l)

dx[m2-p2x(l-x)]ln



m2



p2x(l



x)

(3.6.10)

471/i

and

w-s?^+0^- 1

a2 6



(3.6.11) (3.6.12)

647T d-6

The counterterm Sg for the coupling can also be calculated. From the graph of

Fig. 3.6.1,

we

find (Macfarlane & Woo (1974)) 7

Sg

Fig.

_



we



64n3(d-6)



oto5).

3.6.1. One-loop vertex graph in

3.6.3 Renormalization When



,.3-d/2_ n

(3.6.13)

</>3 theory.

group and minimal subtraction

discuss the renormalization group in Chapter 7,

we

will focus

on

particular subgroup. The transformations in this subgroup consist of multiplying /i by a factor and making compensating changes in the renormalized coupling and mass. As a group it is trivial being a representation of the group of positive real numbers under multiplication. What is non-trivial is the way in which it is represented in relation to the parametrization of the theory by a renormalized coupling and mass. The renormalization group can be exploited in calculating high-energy one



full treatment will be made in Chapter 7, the basic idea can be seen by examining the one-loop self-energy. Let p2 get large (with m and \i fixed), and consider the propagator defined using minimal 6: subtraction at d behavior. While





p2

i-

7687t3

In

+ const.



O(04)

(3.6.14)

3.7 Coordinate space To be able to make use of



perturbation expansion

order corrections small. But this is not

so

59

keep higher- large. The The theory \p2\.

we must

in (3.6.14) if

p2

is too

large correction can be avoided by setting fi2 to be of order is unchanged if we make suitable changes in g, in m, and in

the scale of the

field. We will learn how to do this in Chapter 7, with the result that the

corrections

are

effectively

moved from

perturbation series to the lowest-order

higher-order graphs.

large

terms in the

3.6.4 Massless theories

(for simplicity) to the self-energy of the four-dimensional the limit m -> 0. If we use mass-shell subtraction, we have Consider theory. Let

us return

(3.6.3), which diverges as m2h-+0. The divergence is an artifact of the mass-shell scheme, for which ,2

(*i

Sm*=2(!^r(2" d/2)mdp^Ldx(1"x+x2)dl2 =

^r^^-ln(mp\)

In addition to the

ln(m*h)

pole needed



72-

finite(asmph-0,^4)l.

to cancel the UV

(3.6.15)

divergence, there is



term.

Physically what happens is that in a massless theory there are long-range forces. These mean that separated particles are never completely free of each others' influence. Thus, for example, the singularity in the propagator is not a simple pole, for the self-energy (with MS subtraction) is, from (3.6.5),

s-Ks?)**-2}

(3616>

The mass-shell renormalization prescription relies on the assumption of a simple propagator pole to generate counterterms, so it must fail. However, the nature of the propagator's singularity is an infra-red problem, so it is irrelevant to the question of whether

an

ultra-violet divergence

can

be

renormalized. Some other renormalization scheme, like minimal subtraction, must be used in the massless theory.

3.7 Coordinate space A good way to understand the infinite renormalizations is to work in coordinate space, as was emphasized by Bogoliubov & Shirkov (1980). This

point of view is especially useful in treating field theories

at

finite

60

Basic examples

temperature or on a curved space-time background, as we will see in Chapter 11. There we will see why the counterterms are the same as at zero

space-time. ordinary calculations, it is cumbersome to work in coordinate space, because the propagator SF(x) for a free massive field is a Bessel function. In momentum space the propagator is simple: i/(p2 m2 +ie). However the asymptotic behavior of SF(x) as x-?0 is simple. temperature in flat

For most



We have

Sf



4^/2(_x2)d/2-i

The one-loop correction

G2,z





where E^z





(92/2)

\ddz

+less singular

to the propagator

Udz \d"wST(x

\ddwSF(x z)!^ -



W)2SF(w



w)SF{w

0.

(3.7.1)

is

G2{x,y)

z)SF(z



as x -+





y) (3.7.2)

y),

w) is the self-energy in coordinate space:

±1(z-w)=-(g2/2)SF(z-w)2.

(3.7.3)

This is singular at z w, and causes a logarithmic divergence in (3.7.2), where we integrate over all z and w. The fact that the divergence is from the =

region

z ̄w means

that it is in fact



<5-function:

G2a=\ddzSF(x-z)SF(z-y)

By Wick rotating the w°-integral and using the following result (next chapter) for the ^-dimensional integral of a Lorentz-invariant function

\ddwf(w2) we



i[2nd/2/r(d/2)]

pdvyw^-Y(w2),

(3.7.5)

find G22

as





-?



tddzSF(x



z)SF(z



y)l6Jg{4_d)



finite,

(3.7.6)

4.

Evidently, the divergence is cancelled by adding



(5-function to

t.l:

61

3.7 Coordinate space which is

exactly

the

mass counterterm

computed earlier by

momentum-

space methods. The

important point, which

is in fact true for

arbitrary graph,

an

is that

any UV divergence comes from a region in coordinate space where several interactions occur very close to each other. The divergence can then be

cancelled by

a counterterm

interactions. If the

which is

divergence

is



worse

(5-function in the positions of these logarithmic, then the counter-

than

term will include derivatives of the (5-function. In any event the fact that it is a (5-function means that the counterterm can be included as a local

locality means that it is a product of fields at point. Since the singularity at x 0 of the free propagator SF(x) is independent of the boundary conditions used to define it, we should expect, for example,

interaction in the action. The

the

same



that the counterterms used in thermal field theory

temperature. At state, and the

temperature, the

are

the

same as at zero

replaced by a mixed The conditions for momentum-space boundary SF(x) change. non-zero

vacuum

is

proof that the counterterms are temperature-independent is therefore made coordinate-space proof is unchanged.

difficult, but the

4 Dimensional

We have

seen

how convenient it is to

perturbation theory by date,

no-one

regularization regulate the

UV

divergences

of

continuation in the dimension of space-time. To use the method in the complete theory. But

has shown how to

in perturbation theory, as we will now demonstrate, it is consistent and well-defined. Now all results obtained by this method can be obtained by

other, more physical methods (say, a lattice regulator). But frequently much labor is involved. This is not a triviality, for in complicated situations, especially in gauge theories, it enables us to handle the technicalities of renormalization in a simple way. The idea of dimensional continuation has been used for a long time in statistical mechanics (see, for example, Fisher & Gaunt (1964)). It became more

very prominent when Wilson & Fisher (1972) discovered the

a-expansion and applied it to field-theoretic methods in statistical mechanics (Wilson (1973), Mack (1972), and Wilson & Kogut (1974)). In the a-expansion one e spatial dimensions, and expands in powers of e. At the same works in 4 a in time, purely field-theoretic context, a need arose to find a way of non-abelian regulating gauge theories that preserved gauge invariance and —

Poincare invariance. This led to dimensional

regularization ('t

Hooft &

Veltman (1972a), Bollini & Giambiagi (1972), Cicuta & Montaldi (1972), and Ashmore (1972)). Speer & Westwater (1971) had actually discovered the method earlier, but their paper is considerably more abstract, and had not attracted much attention.

integer non-integer integration dimension, d, cannot be taken completely literally. Either it is a set of purely formal rules for obtaining answers or it is an operation that is not literally integration in d dimensions, but only behaves in many respects as if it were Now vector spaces either have infinite dimension or a finite

dimension.

So the concept of

integration

in d dimensions. It is not sufficient to treat it

on a

space of finite

only

as a set

of

formal rules (even though that is what it becomes in practice), because one must know that the rules are consistent with one another and with the carries out

integrals. To show that explicit definition.

algebraic manipulations

one

inconsistencies

we must construct an

can

arise,

62

on

no

Dimensional regularization There

63

three issues to address: (1) uniqueness, (2) existence, and (3) the possibility of constructing

are

properties. Uniqueness is necessary, to avoid

two definitions, each definition being self-consistent but giving different results from the other definition. Existence, shown by construction of an explicit definition, is necessary to prove that no inconsistencies arise. Once having

seen

cannot

just

that integration in non-integer dimension can be defined, we that all properties associated with ordinary integration

assume

true; indeed they need not be. So we also have to prove those properties which

are

true. We

is

also must prove that the results agree with

we

need and which

are

ordinary integration if d

an integer. These considerations are quite non-trivial, as can be seen by considering,

for example, the

anomaly in the Ward identity for the axial current t'ie 8au8e model (2.11.7). If the fermion masses are zero, naive application of the fermion equations of motion shows that the

j(5)= ipy^ys1!' then





current is conserved: d

jf5)



0. In fact, the current is not conserved,

as

by Adler (1969,1970) and Bell & Jackiw (1969). A counter term can be added to jf5) to make it conserved, but only at the expense of removing its gauge invariance. shown

Among the objects to be extended to d dimensions are the Dirac matrices (y" and y5). If we assumed the obvious generalization of their anticom- mutation relations, then for all values of d

{yV/v}



But then

we



would have

2<TU

{y5,y"}=0,

y52

we

i.





(4.0.1)

would be able to derive the false result that the anomaly for the

gauge-invariant axial current is zero. So there has to be an inconsistency ('t Hooft & Veltman (1972a)). More complicated problems in a similar vein arise when treating supersymmetric theories (Jones & LeVeille (1982)). In this chapter we will start by stating the axioms for d-dimensional integration given by Wilson (1973). These are sufficient to prove uniqueness. Our calculation of a one-loop graph in Section 3.5 was in fact a

uniqueness proof for one particular integral. Then we will explicit definition of ^-dimensional integration. The vector

realization of the construct

space

on

an

which

we

work is in fact infinite dimensional.

Unfortunately, the definition gives a divergent result in most cases, so we a powerful enough extension (Section 4.2). We then

will next have to find

some standard properties (Section 4.3). One particular result involves finding a definition of the metric tensor on an infinite-dimensional

prove

Dimensional regularization

64

space such that its trace is d rather than infinity. Then we will be in a position to derive some useful formulae

(Sections 4.4 graph calculations. Finally we will show how to define Dirac matrices; this is obviously important if we are to be able to calculate consistently graphs containing the Adler-Bell-Jackiw anomaly. The utility of a precise definition such as we give is that if inconsistencies and

4.5) for use in

arise at

some

Feynman

stage, then

always

one can

go back to first

principles

to

discover the error.

4.1 Let d be a

complex number. We wish to define an operation integration over a ^-dimensional space:

that

J-

regard

as

Definition and axioms

/'(p) is

any

given function of a

may

(4.1.1)

d'p/(p).

Here

we

vector p,

which is in the d-dimensional

space. We will suppose that the space is Euclidean. (Minkowski space is l)-dimensional regarded as a one-dimensional time together with a (d —

Euclidean space.) Following Wilson (1973) we definition in which the space is actually infinite

will

give

an

explicit

dimensional; it is the

integration operation that gives the dimensionality. Making d a positive integer n will effectively insert a (5-function in the integration that will force all vectors involved in defining the function /(p) to lie in some n- dimensional subspace. What properties must we impose on a functional of/ in order to regard it as d-dimensional integration? The following properties or axioms (due to Wilson (1973))

are

natural and

necessary in

are

applications

to

Feynman

graphs: (1) Linearity: For

any

complex

Jddp[a/(P) (2) Scaling:



numbers

MP)]

For any number



and b

jd V(P) jddpflr(p). + b

(4.1.2)



fd-p/(sp) (3)







s"-|<lV(P)-

(4.1.3)

Translation in variance: For any vector q

(ddp/(P We will also

require



q)



fddp/(P).

rotational covariance of

our

results.

(4-1-4)

65

4.1 Definition and axioms

Linearity is invariance

are

true of any

integration,

basic properties of

while translation and rotation

Euclidean space, and the scaling



property embodies the ^-dimensionality. Not only are the above three axioms necessary, but they also ensure that integration is unique, aside from an overall normalization (Wilson (1973)). In fact, they determine the usual integration measure in an integer- dimensional space (again up to normalization). The proof is simple: Use linearity to expand/(p) in terms of a set of basis functions. Choose a basis such as the functions

/s,q(P)

(4-1-5)

exp[-52(p + q)2].



Then the integral of a basis function can be written in terms of the integral of one

single function:

The integral of this

[ddP/M one

(P)



s ̄d

(4-1.6)

function sets the normalization. It is natural to

require that the value be the usual write

one

I<

can

|dVxp(-p2).

ddlpd',2qexp(-p2-q2)

in integer dimensions and that



we

(4.1.7)

d^'^kexpt-k2).

J'

Thus the normalization is given by

|ddpexp(-p2) An abstract uniqueness theorem is



7r''/2.

(4.1.8)

not sufficient for us. We also need an

explicit formula so that a ^-dimensional integral can be written as a sequence of ordinary integrals. This will be important in allowing us to prove standard properties of the integration. In addition it ensures that there exists



self-consistent definition. It is

consistent definition exists; the



priori possible that

uniqueness theorem only applies

no

if the

integration operation exists. A function /(p) that we integrate could in principle be any function of the components of its vector argument. However,

meaning of the components of, say, soon see

that there

will work with

are

in fact

a vector

infinitely

do not, a priori, know the in 3.99 dimensions. We will

we

many components. In

rotationally covariant functions.

So

we

will

practice, we that/is

assume

function of a finite set of vectors: p, q2,..., q^ say. For example, scalar function is a function only of scalar products a tensor





/(f>2,p-qi,4,2,.--)-



(4.1.9)

66

Dimensional regularization an ordinary function of scalar numbers, rather than some complicated kind of function. Of course the values of the scalar

Thus / is in fact more

products lie in restricted ranges. Thus:

q2>0,

k-q*|<qX2. A

tensor function is obtained

(4.1.10)

by writing explicit 8ij, with scalar coefficients.

tensors in terms of the

vectors p, qi,..., q^ and of the metric tensor

example,

we

might

/I'iP,q) Such functions

are



^y/fl(P^p?q,q2)

vectors as we

in



an



the most general that

later how to handle the

matrices.) To give

antisymmetric

ordinary

(4.1.11)

^y6(p2,p?q,q2).

we

need to consider. (We will

tensor

realization of the objects p,q1?...,

will show in

For

have

eKA/iV

and

we assume

that they

vector space. The space must be infinite

So

a moment.

are

dimensional,

define the vectors each to be

we

see

the Dirac y-

infinite

an

sequence of components, p (p\p2,...)> just as we can define dimensional vector V as a sequence of three components (V1, V2, metric is given by: =



three-

V3).

The

an integral with, say, regulator for a physical theory in a space- time of dimension d0 4, but also as a regulator for a model theory in any 5 or 6 or The vectors qt,q2,... in (4.1.9) can higher dimension, e.g., d0 be thought of as momenta of external particles, and our vector space must be large enough to accommodate d0 linearly independent momenta. Since d0 is arbitrary, we are forced to infinite dimension.

The d



reason

3.99

can

for the infinite dimensionality is that

be used not

only as







To define the d-dimensional integral of a scalar function, we find a finite- dimensional subspace containing all the q^'s. Then we write p as a

component Pn in this space and P





orthogonal component

an

pT:

P|| +Pt

i>je,.

(4.1.12)

+ pT.

The 'parallel space', in which lie the q^s, is spanned by an orthonormal basis

1,..., J). We define the integral over p to be the ordinary J- e,. (with j J dimensional integral over Pu performed after integration in d =



dimensions

over pT:

Jddp/(P) [dp1 =





dpJ

(dd-W(p).

(4-1.13)

4.1 Since

f(p) does

not

depend

Definition and

the direction of pT

on

d*-J?Tf(p)

we now can

define

(4.1.14)

dPjPdT-J-lf(V\

Kd



67

axioms



Kv (with v d J) is effectively the area of the surface of a hypersphere in v dimensions. The value of Kv is obtained by considering the special case where/is chosen to be a Gaussian see (4.1.8) with the result Here









Hence

we

have



definition of d-dimensional integration in terms of

ordinary integration:

l"dpm-m^ikh"J>* ̄J ̄,/,p)

(4U6)

We must check that the result is independent of the choice of the subspace of

the P||. We must extend the definition to handle the divergences at pT= 0 when d is small, which we will do in Section 4.2. Then in Section 4.3 we will prove

important properties of our definition. But first there

are a

couple

of

details to clear up. The J-dimensional subspace of p,|'s is chosen subject only to the requirement that it include all q .'s. So it is possible to extend the space to include extra dimensions. To show this has no effect on the value of the

integral

we must prove roo

KA JO for any



dpp-lg{p2)=\J

function,

true since the

g, which

right-hand

27r(v-D/2

where

AkKy_A dpjpr'gipf + k2) -

depends



(4.1.17)

JO



on a

scalar argument. This

equation

is

side is

iJ>'"-w,!o dxx(v-3)/2(l-x)-1/2,

r((v-i)/2), p* xp2 and k2 =

r qo

oo

(1



(4.1.18)

x)p2.

'parallel' subspace have no effect on both spaces to a common larger extend integral, merely is that The there sole problem may be a divergence in (4.1.18) at space. x 0; this we will cover by Section 4.2. To show that different choices of the

the value of the

we



Up till

fij"(v\

now we we



scalar function. If it is

component-by-component.

J-

component

have supposed /(p) is

work

d"p/12(p),

To

define,

a tensor

say,

the

Dimensional regularization

68

take the parallel space to include the 1- and 2-directions and any vectors which/17 depends. Then we proceed as before. q^ For example, suppose/°(p) pip*g{¥2\ where g is a scalar function. Then we

on



j/12(p)d<p JdpMp2 [d'-Wp2^1)2 =

while

dV n(p2)







(p2)2 + p2]



0,

d'-Wi^Pjj+P2)

dp oo

cir(d/2)

^yppd+it9(P2)

-h

"W29(P2l

Generalizing this result,

we see

that

JdVp^(p2) ^[dV2^(p2). =

More

general

cases are

treated in Section 4.3.

4.2 Continuation to small d The convergence of the definition (4.1.16) is ^-dependent at pT 0 and oo. It improves at pT oo when d at 0 it but pT smaller, improves pT gets when d gets bigger. Even for a function that decreases exponentially at large =







p, and that is analytic for finite p, the defining integral has a divergence if the transverse space has a dimension d J < 0; this is forced to happen if d is —

negative or zero. So our first task in this section is to find an explicit formula for the continuation of (4.1.16) to arbitrarily negative d. We will see that the pT-integral has poles whenever (d J)/2 is zero or a negative integer, but that these are cancelled by the zeros in 1/T((d J)/2). We will then be able to adopt the resulting formula as a definition of the —



^-dimensional integral of a function for which (4.1.16) converges for value of d. An example of such a function is /(P) The

q2,



(qi+P)2 + (q2



P)2

+ ?2'

and

parallel space must be at least two-dimensional to accommodate q l

so we may set J

only if d > if d < 2.







no

2. Then the transverse

integral

converges at pT

2, while the complete integral converges

at p



oo





only

69

4.2 Continuation to small d

ddp/(p)

We will have a definition that defines

for all small

enough

d. For larger values of d

we define the integral by analytic continuation. In will there be ultra-violet general poles at certain values of d just as in the we in Sections 3.5 and 3.6. Feynman graph computed -

explicitly define the continuation to small d, it is sufficient to consider function/(p2). Let us assume that/->0 rapidly enough as p-> oo that To



dV(P2) converges at p

-?

oo

for

some

2nd

(4.2.1)

dpp'-W



T(d/2).

positive value of d.

We also assume that/(p2) is

0. Then (4.2.1) converges and is analytic in d for some range analytic at p 0 < Re d < dmax. We define the integral for all other values of d by analytic =

continuation in d. Explicit formulae for the continuation to smaller d's are constructed by adding and subtracting the leading behavior at p 0. For -?

example, the following formula gives the integral in the range <dm

2 < Red



2^ ddp/(p2)

27Td/2



dPpd-\f(p2)

'fW){,



dpp--1[/(p2)-/(0)]+/(0)C-/d^



This is independent of the arbitrary When



2 < Re d < 0

we

d'p/(P2) at d



2nil2 =

0 the zero in

rw2)

-?

oc

to obtain

(4.2.3)

dp/-1[/(P2)-/(0)],

l/r(d/2) is cancelled by the pole

term to

fd°p/(p2)=/(0). We extend this procedure

positive integer

(4.2.2)

constant C.

may let C

!-

while



to continue to



give (4.2.4)

2/



2 < Re d <



2/ for any

/:

ddp/(P2)

2nd'2

f00



r(d/2).

dppd-l{f(p2)-m-p2m- ̄ ...-(P2)r\o)/i\},

f|d-2ip/(p2)



This equation gives

(4.2.5)

(-7r)-'/<'?(0).

us

the

integral

when



2/



2 < Re d < —21

on

the

70

Dimensional regularization

assumption that the original formula (4.2.1) converges when d is just greater than zero. Suppose now that (4.2.1) diverges at p oo for all positive values of d, but that/is power behaved as p -? oo. Then it is sensible to adopt (4.2.5) as the definition of the integral. This particular definition is very important =

since that

we

are

will

dimensional continuation to regulate Feynman graphs

use

ultra-violet divergent at d

Feynman graph frequently has





4. The definition

(4.1.16) applied

number d

negative



to a

4 of transverse

complete pT-integral with d

dimensions in order to ensure ultra-violet convergence of the

integral. Then replaced by d

we —

may

apply

the definition (4.2.5) to the

J.

Another obstacle to continuation in d is sometimes that

/(p2)

is not

0 but has a power-law singularity there. We may generalize analytic at p2 of the derivation (4.2.5) to write down a formula for the continuation of the =

integral. An example of the

use

of (4.2.5)

f(p2) where A /



The

(p2



definition is

A)/(p2

and B are numbers. The definition

1, (4.2.5) gives

integral

can

us a

be

jddp(p2 which



as a

can



A)/(p2



B)



B\



(4.2.1) diverges for all d, but with

definition valid for

explicitly computed

given by choosing



to

2 < Re d < 0:

give:

(7iBf2(A/B



1)T(1



d/2),

be continued to all d.

Suppose /has



power-law singularity, J

(P



q)2(p2



as

for

example

m2)*

The definition (4.1.16) of the integral of this function converges if 2 < d < 4. To continue it to lower d we must subtract the power behavior at p q, we can Then in 0 or at for at did as we (4.2.2). p singularities p1 0, just =







define the integral of say

(p



p2 q)2(p2 + m2)'

One result of all these definitions is that the

integral

of



power of p is

71

4.2 Continuation to small d zero:

(4.2.6)

d?p(P2r=o

for any value of a (integer or not). It should not be thought that there is any (4.2.6). It follows from (a) the explicit continuation of (4.1.16) to

choice in small

d, and (b) application of the continued formula

lddP(P2r.

Consistency of the formalism



(P2)7(P2



?2)-

Then when

ddp/(p2) If

linearity is

also





2a

requires (4.2.6). 2 <d <





2a

as a

definition of

For suppose we

that/(p)

have

d'p[/(p2)-(p2)7m2].

(4.2.6). singularity differently, by a function that is not But then, for example, the simplification obtained in (4.2.2) would no longer occur.

to be true then we have

We could subtract out the

just



power of p.

by taking



-?

oo

Observe that if in the first of

positive-definite function,

our

definitions (4.1.16), we take / to be a positive. But when the integral

then the integral is

is continued away from the

region where this definition converges, then the subtraction terms mean that the integrand is no longer positive definite, so that the integral need not be positive. At the end of Section 4.1, we proved that the value of a ^-dimensional integral does not depend on how we split the integral into an ordinary integral over some integer-dimensional 'parallel space' and a spherically symmetric integral over the remaining dimensions. We let J be the dimension of the parallel space. Then the proof consists of examining what happens when J is increased by one. Ultimately we had to prove (4.1.17), which is were no

not



property of ordinary integrals. We assumed d

subtraction terms. To generalize the result to the

positive,

we must



J,

case

so

that there

that d



J is

prove that [J/2-d/2]

K.d-j

A-J- 1

dpp1

-K'-'-'\

f(p2)- d/c

Z dpxPi

f 0

/w(0)p2-/n!

f(k2+P?)

[(J+l-d)/2] -



/■(?), (k2)p2n/n\

(4.2.7)

n = 0

symbol [d\ denotes the largest integer smaller than a. To prove the equation we change variables on the right-hand side to x and p, where Here the

72

p2-

Dimensional =

xp2

and k2

(1





x)p2.

regularization

For the

right-hand side

get

dPP>J-J-\

dxx(J-J-3)2(1_x)-l/2

K-d-j-i

we

[{J+l ̄d)!2)

f(p2)- =

^-,-1

dxX?-J-^2(l-X)-^2\



dP2(Pn2\di2-J/2-



JO



[j/2-d/2]

f(p2) ̄

fn\(l-x)p2)xnp2nln\



/(n)(0)p27?!



n = 0



[(J-c/)/2]

nn2n

[(J+l-d)/2]



n = 0





/i _

\j-n

2j





(4.2.8) Here we have added and subtracted l(J-d)/2] n = 0

so

that the

p2 of pd ̄J ̄2 times each scaling p2 by (1 x), we get:

integral

convergent. After

2iVd-J-

over

dxx(d-J-3)/2(1 _x)-l/2



square bracket term is





dp2(p2)d/2-J/2-1



JO [J/2-d/2]

f(p2)-

Z n



/(W)(0)P2^!



[(J+l-d)/2]xn/j _xjJ/2-d/2-n

2n

r.2j

[(J-d)/2]

/("V)-

/0)(°)



(n-j)l

Integration by parts in the p-integral gives Kd_J_l



dxx^-^l-xT^x [(J+l-<i)/2]



1-

/I

x"(l-x)J'2-d/2-"

?=0

[J/2-d/2] A-J-l

dpp

n = 0



1/2-

nl(l+J/2

-rf/2-n)lJ

4.3 Properties

73

x-integral is used to show that it equals r(l/2)r[(d-J-l)/2]/r[(d-J)/2], from which the required result

An

integration by parts

on

the

follows.

4.3 Properties Property 1. Axioms: The definitions (4.1.16) and (4.2.5) satisfy Wilson's axioms (4.1.2), etc., for ^-dimensional integration.

Proof. We reduced ^-dimensional integration to ordinary integration so linearity follows from linearity of ordinary integration. We must choose the p|| space to be large enough that it is the same for both functions / and g in (4.1.2). Our explicit continuation (4.2.5) to arbitrary negative d ensures that reducing the dimension of the transverse space is no problem. Scaling and rotation covariance are explicit properties of all our definitions. Translation in variance is valid for ordinary integration, so it follows from definition (4.1.16) provided the pu space is big enough to include the vector q used in the axiom

(4.1.4).

Property 2.

U (p2)a =n.2Md+2*-2Pn?+d/mp-*-d/2) r(<*/2)r(jB) J V + M2/ Proof. Immediate from (4.2.5). Note that this implies that the power

of

p2

is zero, since

T(f})



l/p

as





,J

integral of a

/?-?().

Property la.

|dJp(p2)*



0.

(4.3.1a)

Proof. Already done. Property 3.



dV(P,q,-)=

Proof. Contract with

kp^/(p,q,..).

(4.3.2)

<5q which projects out the derivative with respect to the component of q in the <5q direction. Then make the parallel space (of p|| 's) big enough to include dq and use (4.3.2) on ordinary integrals. This is true for all Sq. a vector

74

Dimensional

Property

regularization

4.

&% Proof and definition

(4.3.3)

ofdtj. Now 3ij is defined to be the component form of

a contravariant tensor with 8iJ



1 if / =j and

definition of the covariant tensor

thing. This gives

d.



3^5^



co.

5tj is

zero

otherwise. The obvious

the inverse matrix, i.e., the

as

However in

an

same

infinite-dimensional space,

there is space for a different definition. A contravariant tensor may be defined by specifying its components. But a covariant tensor co is fundamentally a linear function acting on covariant tensors: co(T).We

can

write co(T)=

only if the sum converges. symbolize by 8tj) to be rotation

(DtjTij

We need the covariant 5 (which

we

invariant, and to give S(T)= Tli whenever the sum exists. We would also like contraction with

5tj

to commute with

*u ddVPiPif(p2)



integration.

Sij "

and



5ij d>y/(P2) Since

we

have

us

infinite sum,

dW/(P2)/<*j

(4.3.4)

d>2/(P2)-

we cannot

immediately apply linearity

to

define am

dT{d/2)

ddplTijP,^(p2-l).

71

£T"

Whenever

But if the

example

equation.

prove this

Let

an



For

sum

converges, this definition

diverges, then it is possible

(43.5)

gives

to

get



finite value for 3(1). In

particular,

mJJwvj^n'-i)-*,

as

required.

rotationally invariant. Commutation of will now be a consequence of integration dtj two integrals which we will prove later.

The definition is

contraction with

commutativity

of

and



Property 5. Integration by parts:

|dV/(p)/^



0.

(4-3.6)

4.3

75

Properties

Proof. Work component-by-component. Contract with

an

arbitrary

vector k:

integral, we must put k in the parallel space, and we can of (4.3.6) for ordinary integration in the space parallel to k. proof

Then, to use

the

define this

{'

Property 6. To define integration

over two

(or more) variables:

|d-pdV(p,k;qi,-q?) we must

rules

choose to calculate

one

already stated.

integral

For this definition to be sensible

the order of

then the other,

need the result to be

we

to the

independent

of

integration: ddP dV



JddkjdV.

(4.3.7)

We could also allow the dimensions of the p- and different.

according

Then

of order of

exchange

integration

k-integrals

to

be

j'dcipjdd'k->j'dd'k{ddp.

is allowed only if d d\ or if / is independent of pk. Proof. It is sufficient to consider the case that there =

/=/(p2,p-k,k2). (If

there

q/s,

are

then

we

q/s,

so

that



finite-dimensional

we

apply the theorem

take out

integral for both k and p which spans all q/s and then

are no

remaining dimensions.) The left-hand side of (4.3.7) is

to the

4^-1/2

roc

dp

r(d/2)r((d-i)/2)

foe

/*oo

dkA

dkTp-,-lk"T-2f(p\kip,k2l+k2T), (4.3.7L)

while the right-hand side is /■ oo

f oo

4nd ̄l!2

r(d/2)r((d-i)/2).Jo

d/c



f*

dp, -

oo

| dp^-^-'/fPi+pip,*,*2).



(4.3.7R)

Here kl is the component of k parallel to p, while pt is the component of p parallel to k. Change variables to, say, p2, k2, and z p-k/(pk) Pi/ \/(P2 + P?) kJ \/(k2 + *r X with the result that both (4.3.7L) and =





(4.3.7R)

are

equal

4*d

1/2



to /??

'-1)/2)J0 dpp"

WI2)T({d

dkk*-

dz(l-z2f-3v2f(p\pkz,k2). (4.3.7C)

76

Dimensional regularization

The theorem is thus

proved in

the case that

(4.3.7L) and (4.3.7R)

are

both

convergent. Note that it is not a trivial consequence of the corresponding result for integer-dimensional integration. If the dimensions of the integrations are not the same, then let the k integral have dimension d'. The left-hand side gives

47C<d

+ d'-l)/2

T(d/2)r((d'



l)/2)

pd^-1 pdfc^-1 r

d2(i-z2)<d'-3>'2/,

which in general is not the same as the corresponding expression for the right-hand side. But if / is independent of z, then the z-integral can be

!■

computed

The result is

explicitly.

A-(d+d')!2

foe

&kk*-lf(p2,k2). 0

(4.3.8)

problem is

that if d is not

positive, we must make subtractions as in clearly asymmetric between the two orders (4.3.7L) and (4.3.7R) of performing the original integral (with now d' d);in practice, d will be the number of dimensions transverse to the external vectors A

(4.2.5). These

are



applications to Feynman graphs d will therefore be negative in regulate UV divergences. So we must use (4.2.5) to define the integrals. Then (4.3.7) does not give (4.3.7L), (4.3.7R) and (4.3.7C). We solve this problem by defining an auxiliary integral with a

q!,..., qN. In

order

to

convergence

factor,

say

I(a9d)- Assume



/ is power-behaved

d?V(p,k)exp [ at



a(p2



(4.3.9)

k2)].

or R) is UV and (4.3.7R) are IR (4.3.7L) The function I(a, d) is analytic in a and d.

infinity.

Then for all d, (4.3.7L

convergent. Moreover, if d > 1 then both convergent without subtractions.

Continue down to small enough d that (4.3.7) is UV convergent. Then I{a,d) is given both by (4.3.7L) and (4.3.7R) with / replaced by a(p2 + k2)], and with subtractions made. Now set a 0 to prove / exp [ =



the theorem (4.3.7). Property 7. ddk

d'p/(p2



k2)



dV(q2).

(4.3.10)

Proof. Since / is independent of pk, the previous theorem shows that the left-hand side is

independent

of the order of

integration,

even

if

the

11

4.3 Properties dimensions of the p- and

change variables

Property

to q

k-integrals

(p2





k2)1'2

are

and



different. Then =

use

(4.3.9) and

p2/q2.

8. w v sap ) dV-pW)





with

T1-1'

f'. \Tli-i<At[g], ^iS°dd'

[5ili25i3i4'--5i-*-it +



+ all

if t

rn

(4-3.11)

is even,

'?-

permutations

of the

(4.3.12)

fs]/t!,

and

,W

T(d/2)r(t/2 + l/2) ddp(p2)'/2<7(P2) T(\fl)T(dl2 +1/2) 2nd/2r(t/2 +1/2) [*

-r(i/2)rw2 +

,/2)J0d^""'"(^ lx





<43'13)

Proo^. If r is odd, antisymmetry of the integrand under p-> p makes integral over the 'parallel' space zero. Antisymmetry under reversal of one component of p, symmetry under —

the

permutations of the fs, and rotation invariance give the general form (4.3.11) and (4.3.12). Computation of one component (say, i{ i2 =

it













1) then gives (4.3.13).

Examples.

IdVpWt-aW U'ppW).

Property

9. Consider an

integral

/(Pi,...,P,)= which is

UV convergent

(4.3.14)

fdV(k,Pi,...,Pj)

by power-counting at





4; that is /

(4-3-16) =

0(l/k4+a) as

infinity in any direction, for some positive number a. Then the is analytic in d and in the parameters p., when d is close to four, if the

k goes to

integral integrand

is

analytic. If

integral at d

integral of/.



the

pf's

lie in the first four dimensions, then the

4 has the same value as the

ordinary four-dimensional

78

Dimensional regularization

Example Suppose / has the form

/=/(R5p1,...5pJ)exp(-Ak2 where for

any vector

let 7 be the

we

let





be its projection onto the first four

be its

definitions of ^-dimensional integration,

d£i"4kexp(-Ak2







have



2Ap-k)

I(7i/A)di2-2exp(Av2).

manifestly analytic

in d and p. If

7 Notice that there is

we

(d'k/

/=

This is

(4.3.17)

2Ak-p),

projection onto the remaining dimensions. Then ordinary four-dimensional integral of/. By use of our

dimensions and let we





no

dimensions. However, if



(4.3.18)

we set





4, the integral becomes

7exp(^p2).

restriction

on

p, even

though





0 in four

let d->4 and p->0, the limit is smooth.

we

Proof of Property 9. The proof is easily made by examining the definition of the ^-dimensional integral in terms of ordinary integrals. As usual we divide the space into a finite-dimensional parallel space big enough to contain P!,...,Pj, and into a transverse space containing the remaining

dimensions.

It is convenient to choose the

parallel

space to have

an

odd

number 2N + 1 of dimensions. Then: {d-l):2-N

lT((d-l)/2-N)

x{F(kl9k2,...9k2N+l9k*)-



dk2T(k2Tyd ̄

Fi?\kl9...9k2N+lMlV*d

(4.3.19)

Here we have used F to denote / considered as a function of the first 2N + 1 components of k and of k2. Since / is an analytic function of k, it can be expanded in powers of k2. As required by the definition, we have subtracted offa power series in k\, to

give convergence at fc7 respect to



0. We use F(n) to denote the nth derivative of F with

k\. The integrand

convergence at kT



of the k2 integral behaves

0 if d > 3. Since

/ is analytic,

as

kd- ̄5,

have

so we

there are

no

other

singularities at finite k, and the only other possible source of a divergence is from large k. The subtractions do not introduce a divergence provided that

d<

5. Moreover,

there

are no

we

have assumed that

other large k

divergences

/= 0(l/fc4+fl)

as

fc-> oo,

so

that

when d is close to four. Hence (4.3.19)

79

4.3 Properties

analytic in a neighborhood of d 4. Now the integral will depend on the p.'s only through the Lorentz scalars Va'Vb (with 1 < a < b < J). To determine this dependence, it is sufficient to keep the p£'s within some fixed J-dimensional subspace. Since (4.3.19) is a perfectly finite integral, it is an analytic function of the p-'s. converges and is



To determine the value of the

the first four dimensions,

integral when d



4 and when the

p^'s are in

the freedom to vary the dimension of the in the definition (4.3.19). Let us now make it four 'parallel' space dimensional. We will obtain an integral of the form: /=



we use

d/c1...d/c4T^-^ pdfcfe''-5/(/c1,/c2,/c3,/c4,P)

if d>4. When

we

let d->4, the

resulting divergence cancels the

I(d as



4)=

required. alternatively continue from f



The

2nd/2 ̄2

over



Idk.dk^k.dkJik^k^k^k^Ol

We may /

k is singular at fc of the inverse T-function to

integral zero

f00

Jd/c,...d^Y^7^^2) J

d<4

0; the

give (4.3.21)

using

dfc^"5C/(fci>- -,^,^2) /(^i, -





(4.3.20)



^4,0)]. (4.3.22)

singularity at k 0 is cancelled, but as(i->4we gives the same result (4.3.21). =

get



divergence

at

k= oc which

Comment In this proof we used the freedom to alter the dimension of the 4, it was parallel space. To show that the integral is well-behaved at d =

convenient to choose the parallel space to have an odd dimension. But to 4, it was convenient to choose the parallel compute the actual value at d =

space to have an equivalence in a

dimension, specifically, four. It is instructive to see the simple non-trivial case. (The general case was summarized even

at the end of Section 4.2.) Suppose 3 < d < 4. Then define nm-2 7? =run n

r{d/2 foo

/2





2) J



dk\B)d^f{k^k2^k^\

-(<*-5)/2

(4.3.23)

foo

J d/C5r(^-5))J0dfcT2(fc')(d"7)/2[/(/Cl'---'fe4'fc'



fcT2)

-f(ki,...,k4,k25)l

(4.3.24)

80

Dimensional regularization

Our definition of the ^-dimensional integral of

dfc1...dfc4/1

/= J so we must prove

To do this

(1



x)k2

we





tells

that

us

dfcj ...d/c4/2,

(4.3.25)

00



that lx= I2. change variables in /2, by setting k2



xk2 and

k\



to obtain:

wW-5)/2

r(d/2-5/2) Jo

Jo

1/2^/2-7/2 dk2(k2)dl2-3(l-x)-1!2x'

xlf(kl,...,k4,k2)-f(k1,...,k4,K2(l-xm

rr(d-5)/2

r(d/2-5/2)JO

dx\ dH2kd-6(l-x)-ll2xdl2-7/2x J0

x{[f(ku...,kA,K2)-f(k1,...,kM -U'(k1,...,k4,H2(l-x))-f(kl,...,k4,0)]}. In the last line we subtracted and added

integrate seperately each

(4.3.26)

f(kt,...,k4,0), so that we can particular, we have

term in square brackets. In

dfc2^-6[/(/c1,...,/c4,P(l-x))-/(/c1,...,/c4)0)] =

(!-*) 2-d/2

Comparison with

the definition

i2_7i-1/2r(d/2-2) f1 I,

dk2kd-6lf(kl,...,k^2)-f(kl,...ik^0)].

T(d/2-5/2)

(4.3.23) of /x

shows that

dx[(l-x)-1/2xd/2"7/2 -x^2-7/2(l-x)3/2"d/2].

The integral is in fact the analytic continuation from d > so

that it

10.

Multiple integrals integral

function, result lx =/2

Property

equals F(d/2-5/2)r(l/2)/r(d/2-2). follows. are correct at







(4.3.27)

of a

The

beta-

required

4.

Consider the

/(P15...,PN)=

\d%...ddkLf(kl9...9kL^l9...9vN).

(4.3.28)

This might represent a Feynman graph with N external lines and L loops. Then the p-'s and k.'s represent momentum vectors. Suppose that at d 4 the integral is completely convergent in particular that there are no ultraviolet (fc-> oo) divergences or subdivergences. If we restrict the pf's to the =



first four dimensions and set d

integral of /.



4, then / is the ordinary four-dimensional

4.4 Formulae

for Minkowski

81

space

Proof. For each vector v, we define the projections v and v, onto the physical and unphysical dimensions, as before. The result to be proved is that if Pj



pf for each of the p.'s then / defined

as

the limit

as

d -? 4 of the

dimensionally regulated integral ordinary integral. We each over a into an four-dimensional split integral integral ordinary kz over kt and a (d 4)-dimensional integral over kf. The result to be proved is is identical to the

can



then that

limjd'-4^...^^^ as

(4.3.29)

d->4. This formula is

proved by doing all but the integral over ft,.

Let the result

be/(1): /(l)(k1)= its only dependence (4.3.29) is

on

kt

idd-%...dd ̄*kLf;

is via its

length.

(4.3.30)

We then have that the left-hand

side of

by use of the Property 9. dimensions. We

can

dependence on kj

is

on

its first four

|d-4k3...d--4kL/(kl,k2,k3,...).

/(1)(0)= Another L—2

Notice that the

then repeat this process to show that

repetitions give (4.3.29),

from which the desired property

follows.

4.4 Formulae for Minkowski space In this section

we

derive



collection of results that

are

useful for Feynman

graph calculations.

4.4.1 To convert an

Schwinger parameters

arbitrary graph in d dimensions to using



parametric integral,

we

first rewrite each propagator

Then

we

perform the momentum integrals.

Since all Feynman

graphs are of

82

regularization

Dimensional

polynomial in momenta times a product of simple scalar propagators, we only have to calculate d-dimensional integrals of the form: the form of



I^"?(A,B)

ddkk>"



/c"-exp [

...







Ak2

2Bk)l



(4.4.2)

on the parameters introduced by (4.4.1), while B" these depends parameters and also linearly on the other momenta (both external and loop momenta).

Here A

depends only on

By linearity

we can

'1

uses



■=n 2dBUl J=1 v_??/ljV

It

(This

find ln by differentiating /0:

linearity

translation fc"-?fc"

ddkexp{Ak2



i(n/Af2 exp(

\(n/A)d/2exp(

4\ddkk>-k>'kvexp(Ak2

1 A





the

i(n/Af2e-B2::A



2B-k)

B2/A)(











(4.4.4)

2B-k)



d"/c/c"/cvexp(^2

Ix3"v=



70 by using

i(n/A)dl2exp(-B2/A).

/?= \ddkk"cxp(Ak2 =

(4.4.3)

B"/A, the scaling /c->/c A ̄i/2, and Wick rotation:

Thus



2B-k).

of rf-dimensional integration.) We find —

I0(A,B)= \ddkexp(Ak2

I"2V=









B?/A),

(4.4.5)

2B-fc)

B2/A)(B"BV/A2 -ig"v/A),

(4.4.6)

2B-k)

BXB"BV

(B' g"v +

B"gXv + Bvgx") 2A2

(4.4.7)

\ddkk",kxk"kvcxp( ̄B2/A)

i(it/A)d'2exp(-B2/A)

VBKBXB?BV

(BKBxg"v + five similar) 2A3

A*

}■



{gKXg'lv + two similar)  ̄| 4A2

(4.4.8)

Each of the loop-momentum integrals is performed in this way. At each stage the momenta only appear quadratically and linearly in the exponent.

4.5 Dirac matrices

83

4.4.2 Feynman parameters It is also

common to use

'-£

l/(AB) and its



dx/[Ax + B(\



(4.4.9)

x)]2

generalizations: 1

I\* +

A*&—E'

p+ ?■?&)



r(a)r(^)---r(e) x

Jo

dxdy-dzd{\—x





z)



xa-y-l...ze-l (Ax + By + ---Ezf+fi+-+e'

(4.4.10)

Here A, B,...,E represent the denominators of the propagators of a Feynman graph. The resulting momentum integrals have the form

k'"---k''"

C JHf-lh,— lAdh ?v*

]a\-k2_2P.k + cy

Application of (4.4.1) and J0

\d"k/(





k2



the results

2p-k

ind/2(C + p2)"'2 "Ha

J\



Jf

indl2(C + p2Y'2-'(







\ddk k"/( -k2-2p-k





(4.4.4)-(4.4.8), etc., gives

Cf







{(4.411)'



rf/2)/r(a), Q*

p")r(a

d"kk?kv/(-k2-2p-k

(4.4.12)



d/2)/T(a),

(4.4.13)

C)a



ind,2(C + p2f2-xx x

[r(a



d/2)p"p"



T(a







d/2)g?v(C



p2)/2]/r(a).

(4.4.14)

4.5 Dirac matrices The Dirac matrices satisfy the following

properties:

(1) Anticommutation relation: {/,yv}=20*vl.

(4.5.1)

84

Dimensional regularization

Hermiticity:

(2)

/w

*-/ if/x > 1.

When we use dimensional regularization, the Lorentz indices range over an infinite set of values, so we need infinite-dimensional matrices to represent the algebra (4.5.1). We will also need a trace operation:

trl

=/(<*),

representation behaves as if its dimension were f{d). We must require f(d0) to be the usual value at the physical space-time dimension, d d0. Usually this means /(4) 4. that the

so





linear operation on the matrices which we will define later. integer dimension d 2co, the standard representation of the /'s

The trace is

In

an even





has dimension 2°\ However, it is not necessary to choose f(d) 2d/2. The variation / (d) f (d0) is only relevant for a divergent graph, so, by Chapter =



7, any change in f(d) amounts to a renormalization-group transformation. usually convenient to set f(d) f(d0) for all d.

It is



To set up a formalism for dimensionally regularized y-matrices, treat the following issues:

(1)

we must

representation of the anticommutation relations; consistency. for the trace of an arbitrary product of yM's must be

We must exhibit



this will ensure

(2)

The formulae

derived. (3) While a knowledge of the /'s alone is sufficient for QCD and QED, we must show how to define a y5 so that we can treat chiral symmetries.

This will also give

us a

definition of the

antisymmetric

tensor

eKA/iV.

following construction gives a representation: Let co be a positive integer, and suppose inductively that we have defined a 2W dimensional representation yfw) of the algebra (4.5.1) for 1 by 1. We define the infinite dimensional y* for 0 < \i < 2co 0 < fi < 2oj having a sequence of y^'s down the diagonal, and zeros everywhere else: The





VM





\i



2a> +



0 vy

o vC.. no?





(4-5-3)



higher representation yf(0+1} of dimension 2<° *, 1. In order that (4.5.3) apply independently of co, we must

We will construct the next

with 0

bu —  ̄

85

4.5 Dirac matrices choose

This satisfies the anticommutation relations (4.5.1) and the hermiticity relation (4.5.2), provided that 0 < fi, v < 2co 1. Our task then is to find —

f°r ^ 2w and 2co + 1. Notice that the induction starts with

yfa>+1)





*°?=

0\



co



1. We can define

_/

0 1

-i)> y^-\-\

Given the 2a)-dimensional

?0)-



n_i o

representation yfa

H-5-4)



_,n



we

define another matrix

Observe that

&> Also, when at co at d





y<?>

?<?>,

2, we have f







{iwX>}=o.

1>

(4-5-6)

y5, in the usual notation for Dirac matrices

4. We define

/,w+1,"V-y?.,



yivi=(-° ifr)- It is easy

to

check

that

(4-5-7>

(4.5.1) and (4.5.2)

are

satisfied for 0</i,

v£ 2(o+l.

explicit representation of the Dirac matrices for any u>, (4.5.3). Standard manipulations involving the anticommutation relations are valid independently of d. Two useful results are: We

now have an

and for the infinite-dimensional case, because of

rtv=T{y",y?}i=<i /W?





rfi,

(4-5.8)

20?vy"-/y?yv (4.5.9)

(2-d)yv.

We also need traces of y-matrices, in a



graphs

with fermion loops. The trace of

atv(A)

+ b

matrix is linear :

ti(aA



bB)



(4.5.10)

tr(fl),

and is cyclic: tr(AB) Here A and B

are

any



(4.5.11)

tr(BA).

product of y-matrices, and



and b

are any

numbers.

86

Dimensional regularization These

with the value of tr 1, define the trace of any

properties, together

linear combination of products of For example, tr

(y"yv)



tr

(//)

(cyclicity)



tr







so we

y-matrices.



/y



tr

(y"yv)

2¥T 1)



(anticommutation)

Vvtr

(linearity),



have the usual result tr(yY)



^vtrl.

(4.5.12)

0KVV + 0KV")tr 1.

(4.5.13)

Similarly tr

The trace of the

(yKyVyv) product

(0KVV



of

an



odd number of

y-matrices

is

zero.

For

example dtryA

=tr =



so

tryA



[yKyKyk)

-tr(yY7j + 2tryA

-tr(yKyV)-h2tryA,

0.

possible to make a more constructive definition of the trace, along the lines of (4.3.5). It is necessary to check consistency. We can find a formula for the trace of any number of y-matrices generalizing (4.5.13). It is true for any finite-dimensional representation, yfa, so it agrees with the algebraic properties. Linearity defines the trace of more general products. with the trace. This We must also check that contracting with g^ commutes can be checked directly. A possible explicit definition of the trace of a matrix with components It should be



Mijis trM,. This definition

of



(trl)limi N-ooiV

£ Mjy

(4.5.14)

j=l

exploits the fact that each matrix y" is an infinite set of copies

finite-dimensional

independent



of d, the

yfa strung along the diagonal. Since the y^'s

only possible rf-dependence is in the choice

are

of the

value of tr 1.

4.6

y5and

ekx^v

dimensions, y5 iy°y1y2y3 and eK>fiX is Lorentz-invariant tensor with e0123 1. We need

In four







totally antisymmetric example, to define

y5, for

87

4.6 y5 and eK^v the

current

axial

ie^^yVV/^!,

and

we







i£*^v tr 1

appropriate definition changes

Instead of y5 £01

£-tensor





we

have

£10, £00

To continue

f(1) £u



in

comes

because

y5



have the trace formula:

tr y5yKyYyv The

The

i/sy^y^.





y°y19 and





when

ieKXflv tr 1. we go

instead of eKX/lv

to

we

two

dimensions:

have £MV, for which

U.

dimensionally,

might expect

we

y5 to

satisfy

as in four dimensions. But then, as we will see in Chapter 13, the only consistent result for y5 is that it has zero trace when multiplied by any string of y"'s. Thus we do not have a regularization involving the usual y5.

just

writing

A consistent definition is obtained by

ys

!1 -



if

1 if

iyoyy y3

(K??fiv) is (KAfiv) is



iryrvw4!>

an even an

odd

permutation of (0123), permutation of (0123),

(4-6-1)

(4.6.2)

0 otherwise.

This definition is not Lorentz invariant

on

the full space, but only

on

the

first four dimensions. We have

{y5,/}=0, |j5> V"]

(7s)2



0,1,2,3,



0,

otherwise,



1,

/5=y5-

(4-6.3)

a nuisance, but it does give the correct Hooft & Veltman anomaly ft (1972a)).

The lack of full Lorentz in variance is axial

if ^

Renormalization

chapter we come to the general theory of renormalization. The basic difficulty is that a graph may not only possess an overall divergence. It may have in addition many subdivergences which can be nested or can overlap in very complicated ways. Most of our effort must go to disentangling these complications. We will begin by investigating some simple graphs. These will show us In this

how to set up the formalism in the general case. The ultimate result is the forest formula of Zimmermann (1969). Contrary to its reputation, this is not

procedure, designed for pedantically rigorous treatments. Rather, the forest formula is merely a general way of writing down what is in an

esoteric

fact the natural and obvious way of extracting the divergences from any integral. Its power is demonstrated by the ease of treating overlapping

the handling of which is normally considered the bete noire of renormalization theory.

divergences,

The forest formula is applied to individual Feynman graphs. It extracts graph by subtracting its overall divergence and its

the finite part of a

subdivergences. We will, of course, need to show that the subtractions can be implemented as actual counterterms in the Lagrangian. We will also show that the counterterms are local, i.e., polynomial in momentum. An important advantage of using the forest formula to obtain the finite part of each graph, rather than working directly with counterterms in the Lagrangian, is that the procedure applies to more general situations. As we will

see

more

in Chapter 6, it will enable

important

momenta of



case

is the

us to

renormalize composite operators. A

computation of asymptotic behavior as external

Green's function get large. For this case, the forest formula

good derivation of Wilson's operator-product expansion, which we will discuss in Chapter 10. Let the value of a Feynman graph be written as:

permits



/?

U(G)(Pl,...,pN)= d"fe,...d'ifeL/(p1(...,plv;/c1,...fct)- Here pl9...,pN are the external momenta, and

k{,...,kL

are

(5-0.1) the loop

5.7

Divergences

89

and subdivergences

large-fc behavior divergences. Moreover, the very same techniques can be used to

momenta. Renormalization is removal of that part of the

that

causes

extract the behavior for large p as product expansion in Chapter 10. -

we

will

see

when

we treat

the operator

Although Weinberg's (1960) theorem tells us the power-counting applicable to either kind of asymptotic behavior, it does not tell us how to organize it. In particular it was only much later that Wilson (1969)

expansion, which is the important tool in computing asymptotic behavior, for example in deep-inelastic scattering see Chapter 14. Many generalizations have been made see Mueller (1981) for a review. These are phenomenologically very important, and the method by which they are proved is close to that for Wilson's expansion. formulated his operator product





5.1

Divergences

The idea of renormalization

and

subdivergences

theory is that ultra-violet divergences

of



field theory are to be cancelled by renormalizations of the parameters of the theory. We propose to prove this in perturbation theory. The use of

perturbation theory implies that we expand the counterterms in the action in powers of the renormalized coupling, g, thereby generating extra graphs with these counterterms as some of the vertices. To avoid superfluous technicalities, space-time.

we

will consider the

case

of

</>3 theory

in six-dimensional

was discovered by Shirkov & and (1955, 1956, 1980) Bogoliubov Bogoliubov & Parasiuk (1957), and we shall follow their approach. The first step is to decompose the

A very efficient way to understand renormalization

Lagrangian

as

follows: ^

Here

i/(p2

S£0 -



^0



^b



(5.1.1)

^ct.

is the free Lagrangian used to generate the free propagator

m2 + ie) in perturbation theory:

(5.1.2)

J?0=(d<t>)2/2-m2(f)2/2, with if b



being the renormalized mass. The rest of the Lagrangian, <£, JS?ct, is the interaction, and consists of two terms. The first, which we will





call the basic interaction, is

(5.1.3)

J?b=-g<P3/3\, where g is the renormalized coupling. The second term, J^ct, counterterm Lagrangian.

Consider graphs generated by

we

will call the

the basic interaction. These have UV

90

Renormalization

divergences

which

are

be cancelled

to

by graphs with some Lagrangian

of their

interaction vertices taken from the counterterm

JS?rt (The



(5Z(<30)2/2



5m2(j)2/2



5gcj)3/3!



cj> is needed to cancel tadpole graphs

(5.1.4)

dhcj).

Figs. below.) In order to give meaning to 5Z, dm2, 5g, and 5h, we must impose an ultra-violet cut-off. We will use dimensional regularization in the term linear in



see

5.1.4 and

5.1.5

following sections. The key to the method that we use is to realize that each of the three terms

Lagrangian should not be considered as a single quantity. Rather it is to be considered as a sum of terms, each of them cancelling the overall divergence in one particular graph generated by the basic interaction. For example, the self-energy graph, Fig. 3.1.1, gives a contribution SXZ to SZ, and a term d^m2 to dm2. Our calculation of this in the counterterm

graph led

to

(3.5.7),

so

with minimal subtraction

51Z



^m2

^/[3847r3(rf-6)],



Then c5Z=

we

have





32m2/[647r3(d-6)].j £ 5GZ

graphs G









81Z+--;

with similar formulae for the other counterterms.

utility of this idea by examining graphs like those in Fig. 3.2.1 Graphs like Fig. 3.2.2 contain vertices corresponding to the counterterm bxm2 (and SlZ). Such graphs are all generated by taking graphs like Fig. 3.2.1 with no counterterms and finding where Fig. 3.1.1 occurs as a subgraph. Substitution of the counterterm for one or more of We saw the

and Fig. 3.2.2.

these subgraphs gives the graphs with counterterm vertices. This leads to the idea that we consider by itself the renormalization of

single graph generated from

the basic

Lagrangian.

We add to it

a set



of

counterterm graphs to give a finite result. Only as a separate step do we recognize that the counterterm vertices are, in fact, generated from a piece of an interaction

Lagrangian. graph-by-graph method is probably the most powerful approach to understanding not only the problem of ultra-violet divergences but also many other problems in asymptotic behavior. Even so, it is not at all trivial The

to

ensure

steps

that the renormalization program can be carried out. The essential

are:

(1) To find the regions in the ultra-violet divergences.

space of

loop

momenta of a

graph that give

5.7

Divergences and subdivergences

(2) To show how to generate



series of counterterm

91

graphs

for



given

basic graph. (3)

(i.e., polynomial in

To show that the counterterm vertices are local

momenta). (4)

To find the conditions under which the counterterm vertices amount

only

to renormalizations of the parameters of the

Lagrangian.

complications in carrying out this program arise when one treats the divergence of a graph which has a divergent subgraph. To understand why there is a difficulty, we will examine the graphs of order g* for the full propagator Figs. 5.1.1-5.1.3. The

case of the



—O—O—+—O—*-



-*—0-+



11

(a)

(b)



11

id)

(c)

Fig. 5.1.1. A two-loop graph for the propagator in counterterm graphs.

(j>3 theory, together

with its





l^^^



(?)

*2

J^^ (?

(c)

Fig. 5.1.2. A two-loop graph for the propagator

in

counterterm graphs.

<D-

<j>3 theory, together

with its



(a)

(c)

(b)

(d)

Fig. 5.1.3. A two-loop graph for the propagator in (j>3 theory, together with its counterterm graphs.

We ignore the graphs with tadpoles, such

divergent

and need

condition that zero

<O|0|O>

(e.g., Fig. 5.1.5),

Let

a counterterm

us return to

there is one basic

Sh(p.

We

as

can

use a

vanishes. Then the total of the

so we

5.1.4. These

Fig.

are

renormalization

tadpole graphs

is

omit any graphs containing them.

graphs listed in Figs. 5.1.1-5.1.3. In each set graph and a set of counterterm graphs. Ultra-violet

the sets of



¥5

Fig. 5.1.4. A tadpole graph, together with its counterterm.

92

Renormalization <o|0|o>=







—<D



Fig.





—CV—£>+—*

5.1.5. Graphs to

0(g3)

for



<O|0|O>.

divergences

involve a loop momentum that gets large, so the divergences confined to one-particle-irreducible subgraphs. The simplest always case is Fig. 5.1.1, where the basic graph has two insertions of the one-loop self-energy. It is made finite by adding the graphs with one or both of the self- are

energy subgraphs replaced by a counterterm. (We use a cross to denote a counterter m in a graph, and we use the label' 1' for the counterterm of the one-

loop self-energy.) In Fig. 5.1.2, the basic graph is more complicated. We will treat it in detail in Section 5.2, and we merely summarize the results here. It has two UV divergences. The first comes from letting both loop momenta k and / go to infinity; we call this the overall divergence. But there is also a divergence where the momentum in the outer loop stays finite. This is an example of a subdivergence. Its existence, as we will see, implies that there is a term

proportional

to

p2 In (p2)

in the

divergence

of the basic

self-energy graph.

This cannot be cancelled by any local counterterm. However there is also a graph with a counterterm to the subgraph. This graph is Fig. 5.1.2(h). After we

add the two

graphs, the non-local divergence cancels. The

overall

divergence in we use

the result is then cancelled by a local counterterm, for which the label '2\ We implement this as a counterterm in the action by

inserting terms S2Z and 52m2 into the complete counterterms 5Z and dm2. The pattern is simple. We consider as a single finite entity one basic graph together with counterterms for its subdivergences and for the overall divergence. Another case is shown in Fig. 5.1.3. There are two subdivergences, each corresponding to a vertex subgraph. The one-loop vertex is logarithmically divergent and is made finite by renormalizing the coupling (Fig. 3.6.1). Since the two divergent subgraphs overlap, the counterterm graphs are generated by replacing one but not both of the vertex graphs by its

counterterm. The overall divergence is then local and is cancelled by 34Z and S4m2.

counterterms

5.2

Two-loop self-energy

Our first task in Section 5.2 will be to

generalize the argument

we

in

verify

(</>3)6

93

the above statements. To

will then observe that

power-counting

as

in

Section 3.3 determines the strength of the overall divergence. To prove that the presence of subdivergences does not affect the form of the overall counterterm, remove

we

will differentiate with respect to external momenta to divergence. Then we will be able to construct an

the overall

inductive proof that if subdivergences have been cancelled by counterterms then the overall divergence is local and its strength is determined by simple

power-counting. We will also show how to

disentangle the combinatoric problems when subdivergences Finally, we will discuss Weinberg's theorem. This theorem tells us exactly which regions of momentum we must consider. In practice one is very simple-minded about locating UV divergences. For example, we stated that the regions giving divergences for Fig. 5.1.2are: (a) k and / going to infinity together, and (b) I going to infinity, with k fixed. In each region, all the momenta get large in a particular 1PI subgraph that is divergent by power-counting. Weinberg's theorem tells us that these are the only regions we have to consider explicitly. In the case of Fig. 5.1.2 there is another region that is important, where / goes to infinity with k also going to infinity, but much more slowly. This region interpolates between the other two, but in fact does not need to be treated as a separate are

nested.

case.

5.2

In this section

Two-loop self-energy

in

(¥3)6

will

explain the properties of overall divergences and subdivergences by computing the two-loop self-energy graphs, Figs 5.1.2 and 5.1.3, in 03 theory at space-time dimension d 6. We will again use we



dimensional regularization, and will need the values of the one-loop counterterms in order to cancel subdivergences. The one-loop self-energy was considered in Section 3.6.2, where we found

that the counterterms needed

one-loop (cf., (3.6.13)) the

vertex

given by (5.1.5). We can also compute graph, Fig. 3.6.1, with the resulting counterterm being ?3ff

It is worth



were

M3"-/V/[647c3(d-6)].

noting that this implies a

(5.2.1)

value for the one-loop term in the bare

coupling: 0o





iy-d/20+<530 + 0(<75)]Z-3/2 gn3-"'2{l +ig2/[64n3(d-6)-\



0{g*)}.

(5.2.2)

94

Renormalization 5.2.1 Fig. 5.1.2

In order to be able to compute Fig. 5.1.2 in closed form massless theory. The value of the

**2a"

graph

we

work with the

is then

(2n)2d 1



'[(p k)2 ie](fc2 ifiT2(/2 is)[(fc 02 2Jdd/cdJ/r^_L^2_L.o-|^2_L.^2^72_L.^r^_ +









, A2 +

ie]

.nl.

(5.2.3)

The inner loop is easily computed in terms of T-functions: 1

K2 dd/7277  ̄l2{k-l)2



now



d/2)

dx[

fc2x(l





x)]d/2

-2





We

i^/2r(2

i^2r(2



*A d/2)r{?/.2 r(d 2)





*2)?2"2.

have



ld'^iY-?[-(p+*)■]?

(5.2.5)



The denominators

can



be combined by

Feynman parameter:

r(4-rf/2)J0dX[^x

after which the /c-integral L2°



r(5-d/2)r,

^4-<,/2B



(5.2.4)

can

\64n3) 2\4nn2J -

d/2)T(5

T(d





3-d/2

B(l-x)]5-*2'



2)r(4



d)r(d/2 -



l)3r(rf

rf/2)r(3rf/2





4)

5)

^)2y(4^)"6r(2-rf/2)r(5-^w The overall ultra-violet

(5"2"6)

be performed. The result is

-(jLYeL(zpLY ̄6 T(2



divergence is contained

<5-2-7)

in the factor T(5

d). Observe that the argument of this T-function is exactly minus half times the degree of divergence. The subdivergence is contained in the factor T(2 d/2); this is the same as we calculated in Chapter 3. -



Before there

are

we

discuss further the UV divergences we should observe that divergences. These come from the existence of long-

also infra-red

range forces in



theory with massless fields.

In momentum space,

they

5.2

Two-loop self-energy

in

(^3)6

95

divergences when some momenta go to zero. For example, if integral over the momentum through any propagator has a divergence at zero momentum:

appear as d < 2 the

ddq/q2



constant/^

This accounts for the factor

divergence problems,

at k

so we



its

T(^/2-l)3.

ignore the IR divergence.

expand Z2a in dependence on p2: us

1,2(2

2)

at d close to 2.

When d<4 there is also

0 with / and p fixed. Our

divergences would go away, but Now let



g2 \Vf

\64ttV

we

would not have

powers of d





only

If we used

36((<*-6)2+<*-6



concern

-—=-

Aw2

is with UV



massive field, the IR

an

explicit formula for

6 to exhibit its

In I



divergences,

I -|- constant

+iln2( t-P ̄2 ) + constant In ( —^ j + constant + 0(d



6)

and



1. (5.2.8)

The double pole at d

logarithm in the finite part are both reflections of the fact of having a subdivergence. The p-dependence is a power of p2 times a polynomial in In( p2). This is a characteristic feature =

6 and the double



of massless theories. The simple pole has

in p. Consequently, it cannot be cancelled by any local counterterm. It is easy to see that this is caused by the presence of the subdivergence. The a

coefficient that is

not

polynomial

comes from the region where the loop momentum of the inner loop goes to infinity while the momentum k in the outer loop remains finite. Integrating over finite k gives a logarithm of p times the divergent part of the inner loop. We have already introduced into the Lagrangian a

subdivergence

counterterm

for the inner loop,

which this

counterterm

so

appears

that there is

in

such





graph, Fig. 5.1.2(h),

way

as

to

cancel

in

the

subdivergence. Therefore the sum of Figs. 5.1.2(a) and (b) should have no subdivergence, but only an overall divergence. This can be cancelled by a local counterterm We will prove this in Section 5.2.2 by differentiating three times with respect to the external momentum pfi; this gives a result which has negative degree of divergence, i.e., there is no overall divergence.

(i.e., a polynomial in p).

Since the subdivergence is cancelled, there is no subdivergence whatever, so the counterterm must be quadratic in p. We represent this by Fig. 5.1.2(c).

96

Renormalization

In Section 5.2.2

we

will make explicit this proof of locality of the

counterterms. Here

we

case that

will





verify

the above statements

by explicit calculations.

In our

0, the value of Fig. 5.1.2(h) is ffV-' |\?. d"k

Z?=-i8lzZ-Z-r- (2nf e2 V

64n3)



k2

(k2)2(p + k)2

(-p2\r(2-d/2)r(d/2-i)2/-P2y2-3 \ 6 J (d-6)T(d-2) \4nfx2 )

g2 VP2f

-2



647i3J 36{(d-6)2 —£ln2(

-—f



d-6

In (

-—j

) -|- constant

j + constant In ( —^ ) + constant + 0(d- 6) (5.2.9)

The non-local divergence disappears when

we

add this

graph

to

Z2fl,

with

the result v

+v



( g2 Vp2 S

+iln2i2 ( —^-







constant

) + constant In

\4V/

f —V-^ ) + constant + 0(d \4V/



6) }.



(5.2.10) The non-local divergence has cancelled, as promised. However, the double pole and, in the finite part, the double logarithm have not cancelled, even though it is evident from our calculation that they are associated with the subdivergence nested inside the overall divergence. This is a general phenomenon. Indeed we will see in Chapter 7, where we discuss the renormalization group, that the coefficients of the double pole and of the double logarithm could have been predicted from the one-loop counter- terms without any explicit two-loop calculations. Since the non-local divergences have now cancelled, the overall

divergence

Then

can

we

be cancelled by

obtain at d







choosing



wave-function counterterm

finite result, which

we term

the renormalized

5.2

value of

Fig.

Two-loop self-energy

(^3)6

97

5.1.2:

E<2MRS)=S2fl + Z26 + (E2c =

in



-p2<52Z)

{<£?] fe {>2(j¥)

+ constant ln

{i¥) constant[ +

(5.2.12) 5.2.2 Differentiation

with respect to external momenta

graph Fig. 5.1.2 has an overall divergence which is local, only after we have subtracted the subdivergence. In we will need to show that the counterterm of a 1PI graph is a general in its external momenta with degree equal to the degree of polynomial divergence. Our argument (following Caswell & Kennedy (1982)) depends on differentiating with respect to external momenta. In this subsection we will apply the argument to Fig. 5.1.2, emphasizing its generality. Then in the next subsection we will apply it to Fig. 5.1.3. Even though that graph has an overlapping divergence, traditionally considered We

saw

that the

but that it is local



hard

problem, we will see that our method works as easily for this graph as

for Fig. 5.1.2.

Fig. 5.1.2(a) three times with respect to p", to make divergence negative. We represent the result pictorially by

We first differentiate its

of

degree Fig. 5.2.1, where each dot indicates one differentiation with respect to p. Similarly, differentiating Fig. 5.1.2(h) three times gives Fig. 5.2.2. Now Fig. 5.2.2 cancels the subdivergence in Fig. 5.2.1, and there is no overall divergence, so their sum is finite. Thus the third derivative of the sum of Figs. 5.1.2(a) and (b) is finite. So the overall counterterm is quadratic in p. Lorentz in variance forces it to be of the form A(d)p2 + B(d). We glibly asserted that Fig. 5.2.1 plus Fig. 5.2.2 is finite. This statement is not as obvious as it seems. Let us prove it. We Wick-rotate the integrations over k and /in Fig. 5.2.1, and consider regions of the integral that might give a UV divergence. If k and / go to infinity at the same rate, then there is no

<2>



of differentiating Fig. 5.2.1. Result Fig. 5.1.2(a) three times with respect to

Fig. 5.2.2. Result of differentiating Fig. 5.1.2(6) three times with respect to

its external momentum.

its external momentum.

98

Renormalization

divergence, because the degree of divergence is negative. If / goes to infinity with k fixed, there is a divergence, but it is cancelled by the counterterm graph, Fig. 5.2.2. The remaining significant possibility is that both k and / go to infinity, but that k is much less than /. The ratios of different components of either one of k

or



are

finite,

so

we

contribution from this

may summarize the order of

region

of the

magnitude

as:

finite

AIL

dfcfe"

(5.2.13)

divergent contribution, if / is of order k3!2. We must add Fig. 5.2.2, which, as we will show, cancels this new divergence. Observe that the counterterm was arranged to cancel the divergence when / goes to infinity with k fixed, rather than when k is large, as we now have. Let us expand the inner loop of Fig. 5.2.1 in powers of k when / > k, up to its degree of divergence, which is quadratic. The coefficients of these powers are integrals over all /, restricted to / P- k. The divergences in the coefficients are cancelled by Fig. 5.2.2, and we have the following estimates of the sizes This gives



of the coefficients: coefficient of k°

finite<



finite<



0{k\



All

i;

coefficient of /c1



All

The

divergence P*



divergence

> +

divergence



Al/l



0(ln(/c)).

of Figs. 5.2.1 and 5.2.2 in the region k-* oo, with / than k, is then of order

sum

bigger

finite

All

0(k\

coefficient of k2= finite < =



dkk ̄2\n{k).

(5.2.14)

possibly

much

(5.2.15)

The power of k is the same as is given by the overall degree of divergence, but there is an extra logarithm. We get a finite result, as claimed. The higher-

than-quadratic terms in the expansion of the loop in powers of k give no divergence at all. What has happened? The divergence for l>k>\ could only occur

5.2

Two-loop self-energy

in

((/>3)6

99

because the interior loop was itself divergent. The fact that we sent k to infinity merely suppressed this divergence somewhat. Suppose we neglect k in the integral for the inside loop. Then the counterterm is in effect the negative of the integral over / of the loop from finite / to infinity. But in the region we are considering for Fig. 5.2.1, we are restricting / to be much bigger than fc, which is itself getting large. So if we neglect k in the loop, then we are

left with finite +

d/ (integrand of loop with k neglected).

(5.2.16)

Furthermore, we expand the loop in powers of k, to uncover the sub-leading divergences. Each extra explicit power of k in the expansion compensates for the lowering of the divergence. The quadratic term multiplies jd///, so

giving an extra logarithm (but not a power). The key step in the proof is to perform the integral with momentum / first. We have shown

whether

or not



divergence

the

larger

that, for the purpose of determining

occurs, we need

only consider

as

distinct

regions: (1) fc, / -? oo at the same rate, and (2) / -> oc with k finite. (We might also try k -? oo with / finite, but the subgraph with the lines carrying the loop momentum k has negative degree of divergence, so we get no divergence from that region.) The region fc, /-? oo, with k ^ /, is schizoid: it can be considered as essentially part of either of the two regions (1) and (2) that we

have just defined. As region (2), the divergence is cancelled by a counterterm when /-? oo, with k large but fixed. As region (1), the final integral over k is finite, and the only sign of this intermediate

case

is the extra

logarithm in the

integrand. 5.2.3 Fig. 5.1.3

by considering the example of Fig. 5.1.3. kid —6 the graph (a) quadratic divergence. It also has a logarithmic subdivergence when either of the loop momenta k or / gets large. The subdivergences are cancelled by vertex corrections, which are shown in graphs (b) and (c). We must prove that the overall counterterm (d) is quadratic in p. Conventionally, this graph is regarded as a difficulty in the theory of renormalization, for it contains an overlapping divergence. That is to say, one of the lines is common to both subdivergences. This is seen as a problem (Bjorken & Dr^ll (1966)) if one tries to write the graph as an insertion of a We conclude this section has

an

overall

renormalized vertex, Fig. 5.2.3, in the one-loop self-energy. The graph (b)

100

Renormalization p+



Fig. 5.2.3. Illustrating the problem of overlapping divergences, as

seen

in Fig. 5.1.3.

?■ ra>-oo: -

<i>-cixi> (*)

scek-jcd-

Fig. 5.2.4. Result of differentiating Fig. 5.1.3 three times with respect

to its external

momentum.

for the counterterm to one of the

subdivergences is not of this happen in our first example,

form. The

Fig. 5.1.2. of trick to with three times However, respect p works differentiating as well for Fig. 5.1.3 as it did for Fig. 5.1.2. For the sum of (a), (b\ and (c), we find Fig. 5.2.4. The point is that differentiating either of the subgraphs

corresponding difficulty

does not

our

makes it convergent, while the counterterms for the subdivergences are independent of momenta. We get terms (a) and (b\ which have re- normalized subgraphs, and graphs (c) and (d), which have no sub- divergences at all. None of the graphs has calculation of the overall counterterms is left

an

overall

as an

divergence.

The

exercise for the reader.

The correct result is (Macfarlane & Woo (1974)):

'?*-(£ 3*m if

we use





6(d-6Y

3(d-6)J*

-1,64;? '-(i&H"^-^}

minimal subtraction.

,"")

5.3 Renormalization

5.3 Renormalization of

We

have

seen

that, in order

101

of Feynman graphs Feynman graphs

to construct a sensible

(i.e., local)

counterterm

Feynman graph, one must first subtract ofifits subdivergences. This is natural since the subtractions for subdivergences are automatically generated from having the counterterms be definite pieces of the interaction for a (1 PI)

Lagrangian. Without subtraction of the subdivergences, the divergence of a graph need not be local. It may even have a power of momentum greater than the degree of divergence; an obvious case of this is a graph that is finite according to naive power-counting but that has a subdivergence. It is therefore useful to devise a procedure for starting with a basic Feynman graph G, constructing a set of counterterm graphs, and thereby obtaining a finite renormalized value R(G): R(G)=U(G) + S(G).

(5.3.1)

Here (7(G) is the 'unrenormalized' value of the basic graph (which diverges as the UV cut-off is removed), and S(G) is the subtraction the sum of the -

counterterm

graphs.

we use to construct S(G) is very general. It applies to the behavior of any integral as one or more parameters approach a asymptotic

The strategy

value. In field theory it can be applied not only to the renormalization problem but also to the calculation of the asymptotic

limiting

behavior of a Green's function

as some

but not all of its external momenta

get large. (A standard example which we will treat in Chapter 10 is the operator product expansion of Wilson (1969)). The procedure that we use for renormalization was first developed by

Bogoliubov and Parasiuk (see Bogoliubov & Shirkov (1980)), with corrections due to Hepp (1966). Their construction was recursive and has the acronym BPH. Zimmermann (1969) showed how to solve the recursion the result being called the forest formula. All these authors used -

zero-momentum subtractions. Zimmermann (1970, 1973a) showed moreover that there is then no need to use an explicit UV cut-off. He applied

the algorithm for computing R(G) directly to the integrand rather than to the integral; this formulation has the title BPHZ. It is not necessary to use subtractions. For example Speer (1974), Collins (1975b), Breitenlohner & Maison (1977a, b, c) showed how to make the same ideas work using minimal subtraction.

zero-momentum

Our treatment will aim at showing the underlying simplicity of the methods and their power to demystify renormalization theory. We will see that the methods do not depend on use of a particular renormalization

prescription,

even

though

we

will often

use

minimal subtraction.

102

Renormalization

a graph G. (A graph define by specifying its set of vertices and lines, each line joining two vertices and each vertex attached to at least one line.) We write the graph's

We will examine the structure of the subtractions for

we

unrenormalized

value

?^g(Pi

as

Pn)=

\d%...ddkLIG(Pl,...,pN;ku...,kL).

(5.3.2)

Here we let L be the number of loops and N be the number of vertices. The

external momenta at the vertices function there is

fields, but at

an

an

are p.. In a

Feynman graph for



Green's

external momentum at the vertices for the external

interaction vertex,

we

5.3.1 One-particle-irreducible

have p.

0.



graph with

no

subdivergences

The simplest case is a one-particle-irreducible (1PI) graph with no subdivergences. Then the only possible divergence is an overall divergence where the momenta renormalize the

all the lines get large simultaneously. We may by subtracting an overall counterterm:

on

graph

R(G)

(5.3.3)

U(G)-ToU(G).



some operation that extracts the divergence of U(G). It whatever renormalization prescription that we choose to use. implements For example, we might use minimal subtraction. In that case T takes the Laurent expansion of (7(G) about d d0, and picks out the pole terms. (We

Here T denotes



let the physical space-time dimension be d0; i.e., d0

d0



6 for the

c/>3

model

we

used in the

of two notations for the action of T or

T(G). Both will We

could

mean

the



4 for the real world,

previous sections.)

on an

We will

use

or

either

unrenormalized object: 7°[/(G)

same.

use zero-momentum

subtractions. In that

case T

picks

out the

5(G) in the Taylor expansion of (7(G) about zero Here 3(G) is, as usual, the degree of divergence. There are many

terms up to order momentum.

possibilities. In our work, we will use the minimal subtraction scheme. Then, for example, the one-loop self energy in (c/>3)6 gives

other

To

^ 2



,,6-d

ddk

(27i)d (k2-m2)[(p + k)2-m2]

pole part of dx

\^-i--T(2 (1287T

m2 —p2x(l

4nfi2

g2 (p2/6-m2) 64tt3 (d 6) -







x)

dfl)



ld/2- 3

[m2-p2x(l-x)] (5.3.4)

5.3 Renormalization The

one-loop vertex, Fig. 3.6.1, gives dk

fi9-^2

■MS

)' (k2

m2) [{p + k)2



pole part of



dx o

-W"2

\ —^f-3

( Jo

dy

m2

Observe that in this last

^3-d/2





m2] [(p + 9 + fc)2

T(3

q2xy





d/2)



m2]



(p2x + (p + q)2y)(l -x-y) 4nfi2

ld/2-3

gV""



part

103

of Feynman graphs

Y^jg js

an

exampie

of (7(G) to have the

0f



same

define the

pole to come with a factor rule that one must define the pole general dimension as (7(G), for all d.

case we

5.3.2 General

case

general we not only have to handle the case of an overall divergence, but also the case that subdivergences are nested within the overall divergences.

In

is exemplified by the propagator with two self-energy (Fig. 5.1.1), where within one graph there are two subgraphs which can diverge independently. As we saw from examples, we must subtract off subdivergences before finding the overall divergence. In view of the complications when the subdivergences themselves have subdivergences, etc. {ad nauseam), we must Another

case

insertions

be extremely precise is

helpful

to have a

mathematics.

as to

what is to be done. This is what

we

will

now

do. It

specific non-trivial example in mind, to make sense of the

Such

examples

are

treated in

subsection 5.3.3

and in

Section 5.4. The reader should try to read these sections concurrently with the general treatment in this section. First, let us define a specific divergence as being the divergence occurring

when

the

loop momenta on

a certain set of lines get

on other lines and the external momenta

given

set of lines has a

divergence

staying

big, with the

momenta

finite. Whether

or not a

associated with it is determined by power-

counting. A divergence is thus associated with a certain subgraph. (At this stage, we do not require that the subgraph be connected.)

diverges when all its lines get large loop momenta, it is said to divergence. A one-particle-reducible graph (like any of have an overall divergence some lines are not a part of cannot Fig. 3.2.1)

If a graph



have an overall



any loop. All other divergences involve a smaller subset of the lines. They, of course, overall

are

called subdivergences. Every subdivergence of

divergence

of

one

of its

subgraphs.



graph

is the

104

Renormalization

Fig. 5.1.1(a)

Observe that

subdivergences.

These

has

no

overall divergence, but has three

from the regions in the integration

over loop the left-hand has (1) loop large momentum, (2) the right- hand loop has large momentum, and (3) both loops have large momenta. To renormalize a graph G we assume that we know how to renormalize

come

momenta where:

its

subdivergences,

and

we

then let R(G) be the unrenormalized value of G

with subtractions made to cancel the subdivergences. Then the only remaining divergence that is possible is an overall divergence. So we define an

overall counterterm: (5.3.6)

C(G)= -ToR(G)

to R{G) the same subtraction operator T as we discussed if there is no overall divergence (e.g., if G is one-particle-reducible) earlier; then C(G) is zero. In any event the renormalized value of G is defined as

by applying

R(G)



R(G)



C(G).

(5.3.7)

The definitions (5.3.6) and (5.3.7) give us R(G) provided that we know how to subtract

subdivergences.

This is essentially

a matter

of renormalizing

smaller graphs; we will construct R(G) in a moment. Once this, we will have a recursive definition of R(G): successive

we

have done

application of us to smaller and to smaller (5.3.7) graphs with subgraphs ultimately brings no subdivergences. These we know how to renormalize. Now let us define R(G\ which is to be U(G) with subdivergences subtracted. For the case of a graph with no subdivergences we must define R(G) For



larger graph



we

U{G)

(if G has

no

subdivergences).

(5.3.8a)

define R{G)



U(G)+

(5.3.8b)

X Cy(G).

We sum over all subgraphs y of G, other than G itself, as indicated by the notation y £ G. The other new notation Cy(G) means that we replace the subgraph y by its overall counterterm, as defined by (5.3.6), i.e., f ^

To make only

over

could of



[0

ToR(y),

if

y has an overall

if y has

no

divergence

overall divergence

(5.3.9)

simple formula, we write the sum as being over all y's rather than divergent y's; then the Cy(G) is zero if y is not overall divergent. We course restrict the sum only to those subgraphs that have an



overall divergence. One tricky point in the above equations arises in defining C(y) for a disconnected subgraph y. An example is the subgraph of Fig. 5.1.1(a) consisting of the two self-energy loops. We will discuss this next.

105

of Feynman graphs

5.3 Renormalization

Application of general formulae

5.3.3

Equations (5.3.6) to (5.3.9) give a definition of R(G). Let us see how they apply to simple examples. For a 1PI graph with no subdi vergences they just reproduce R(G) (7(G) To U(G). Next consider a graph like Fig. 3.2.1(a) or (c), whose only divergence is a subgraph with no further subdivergences. Then there is no overall divergence, so by (5.3.7) =



R(G) There is only

one

subdivergence,



(5.3.8) collapses

so

R(G)=U(G) where y is the

(5.3.10)

R(G).

divergent subgraph.



to

give

(5.3.11)

Cy{G)9

Here

is the full

Cy(G)

graph

with y

To U(y). We reproduce the obvious result. There is one graph like Fig. 3.2.2(a) or (c). We now look at a graph with two or more divergent subgraphs which do not intersect and which have no subdivergences. It is sufficient to consider G to be Fig. 5.1.1(a). There is no overall divergence, so again R(G) R(G). Let y1 and y2 be the self-energy bubbles. Then the subdivergences correspond to the three subgraphs yl9 y2, and yl^jy2. (Here yx uy2 means, as usual, the union of yx and y2.) So

replaced

by



counterterm



R(G)



R(G)

Evidently Cyi(G) —



U(G)



C7l(G) + CJG) + CluT2(G).

(5.3.12)

is just

U(G) with yx replaced by its counterterm, To [/(yj; and similarly for y2. But what is C, (G)? It corresponds to a subtraction for y} \jy2 for the region where all loop u.,

momenta are

large.

regions where only

C(y^y2)= Here

we

But

we must

subtract from it the counterterms for the

one momentum



is large:

To[U{y1)U{y2) + C(yl)U(y2)+ C/(yi)C(y2)]. (5.3.13)

used the fact that yx\jy2 is disconnected,

U(y1^y2) To work out (5.3.13),

we must

to act independently

on

To[U(yi)U(y2)] To[C(yi)U(y2)]



so

that

(5.3.14)

U(yl)U(y2).

define T when

acting on a disconnected graph

its components. Thus: =



[Tot/(yi)][Tol/(y2)] [ToC(yi)][To t/(y2)]





C{y,)C{y2l -

C{y,)C(y2\

(5.3.15) (5.3.16)

etc. We used the property that To

To

U(yt), i.e.,

the

pole part

C(y ■). Furthermore, To [To U(y ?)] U{y() a pole part is itself. We therefore find

of







106

Renormalization

that

(5.3.17)

C(y1uyJ=[ToU(y1)-]lT°V(y2)l so

that

we

reproduce the

counterterm

The above procedure generalizes to

excessively complicated, but

graph Fig. 5.1.1(d). arbitrary graph.

It may appear it allows the smoothest way of defining R(G). an

we observe that our definitions (5.3.6)—(5.3.10) exactly reproduce results for the two-loop graphs like Figs. 5.1.2 and 5.1.3.

Finally, our

5.3.4 Summary In this section

we

have proved very little. We have set up



series of

definitions that state exactly what we mean by the renormalization of a Feynman graph. The notation we have introduced will be important in

making proofs. What we will need to prove is that the overall counterterms are local and of a degree in momentum given by naive power-counting. We will also show how to solve the recursion to find an explicit formula due to Zimmermann (1969). 5.4 Three-loop example

three-loop self energy graph of Fig. 5.4.1 in 03 theory in six dimensions is an example of a graph with nested and multiply overlapping divergences. We call it G. Its divergent subgraphs are: The

yx



y2



y3



y4



ys



{lines {lines {lines {lines

carrying carrying carrying carrying

loop loop loop loop

momentum momentum

momenta k momenta q

fe}, q}, and/or /}, and/or /}, (5A1)

yivy2>

The first four of these are connected IPI vertex

two unconnected vertex

graphs;

the last is

a set

graphs. According to our definitions of Section

of

5.3

we have

R(G)





R(G)



C(G)

R(G)-Tc[R(G)l

Fig. 5.4.1. Three-loop self-energy graph in 03 theory.

(5.4.2)

107

Three-loop example

5.4

This

equation states that we first subtract subdivergences and then take off the overall divergence. To define

R(G)

subtract

we

R(G)



this

equation

G+

i Cyi(G)

subgraphs

(5.4.3)

?Cfo)

5.4.2. The notation with the vertical bar in

we take G and replace y. by the corresponding figure, the labels 1,...,4 signify which of the been replaced by its counterterm.

denotes that

C(yt).

counterterm

Fig.

as

R(G\

subdivergences:

-G+IG| We represent this

to obtain

In the

y!,..., y4 has

R(G)



+ '4



Fig. 5.4.2. Subtraction of subdivergences of Fig. 5.4.1. Only yl and y2 have

C(y J and C(y2) are the But we have still to define C(yt) for i 3,4,

further subdivergences,

no

ordinary one-loop counterterms.

so



5: C(y3)=

C(y4)

c(y5)







-^bs-ysly^ny^l To[y4 y4|y2_T(V2)], -



\yxy2 T{yi)y2 [-Tf(yi)][-r(ya)]. -To





yt

T(y2) ̄\ (5.4.4)

by combining (5.4.2)-(5.4.4). If we represent 1 PI graph by enclosing it in a box, we can write

The overall result is obtained

the effect of applying T to



as shown in Fig. 5.4.3. There are sixteen terms in all. The first eight represent (7(G) minus its subdivergences, and the last eight form the subtraction for the overall divergence. The expansion of R(G) represented in

R(G)

Fig. 5.4.3 is

an

example of

the forest formula, to be discussed in the next

section. As

we saw

by examining two-loop graphs,

the overall counterterm for



first subtract off subdivergences. Otherwise graph there would be divergent contributions from where some but not all is non-local unless

we

108

Renormalization

R(G)-

<n> (?)

(b)

_M-

(d)

(e)

(f)

(9)

(A)

!L

| Previous 1 18 graphsj Fig.

5.4.3. Renormalization of Fig. 5.4.1.

subgraphs have large loop momenta. There would also be divergent contributions from the region where all loop momenta get large but at Fig. 5.4.3 explicitly.

different rates. We can check from

for R{G). Let

occur

Let

us

us

do this

that

none

of these

problems

show that the overall counterterm for G is local. We differentiate

R(G) three times with respect to the external momentum p and show that the result is finite. Given the momentum routing of Fig. 5.4.1, there are three lines to differentiate: p + fc, p + /, p + q. Here we have used the momentum by the line as a label for the line. Differentiating the original graph

carried

gives

ten terms, where the three derivatives are

applied

to any combination

of the three lines. One term is where we differentiate p + k three times

5.4.4). Although

(Fig.

overall divergence, there remain sub- examine the corresponding differentiations applied

there is then

no

divergences, so we must to the counterterm graphs (b)-(h). We must regard the derivatives as acting on these graphs before divergences are computed (by the operation

symbolized by

the box).

subgraphs yx and y3 completely finite, by divergence and y3's only subdivergence. Hence

The differentiation makes the

removing

both their overall

Fig. 5.4.4. One of the terms obtained by differentiating Fig. 5.4.1 three times with respect to its external momentum.

5.5 Forest formula

109

the counterterm

graphs (b\ [d\ (e\ and (h) are zero after differentiation. This graphs (c), (/), and (g). These cancel the subdivergences of (a) coming the subgraphs y2 and y4, which are unaffected by the differentiation.

leaves from

We may examine the other nine terms in d3R(G)/dp3 find that in fact d3R(G)/dp3 is finite, as claimed.

similarly,

and

we

5.5 Forest formula 5.5.1 Formula Zimmermann (1969, 1970) gave an explicit solution of the recursive definition of the renormalized value R(G) of a graph G. The general idea can be gleaned from the example we examined in the previous section. There the recursion generated a series of sixteen terms. One was the original graph, and the others had the subtraction operation T applied one or more times.

example, in graph (e) we first replace the left-most loop y x by T(y J, with a result we can write as Tyi(G). We then take the subgraph equivalent to y3, it viz. and replace by the result of acting with T. This gives graph (e). Tyi(y3\ The sum of the two graphs (d) and (e) is used as the subtraction for the subdivergences of G associated with y3. Each of the sixteen terms is pictured as the original graph with some number of connected IPI subgraphs enclosed in boxes to indicate application of T to the subgraph. Each term can be specified by giving its set of boxed subgraphs. Each such set is called a for est. The subgraphs which form a particular forest are either disjoint or nested: they are said to be non- For

overlapping. There set

are

The set of all

called #"(G). The first

eight

are

(in

theory notation):

(a) the empty (c) (d) (?) if) to) (h)

possible forests for G is occurring in Fig. 5.4.3.

sixteen forests

set

0,

{y2}, {y3}, {Vi.Vs}. (U {^.yJ. {y?ya}.

These do

not contain the whole

graph; they

are

called the normal forests.

The other eight forests in Fig. 5.4.3 consist of one of the above eight forests to which is added as another element the complete graph G. A forest of G

110

Renormalization

containing

G is called

a.

full forest. The distinction between

normal and full

forests is that the normal forests subtract off the subdivergences, and the full forests combine to subtract the overall divergence. Not all forests U

occur

in Fig. 5.4.3; for

{subgraph consisting



example, the

of lines carrying loop momentum

does not appear. Such forests contain at least

subgraph

as an

forest

one

/}

overall convergent

element.

Inspection of Fig. 5.4.3 shows that



*(G)=

(5.5.1)

ri(-r,)G.

C7e.^(G) yeU Here

the

sum

is

over

all forests U of G. The operator

Note that for nested

Ty replaces y by T(y).

y's the 7,,'s should be applied inside

to outside.

Equation (5.5.1) is called the forest formula; it is due to Zimmermann (1969). Suppose we compute R{G) for an arbitrary graph G by using (5.5.1). Then,

will prove shortly, the result is the recursive definition of R(G) given in Section 5.3. as we

It is convenient to let the

sum over

forests be

over

if

we

used the

all forests rather than

superficially divergent; the reason for doing this is that we

of subgraphs that

those

only consisting extra forests give a zero contribution. The over

same as

are

will sometimes wish to change the definition of T so that we make subtractions for some convergent graphs, as well as for divergent graphs.

example, such

redefinition will be the key to proving the operator product expansion in Chapter 10. We now have both a recursive and a non-recursive definition for the For



renormalization of a Feynman graph. It will prove very useful to have both proofs will need different forms of the

definitions available. Different

definition. In particular, proofs

graph

will naturally

use

by induction

on

the number of loops of



the recursive definition.

5.5.2 Proof The proof of the forest formula is elementary, but somewhat involved. We first use (5.5.1), and the following equations:

C7e^(G) yeV

C-TGoR(G\ C(G) =< n[c(ft)]> (0



*)

if G is 1 PI,

if G is



disjoint union of

otherwise,

1PI

yt%\ (5.5.3)



5.5 Forest formula

111

R(G), and C(G). Here #(G) is the set of normal forests (i.e., those that do not contain G). These definitions are correct for the graph Fig. 5.4.1, as can be seen by inspection of Fig. 5.4.3. Since the recursive definitions uniquely give R(G\ R(G\ and C(G) in as definitions of R(G\ of G

operation Ty, it suffices to show that (5.5.1)—(5.5.3) satisfy the (5.3.6)-(5.3.9). Notice first that (5.3.6) and (5.3.9) are the same, really except for being applied to different graphs. If R{G) given by (5.5.2) is correct, and if subgraphs are correctly renormalized, then (5.5.3) is equivalent to our original definition (5.3.6) of terms of the

recursion relations

C(G). Moreover, suppose that G is connected and one-particle-irreducible. Now each forest of G is either a normal forest, that is, a forest of which G is not

element,

an

or

it is



adjoined G. Then the of (5.5.1)—(5.5.3) for such a

normal forest to which is

(5.3.7) for R(G) is a direct consequence graph. If G is not a union of 1PI graphs, then there is no overall divergence,

formula

and again (5.3.7) holds. So it remains to prove the

following:

(1) R(G) is correct when G is a disjoint union of more than one 1PI graph. (Note that this case occurs in renormalizing the graph of Fig. 3.2.1(h), as in Section 5.3.)

we saw

(2) R(G) is correct, i.e., it satisfies (5.3.8), for

If

G is

forests,

disjoint



one

general graph.



union of 1PI graphs y., then each forest is

for each component. Then R(G)



[ ̄] R(y.),

as we



union of

should expect.



The

problem

where

we

is that this is not

make

an

manifestly

definition,

overall subtraction for G. We dealt with this problem

between (5.3.11) and (5.3.17). Our proof of (5.3.8) is by induction

graph has

true in the recursive

on

the size of G. Now



one-loop 1PI

non-trivial subgraphs, so its only normal forest is the empty set. Then formula (5.5.2) collapses to R(G) U(G\ just as it should. This no



enables

us to start

the induction.

(5.3.Sb). For this, observe that each forest U has a biggest subgraphs Ml9...,Mj. Each M, is contained in no unique bigger subgraph in U, and each y e U is contained in some Mt. The existence and uniqueness of this set of M's is seen by considering pairs y., yk of It remains to prove set of

elements of U.

Since y. and yk are

non-overlapping,

there

possibilities: (1)

y.

(2) yk

(3)



yk, in which



y., in which case remove yk from further consideration.

y.nyk



case remove y.

0, in which

case

from further consideration.

leave both in.

are

three

112

Renormalization

Repeat until

no

further eliminations

possible; then the result is the set of

are

Mf's. The forest U is the union of a full forest,

definition (5.5.2) of R(G)

R(G)





Here the first term





??

our



(-Tn)-U(-Tyj]Gl

n yi

comes

for each Mt. We can write

\f\(-TM)



G+



one

as

from the case in (5.5.2) that U



(5.5.4) 0, and the sum in

the second term is over non-empty sets of disjoint 1PI graphs M{ excepting G. By setting y the case that Mt M1u...uMn and using (5.5.3) to =

determine

C(M{



u... u

Mn\

we

find (5.3.%b).

5.6 Relation to if We have

how to renormalize

making doing this came from consideration of examples in which the subtractions were generated by counterterms in the interaction Lagrangian. We will now show that this is a

seen

series of subtractions.

The

an

individual Feynman graph by

motivation for

true to all orders. We will

assume the natural result (to be proved later) that the polynomial degree of the overall counterterm of a graph is given by its

degree of divergence, just as for low-order graphs. First, we must make precise the result that we will prove. For each graph G, we have constructed its overall counterterm C(G). Since this polynomial

in the external momenta of G, it

derived from

N(G) is



an

can

be written

as

1PI is



the vertex

interaction term D(G)/N(G) in the Lagrangian j£f. Here same sort as the 3! that appears with the 03

symmetry factor of the

interaction term in if. Each power of

corresponds point graph and

entering D(G) corresponding field. If G is an n- lines corresponds to the same type of a

momentum

to i times a derivative of the

each of its external

field, then N(G) is n\. If there number of lines of type i

are a

entering

number of different fields and

n.

is the

G then

iV(G)=n^!-

(5-61)



graph for a Green's function, the forest formula gives a set of counterterms. We will demonstrate that the set of counterterm with graphs is generated from the counterterm vertices in the interaction graphs For each

Lagrangian. Consider '(f)3 theory. if



As before, jsf 0



seh

we

write the

+ <? ct



se 0 +

Lagrangian <ev

as:

(5.6.2)

113

5.6 Relation to ¥£ The free

Lagrangian SeQ



(d(j))2/2

(5.6.3)

m2(j)2/2



generates the propagator, while the interaction !£x consists of two terms, ifb and ifct. The basic interaction is

(5.6.4)

ifb=-^3 ̄d/2(/>3/3!, and

ifct is the divergences:

counterterm

Lagrangian

used to cancel the ultra-violet

(5.6.5)

<?ct=ZD(G)/N(G). G

sum

is

all 1PI

graphs. Those that have no overall divergence for these D(G) 0. Each 1PI graph that has an counterterm; generate overall divergence generates a term in (5.6.5). The formulae (5.6.2) and Here the

over

no



(5.6.5) apply in any theory. Since (5.6.5) applies to any than

six

dimensions.

theory, it applies in particular to 03 in higher

Thus

it

renormalizable theory. But the

enables

sum must

us

renormalize

to



non-

include counterterms D(G) with

arbitrarily large number of powers of momentum and with an arbitrarily large number of external lines for G. It is only in six or fewer dimensions that

an

the counterterms have the if 0



same

form

as terms

in the basic

03 Lagrangian

£ch.

Now that

part of the

we

have

proof is to

developed

We will prove that the



convenient notation, the most difficult

that the combinatorial factors

come out right. defined and by (5.6.2) (5.6.5) gives the Lagrangian

ensure

same renormalized Green's functions as those generated by our recursive definition in Section 5.3 (and therefore the identically same Green's functions as given in Section 5.5 by the forest formula). The proof will be

given for

03 theory

in six

or

fewer dimensions, but it easily generalizes.

Consider the full AT-point Green's function GN at order gp. It is sufficient only with connected graphs. If the theory is renormalizable (as we

to work

will prove in Section 5.7), then the Seci



SZ(d(t>)2/2



sum

of counterterms has the form:

dm2(j)2/2



Sg(f>3/3!.

(5.6.6)

with (by (5.6.5)) -3Z=



[Coefficient



[Coefficient of ip° in C{G)\



C(G)/i.

2-point G

-dm2=

2-point G

-5g= We

ignore

the

3-point O*

of

tadpoles, yet again. The term



ip2

in

C{G)\

(5.6.7) of order

gp in the perturbation

114

Renormalization

expansion of GN has vertices generated by the different terms in if b -I- if ct. There will be graphs with all of their vertices being the basic interaction if b. Let the set of these be called B. The other graphs will contain one or more of the counterterm vertices generated by (5.6.6). Each counterterm

decomposed into

a sum

of terms

by applying (5.6.7)

vertex. Each term has each of the counterterm vertices

divergence of some graph. corresponds to a unique basic graph overall

So

we



replaced by

Then in the result, each term

the



b(T)eB.

Z(g+ G

On the other hand, G

then be

have

G?

graphs

can

at each counterterm

we





b(T)







(5.6.8)



have constructed the renormalization of each of the

by writing R(G)

Each of the terms



£ Cy(G).

G+

(5.6.9)

(5.6.8) is constructed by replacing each of a set of one or more disjoint 1PI subgraphs yl9...9yj by its counterterm given by On in (5.6.9) with identifying y iDy{G). ylvy2"'vyp we expect that Tin

G?=EK(G).

(5.6.10)



This result would be obvious,

were it not that the symmetry factors do not match manifestly up. The problem is illustrated by Fig. 5.6.1. There the basic graph is (a\ and the complete set of subtractions needed to renormalize it consists of (b)-(e).

Now the symmetry factor for (a) is 1/8: There is a factor 1/2 for each self- energy graph and an overall 1/2 for the top-bottom symmetry of the whole

graph.

Each of the subtractions

(b)

and

the remaining factor 1/2 goes into the

graphs (b) and (c)

are

(c) has



symmetry factor 1/4, since

counterterm for the

self-energy.

equal.

(a)

(6)

o-o



M)

(c)

Fig. 5.6.1. Renormalization of

-*



graph,

to illustrate symmetry factors.

Both

5.6 Relation to<£

115

Considered as Feynman graphs, these are the same graph, but with symmetry factor 1/2. So they give one term with factor 1/2 in (5.6.8) (derived from the Lagrangian), while in (5.6.9) (from the recursion formula) there are two equal terms with factor 1/4. The end result is the same. We must

consider

(b)

(c)

and

distinct graphs when defining R(G\ since they we must take

as

correspond to different regions of loop-momentum space each momentum variable to be distinguishable.

general proof is tedious. The symmetry factor of a graph G 1//V(G), where N(G) is the dimension of the graph's symmetry group. So To construct

is



we



write G

where the overbar indicates

Similarly

we

define C(G)



computation ignoring all symmetry factors.

by C(G)





G+

we

graph

G is

Y Cy(G)

N(G) In the last line

(5.6.12)

C(G)/N(G).

Now the renormalized value of a

R(G)

(5.6.11)

G/N(G\

N(G)

n*.

C,(G)

a disjoint union of 1PI graphs y,, explicitly indicated the symmetry factors 1/Nt yt which is replaced by its overall counterterm. For a u y( the symmetry groups of the y;'s are a commuting set

have observed that y is

y2,.... Moreover, we have

for each

(5.6.13)

l/N(yt) given subgraph y of subgroups of





JV(G)/J ̄[ AT,

the symmetry group of G. Therefore the quantity

must be an

integer.



Next, consider the Green's functions generated by the Lagrangian (5.6.2), as

in (5.6.8),

GN=Y<-^— G+Y'—-—C (G) 1.

(5.6.14)

graph containing one or more counter- terms is generated from a basic graph by replacing some 1PI subgraphs y l,..., yi by counterterms. We write y as the union of the yf's. Then we let G/y be the graph resulting from substituting counterterms for the yf's. By the Here

we

have observed that each

116

Renormalization

Lagrangian, the result

definition of the counterterm

is the

same as

Cy(G),

aside from symmetry factors. The prime on the £' indicates that only y's giving distinct Feynman graphs are considered. (Thus, for example,

Figs. 5.6.1(h)

and

K(G, y) then

(c) =

are not

[number

we must prove

counted

separately.) Thus, if

of graphs / for which G/y



we

define

G//],

that X(G,y)







[m

(5.6.15)



[N(G/y)l

language of group theory. The of the right-hand side of (5.6.15) is the product of dimensions denominator of commuting subgroups of the symmetry group of G. (Note that, for It is easiest to couch this final step in the

example, two (ft3 counterterms generated by different self-energy subgraphs are counted as different.) These subgraphs generate another subgroup, of which the set of cosets in the symmetry group of G has exactly the dimension of the right-hand side of (5.6.15). But, concretely, each coset corresponds to one of the graphs counted by K(G, y).

5.7

Renormalizability

5.7.1 Renormalizability and non-renormalizability

explain the properties of renormalizability, non- renormalizability, and super-renormalizability of a field theory. We do this In this section

we

first for every order of perturbation theory, and then we consider to what extent the properties are true beyond perturbation theory, for the complete

theory. The method in perturbation theory is power-counting dimensional

analysis.

Consider first

and

(ft3 theory in a space-time of integer dimension d0. We have

how to renormalize it to get finite Green's functions by adding counterterms (5.6.5) to the Lagrangian. Each counterterm is a polynomial seen

in the field

(ft

and its derivatives. The theory is called renormalizable if the are proportional to the terms (d(ft)2, (ft2, and (ft3

only counterterms needed

present in the original Lagrangian if 0 that the

Lagrangian

ifb. This is equivalent

to

saying

has the form

if



(d(ft0)2/2



m2<ft20/2



0o</>o/3!,

(5.7.1)

(ft0 is Z1/20. The bare mass m0, the bare coupling g0, and Z each have singular behavior as the renormalization field-strength

where the bare field the

-f

5.7

Renormalizability

117

regulator is removed. A linear term Sh(f) is needed as well. We as being present in the original Lagrangian. In any event it is a only single extra coupling. It can be ignored if we impose the renormalization condition that <0|(/>|0> =0 to determine 8h, and if we ultra-violet

may regard it

ignore tadpole graphs. We generalize to an arbitrary theory by calling a theory renormalizable if the Green's functions of its elementary fields can be made finite by rescaling the fields (in a cut-off dependent way) and by making some suitable cut-off dependent change in the couplings and masses. In perturbation theory we determine whether

or

we

not

have

renormalizability by examining possible values of the degree of IPI for the For divergence 3(G) graphs. every graph G a counterterm is needed if the

3(G) >0. As we will prove in Section 5.8 the counterterm C(G) is polynomial of degree 3(G) in the external momenta, and provided we use





scheme like dimensional regularization that preserves Poincare invariance, the counterterms are Poincare invariant. Let

us

now

determine whether

or

not

<j)3 theory

space-time dimensions momentum (in space)

dimensions is renormalizable. In d

graphs

have dimension

d(GN)



N+

3(GN)

d + (1



the

space-time N-point IPI

d-Nd/2.

Then (by (3.3.12)) the degree of divergence of a graph =

in d

d/2)N



(d/2



for

GN

at order

gp is

(5.7.2)

3)P.

Note that the minimum value of P to have a one-loop connected graph is N. Inspection of (5.7.2) shows that if d > 6 then, for any value of N, there can be made AT-point graphs with arbitrarily high degree of divergence by going to large enough order in g. The theory is therefore not renormalizable if d >

6, and the non-renormalizability is



direct consequence of the negative

dimension of g.

'-?-0—O—CD

'-'-0—£)--<D-' and tadpole graphs to 4 loops

Fig. 5.7.1. All the graphs with overall divergences in <j)3 theory at dimensions where it is super-renormalizable.

those

space-time

118

Renormalization

6, only the one-, two-, and three-point functions are divergent, with degree of divergence 4,2, and 0, respectively. The permissible counterterms are just terms of the form of those in j£?0 + j£?b, so the theory is renormaiizable if d 6. Moreover, there is a divergence in every order of g (except for tree graphs, of course). ltd





3, 4, or 5, then only a finite set of graphs, illustrated in Fig. 5.7.1, have overall divergences, and renormalization is needed only for the mass and for the tadpole coupling. Again we have renormalizability. If d



5.7.2 Cosmological term Strictly

speaking,

we

should also consider Feynman graphs with

no

bubbles. They generate the energy of and the vacuum, density normally are ignored. But in gravitational be cannot physics, they ignored. Counterterms for such graphs (present in external lines. These

are

(j)3 theory whenever d



the

vacuum

2) are proportional to the unit operator. They are a cosmological constant. A counterterm is even needed for free-field theory where the divergence is conventionally removed by normal-ordering (see, for example, Bjorken & Drell (1966)). We see that normal-ordering is nothing but a primitive form

renormalization of what in General Relativity is the -

of renormalization.

5.7.3 Degrees of renormalizability It is convenient to

(1) Finite:

distinguish

no counterterms

three types of renormaiizable theory:

needed at all.

(2) Super-renormalizable: only



finite set of graphs need overall counter-

terms. (3) Strictly renormaiizable: infinitely many graphs need overall counter- terms. (But note that they only renormalize a finite set of terms in the basic Lagrangian, since we assumed renormalizability of the theory.) or super-renormalizability normally occur when all the in the basic Lagrangian have positive dimension. couplings Note that in a super-renormalizable theory, the number of divergent basic graphs is infinite. For example, even if there is only one graph y with an overall divergence, any graph containing y as a subgraph is divergent.

Finiteness

However all such graphs become finite after adding to y its counterterm,

so

one counterterm, C(y), appears in the Lagrangian. Mathematical physicists (see Glimm & Jaffe (1981)) have investigated renormalizability beyond perturbation theory. This is important, since

only

5.7

perturbation series

119

Renormalizability

in general

asymptotic series rather than convergent the perturbation series to obtain the Even has it been proved for many super- so, complete theory. renormalizable theories that perturbation theory gives an exactly correct account of the divergences. (A much investigated case is 04 theory in two and in three space-time dimensions.) In a super-renormalizable theory, the series for a bare mass or coupling in terms of the renormalizable quantities has a finite number of terms. Therefore the series converges, and one only has to prove that (a) perturbation theory is asymptotic to the true theory, and (b) there are no terms like exp ( l/g) in the bare masses or couplings that are smaller than series. Thus

are

one cannot

simply

sum



any power of g. The rigorous proof amounts to showing that in summing the perturbation series to a finite order, the error is correctly estimated by the first term omitted. In particular, this

applies to the existence of any ultra-violet possible divergence. Rigorous proofs are not yet available for any strictly renormalizable

theory. One difficulty is obvious: the series for, say, the bare coupling, g, is an infinite series, each term of which diverges as the UV cut-off is removed.

presumably asymptotic rather than convergent, one obtain directly any information about renormalization in the full error obtained in using a truncated form of the series is of the theory: the Since the series is

cannot

order of the first term omitted, and that is always divergent. It might even appear that perturbation theory has no light at all to shed on the as we

of renormalizability of the full theory. This is in fact not so, when we discuss the renormalization group in Chapter 7. If

question will

see

the theory has the property called asymptotic freedom then a series of suitable redefinitions of g allows short-distance phenomena to be computed

reliably.

In

particular the

UV

divergences

can

be

computed

in terms of

weak coupling series without divergent coefficients. It is sensible to conjecture that a suitably refined analysis can be made to obtain rigorous bounds of the errors so that the perturbative results correctly give the divergences. Monte-Carlo studies of the functional integral (Creutz & Moriarty (1982)) support this conjecture. In four dimensions, only certain

non-abelian gauge theories (including QCD) are asymptotically free (Gross

(1976)). We will also

04

see

in

Chapter

7 that in

non-asymptotically

free theories,

and QED in four dimensions, perturbation theory cannot reliably short-distance phenomena. There are, in fact, indications (Symanzik (1982)) that the full c/>4 theory is not renormalizable, contrary to the situation order-by-order in perturbation theory. like

describe

120

Renormalization 5.1 A

For theories which

are not

(1) There is only



perturbation theory, there are following:

renormalizable in

possibilities. Among

many

Non-renormalizability them

are

the

finite set of 1PI Green's functions which have overall

divergences, A typical

dimensions when



in six

or

fewer

space-time

Lagrangian,

the basic

<*?

03 theory

is

case

(d(j))2/2



m2cj)2/2



(5.7.3)

#3/3!,

linear in <\>. The one-, two-, and three-point functions have divergences, but there is no term h<\> whose coupling can be re- has

no term

normalized to cancel the divergence of the tadpole graphs. However, addition of such a term generates a renormalizable theory. More

generally, suppose

we

have

functions. A renormalizable



overall-divergent Green's generated by adding a finite set of

finite set of

theory

is

extra interactions. (2) There is

an infinite set of Green's functions with overall divergences. However, for all but a finite set of the Green's functions, the divergences cancel after summing over all graphs of a given order. (There are no known cases of this.) case 2, except that the divergences cancel only for the S-matrix, rather than for all off-shell Green's functions. An important case is a

(3) As for

spontaneously broken

gauge

theory, when

it is

quantized

in its

unitary

gauge.

(4) The theory is made renormalizable by going beyond perturbation theory in some systematic and sensible way. One case (as in the Gross- Neveu (1974) model see Gross (1976)) is of a theory that is strictly renormalizable and asymptotically free for some dimension d d0, and -



that is considered in

(5)

As for case

some

dimension d slightly greater than

is the Yang-Mills theory with extra terms destroy

(6)

d0.

1, except that the extra terms make physical nonsense. A case a mass term

in Feynman gauge. Then the

unitarity ('t Hooft (1971a)).

None of the above.

Roughly speaking, there are no general rules.

Each

case must

be handled

separately. Only for the last two cases (5 and 6) should a theory be called non-renormalizable. A fundamental theory should be renormalizable, for

actually infinite or they are finite, infinite set of parameters is needed to specify the finite parts of the

otherwise either

but

an

physical quantities

are

counterterins. Nevertheless,

a statement

that



particular theory is non-renormalizable

5.7

121

Renormalizability

is really a statement of ignorance: nobody has found a way to construct a physically sensible version of the theory. (Cases 1 to 5 are where somebody has found a way.) In practice, when a theory is labelled non-renormalizable, what is usually meant is that the theory is not renormalizable order-by- order in perturbation theory; such a statement can be proved by calculating a

finite number of graphs. Within the usual functional-integral

only has the complete 3, but it has been

04 theory

approach (with a lattice cut-off), not 2 and been proved renormalizable for d =

proved non-renormalizable

5.7.5 Relation of renormalizability

for d

> 4

to dimension

(Aizenman (1981)).

of coupling

perturbative renormalizability of a theory of scalar fields, we the generalize argument leading to (5.7.2). The argument will apply when no has coupling negative dimension. Renormalizability will hold with possibly the addition of extra interactions (like the h<\> term in 03 theory) whose coefficients have non-negative dimension. Our proof will easily generalize to theories with fermion and gauge fields. The problems we will encounter in gauge theories will all be to do with the question of whether these extra terms are compatible with the gauge invariance. But we will leave these questions to Chapter 12. Let a general term in <£ or j£?ct be written schematically as To prove

(coupling /')(derivative)-4 (field)*.

(5.7.4)

The vertex generated by this term is one possible graph for the 1PI Green's TN with N external lines. Thus the dimension of TN satisfies

function

d(TN)



d(f)

(5.7.5)

+ A.

Since coupling has negative dimension, the degree of divergence of any graph for TN is at most d(TN), as we saw from examples in Section 3.3.3, and as we will prove in Section 5.8. That is, the degree of divergence d(TN) satisfies no

(5.7.6)

d(rN)<d(rNi with equality only

graphs all of whose couplings have zero dimension. N-point graphs, we add counterterms of the form S(TN) derivatives. So the possible counterterms satisfy

for

To renormalize the (5.7.4) with at most

d(f) The last



d(TN) -A> 5(TN)

-A>0.

inequality follows since a counterterm with A

(5.7.7)

derivatives is needed

only if the degree of divergence is at least A. From (5.7.7) it follows that

we

122

Renormalization

need

no

couplings of negative dimension, given

original Lagrangian. Some of the couplings generated the

original Lagrangian dimension. But the number

even

because only

of

none are

present in the

be present in couplings of negative

as counterterms may not

if it contains

new

that

couplings

no

needed is nevertheless finite,

finite set of counterterms satisfy (5.7.7).



5.7.6 Non-renormalizable theories of physics From the discussion above, it is natural to conclude that



theory of physics

should be renormalizable. In fact, the strong, electromagnetic, and weak interactions appear to be described by a renormalizable theory. This theory

is



combination of quantum chromodynamics for strong interactions and

the Weinberg-Salam theory for weak interactions. Around 1970 there was a revolution in the theory of weak interactions when it was discovered that non-abelian gauge theories are renormalizable. It is precisely one of these theories that was found to be necessary to

construct

a renormalizable theory of weak interactions in agreement with experiment. See Beg & Sirlin (1982) for a historical review.

Unfortunately, this progress has not extended to gravity. Einstein's theory of general relativity is non-renormalizable, after quantization, and alternative. (This situation exists despite many to significant attempts improve it Hawking & Israel (1979).) It is a mistake to suppose that non-renormalizable theories should be there is

no very

promising



banished from consideration. interactions

were

Remember that for many years weak the four-fermion' theory,

successfully calculated using

which is non-renormalizable. For most purposes, weak interactions could be adequately treated in the lowest order of perturbation theory, where no

renormalization is needed. But the non-renormalizability of higher-order calculations raised the question of consistency of the theory: is it legitimate even an approximation to a nonsensical (i.e., non-existent) the results of calculations mean anything? The same questions Will theory?

to calculate

gravity. There, the classical theory of general relativity successful, but the quantized theory is badly non-renormalizable.

arise in

is very

We must therefore understand how and why we may use non- renormalizable theories in physics. Now, to perform consistent calculations in any theory which contains ultra-violet divergences, we must impose an ultra-violet cut-off, M, of some sort.

In the

case

of



we can take M to infinity and insensitive to the cut-off. Another related

renormalizable theory

obtain finite results that

are

5.7

123

Renormalizability

property of

a renormalizable theory is the decoupling theorem of Appelquist and Carazzone, which we will discuss in Chapter 8. This

theorem

applies

to a renormalizable

theory which contains fields whose

bigger than the energies of the scattering processes under consideration. The theorem states that the heavy fields can be deleted with masses are much

only



small effect (suppressed by

sections,



power of the heavy mass)

etc. The hallmark of a renormalizable

on cross-

theory is in fact that it is

complete in itself. It contains no direct indications of whether it is only part of a larger and more complete theory. These statements are false for a non-renormalizable theory. Consider the old four-fermion theory of weak interactions. Its coupling is G ^10 ̄5 GeV2. We cannot take the UV cut-off arbitrarily large, for an nth order graph has a divergence of order 2n\ it behaves like (M2G)n for large cut-off M. Counterterms to make the graph finite need



correspondingly large

finite number of counterterms

number of derivatives, but

only

available. Hence

take the cut-off to infinity, and if

we cannot



are

we want

insensitivity to the cut-off we must take M <G ̄l/2. Moreover, the energy, E, of the process under consideration must be much less than M, otherwise the calculation is dominated by details of the cut-off procedure. In other words, the four-fermion interaction is a good approximation to physics only if E

^ M ^ G ̄1/2. The minimum

possible relative error of calculations

is of the order of the maximum of M2G and

E2/M2.

Now,

it is always possible in principle to do experiments at arbitrarily energy. So the applicability of four-fermion theory at low energies

high 1/2 300 GeV there is new physics. implies that at energies rather below G" That is, the four-fermion theory becomes incorrect at that energy. The last fifteen years of weak interaction physics confirms this. (See, for example, Bjorken (1982)) For gravity, the corresponding energy scale is the Planck mass, of the order of 1019GeV. This is extremely far beyond the range of normal  ̄

accelerator experiments. Evidence for phenomena on such an energy scale come from much more esoteric observations. Examples might be

must

found in certain areas of the cosmology of the early universe, or from seeing the decay of



proton (Langacker (1981)).

In any case,

a non-renormalizable theory contains indications that it cannot describe all phenomena. It contains the seeds of its own destruction as a viable theory of a field of physics. So, given a successful non-

renormalizable theory, one must ask the following questions: (1) 'Of which complete theory is it a part?' (2) 'How is it related to that theory?' An

more

example is given by

the relation between the

Weinberg-Salam theory,

124

Renormalization

X-X.

Fig. 5.7.2. W-boson exchange gives

an

effective four-point interaction at low

energies.

which is

renormalizable, and the four-fermion theory, which is not. The theory arises as an approximation to W-boson exchange at

four-fermion low energy

(Fig. 5.7.2).

One replaces the propagator

il{l2-m2w\ for the W-boson,

by i/(-0-

The graph is suppressed by factor of at least E2jm\y compared to photon exchange. It gives an example of the decoupling theorem: the heavy particles have small effects at low energies. The only reason we can see such effects is the high degree of symmetry of the strong and electromagnetic interactions. These interactions conserve P,

C, T, and the number of each flavor of quark and of each flavor of lepton. Weak-interaction amplitudes are much smaller than strong-interaction or for similar processes, and are therefore normally invisible. But there are many processes that are completely forbidden in the absence of weak interactions; for these, any weak-interaction cross-

electromagnetic amplitudes

section,

no matter

how small, is all there is.

So one important way in which a non-renormalizable theory arises is as a low-energy approximation to a renormalizable theory in a process that is forbidden in the absence of the heavy fields. The heavy fields effectively give a

cut-off

on

the non-renormalizable theory. Then, for example, the four-

fermion coupling G is computable in terms of the underlying theory via formula like





constant

#2/m^ + higher



order corrections in g.

Here g is the dimensionsless coupling of the Weinberg-Salam theory. One manifest characteristic of this non-renormalizable theory is the weakness of

higher power of G is a higher inverse power leading power of mw as mw gets large, so it is in

its interactions. Also note that of

m^r.

We are

taking

the



general incorrect to calculate in the non-renormalizable theory beyond lowest order. Higher-order calculations must be done in the full theory. A slightly different situation arises in gravity. There one must perform calculations beyond tree approximation, since

gravitationally bound states

5.8 Proof of locality like

the

solar

system

are

of counter terms

125

formed by multiple exchange of gravitons.

generated involving higher-derivative interactions (e.g., The R2, R*v. etc.). ambiguity in the finite parts of these counterterms gives an uncertainty in the Green's functions. However the uncertainty is a power of momentum divided by some large mass scale, and is negligible for low- momentum-transfer processes. In weak interactions, the size of the higher- Counterterms

are

order corrections is of the

the calculations, but in

same

gravity

order of magnitude

this is not

so

as

the intrinsic

because of the

error

zero mass

in

of the

graviton. Another difference is that

gravity

is actually the strongest of the four

fundamental interactions when considered contrast,

on

atomic

or

on



large enough scale.

In

molecular scales, it is the other three interactions

by far the strongest. However, the strong and weak interactions finite range, so that they are essentially zero outside the nucleus. Particles can have both signs of electric charge, so that bulk matter, if

that

are

have



charged, tends to attract charge of the opposite sign to it. Bulk matter is therefore generally neutral. But gravity couples to mass or energy, so it is always attractive. Hence gravity wins out as the strongest interaction for large enough assemblages of matter. However at nuclear and atomic scales, it is negligible by a factor of about 1040 compared to the other interactions. Let

us

between

summarize

by restating

renormalizable

and

the key conclusion about the distinction theories. A non-

non-renormalizable

renormalizable theory considered at low energy gives some indications that at high enough energies it must break down, and cannot be a complete

theory. A renormalizable theory gives

no

such indication.

5.8 Proof of locality of counterterms; Weinberg's theorem

examples, we saw that the counterterm C(G) of a graph G is a polynomial in its external momenta, of degree equal to its overall degree of divergence 3(G). This is a general property, as we will now prove. The original proof of this theorem and some related results is due to In

our

Weinberg (1960); a simpler proof was given by Hahn & Zimmermann (1968). It is useful to distinguish three results: (1) Suppose that



1PI

graph

G and all its 1PI

subgraphs

have

negative

degree of divergence. Then the graph is finite. That the degrees of divergence of the graph and subgraphs are negative means that there is when all or some of the loop momenta go to infinity the other momenta finite. The problem is to eliminate the with together, of a divergence from more exotic scalings. possibility no

divergence

126

Renormalization

Suppose that a 1PI graph G has negative degree of divergence, but that it might have subdivergences. Then the graph is finite if we first subtract

(2)

off subdivergences. More simply, if 3(G) < 0 then R(G) is finite. (3) If a 1PI graph G has degree of divergence 5(G), then its overall counterterm C(G) is polynomial in the external momenta of G of degree 5(G). Property (1) is a trivial case of (2). We will reduce (3) to (2) by the same differentiation method as we used in Section 5.2.2. The proofs will be by induction. This naturally suggests that renormalization R(G) of G. One

generalization

we use

the recursive definition of the

is useful. It is that the renormalization

may be chosen so that result

prescription (3) reads ft Hooft (1973), Weinberg (1973), and

Collins (1974)):

graph G has degree of divergence 5(G), then its overall C(G) is polynomial in the external momenta of G and in the massive parameters in the Lagrangian. (The parameters in question are the masses of fermions and the squared masses of bosons.) The dimensions of the terms in the polynomial are at most 3(G).

(3') If



1PI

counterterm

5.8.1 Degree of counterterms equals degree of divergence

Property (3), that the overall counterterm C(G) of a 1PI is graph polynomial in the external momenta of degree 3(G). We will do this assuming Property (2), that a graph with its subdivergences subtracted is finite if its degree of divergence is negative. Let G be a 1PI graph with degree of divergence 3(G) > 0. We will consider R(G\ which is G plus counterterms We first prove

for

its

subdivergences. Following

Caswell & Kennedy (1982), let

us

differentiate the graph 3(G) + 1 times with respect to external momenta. This produces a result that has negative degree of divergence. We differentiate not only the graph G, but also its various counterterm graphs

Cy(G). we

The aim is to show that the result is actually convergent. To do this graphs are the correct

will show that the differentiated counterterm

counterterm

graphs

for the differentiated

obvious, but there are

some

subtleties,

original graph. This may will give the details.

sound

so we

Let the external momenta of G be pu...,pn. Its renormalized value is R(G)=G+ =



R(G)

X Cy(G) +

C(G)

R(G)-TGoR(G).

(5.8.1)

5.8

Proof of locality of counterterms

127

respect to one of the external momenta, dx denote any A-fold differentiation with respect to the external

Let d denote differentiation with

and let

dxR(G) is finite if X > 5(G). (It is clear from naive power-counting that S(dxG) S(G)- X.) Suppose we ensure that differentiation commutes with the basic subtraction operator Ty. This amounts to imposing a very natural relation

momenta. The property to be shown is that



between the finite parts of, for example, Ty(dG) and Ty(G). (The relation is satisfied by the pole-part subtractions, but it is possible to choose exotic

prescriptions

renormalization

graph

not

satisfying

hypothesis.)

Then for any

y we have

dC(y) It follows that, for the

point

here is that





we

have

(5.8.3)

R(dxG).

differentiation when acting

number of terms, in each of which

differentiated. It is



(5.8.2)

C(dy).



original graph,

dxR(G) The

the

one

on a

of the propagators

graph gives or



vertices is

simple generalization of the argument given in specific graphs that the counterterms for

Sections 5.2.2 and 5.2.3 for

subgraphs Now

of G

R(dxG)

are

the correct

is the

sum

of

after differentiation in (5.8.3). graph dxG that has negative degree of

ones a

divergence and the counterterm graphs for its subdivergences. Hence by Property (2) it is finite, so that we may choose the subtraction operator T to give zero. Therefore the counterterm in (5.8.1) for the undifferentiated graph is polynomial of degree 3(G) in the external momenta. The same argument (Collins (1974)) also shows that counterterms are in mass. Here it is necessary to note that differentiation with m2 to does not automatically reduce the degree of divergence. This respect only happens if counterterms for subdivergences are polynomial. If a

polynomials

counterterm

has



piece proportional

to In

(m2)

then

differentiating

with

respect to m2 leaves the degree of divergence unchanged. The proof merely demonstrates that it is always possible to choose counterterms to be

polynomial

in

m2;

it is not compulsory.

5.8.2 R(G) is finite if 5{G) is negative evident in one-loop examples that a graph with degree of divergence 3 is renormalized by a local counterterm of degree 5 in the external It

was

momenta. To

generalize the

result to

an

arbitrary graph, we constructed a computing the counterterm for

renormalization procedure which involved

128

Renormalization

1PI graph only after subtracting subdivergences. We differentiated the graph S + 1 times with respect to its external momenta to prove its counterterm to be local and of degree S. This proof relied on assuming the following statement: a

If a graph has negative degree of divergence and has its subdivergences subtracted according to the rules, then it is finite. More briefly, if 5(T) < 0

then R(T) is finite. This statement sounds extremely plausible. It is nevertheless in need of proof. We have to

ensure

that the subtraction procedure actually

removing the subdivergences. (That is, there are no induced spurious divergences by the procedure.) In addition, we normally the only consider divergences as arising from regions in which some loop momenta go to infinity, all at the same rate; this generates the usual power- counting. It is necessary to eliminate more exotic possibilities. The most important problem, which is the one we will examine, is to treat accomplishes its purpose of

case that a collection of loop momenta go to infinity, but at different rates. In Section 5.2.2, we examined the special case of Fig. 5.2.1. The general case is very similar. Inductively, we assume that properties (1) to (3),

the

listed

graphs than the of the graph integration over regions loop momenta where all or some momenta go to infinity, not necessarily at the same rate. We will eliminate them as possible sources of additional divergences. If all the momenta go to infinity together, then the negative overall degree of divergence means that there is no actual divergence from at the

beginning of Section

5.8,

are true

for all smaller

G under consideration. We consider

this region. If some momenta stay finite while the others go to infinity (not necessarily at the same rate), then let y be the subgraph consisting of all those lines with the large momenta. Our inductive

hypothesis

ensures

that all the

divergences are cancelled by counterterms for subgraphs. The remaining case is that all of the loop momenta go

resulting

to infinity, but

again not at the same rate. Let k denote the components of the smallest momenta, and let / denote the rest. (Our notation is meant to copy that used

is the proof.) Let y be the subgraph consisting of all the loop momenta /. It may be a single 1PI graph or a union of 1PI subgraphs. Let these 1PI subgraphs be yl9...,yL.

for Fig. 5.2.1, and those lines

disjoint

so

carrying

Expand each subgraph in powers of its external momenta up to its degree of divergence. The remainder for each subgraph is really a graph with negative overall degree of divergence; the contribution vanishes as / goes to infinity, so we

should have set /



0(k). The expanded

terms contribute

just

as

they

5.8 Proof of locality

of counter terms

129

Fig. 5.2.1. After subtraction of divergences we have a factor in the integral over k corresponding to the dimension of the subgraph. We gloss over here some of the important details, notably what happens to the value of a general subgraph when some of its external momenta get large. But the main lines of the argument should be apparent. The structure of the proof is the same as in subsection 5.2.2 and the result did for

is the

same.

5.8.3 Asymptotic behavior

Weinberg (1960) not only proved the convergence theorem stated above with more complete rigor, but he also investigated what happens when several of the external momenta pt, p2,..., pn of a graph y get large in the Euclidean region. They are assumed all to be of an order Q, with the ratios

Pj/Q fixed as Q

-?

None of the

oo.

sums

of subsets

ofpf/Q vanish. Weinberg

then states how to find the asymptotic behavior:

(1) Consider any subgraph y connected to all the lines carrying the large momenta. Let all the loop momenta of y be of order Q. Compute the power of Q: Qay. (2) Look at all such subgraphs. Let ay have a maximum value a. (a) If there is a unique graph with this maximum power, then T oc Qa

Q^oc. (b) If there such





are

several

subgraphs

with ay a, then let N be the number of The asymptotic behavior is:

subgraphs.



Qa[A0B0 + AlBl\nQ + A2B2(\nQ)2 + 0(Qa ̄ly

Here the A?s

and the B.'s

are

are

as



--

A^.B^^lnQf-^ (5.8.4)

functions of those momenta that

functions of the finite quantities

are

fixed

as

Q->

oo

Pj/Q.

inductively in the guts of the convergence theorem proof is similar. It is not obvious that this part of Weinberg's theorem is of much use for physics, other than for its part in this convergence proof, since the asymptotic behavior is of Euclidean momenta. However, in the deep- inelastic scattering of a lepton on a hadron, there is a photon or a weak This theorem is needed

proved in

the last subsection. Its

interaction boson that is far off-shell. The momentum carried by the boson is effectively Euclidean, and Weinberg's theorem applies. We will see this in

Chapter 14. There are also generalizations to other intrinsically Minkowskian situations (e.g. Amati, Petronzio & Veneziano (1978), Ellis et

130

Renormalization

al (1979),

Libby & Sterman (1978), Mueller (1978,1981), Stirling (1978), and are beyond the scope of the present book.

Buras (1981)). These

5.9 Oversubtractions We showed how to renormalize a Feynman graph by making subtractions for the divergent subgraphs and for the overall divergence of the graph. It is possible, however, to make subtractions on graphs that are not divergent. Subtractions can also be made with a higher degree polynomial in the external momenta than called for by the degree of divergence. Either of these cases is called oversubtraction. Now, the general form of the renormalization, either by the recursive method or by the forest formula, did not specify the exact form of the subtraction operator T. So oversubtractions can be made without changing the general formalism.

There

wish to

important uses for oversubtractions. The first is when we 'physical values' of masses or couplings as the renormalized

are two use

parameters. We will discuss this in

a moment.

The second

operator product expansions. There, subtractions are cancel UV divergences but also to extract asymptotic

use

is to construct

made not only to behavior

as some

external momenta get large. We will discuss this later in Chapter 10.

5.9.1 Mass-shell renormalization and oversubtraction We have considered renormalization

divergences. Another point

of view

as

comes

of removing from the observation that one the

procedure

coupling parameters that appear as Lagrangian. example, consider a theory where each field has a corresponding single-particle state. Then the masses that are measured are those of the single particles, and it is often sensible to cannot

observe directly the

mass

For

coefficients in the

parametrize

and

the theory in terms of these

masses.

Similar remarks

can

be

applied to couplings. (Thus in QED one normally parametrizes the theory by the electron's mass and charge, defined by the long-range part of its electric field.) It can also be convenient to rescale the fields so that each propagator has



pole

of unit residue.

simple renormalizable theory like 03 in six dimensions the renormalizations to accomplish such a mass-shell parametrization are In



precisely those necessary to cancel the UV divergences. Thus we may define the subtraction operator applied to a self-energy graph Z(p2) to be

T(ph)o2V)



Z(mp\) + (p2



mp2h)I'(mp2h)>

t5-9-1)

5,9 Oversubtractions that the inverse propagator satisfies

so



as

131

p2

-?

i[P2

m2h.







We use the

£ph(P2)]





subscript 'ph'

i(P2



O + 0(p2



mp\)2,

to indicate renormalization

(5.9.2)

according

to the mass-shell scheme. Of course, mass-shell renormalization is

renormalization prescriptions that

we

may

only

use to

one

out

of many

cancel UV divergences. But

Consider, example, (ft3 theory again, but now in four dimensions. We may continue to use (5.9.1) and (5.9.2) for the renormalization of the propagator so we have a 'physical' parametrization. But all except the one-loop self- no have so counterterms are all the wave-function energy graph divergence, we

may also choose to renormalize in the absence of divergences.

as an

finite and all but

one

of the

mass counterterms are

of the renormalization procedure

as

finite. The combinatorics

described earlier all work unchanged.

5.9.2 Remarks One important technical problem is to check that the

over subtracted

and

the normally subtracted theories differ only by a reparametrization. This can be done by the methods which we will describe in Sections 7.1 and 7.2. In the previous subsection 5.9.1, we took the point of view that renormalization is the process of reparametrizing the theory in terms of 'physical' quantities. It should be noted that this is not always a useful point of view. In the first place, other renormalization prescriptions are more convenient for handling certain types of calculation. In the second place, there may be infra-red divergences that make the mass-shell structure of a

theory

not what one would

naively expect:

thus in

QED

the electron's

propagator does not have a simple pole. And, finally, in some theories there are many more particles and couplings than independent parameters. This

is very

common

in gauge theories.

5.9.3 Oversubtraction

on

IPR

graphs

The aim of oversubtraction, generally, is to impose some condition Green's functions. So far, we have assumed the condition to be imposed

graphs, since those are divergences. However, consider (ft3

the

1PI



Let

us



(c<p)2/2



the +

zero

needing

counterterms for

(ft4 theory:

m2cft2/2 -f(ft3/6

choose to renormalize at

ones

on on



#724 + counterterms.

(5.9.3)

external momentum. Thus the self-

132

Renormalization

IHIX (b)

(a)

(c)

{d)

Fig. 5.9.1. Subtraction of one-particle-reducible subgraphs. energy I and the

three-point

l(p2=0) ̄(p2

IPI function =

dp2

ro)(Pi

Following from function

our



Pi

0)

P%







earlier work

r(4) by requiring

the

sum

r(3) satisfy

0,

0)



we

lowest order





if

(5.9.4)

might renormalize the four-point graphs to be equal to their

of the IPI

lowest order value at zero external momentum. However it is also sensible

impose instead the condition on the amputated four-point function T^. graphs are IPI only in the four external lines. (We should amputate the graphs since the counterterm vertex will have attached to it external propagators.) Thus in addition to the three tree graphs of Fig. 5.9.1 (a)-(c\ we require the counterterm, Fig. 5.9.1(d): Our general method of renormalization tells us that whenever we have a basic graph containing one of the graphs (a), (b\ or (c) in Fig. 5.9.1 as a subgraph, there will be counterterm graphs in which this subgraph is replaced by the counterterm vertex (d). These counterterm graphs may be divergent even when the basic graph is finite. An example is shown in Fig. 5.9.2. In Fig. 5.9.2(a) if we impose the renormalization condition on the IPI functions only the graph (b) occurs as counterterm;(a) plus (b) is finite. If to

These

impose the condition on amputated graphs we immediately meet graph (c) where the line A is replaced by its 1/3 share of the counterterm Fig. we

5.9.1(d). But

graph (c) has a divergence, so we must renormalize it by a three-point subgraph consisting of the line A and the loop B. This is

counterterm to the

[a (b)

[a



(a)

(b)

"TV

^T

(c)

W)

Fig. 5.9.2. The subtractions of Fig. 5.9.1, inside



bigger graph.

5.10 Renormalization without regulators

133

graph (d). Note that the graph consisting of line A and loop B has a subdivergence, but no overall divergence. Even so, the overall counter- term in (d) is divergent. It is not difficult to see that the extra counterterms needed to impose the renormalization condition on the 1PR amputated graphs do not change shown in

our

results

on

renormalization. The instructions for renormalization in

Sections 5.3 and 5.5

can

be used provided only that

subgraph' by 'amputated subgraph'. The use of subtractions on 1PR graphs However it is



is too

we

replace the term '1PI

baroque for

device that is useful for discussing the large

and the operator-product expansion (Chapters

normal

mass

use.

expansion

8 and 10).

5.10 Renormalization without regulators: the BPHZ scheme

setting up the renormalization procedure in Sections 5.3 and 5.5 we were careful not to use a specific definition of the subtraction operation. This was In

allow for the choice of one out of the infinitely many possible renormalization prescriptions. An obvious one is the mass-shell subtraction procedure indicated in the last section. Another is the minimal subtraction procedure to be defined precisely in Section 5.11; we have already made much use of it. In this section we will explain the method of Zimmermann (1969), in which the subtractions are applied directly to the Feynman integrand, so that no regulator need be used. The starting point is the method due to Bogoliubov & Parasiuk (1957) and Hepp (1966), called the BPH scheme. They observed that the overall counterterm for a graph T is a polynomial of degree d(T), its degree of divergence. So they defined the subtraction operator T(T) to be the terms up to order S(T) in the Taylor expansion of T about zero external momentum. For example, consider the one-loop self-energy graph to

Fig.

3.1.1 in

03 theory in

six dimensions. After dimensional

regularization

its unrenormalized value is

Za(p\d)



2(An 412 v

The terms up to order



dx[m2-p2x(l-x)f/2-2.

T{2-dl2)



p2



(5.10.1)



in its Taylor

expansion

about p





\ldxmd-4[l-(d/2-2)p2x(l-x)/m2l

are

(5.10.2)

134

Renormalization

The renormalized value of the

g2 1287T3

dx{ [m2



p2x(l

graph





is

Fa

x)]In [1





T°ra.

p2x(l



So at d

x)/m2]





6 this is

p2x(l



x)}.

(5.10.3) Zimmermann's (1969) achievement was to realize that this construction can be applied directly to the integrand. Subdivergences are subtracted with the aid of his forest formula. Then the result is an integral which, according to power-counting, has no UV divergences. The integral therefore has in fact no divergences (Hahn & Zimmermann (1968), Zimmermann (1968)). This method is called BPHZ renormalization. In Section 3.4, we applied this method to the above graph, with the result (3.4.7). It can be explicitly calculated by putting all the terms over a common denominator and then using standard parametric methods. The result

agrees with (5.10.3). An example involving



subdivergence is given by Fig. integral be

5.10.1 for

04

in

four dimensions. Let the renormalized rBPHZ 1

->3

2(2tt)8

(5.10.4)

d4/cd4//(Pl,p2,p3,p4,/c,/).

Pz

Fig. 5.10.1. A two-loop Then

we

will construct the

The unrenormalized



1 =

vertex

(I2



integrand integrand is



m2) [(*



I)2

graph in 04 theory.

/.



1 -

m2] [(*



p3)2



m2] [(*



p4)2 m2] -

Subtraction of the sole subdivergence gives R(U)  ̄U ̄

(/2-m2)2[(/c + p3)2-m2][(/c-p4)2-m2] -2k-l-k2



{I2



m2)2 [(* + I)2



m2] [(/c + p3)2



m2] [(k



p4)2



m2]' (5.10.6)

5.11 Minimal sub traction Then the overall





_

R(U) (2k-l

k2){(k2



(I2

divergence

R(U)-[R(U)]\pi=P2





135

is subtracted to give =

P3=P4





m2)[p2 + p\ + 2/c-(p3 p4)] + (p2 + 2k-p,)(p2 2/rp4)} m2)2(k2 m2)2[(/c + I)2 m2] [(* + p3)2 m2] [(* p4)2 m2] -

















(5.10.7) The BPHZ scheme has (1) It is

applied

number of advantages:



integrand

to the

and generates



without requiring any regularization. Thus it exhibits the fact that the properties of (2) theory do not depend

Mathematically it

(3)

convergent integral a

renormalized field

which UV regulator is used. is rather elegant. In particular there is on

no

need to

required to have a theorem that tells us that a graph that is convergent according to the naive criteria is actually convergent. (4) It allows a very simple proof of the operator-product expansion. discuss directly the

There

(1)

are a

divergences

number of

of Feynman

graphs;

it is only

disadvantages:

It is not the best scheme for theories

complicated symmetries,

(especially

gauge

theories) with

where relations between counterterms have to

be preserved; the scheme does not allow direct computation of

(2)



of the value

divergence.

The subtractions

in

are

made at

zero momentum

and therefore

are

massless theory. divergent (3) When the scheme is generalized to handle massless theories, it becomes much more complicated (Lowenstein, Weinstein & Zimmermann, infrared

1974a,



b). 5.11 Minimal subtraction 5.11.1

Definition

proved (Speer (1974) and Breitenlohner & Maison (1977a, b, c)) when dimensional regularization is used, the UV divergences of that, Feynman graphs appear as poles at isolated values of the space-time dimension d. Minimal subtraction ft Hooft (1973)) the MS scheme consists of defining the counterterms to be poles at the physical value of d, 4. We have already used this scheme, in Chapter 3. Our purpose in this d section is to make precise the definition of minimal subtraction. It can be







The main

depends

on

complication d,

so

that

is that bare couplings have a dimension that introduce the unit of mass fi, as follows:

we must

136

Renormalization

(1) Consider in turn the coefficient g. + dgt of each term in if. Let the dimension of g. be a{ + bf(4 d). Then we replace gx + (5#- by ^,(4_d) -

Thus the renormalized

(Gi + ^/)-

coupling

gt and the counterterm

both have dimension ai9 independently of d. (2) Let T be a 1PI graph to which it is desired to operator T. Let the dimension of T be A + B(d

apply

6gi

subtraction



4), and suppose the just explained. Then we define -

couplings all contain T(T) The We



powers of \i as

/("-4){pole

part of

(/(4"d)r)

at d



4}.

pole part is obtained by making a Laurent expansion about d 4. have arranged to take the pole part of a function whose dimension =

does not depend

on

d.

talking about a theory in a different number of physical dimensions than four. For example, we might be in 03 theory in six dimensions. Then the '4' in the above formulae is replaced by the correct

(3) Suppose

physical

we are

value.

For a simple graph with no subdivergences, like the one-loop self-energy in (5.10.1), this prescription amounts to subtracting the pole:

Zr)(d



6)



\im\^^r(2-d/2){1dx[m2-p2x(l-x)r12-2 Af*n) Jo

d^e(

g2 1287T3

\[yE-l-ln(4n)](m2-ip2) ^



m2

2^ dx[m2-p2x(l-x)]ln





p2x(l —x) it

graphs with subdivergences, the subdivergences removing the overall pole. The advantages of the scheme are:

For

(5.11.1)

must of course be

subtracted before

(1)

It

automatically preserves complicated symmetries.

The

exceptions are by

chiral symmetries and the like, which in general cannot be preserved

quantization see Chapter 13. (2) It has no problems with massless theories. In fact, dimensional continuation regulates both IR and UV divergences, thus removing the need for a separate IR cut-off. -

(3) Calculations

are

very convenient.

(4) Computation of the divergent part of



Feynman graph



needed for

5.11 Minimal sub traction

renormalization-group

calculations

137

is almost trivial at the



one-loop

level.

disadvantages of minimal subtraction

Some

(1) (2)

unphysical. proof of the operator-product expansion is made harder than in the

It is

The

are:

BPHZ scheme.

5.11.2 MS renormalization

QCD, where it has become standard. Another disadvantage that then appears is that minimal subtraction tends to produce large coefficients in the perturbation expansion. These are primarily due to the In {An) yE 1.95 term such as appears in (5.11.1). It has become conventional to work with a modified The MS scheme has found much

use

in work

especially





scheme, called the MS scheme

(Bardeen, Buras, Duke &

Here the \i of the MS scheme is written

we

Muta

(1978)).

as



Then

on

(5.11.2)

;0-38/i.

have, instead of (5.11.1), the cleaner form

1287T m2

dx[m2-p2x(l-x)]ln





p2x(l



x)

(5.11.3)

fi2

5.11.3 Minimal subtraction with other regulators Minimal subtraction could also be applied with other UV cut-offs. For example, if a lattice of spacing a is used, then the singular a -? 0 behavior of graph of degree of divergence 3 is

a ̄s[polynomial One

can

therefore define T(T)

as

the

in In

(a)].

singular part of T, with

the

general

form T(D= /?

Z"[lnM]^0;/J+ £ l^{.\n(an)fAxJ. =

After subtraction of masses



a=l P



(5.11.4)

0"

subdivergences, the coefficients Aa^ are polynomials in

and momenta. Note again the appearance of

scheme has found little

use.



unit of

mass.

This



Composite operators

In

Chapter



number of

we met a

equations involving products

of field

operators at the same point. Examples are given by the equations of motion (2.1.10) and the Ward identities (2.7.6). These products we will call When computed directly they have ultra-violet 0(x)(/>(y) makes unambiguous sense if x is not

composite operators.

divergences: the product

equal to y9 but if x equals y then we have 0(x)2, which diverges. Since the equations of motion and the Ward identities express fundamental properties of the theory, it is useful to construct finite, renormalized composite operators with which to express these same properties. It could be argued that there is no need to have renormalized equations of motion. One could say that one only actually needs the equations of motion in the regulated theory, where they are finite. A situation of practical importance where we actually do need renormalized composite operators is the operator-product expansion, to be discussed in Chapter 10. This is used in a phenomenological situation such as deep-inelastic scattering

(Chapter 14)

where we wish to compute the behavior of a Green's function of its external momenta get large. Equivalently, we need to know how a product of operators, like cp(x)(j)(y), behaves as x^y.

when

some

This information is contained in the

operator-product expansion

of

Wilson (1969) which has the form

<t>{x)<t>{y)



Cx(x



y)\



C+2(x



y)[cj)(y)2]







?.

(6.0.1)

symbol [^(x)] denotes the renormalized operator corresponding composite operator A(x). The coefficients C(x y) are onumbers, and each has a subscript which labels the operator that it multiplies. Here the

to an unrenormalized



Therefore in this chapter of

composite operators,

we

show how to renormalize Green's functions

e.g.,

<O|T0(x)(/>(y)(/>2(2)|O>, <O|T0(w)(/>(x)(/>(y)2(/>(2)2|O>, <O|T0(x)2(/>W2|O>. 138

(6.0.2)

6.1

Operator-product expansion

We will first motivate the

use

139

operators by seeing how the low-order graph. Then we will low-order graphs for composite

of composite

operator-product expansion arises in



examine the divergences that appear in operators. We will see that we must expect multiplicative renormalization:

[4>2]



Z^2,

[02] is finite as the UV cut-off is removed, while Z02 is a renormalization factor. The unrenormalized operator (j)(x)2 is when the cut-off is removed. where

(6.0.3)

divergent divergent

These examples will provide motivation to define renormalized composite operators by application to Feynman graphs of the same

R-

defined in Chapter 5. After discussion of a number of technical issues, we will derive some basic properties of the renormalized

operation

that

we

operators, including the equations of motion and the Ward identities.

6.1 Operator-product expansion We will postpone a complete treatment of the operator-product expansion Chapter 10. Here we merely wish to motivate our definition of composite

to

operators with

Fig.

an

6.1.1. Take q^>

example of their

oo

use.

in these graphs to obtain the lowest-order example of the operator-product expansion.

graphs of Fig. 6.1.1 for the four-point function in 03 theory. infinity with px and p2 fixed, and with the ratios of the We expand the graphs in powers of of q fixed. Then \q2 |

Consider the

We let q* go to components

q2

-?

oo.

to find:

Fig. 6.1.1 (a)



(ft)-

(p2-m2)(p22-m2). (q2r

(6.1.1a)

Composite operators

140

Fig. 6.1.1(c)

10



2\2 (Pi-m2)(p22-m2)((Pl+p2)2-m2)_ (q2)





2g'(Pi+P2) <T







(2?2



(Pi

P2)2) + 4g'(Pi+P2







(9:2^2

(6.1.1b) show dependence of of an operator-product expansion like (6.0.1) (after Fourier transformation into momentum space). In (6.1.1 a) the factor in square brackets is in fact the value of the lowest- order graph for the following Green's function:

In each

px and p2 has factorized. We

term the

that this is



now

case

<o|r^(-Pl)^(-p2)02(O)/2|o> d^exp(-iprx-ip2^)<0|T(/>(x)(/>W(/)2(0)/2|0>.

(6.1.2)

We have

not Fourier transformed the 02(O) operator, but have set it at the we If had made the Fourier transform, then we would merely pick up origin.



momentum-conservation 5-function, which

we

To understand the appearance of the operator

do not have in

(6.1.1a).

02(O)/2 in (6.1.2), we may

functional-integral formula for the Green's functions that appear in equation. Since 02(O) means the product of two fields at the same space- time point, such a formula follows from our work in Section 2.2. It is

find



this

<O|T0(x)0(j#a(O)/2|O>



.>K

[dA]A(x)A(y)±A2(0)eis.

The Feynman rules for this Green's function are

the usual

ones

for the Green's function

addition that each graph contains exactly

one

can

(6.1.3)

then be derived. They with the

(0\T(ft(x)(j)(y)\0} special

vertex for the

02(O)/2

operator. The lowest-order graph is shown in Fig. 6.1.2, where the special

cj)2/2

vertex is indicated

by

a cross.

The value of the vertex is unity, for the

explicit factor 1/2 in 02/2 gets cancelled. This happens in exactly the same way as the 1/4! in the c/>4 interaction or the 1/3! in the 03 interaction gets cancelled to leave The operator

composite field).



value



\g for

an

interaction vertex.

first example of a c/>2(0)/2 we this term mean, in general, By is

our

composite operator (or product of elementary



Fig. 6.1.2. Lowest-order graph for two-point function of

02.

6.1

fields

Operator-product expansion

(or their derivatives)

operators that we will We can write

at the same

investigate

Fig. 6.1.1(a) + (b)

point. It is chapter.

141

the

properties of

2iq2



such

in this

^3 <0| 7WPl)0( P2)</>(0)2/2|0>,

(6.1.4a)

-2

Fig. 6.1.1(c)

2ig2 23

id )



(q2

2iq" 3

q2

dx"

/ \





(2m2+D) '

(c)



Fig. 6.1.3. Generation of terms in the operator-product expansion from the graphs of Fig. 6.1.1 (a) and (b).



V<f3?a

q2

<O|T^(p1)^(p2)0(x)|O>|x



W + ( ) ̄



ig

{q2)2

L]

This is illustrated in diagrams by Fig. 6.1.3. In similar fashion, we derive an operator formula for (6.1.1b):

q4 =

(6.1.4ft)

o.

coefficients y ^



derivatives

Fig. 6.1.4. Generation of terms in the operator-product expansion from the graph of Fig. 6.1.1 (c)

Here the form of the first square-bracket factor

means

that

we

need

an

elementary field 0(x), rather than a composite field. The pt's and /?2's in the numerators in the second square-bracket factor have turned themselves into derivatives with respect to x;

we set x



0 at the end.

Equation (6.1.4b)

is illustrated in Fig. 6.1.4. The form of (6.1.4) suggests the following formula:

<0\T^q)ct>m(p^(p2)\0} ^^C^)<O|7^(O)0(Pl)0(p2)|O>.

(6.1.5)



The sum is over a set of local operators (9t. Each of these is either the elementary field 0, one of its derivatives, or a composite operator such as 02. The coefficients C^q) are called the Wilson coefficients. In Chapter 10

generalize this result to all orders. We will have an expansion for any Green's function with large momentum on some of its external lines. Now higher-order corrections to the Green's functions of the composite

we

will

operators such as the 02(O) that appears in (6.1.4a) have ultra-violet divergences beyond those appearing in Green's functions of elementary fields. We will see this in the next section, Section 6.2. To obtain an operator product expansion, like (6.1.5), with finite coefficients, most of this

we

will need to

composite operators. This particular problem chapter.

renormalize the

will occupy

142

Composite operators 6.2 Renormalization of

composite operators: examples

6.2.1 Renormalization of (j)2 The Feynman rules for Green's functions of unrenormalized composite operators can be derived from the functional integral, in the presence of an

ultra-violet cut-off. In coordinate space they are the usual rules, modified only by having several external fields at the same point. For example, we consider

<0|T4,(x)4>0#2(2)/2|0>

(6-2.1)

03 theory in six-dimensional space-time. The connected graphs up to order g2 are shown in Fig. 6.2.1. As before, the vertex for <f>2/2 is denoted by in

a cross.

-o- (?)

(b)

(c)



(d)

Fig. 6.2.1. Renormalization of the operator (j)1. To work in momentum space,







The

we

Fourier transform,

as

usual, and define

<O|T0(p)^)0(z)2/2|O>

|d"xd''>;exp[i(p1-x

lowest-order



p2-3;]<O|T0(x)0(3;)0(z)2/2|O>

graph, Fig. 6.2.1(a), is equal

to



/-<

a ̄(p2-m2

(6.2.2)

tii



ie)(p2-m2



ie)'



o\



Observe that the factor 1/2 in the operator <\>2j2 is cancelled, just like the 1/3! that comes with the interaction vertices.

graphs of Fig. 6.2.1. They are all divergent: Fig. 6.2.1(b) logarithmically divergent, while the remaining graphs, Fig. 6.2.1(c) to (e\ are quadratically divergent (all at d 6). The divergences in the last two graphs, Figs. 6.2.1(d) and {e\ involve self-energy corrections only, so these divergences are cancelled by the usual wave- Let

us

now

turn

to

the one-loop

is



6.2 Renormalization

of composite operators: examples

143

-X—?- (a)

(b)

Fig. 6.2.2. Counterterm graphs for Fig. 6.2.1(d) and (e).

counterterms, Fig. 6.2.2. In these and other graphs in this

function and

mass

chapter,

indicate counterterms by

we



heavy dot

and

an

composite operator by The remaining two graphs have no counterterm from and they are both divergent. For Fig. 6.2.1(b) we get

insertion of



a cross.



(p2



the interaction,



m2 + ie)

ddk

(pj-m2



ie) l

(k2 -m2



m)[(*



Pl)2



m2 +



-T(3-d/2)x (Pi-m2)(p2-m2)64n2  ̄m2



dy\

dx

ie] [(jfc + p2)2



m2 +

ie]





p\y(\ -x-y)- p22x(\ -x-y)-(p1+ p2fxy Anil2 (6.2.4)

and for

Fig.

6.2.1

(c)

we

get

(p?-m2)(p22-m2)[(p1+p2)2-m2] -d/2

2(2*)'





(/c2-m2)[(p1+p2 + fc)2-m2] -flfAt3--'2 (p2-m2)(p2-m2)[(p1+p2)2-m2] "

-^/2"3rn 128,3

^f1. [m2-(Pl+p2)2x(l-x)r2-2

-r(2-./2)jod^

-(4;V-3

(6-2.5)

Note that there is a symmetry factor 1/2 in this last equation. The fact that the

sum

of (6.2.4) and (6.2.5) diverges

means

that the operator

02(O) is

not

finite. in the operator-product expansion we do not need precisely the operator 02. Rather, we need some local operator similar to 02 that is finite. This indicates that we should define a renormalized operator by

For

use

subtraction of the divergences. Let us agree to use minimal subtraction. Then the counterterm graphs are obtained by replacing each divergent loop by

144

Composite operators

(a) Fig.

(b)

6.2.3. Counterterm graphs for Fig. 6.2.1(6) and (c).

pole part, as illustrated in Fig. 6.2.3. Thus the counterterm graph

minus its

for Fig. 6.2.1(h) is r

ip\



m2)(p22

and the counterterm graph for

(P?



m2)(p\



m2)[(Pl



p2)2



m2)647r3(d



(6.2.6)

6)'

Fig. 6.2.1(c) is



m2] \(An\d

6)^  ̄iiPl P^)' +



(6.2.7) The

of the factors of \i is such that the counterterm in curly brackets has exactly the same dimension as the loop to which it is a

positioning

counterterm.

We thus find the renormalized values at d

R(Gb) "

(p?-m2)(p2_m2)647r3l

' ̄x



R(GC)

-{p21y dylJm2 L





To

6:

dx



p22x)(l-x-y)-(pl+p2)2xyTl 4jt)"2

-9

(6. 2.8)

JJ' -9

{j>l-m2)(p2-m2)[{Pl+p2)2-m2-\\mn\

x?j(y-l)[m2-i(p1+p2)2]+ x



In

f'dx^-^



p^xa-x^x

4nfx2

(6.2.9)

interpret these renormalizations we observe that the counterterms are <f>2(0) in (6.2.6) and for (m2 + D/6)</> in (6.2.7). Thus

vertices for

Ga



R(Gb) + R(GC) + R(Gd) + R(Ge) i +





64n3(d ..

9H



J<o|r<ft(Pl)<?(p2)l4>2(0)|0)

6)

d/2-i

64ttV-6)

<0\T¥(Pl)4>{p2)(m2+i\3)<t>(0)\0> + O(g4 (6.2.10)

6.2 Renormalization So what

we are

computing

is

of composite operators: examples a

Green's function of the operator ,.d/2-3

?2

<t>2 + 29



145

^m2 + *□)</> + higher order.

64n\d-6) *a

3,j

(6.2.11) renormalized

use the square brackets on the left-hand side to denote operator. In subsequent sections we will see that this result generalizes to all orders: a renormalized operator [02] can be defined to all orders by a

We



formula of the form:

i[02]=^02 + /id/2-3Z,m20 + //2-3ZfDc/>,

(6.2.12)

where the Z's depend only on g and d. Such renormalized operators were defined by Zimmermann (1973a). He called them normal products and used the notation N[(/>2]. His definition differed from

ours

only in that he

used BPHZ renormalization instead of

minimal subtraction. Observe that in order to obtain

degree of divergence of at least zero, the (6.2.11) or (6.2.12) have dimension original operator 02. This is a general



operators that appear as counterterms in

less than

equal

or

to that of the

phenomenon. Moreover, the only operators of such dimension

are

actually appearing in (6.2.12). We may write the renormalization matrix equation in the following form:

\Za ^->Zbm2 it?-\

Here

we

Ocj>

are











have used the fact that

said to mix with

diagonal elements Zb are



and Zc

cp2



and

D0

are

I I

(/>

].

those as



(6.2.13)

finite. The operators (j) and

under renormalization, because the off-

are non-zero.

needed in the renormalizations,

so

further operators and □</> are said to form a

Moreover,

02, 0

no

closed set under renormalization.

6.2.2 Renormalization of Sometimes

operators.

we



(j)2(x)(j)2{y)

need Green's functions involving two

or more

composite

simple example is

<oin[?p](p)i[>2](o)|o> =

for which

fd'xe-^OlriOTMK^OJlO),

the lowest-order

graph is Fig.

(6.2.14)

6.2.4. The renormalizations of

146

Composite operators

Fig.

(6.2.11)

6.2.4. Lowest-order graph for

<O|r[02]/2[>2]/2|O>.

do not appear until the next order,



2{2n)d) This is ultra-violet



V-m2)[(p

divergent,

even

so



the

graph

has the value

(6'115)

k)2-m2]'

in free-field theory. Even

though

the

(j)2(x) and 02(O) on the right of (6.2.14) are well-defined, we have to

operators

integrate through the point x 0. At x 0 there is a singularity, which is not integrable if d > 4. We may nevertheless define a finite Green's function =

by adding





local counterterm:

<O|n[02]W|[^?2](O)|O>J?



with

C(x) Once



(m2



<O|Ti[^2]Wi[02](O)|O> (6.2.16) -C(x)<0|l|0>.

D/6)d^(x))A 3r, ;N + 0(g2). ,64tt3(^-6)

have used minimal subtraction at d

more we



(6.2.17)

6.

6.3 Definitions We define renormalized Green's functions of composite operators by applying the K-operation to the Feynman graphs, just as we did for Green's functions of elementary fields in Sections 5.3 and 5.5. We will need to show that the counterterms generate multiplicative renormalizations of the

operators (e.g., (6.2.12)). This is similar to what we

we

did in Section 5.6, where

showed that the counterterms in elementary Green's functions

are

starting generated by Lagrangian. with the graph-by-graph renormalization is again to allow a simple treatment of the problems of subdivergences. We do not need a new proof that the counterterms for operator insertions are local; our original proof suffices. Our motivation for

counterterms in the

As before, ones

that

scheme



we

have

are most

see



prescriptions. The subsequent developments are the BPHZ

choice of many renormalization

useful for

Zimmermann

(1973a)



and

minimal

subtraction



see

Breitenlohner & Maison (1977a, b, c) and Collins (1975b). In any case we have a subtraction operator 7(G) which is applied to a graph G (after

removal of subdivergences) in order to extract the divergent part of G. In the

6.3 BPHZ scheme, T(G)

147

Definitions

gives the first 3(G) terms in the Taylor expansion about

external momentum, where d(G) is, as usual, its degree of divergence. In the case of minimal subtraction we must state how the unit of mass \i is

zero

from the examples in Section 6.2, we must always arrange to compute the pole part of a quantity whose dimension does not vary with d. So suppose we need TMS(G) for a 1PI graph G of dimension treated. As

^B(d-d0)

can

Then

be

we

seen

define 7=

/(d-do){pole

part of

(6.3.1)

G^ ̄%

where, as usual, d0 is the physical space-time dimension. The definition of a renormalized Green's function by the R-operation is rather abstract, and counterterm

we

will

show that it amounts to

now

adding

operators. In the one-loop examples of Section 6.2 this

rather obvious. In the

from

renormalization of an

(see

general case, we start arbitrary Green's function

R(G)



was

the formula for

(5.3.7)—(5.3.9)):

G^^Cy(G).

(6.3.2)



Here, as usual, the sum is over all subgraphs y that consist 1PI

subgraphs

yl9...,yn. Each y. is as in Section 5.3.

replaced by

of a set of disjoint

its overall counterterm

vertex, generated

distinguish the various y.'s that occur, according to the composite operator insertions that they contain. Consider, for example, Fig. 6.3.1, which illustrates the renormalization of <0| T(j) (j) (j)2/2\0y in the 03 theory in six dimensions. There is a one-loop subgraph ya for the three-point function. This is renormalized by its counterterm Cya in the Lagrangian. There is also a two-loop subgraph yb which contains the composite operator vertex. (This has a subdivergence, Let

us

number of

which must be subtracted.) The counterterm C,.

where

be considered

as







can



4?. 13}

Fig. 6.3.1. Renormalization of two-loop graph with insertion of composite operator.

148

Composite operators

<2> O (?)

(b)

cO O" (c)

(e)

(d)

Fig. 6.3.2. Renormalization of three-loop graph for

generated by

0(#4)



<0|T[>2]/2|>2]/2|0>.

Za. graph of order g4 for the Green's function <0| 7 ̄[02]/2[</>2]/2|O>. In addition to an overall logarithmic divergence, it has three divergent subgraphs. One subgraph ya is renormalized by a an

Next consider

term in the renormalization factor

Fig. 6.3.2,



in the interaction Lagrangian. The other two subgraphs yb and yc each look like yb of Fig. 6.3.1, and are renormalized by the same counterterm. These two counterterms are generated from counterterm

<O|7i[02](x^ + other terms from

by expanding each Za to 0(#4) and picking out the overall counterterm is obtained. It gives a term

Zb

and

terms for yb.

of

0(g*)

Zc,

(6.3.3)

Finally, the

in the C(x) of

(6.2.16). These arguments generalize easily to arbitrary consider



renormalized Green's function of

one

graphs. It suffices composite operator

(0\TY\4>(x^[A{y)}\0y.

to

(6.3.4)

i=l

We define the renormalization

by the recursive formula (6.3.2), and we wish

to prove that this equals

EZ^Olrn^OBMlO). B

Here

ZAB is



(6.3.5)

i=l

renormalization factor whose value is given by

writing,

in

analogy to (5.6.5),





Ez^=EttT^G). iyiGl B

(6.3.6)



Here G is any 1PI basic graph that includes a vertex for A. It has NG external lines in addition to the vertex for A, and D(G) is the operator corresponding

6.4 Operator mixing to the overall counterterm C(G). The factor

1/JVG!

(5.6.5), to organize the symmetry factors. Consider a graph for (6.3.4). We investigate The

subgraph

149

one

is just like the l/N(G) in

of its counterterms

Cy{G).

graphs yl9...9yn.Ayj that does not contain by C{y), which corresponds to one of the

y consists of 1PI

the vertex for A is replaced

counterterms in if. A y. which does contain the vertex for A must be one of

(6.3.6), so the counterterm C is generated of the counterterm by operators on the right of (6.3.6). Now sum over all graphs G for our Green's function and expand each R(G) by (6.3.2). The sum over 1PI subgraphs y. that correspond to the G's that are summed over in one

independently

the

££ can

subgraphs giving

the counterterms for the operator vertex. The result is

then the desired result

be done

of the

counterterms in

sum

over

(6.3.5). 6.4 Operator mixing

We have seen that a renormalized

composite operator \A\ by

is expressed in

terms of unrenormalized operators

\_A]

^ZABB.



(6.4.1)



In the case of counterterms

[02]

had the

we

that the operators that

saw

same or

lower dimension. Let

were

us now

needed

as

demonstrate

this for the general case. The proof is essentially

dimensional analysis. Let G be a 1PI graph containing a vertex for A and having the same number NB of external lines as a particular operator B. Now B is a product of NB fields with a certain number DB of derivatives. A counterterm can

only be generated if the degree

of divergence 3(G) is at least DB. Now in couplings have non-negative dimension, so

d(G)



dim (G)





renormalizable theory all

dim (couplings) (6.4.2)

<dim(G). On the other hand, since ZABB is dim

But we

only need

(G)





DB

possible + dim

B as a counterterm if

DB

counterterm we have

(6.4.3)

(ZAB). <

3(G),

so

dim(Z^)>0 for every counterterm. This

means

that the maximum dimension of



operator is the dimension of A9 as we wished to prove. In Section 5.8.1 we examined the dependence of counterterms in the

counterterm

150

Composite operators

Lagrangian on mass parameters. Provided we used minimal subtraction this dependence was polynomial, with the mass behaving as a dimensional coupling in determining allowed counterterms. The same argument applies here. The result is that ZAB is a polynomial in masses (and super -

renormalizable couplings) times the inevitable power of the unit of mass p. The coefficients of the polynomial are dimensionless functions of the dimensionless couplings and of d. A typical example of this is given by our see (6.2.11) and calculation in Section 6.2 of the renormalization of [</>2] -

(6.2.12). 6.5 Tensors and minimal subtraction Suppose we tempting to

minimal subtraction to define

use

[(50)2]

[d^dv4>].

and

[(S0)2]=^v[3,05v0]. This supposition is in fact false, that the

It is

suppose that

taking

of

as we

will

(6.5.1) demonstrate. This

now

does not commute with taking general. We will explain the significance of this fact. a trace



means

finite part, in

The lowest-order graph (in 03 theory) for either operator is d<j>2/2 before renormalization is

Fig.

6.5.1.

The 1PI part for

G(P)



2(2n)d J (k2 -^yrjd'fc^2

?H! J L

128tc3 x

For

d^(j)dv(j)/2

we





-i

 ̄* m2]

- -



3 X

4np2

x)[m2 p2x(l x)] d/2)[m2 p2x(\ x)]2}

d/2)p2x(\

(d/2)T(\

 ̄2^ ̄ M2 + k)2 m2)[(p

|d/ 2



{T(2 +

- -











(6.5.2)

have

G,M- ',,,_*, ld

k?(k



V-m2)t(P



p)v

*)2-??2] |d/2

128tt3 x

J0 *l



3 X

Ann2

x)[m2 p2x(l *)] <i/2)p?pvx(l + ^,r(l d/2)[_m2 p2x(l x)]2}.

{T(2















k+p

Fig. 6.5.1. One-loop graph for the operators in (6.5.1).

(6.5.3)

151

6.5 Tensors and minimal subtraction

Manifestly (6.5.4)

g^G^^G for the unrenormalized Green's functions. This has We can renormalize

K[G(p)]

K[<V(P)]





by minimal subtraction.

G(p)

3m* +

-128?3(d_6)C-

G?,(p)

!mV

-12g^_6)[P,Pv(X



K[G(p)]-^K[G?]= The

reason

why

&>*]. (6-5-5fl)



iVP2)

-^(X-!?V

Thus

subtraction of the

be true since

to



-^-^[X-imV

^4)]- (6.5.56) +

foP4]. (6.5.6)

contraction with 0"v does not commute with the is simply that taking the trace introduces d-

pole

dependence. Thus: g^

pole part pole part

of

of

d d-6'

d-6

——^ d-6



pole r

d-6

d-6

We must evidently be careful to specify whether a trace is inside or outside of the renormalization. The need to do this is characteristic of dimensional regularization. Which place to put the trace depends consideration.

on

the

problem under

problem arises whenever we have to consider a tensor of rank at least a trace of Dirac y-matrices except that we choose the trace of the unit Dirac matrix to be independent The

2. (It could also arise in connection with taking

our renormalized operators do not have all that the bare properties operators do. The lack of commutativity of the trace and the finite-part operation is related to a physical effect, that there is

of d.) We have discovered that

the

anomaly in the Ward identity for scale transformations see Callan (1970), Symanzik (1970b) and Brown (1980). If we were to use, say, zero-momentum subtractions (BPH or BPHZ), then the trace and the finite-part operation would commute as can be checked from our example. So it might appear that zero-momentum subtraction provides a better all-purpose definition of renormalized operators than does minimal subtraction. However, some of the properties we will prove when using minimal subtraction now disappear or become more complicated. an





152

Composite operators

example, the equations of motion which we will prove in Section 6.6 are only true if the mass terms are oversubtracted. This turns out to prevent some Ward identities from being true when the simplest renormalized operators are used, whereas they are true in their simplest form when using minimal subtraction. We will prove this in Section 6.6 also. The moral is that one cannot completely eliminate the problems. It is possible (Collins (1975b)) to construct a definition of, say, For

[d^d^/2] which uses minimal subtraction and for [(d</>)2]- This is done by writing tensors in terms

which

components. Thus we write a second-rank tensor

M^

antisymmetric term,

M?v





symmetric traceless term, and

i(M?v" MJ + i(M?v

The subtraction

MVfi





g^ld^c^/l]



of Lorentz-irreducible as

an

the

sum

of

an

invariant term:

{2ld)g^MD + (1/^VM?.

procedure is applied to each term separately. This definition

loses other properties of the renormalized products. For example, conservation of energy and momentum is a consequence of the fact that the bare 0. If we define a energy-momentum tensor 0fiV has zero divergence: <3"0^v renormalized energy-momentum tensor [0 ] by our original definition of =

minimal subtraction, then this is the

same as

redefinitions and it is conserved. But if we

just suggested, then it is

our

0^v

up to allowed

[0 ] by the procedure

not conserved.

6.6

One of

the bare

construct

motivations for

Properties

working

out the

theory of renormalization of

that in Chapter 2 we had proved equations of composite operators motion and Ward identities. These results involved unrenormalized was

composite operators. So

operators,

now

we must prove

that

the

we

have defined renormalized

equations

composite

of motion and Ward identities

expressed in terms of these renormalized operators. This is particularly important for the Ward identities, for these express the symmetry properties

of the theory. In this section

we

will derive

renormalized operators. Some



properties

will be

others will be the actual equations of motion and will be

given

for the

case

that the operators

subtraction. A typical proof of

corresponding equation is

some

true

properties of the purely technical, while Ward identities. Our proof

number of useful

are

equation for

the

renormalized by minimal

starts

by observing that the

unrenormalized

operators.

procedure applied to both sides of the unrenormalized equation, so the main problem is to find the places where the Renormalization is almost the same

renormalization procedure is not identical for the two sides.

6.6

153

Properties

It is always possible to make the theorems false by changes in the renormalization prescription. The point of using minimal subtraction is that it is

universal prescription that preserves almost all of the desirable properties. (The reason is that it amounts, roughly, to defining each counterterm by the requirement 'remove exactly the singularity'.) These a

properties

are

relations between different operators. prescription that preserves most of

The other standard renormalization these relations is the

BPH

(or BPHZ) method of zero-momentum subtraction. In fact, the proofs were first given using the BPHZ prescription (Zimmermann (1973a, b), Lowenstein (1971) and Lam (1972)). However the use

infrared

of minimal subtraction is better for gauge theories because of their singularities. The proofs were given in this case by Collins (1975b) and

Breitenlohner & Maison (1977a, b, c). All these works However, the basic ideas are simple.

Property

rather technical.

1. Linearity.

a[A] where a and b are

equation

are



b[S]



[aA



(6.6.1a)

bB\

numbers, while A and B are composite operators. This interpreted as an equation for Green's functions of the

pure

is to be

operator. That is, if X is any product of renormalized operators or composite), then

a<0| T[A]X\0)



fe<0| T[B]X\0>



(elementary

<0| T[aA + bB]X\0}

(6.6.16)

Proof. This property is almost obvious. If A and B have different numbers legs (e.g. 02 and 04), then there is no simple way of defining the right-hand side of (6.6.1) except as being the left-hand side. But if A and B have the same fields, like (dcj))2 and ¥2, then the Feynman graphs for of external

<0|7\4X|0>, (0\TBX\0y differences are

corresponding

and

<0|T(aA + bB)X\0>

are

the same; the

only in the placement of powers of momentum. The equation to

(6.6.16)

is true for the basic

graphs (i.e.,

without

counterterms). To obtain the renormalized Green's functions we apply the forest formula to each graph. The terms in the forest formula are the same, since the graphs for the three Green's functions

are

the

follows from linearity of the subtraction operators

Ty.

Comments. a

same.

Then (6.6.1)

(1) It is necessary to be pedantic about this proof because: (a) it is

prototype for less trivial cases, and (b) it fails for the case ofzero-momentum

subtractions. The,reason for the failure is that the linear. For example,

[(<30)2]

and

[(<30)2



m2</>2]

Ty operation is then not

need two extra subtractions

154

Composite operators

compared with [02], because the degree of divergence is two higher. can only have

[(Wbph



?2[02]bph



So

we

(6.6.2)

[W)2 + m202]BPH

[02] operator is oversubtracted. The coefficients a and b in (6.6.1) must be independent of d, for otherwise (2)

if the

taking a pole part is non-linear. Furthermore, we cannot use (6.6.1) to show, [(30)2]. As we saw, this equation is in fact false. The proof fails because it can only be applied to the case that we sum over a finite number of operators. It does not automatically apply to infinite for example, ^MV[3M05V0]

summations.



However, when

we

defined dimensional regularization in

Chapter 4, we saw that our vectors and tensors have to have infinitely many components. Property 2.

Differentiation

is distributive. Let A be the

composite operator

A=f\ <t>j(x) where each

is

(j)j

an

elementary field

£"- Again this equation

dA_

is to be

or one

^IUw

-I

ox

interpreted

^<0|T[^(x)]X|0>



of its derivatives. Then

(6.63a)

jfi

for Green's functions:

<0|T[^/^]X|0>.

(6.6.36)

Proof. Let p" be the momentum leaving the Green's function (6.6.3b) at the 3/cbc^gi ves a factor ipM. The point of (6.6.3)

vertex for A. Then the derivative

is

to state that we

finite-part

get the

same



results whether

statement is true for the basic subtraction

Comments.

(1)

take pM inside the

operator

Ty.

We can contract \i with an index in A. Thus

9** Wd*<ft



Q'V^M



[(#)2]

Since the overall derivative merely gives a factor of

or not we

operation. To prove the equation, it is enough to observe that this

introducing

contrast to the

extra case





i/^

^-dependence by contracting

we

have

(6-6-4)

[0 □ </>]? there is no

possibility

with g^. This is in

considered in Section 6.5.

(2) Note that derivatives ordering. Thus:

are

always implicitly taken outside of the

<0lr(<^)(x>itf>2(y)|0>



lim—<0|T^(z)</>(x)i</.2(y)|0>.

time

(6.6.5)

6.6 Properties This

155

gives the simplest Feynman rules in momentum space, with each giving a factor of momentum on the corresponding line. lowest-order graph for the Fourier transform of (6.6.5) is

derivative of a field

The

ddxdVp'x+i9,3;<O|T(05M0)(x)^2(y)|O> =

Property

?*r]?

a counterterm

-m

<6-6'6)

_2v

%)d(k2-m2)[(k-p)2-m2y

Simple equation of motion. Decompose

3.

action and

g)J(2^/i,2

(2n)dd^(p +

the action into



basic

action: ^



^b



^ct,

(6.6.7)

just like the decomposition (5.1.1) of the Lagrangian, except that 9"h includes both the free and interaction terms: 9?h Jddx(if0 + ifb). Then define =

functional derivatives with respect to renormalized fields

^(X)S^W=^)-^^^)' y*\x) We

already know



by

ttA



5<p(x)

&*

with counterterms omitted.

the unrenormalized

we

(6.6.9)

equation of motion (2.5.5)

<0|T<^(x)X|0>=i^<0|T*|0>. Now

(6A8)

(6.6.10)

wish to prove the renormalized equation

<0|T[^(x)]A-|0>=i-^-<0|rX|0>,

(6.6.11)

from which follows the operator equation

[^]=0. Comments.

(6.6.12)

(1) The functional derivatives are to be treated in a purely formal

sense.

(2)

Even

though

[<9%(x)]

is

zero as an

operator, its Green's functions

define them by taking derivatives outside (6.6.11) them inside gives equal-time commutators the time-ordering. Bringing are non-zero

because

we

(Section 2.5). (3) Signs for fermions are easiest to determine by examining the derivation in Chapter 2.

156

Composite operators

Example.

In the

<f>3 theory JS?

we



Z(d<p)2/2

m2Z (j>2/2





g0Z3l2(j)3/6,

have

5%

[^] Then









of (6.6.11)

cases

ZQ4, □ 4>



ti0Z3'2<l>\

(6.6.13)

tf-'Vgtfl

(6.6.14)

m2Z4>



m2<t>





are

<0| T[^(x)]0(y)|O>





m2K0\ T0(x)0(y)|O> -^3"d/2^<O|T[02(x)](/>(y)|O>





nx



(6.6.15)

idid\x ̄y\

<O|r[^(x)][03(y)]0(z)0(w)|O>R 3W?>(x y)<0| T[02(y)](/>(z)(/>(w)|O> =



+ +

i^>(x-z)<O|T[</>3(y)]0(w)|O> i(5(d)(x-w)<O|7[(/)3(y)]0(z)|O>.

(6.6.16)

Pra^/. general proof of (6.6.11) is rather complicated because of the arbitrary number of fields. To show the main points it is sufficient to prove one case, (6.6.15) in </>3 theory. The problem in proving the renormalized from the unrenormalized the Feynman graphs are is that equation equation A

different for the different terms. We write

m2)tf>-i0/i3-d/2</>2.

(6.6.17)

graphs are given in Figs. arising from the action. we evidently have

6.6.1 and 6.6.2. The

-(D



Examples

of low-order

counterterms

are

those

In momentum space

<0|r^o?^)|0>

?>|ryM <?(P)|o>=(p2



m2) +





(p2-m2)<0|T^(p)^)|0>.

|—+ (-O- —-) (-00+^0+0^



Fig. 6.6.1. Low-order graphs for (6.6.15) with free part of action.

(6.6.18)

157

6.6 Properties

<°m^£*jfo)io>

(o-O- +0—) (<i^+o-)

<^+









??

Fig. 6.6.2. Low-order graphs for (6.6.15) with interaction part of action. The

p2



m2 multiplying the free propagator attached

to

0(p)

cancels the

denominator of the propagator. Thus in the graph where the other end of the propagator is ¥(q) we obtain the right-hand side of (6.6.15). In all the remaining graphs the other end of the propagator is

interaction vertex, either



basic interaction

or a counterterm.

In fact

an we

obtain

<O|7\?lnl?0te)|O>

(6.6.19)

where



-(Z- l)n<t> -(m2Z-m2)(f)

-ig0Z3i2<j)2.

(6.6.20)

This is exactly what we must obtain in order that the unrenormalized equation of motion (6.6.10) is true. Notice that because (6.6.18) is finite, so is (6.6.19). We must now prove that the counterterms in

(6.6.20)

are

precisely

those

[^nt J= --£#/i3 ̄d/2[</>2]

that

are

needed

to

renormalized according to

give our

the operator standard

prescription for composite operators. Now, the renormalization prescription is precisely to add to the basic term S*nu<j> —^gfi3 ̄d/2(j)2 a series of counterterm operators whose coefficients =

poles at d 6, so as to make its Green's functions finite. But this is precisely (6.6.20). The relation between the counterterms can be seen from a comparison of Figs. 6.6.1 and 6.6.3. are pure



<^1*¥(P)\0>= +

(0"+





(oo+o—+*o-+*—)

+'" ,

Fig. 6.6.3. Renormalization of Fig. 6.6.2.

158

Composite operators

Property 4. Equation of motion times operator. With the before we have

<0|T[^^(x)]X|0>R



same

notation

i<0|T^(x)^X|0>R,

where A is any product of operators at the

same

point.

as

(6.6.21)

Hence

[A,?g=0.

(6.6.22)

Comments and examples. (1) This property is crucial

to

proving

Ward

identities. (2) All operators appearing on the right of (6.6.21) (3) In 4>3 theory, cases of (6.6.21) are

are to

be renormalized.

<0| T\_ <f> □ 4> m24>2 ^3](x)M#(z)|0> tf(x-y)<0|T4>0#(z)|0> + (y<-+z), <0|T[-^2n^-m2</>3-i^](x)[^2](y)</.(z)|0>R 2i,5(x-y)<0|T[>3]0#(z)|0> + i<5(x-zK0|T[tf>2](y)[,£2](z)|<)>R. -







(6.6.23)



(6-6.24)

Proof. The unrenormalized version of (6.6.21 )follows almost directly from the previous property in its unrenormalized version. Thisin turnfollows from the functional-integral solution of the theory, as shown in Section 2.5. We have

<0|T^(x),^(x)X|0>

-<0l^w3^)l0> i<0lT^xl°>- +

Then

we use

the fact that in dimensional

£(d)(0)=

(6A25)

regularization

|ddpl=0,

(6.6.26)

according to the results in Chapter 4. This enables us to eliminate the SA{x)/d(j)(x) term. The renormalized equation of motion (6.6.21) can be proved by generalizing the method for the previous property. It is enough to consider the example (6.6.23). Low-order graphs for the left-hand side of (6.6.23) are shown in Fig. 6.6.4. The (- □ m2) factor in ?S^O>0 cancels an attached propagator. If the other end attaches to an external field (viz., 0(y) or 0(z)), then we have a contribution to the right-hand side. If it attaches to an -

interaction, then the negative of a contribution with <9^nt is obtained, such

Fig. 6.6.5. Since these manipulations

do not change the

as

one-particle-

6.6

159

Properties

<O\T(<j>yoJ¥(k)¥(q)\0> --(q2-m2) <f

+(k2-m2) *9 p + k + t



{(p2-m2)



[(p

+ fc +

<?)2-m2] P



(k2 + q2 ̄2m2)-

Fig. 6.6.4. Low-order graphs for the left-hand side of (6.6.23), using free action.

<0\n<p^J¥(k)¥(q)\0}

+ cts

Fig. 6.6.5. Low-order graphs for the left-hand side of (6.6.23), using interaction part of action. (ir)reducibility structure, renormalization can be performed without

changing

the result.

Since the

SfQ^ terms

(1) Since □</>



need renormalization two subtleties arise:

#MV<3M</)SV¥,

the

ambiguity about

the

relevant. To preserve the derivation it must renormalization is performed.

placement of #MV is

come

before the

(2) In the BPHZ scheme m2</>2 must be oversubtracted otherwise we cannot use

linearity

to combine



</>□</>

and



m202.

The case of (6.6.24) involves two further subtleties illustrated In the first graph (a) the (-

by Fig.

6.6.6.

m2) multiplies the line coming back to a

□ </>(x) factor. This term gives zero after use of (6.6.26). The second graph (b) has two —

separated by a line. In the basic graph, the q2 m2 factor cancels the propagator to give graph (c), which has a different reducibility structure. 1PI loops



The first two counterterm graphs give the obvious counterterm graphs in Fig. 6.6.6(d). These correspond to the first two counterterms in Fig. 6.6.6(b). But the last counterterm in (b) has two vertices while the corresponding graph vertex. Their operator structure is different: graph (e) is another counterterm to renormalize the x -? y singularity of 04(x) </>2(y). Even so the

(e) has one

160

Composite operators

(k2-m2)



'Q

<■>

(b)

(92-m2)



/y

{xC^O-z W—.

'g^

(0

*^>—

1-



x—?

^ x

(e)

x, y





Fig. 6.6.6. Some graphs for (6.6.24).

two counterterm

graphs

must be

equal. They

are

both



single

free

propagator times a pole part coefficient times a polynomial in momentum times the same power of the unit of mass. They both make the complete Green's function finite. The general proof is rather tedious and

can

be found in Collins (1975b).

scheme, but also works for the BPH(Z) scheme. In the original BPHZ proof, by Lam (1972), of (6.6.21), there is no treatment of this complication, that the counterterms for the two sides are not in manifest correspondence i.e., that the forests are different. This proof was given for the minimal subtraction



Property 5. Ward identities: We will use the notation of Section 2.6 for transformations under potential symmetry operations. Let

Qj^Qj + Sifrj be

an

infinitesimal transformation of the fields under which the basic

Lagrangian transforms

as

-^basic

 ̄ ̄*

^basic



^b



^M *

b*

restricting our attention to the transformations generated by one particular generator of a group. Thus, as compared to Section 2.6, we now drop the index V, which labelled the generators. The subscript 'b' on Ab and Yb indicates that we are considering transformations on the basic Lagrangian (i.e., without counterterms). In the equations below, we will add in the We

are

counterterms by use of our standard renormalization scheme. Note also that in setting up the renormalized Green's functions, in Section 2.8, we defined a free Lagrangian if 0 and a basic interaction Lagrangian jSf b. We now work with their sum:

i?basic i?0 + jS?b. We proved earlier the unrenormalized Ward identity (2.7.6). The =

161

6.6 Properties renormalized Ward

identity

is

^r<o|TC/SW]A'|o>R =

<0| 7[Ab(x)]X|0>R



i<0|T5b4>(x)J^-\0}R.

(6.6.27)

We will have to prove it. From it follows

3?[/6] and

by integration 0

Of course







all

over



d4x<0| T[Ab(x)]X|0>R

0 for

symmetry. The



(6.6.28)

[Ab]

x:



id(0\TX\0)R.

(6.6.29)

current is

^^?J-Cn]- Proof. one

In

defining Yg and Ab,

we

(6-6.30)

have used the basic

with the counterterms omitted. This is because

Lagrangian, i.e., the the operation

we use

symbolized by square brackets to generate the counterterms. The proof of (6.6.27)-(6.6.30) follows the usual proof of Noether's theorem, but using the previously proved properties to write it directly in terms of renormalized operators. First -

we use

[Ab]



[<yg]





From this the Ward

(6.6.22). We exchanged

^L/b]



and

linearity

[djjb]-



distributivity

to obtain

d^l^^js^ J -

yg



[a, yg



<5i?basic]

-I[^^J. j

(6.6.31)

identity (6.6.27) follows by

the

the order of renormalization and

This is

permitted



see our

equation

tracing

of motion

over \i to

write

remarks below (6.6.4).

Comments (1) The theorem appears to give an unrestricted proof of the renormalized Ward identities. This appearance is false, since there are

symmetries that, can and often do have anomalous breaking-see Chapter 13. Such symmetries are dilatation and conformal symmetries and

162

Composite operators

chiral and supersymmetries. The potential for such anomalies can be seen by computing Ab in a regulated theory; it contains a non-zero coefficient which

vanishes

as

d->4. Such is the

chiral

symmetries (where

in the

proof

for conformal transformations and for

case

the transformations involve y5

eKXflv explicitly). Minimal subtraction is then not easily applicable and the properties we used are

or

false.

(2) Corresponding problems appear with any other regulator and with any other renormalization scheme (Piguet & Rouet (1981)). Minimal subtraction confines the problems to cases with anomalies. Property 6. Non-renormalization of current: Consider

an

internal

exact

the symmetry that

symmetry (such Compute the corresponding as

gives electric charge conservation). unrenormalized current;" from the complete

Lagrangian. Now/ contains counterterms Lagrangian. We will now prove that these /



Comments This theorem does not

derived from the counterterm

make / finite and that

(6.6.32)

[)£]?

apply

to

space-time symmetries



see

Callan, Coleman & Jackiw (1970), Freedman, Muzinich & Weinberg (1974), Collins (1976), Brown & Collins (1980) and Joglekar (1976) for the case of the energy-momentum tensor. It also cannot be extended to the

case

of a

non-

conserved current unless the

breaking term has dimension below that of S£ (Symanzik (1970a)). Furthermore, the proof does not apply directly if the transformation 5(j>}\s non-linear in </> It also needs generalization for gauge ?.

theories.

Proof.

Both/

and

Q/g]

consist of the basic current;g plus

some

minimal

subtraction counterterms. The difference

(6.6.33)

z?=r-[Jtt

series of pure pole terms, and we wish to prove it vanishes. Each term has dimension 3 or less (at d 4), since the currents have dimension 3. is





Now both

;£ and [jg] satisfy the

same

Ward identity,

so

dfl(0\Tefl(x)X\0)=0. Thus

5M£M



0, without

use

of equations of motion; any need to

(6.6.34) use

the

equations of motion would give a non-zero right-hand side to (6.6.34). In the absence of gauge fields, it is impossible to construct such a term. The theorem is thus proyed. In the presence of gauge fields, such terms do exist. For example, in

6.7

Differentiation with respect

quantum electrodynamics

proportional

current

we

<3V.PV,

to

to

parameters in ££

163

the electromagnetic field-strength tensor dilAv might have ^oceK^v3K(^^^),

have

a counterterm to

where ¥ ^ is the



With non-abelian gauge fields one but the presence of the £kA/iv indicates a chiral symmetry, which in any case needs special treatment. Moreover, in a non-abelian theory, we must also

d^A^.

take account of the constraints

subject

we

imposed by

The energy-momentum tensor also has

(djA



gauge invariance, which is



will not treat until Chapter 12.

g^O)^2



see

counterterms, like

possible

the references quoted above.

6.7 Differentiation with respect to parameters in S£ Consider Green's function derived from the bare classical ¥£

by using



the functional

GN

Z(dA)2/2

m2BA2/2





Lagrangian

gBA*/4!

integral: =

(0\T<t>(x1)...<t>(xN)\0}

J<

j[dX]A{x1)...A(xN)&* Differentiating with respect

to gB

(6.7.1)

gives

[dA]A(Xl)...A{xJ^\d*yA4(y)y^

80^

[cUJe"-2, J

|J[d^]A(x1)..M(xw)e,^}||[d^]^)JdV40')ei'jr} [dA]e"^ U)





d4KO|T0(xJ...0(xN)f^-<O|^^|O>)|O>.

(6.7.2)

Similar formulae hold for differentiation with respect to the other parameters Z

or

m\.

The renormalized

equivalents of those equations are also useful.

One

use

164

Composite operators

will be to show, within can

perturbation theory, that terms quadratic in the fields Lagrangians without affecting the

be shifted between free and interaction

Green's functions. First consider differentiation with respect to the renormalized coupling, g. We differentiate the renormalized Green's function

(6.7.3)

rL

Applied

to each basic

graph



T in this formula, the differentiation





TT

nocounterterms"

Renormalization of (6.7.4) produces those in

a set

of counterterms

just gives (6.7.4)

isomorphic

to

(6.7.3). So

dg°N-

ddy<O|T{[0(y)4]-<O|[(/>(y)4]|O>}(/>(x1)...(/>(xN)|O>.

4!

(6.7.5) The subtraction of the because no

vacuum

connected to Suppose

some

we



z(d<f>)2/2

and let the free and interaction

<e0

<eh

m\



m\

[04] comes about </>4 vertex in dGN/dg is

of

let the basic Lagrangian be ^basic

with

expectation value

vacuum

bubbles are used in GN. Thus each external line.









m2<t>2/2

Lagrangians

(d<j>)2/2 m2<j)2/2 (z l)(d(f))2/2 (m2



gtpH!





m2. Notice that



we



(6.7.6)

be

m2)(t)2/2



have allowed the

gif/A!,

(6.7.7a) (6.7.7b)

(d(j))2 term to have an

arbitrary coefficient. We choose to put some of the terms quadratic in <j> into so that we can derive an equation for dGN/dz or dGN/dm2 like (6.7.5). Then we will show we can move the quadratic terms to the 5£0 without changing the Green's functions. the interaction,

Differentiation with respect to

JzGn



z or

m2 gives

d"y<0|r{[(o</))2]-<0|[(a^)2]|0>}^(x1)...</)(xJV)|0>) (6.7.8)

_d_

dm2

GJV



^j'd''y<O|r{[?/)();)2]-<O|[0(y)2]|O>}^(x1)..^(xN)|O>. '



(6.7.9)

Differentiation with respect

6.7

However,

Lagrangian.

we

may want to put all of

to parameters in 5£

(<7</>)2z/2

m2(j)2/2



165

into the free

In this case the free propagator is

i/(zp2-m2). (6.7.8) and (6.7.9) remain valid. In this case d/dm2 applied to an unrenormalized graph

We wish to prove that

terms

T gives

a sum over

in which each propagator is differentiated i

d dm2

zp2



m2

zp*

(6.7.10)



This gives us the same result for the unrenormalized graphs as the right-hand side of (6.7.9). Next,

we

differentiate

a counterterm

propagator is differentiated, so that the



graph Cy(r).

Either



i in (6.7.9) gives the basic vertex for

[<t>2(y)]>or a counterterm C(yx) is differentiated. In this second case there is a counterterm graph Cy(dT/dm2) with a term cyjdm2

also

Now

C{dyjdm2)



(6.7.11)

dC(yi)/dm2

in the minimal subtraction scheme. (The reason is that both are defined to be pure poles times p to a power the same for both graphs.) We thus obtain all -

the counterterm graph for the right-hand side of (6.7.9). Similarly (6.7.8) is true if z(d(j>)2 is all in the free Lagrangian We thus see that, for any renormalized parameter



in the

if0. Lagrangian S£,

have

we

-GN CA





\ddy(0\T

*^ ̄

basic

[^

(y)

<J> basic

M O)U(x,)...0(%)IO>

da

(6.7.12) From this result we can see that the Lagrangian (6.7.6) is equivalent to the one with unit kinetic term

£'h

by





(84>')2/2



m'2<t>'2/2

g'4>'4/4!



(6.7.13)

scaling of the field, with 4> m2 g

The proof is to write (6.7.6)



(6.7.14a) (6.7.14b) (6.7.14c)

z-v24>\



zm'2,



z2g'.

as

^baS,c= z(8<f>)2/2



m'2z<t>2/2



g'z2cf>V4!.

(6.7.15)

166

Composite operators

Then differentiation of a renormalized Green's function GN of 0 with respect to z gives

zdGN/dz



zdGN/dz\fixed^g +



m2dGN/dm2 \fixcdZjg + 2gdGN/dg\f ixcdz.m

d4y<0| T{[z(<l<f))2/2







vacuum

d4y(<O|T{[0f5^^]

m2</>2/2

2#4/4!]

expectation value} 0(x1)...0(xw)|O>

vacuum





expectation value}



(6.7.16)

x0(x1)...0(xN)|O?. We now use the

equation of motion (6.6.21) with







to

give (6.7.17)

zdGN/dz=-NGN/2. From this it follows that

i.e.,

<O|T</>(x1)...0(xN)|O>

z-^2<O|T0'(x1)...0'(xN)|O>,



(6.7.18)

exactly as we would expect. The proof is non-trivial only because we are shifting terms between the free and interaction Lagrangians. Thus we must ensure

that counterterms do not go astray.

6.8 Relation of renormalizations of

<f>2

and m2

g2 the renormalization factor Za for <f>2 in (6.2.13) and m\jm2. This relation is as we will now prove. (We are now back in 03 theory at

Observe that at order

(6.2.11) is the inverse of the renormalization factor true to all orders, d

6.) We use the renormalized formula



dm2

GN





I d4y< 0| T< m2^-^;



vacuum

expectation value

> x

X(j)(xl)...(j)(xN)\0} -1

2 But

we



-mz

d4KO|r{[02]-<o|[02]|o>}0(x1)...0(xN)|o>.

also have

(6.8.1) if



{d<j>0)2/2



m2Zm02/2



0o^/6,

(6.8.2)

6,8 Relation of renormalizations of <f>2 and m2 so

167

that

m2d£?/dm2





m2Zm<^/2.

(6.8.3)

Hence 8 dm2





ddy<O|T{^-<O|02|O>}0(x1)...^(xN)|O>. (6.8.4)

Therefore dVM=

d>o^,

(6.8.5)

from which the desired result follows. Generalizations of this method can be found in Brown (1980) and Brown & Collins (1980).

7 Renormalization group

As

we saw

in

Chapter 3,

the renormalization

arbitrariness: the counterterm for



graph

procedure

has considerable

must cancel its

divergence

but

may contain any amount of finite part. A rule for choosing the value of the counterterm we called a renormalization prescription. In one-loop order it

examples that a change in renormalization prescription by a change in the finite, renormalized couplings each to corresponding divergence. Thus a change in renormalization prescription does not change the theory but only the parametrization by was

clear from the

can

be cancelled

renormalized coupling and mass. What is not so easy is to see that this property is true to all orders. This we will show in Section 7.1. The

theory under such transformations is called renormalization-group (RG) invariance. A particularly useful type of change of renormalization prescription is to change the renormalization mass \i. Infinitesimal changes are conveniently described by a differential equation, called the renormalization-group invariance

of

the

equation, which is derived in Section 7.3. This leads to the concept of the effective momentum-dependent coupling. This concept is very useful in calculations of high-energy behavior, as explained in Section 7.4. The

renormalization-group equation are called the renormalization-group coefficients and are important properties of a theory. Various developments of the formalism occupy the remaining

coefficients

in

the

sections. The renormalization group was first discussed by Stueckelberg & Petermann (1953) and by Gell-Mann & Low (1954). Very similar ideas are

applied in statistical physics (Wilson & Kogut (1974)). Many important applications arise because of the asymptotic freedom of QCD. Results of calculations of renormalization-group coefficients can be

recent

found in many places: (1) Gross (1976) lists many one-loop results for theories with scalars and fermions and up to two loops for gauge theories with only fermions.

(2) Cheng, Eichten

& Li (1974) give the ^-function for able theory to one-loop order.

168



general renormaliz-

Change of renormalization prescription

7.1

169

(3) Tarasov, Vladimirov& Zharkov (1980) compute renormalization- group coefficients to three-loop order in gauge theories with fermions

using minimal subtraction. (4) Vladimirov, Kazakov & Tarasov (1979) compute to four-loop order in

</>4 theory. (5) Chetyrkin, Kataev

& Tkachov (1981) and Chetyrkin & Tkachov (1981) compute the anomalous dimension in ¥4 theory at five-loop order. (6) Tkachov (1981) summarizes the methods used for the above calculations. (7) Caswell

& Zanon (1981)

theories at

three-loop 7.1

Change

techniques will

we

alternative, but

all,

we can

will describe

03 theory

prescription

valid for any theory. However, to be

are

in six space-time dimensions. There

are

three

equivalent, forms in which to write the Lagrangian. First of

write it in terms of the bare field </>0: if

(As before,

of renormalization

mainly work with the theory we have been using as a source

specific, of examples, the we

calculations in supersymmetric

Change of parametrization

7.1.1

The

perform

order.

we



(d<j>0)2/2



m20<f>20/2



g0<t>30/6.

(7.1.1a)

ignore the term linear in </>.) The importance of this form is 0O is invariant under change of renormalization

that the bare field

normalization is determined, because it satisfies canonical commutation relations. equal-time

prescription: its When

we

renormalize the theory,

the renormalized field </>



Z ̄

Lagrangian is Se





1/20o.

Z(dcf>)2/2 Z(d<t>)2/2





we

obtain finite Green's functions of

In terms of the renormalized

m20Zcf>2/2



field, the

g0Z^2^/6

m2B(j>2/2 -gB<j>'l6.

(7.1.16)

This is the second of the three forms. In the perturbative theory of renormalization, we wrote the Lagrangian as the sum of a free Lagrangian, a basic interaction Lagrangian, and a counterterm

Lagrangian: =

ifbasic



(7.1.1c)

i?ct.

This is the third form of the Lagrangian. Here,

we

have chosen to define



170 basic

Renormalization group

Lagrangian ^basic



(d<P)2/2



m2<t>2/2



(7.1.2)

fi3-d'2g<p3/6,

m and g are the renormalized mass and coupling. Since we will mostly minimal subtraction, it is sensible to define g to be dimensionless, and therefore to introduce the unit of mass \i. The counterterm Lagrangian is

where use

5£cx



bZ(d(f))2/2



bm2(j)2/2



(7.1.3)

bg^/6,

bm2, and bg are computed as definite functions of and with the aid of some renormalization prescription. g, \i To be concrete, let us use minimal subtraction, so that and the counterterms bZ, m

bZ=



(7.1 Aa)

(6-rf)-V7.(0,m,/i), 00

bm2

Sg



m2



^--/2



(6





d)-Jbj{g9m9ri9

(6

_

7=1

(7.1.46)

d)-jaj{g^m^y

(7.1.4c)

We saw, in Section 5.8, that in fact the coefficients at, bi9 c{ are independent m and \i\ they are functions of the dimensionless coupling g only.

of

However

we

will not

use

this fact at the moment.

The three forms of the Lagrangian listed in (7.1.1) are equivalent if we use any of them in the functional integral, then the same Green's functions -

will result. The coefficients Z, 6 with #,m, and \i fixed. The

m2,, and g0 will be singular when d approaches singularities

will be just

such

as to

give

finite

Green's functions of 0 at d 6. The parametrization of the Green's functions by g, m, and \i is rather arbitrary. Suppose that we change variables to g\ m\ and \i, which are =

given functions of g, m, and \i. These functions may even depend on the regulator, d, provided that the change of variable remains non-singular at d 6. Then we get the same theory, for the collection of Green's functions is GN unchanged. It is just the numerical values of the renormalized mass and coupling and of the unit of mass that have changed. We may even change the scale of the renormalized field by writing </>' (0, where ( is finite. The Green's functions are now different:

some





G'N



ZNGN.

physical observable like the S-matrix. For example, consider an S-matrix element involving N particles. It is obtained from GN by (1) dividing out an external propagator G2(Pi) for But observe that the value of (is irrelevant for



7.1

Change of renormalization prescription

111

each external line and (2) multiplying by z1'2 for each external line, where iz is the residue of the pole of Thus S= =

G2(p). Finally we

let the momenta pt go on-shell.

\gnzn'2/y[G2} I

lim

p^m*h I



i=l

\im{G'N{^2flY\G'2).

(7.1.5)

But

G2 so

the



(7.1.6)

C2G2,

particle pole is at the same position in G2 as in G2 and the new residue

is z'



t2z.

(7.1.7)

Hence

S=\im{G'N(z'V2)N/Y\G'2}. and

the S-matrix is invariant,

as

(7.1.8)

claimed.

7.1.2 Renormalization-prescription dependence

parameters. So there two-parameter collection of physical theories obtained

In the bare Lagrangian (7.1.1a), there

are

only

two

should be only a from it. As we have just seen, the freedom to vary the scale of the renormalized field

cf> in

introduce a third real

the second form of ££, viz.

(7.1.16), does

not

parameter into the physics.

Unfortunately, we appear to have introduced a large and indefinite number of parameters by having to choose one out of the infinitely many possible renormalization prescriptions. One might suppose that in different renormalization prescriptions, the singular behavior of m0 and g0 as d-^6 could be different in such a way that one picks up different phases of the theory. In fact, this is not so. We will show that a change of renormalization prescription is one of the reparametrizations discussed in the previous subsection 7.1.1. This is the property group invariance. Even within



we

have defined

as

single renormalization prescription, we

renormalization-

have introduced

third parameter, the unit of mass, p. Notice that the basic

Lagrangian



does

and g separately, but only on the combination p3  ̄dl2g\ but notice also that this property is not true for, say, the renormalized Green's

not depend

on p

functions at one-loop order. However, renormalization



change of p is

in effect a

prescription. Indeed, prescription the requirement that p and

the renormalization

we

could include in

our m

change

in

definition of have



fixed

172

Renormalization group

ratio;

we

would still have two free parameters. A change in this ratio is then

change in renormalization prescription. Our proof of renormalization- group invariance will, in fact, only explicitly cover the case of a change in p. The more general case will be essentially the same. We will prove that if we change p to \i', then the physics is unchanged, provided that we choose suitable new values, g' and m', for the renormalized coupling and mass. That is, an S-matrix element S(g9m,p) satisfies a

S(g9m9fi) The bare The

S(g'9m\fif).

and coupling m0 and g0 are renormalized field 0' with the

mass

new

parameters is not the

where (is functions





same as

similarly invariant. new values (g\m\p)

of the

with the old values but is related by

finite function of g9 m, \i and p. Hence the renormalized Green's

satisfy GN(p1,-..,pN;g,m9p)





It will be

convenient, in

our

</>',

proof,

(7.1.9)

compute Green's functions of the

to

value p! of the unit of mass. Now, in terms the Lagrangian is

original field 0, but with the of the new field

t-NGN(pl9...9pN;g'9m,9fi') CNG'N.

<£



new

Z'd(f)'2/2



m'B2<j),2/2



g'^/6.

(7.1.10)

Here we write Z'



m'2



9B



Z{g'9m'lp'9d)9

(7.1.11a)

m2B{g'9m'9p'9d)9

(7.1.11b)

9B(g'^'>d).

(7.1.11c)

are the same functions of the new renormalized parameters m\ g\ and the original bare parameters were of the old renormalized parameters \x and m, g, p. To get the Green's functions of the original field 0 but with the new value of the unit of mass, we substitute £0 for 0' to obtain

These as

S£



Z(0',my//,rf)C2#72



m2B{g\m'WM2<b2

-gB(g\m\fi\d)C3(t>3/6. 7.1.3 Low-order Let

(7.1.12)

examples

remind ourselves how the changes in g and m are obtained at one- order., We must start at tree approximation, where we ignore all

us

loop

counterterms. In order that the two formulae

(7.1.16) and (7.1.12) for

5£ be

Change of renormaliza tion p rescript ion

7.1 the same in tree

approximation when 0'



0n.w

m'



m,







we

change

\i to

p,\

17 3

we must

write

(/V/O3"d/20,

1. (All tree

approximation.)

(7.1.13)

6 we have g' g to lowest order. We distinguish g' and defined to be exactly the value of g' in tree approximation. In higher-order calculations, g' gets corrections, but #ncw will be defined to be Note that at d

0new: 0new





ls

(\il\i? ̄dllg always. To treat

higher-order corrections, 5£



write the

we

Lagrangian (7.1.12)

(8<t>)2/2 m2<t>2/2 ii*-*2[g(iilitf-'l2Wl6 + d'Z(d(f))2/2 d'm2(f)2/2 (5Vy</>76. -

as







(7.1.14)

Our strategy will be to express the counterterms in (7.1.14) as minimal subtraction counterterms plus some new finite pieces. The finite pieces will

accomplish the change of parametrization. First we consider the one-loop graph for the

self-energy, Fig.

3.1.1. The

*)r/2-2(4V)3-'/2-

(7.1.15)

unrenormalized value is

r? This



'g2?-)20 ̄3rf/2) Jf 12S7T

is

invariant

g(li/fi')3 ̄d/2).

If

dx

1>2



P2**1





under

we use

the

unit of

(/*, m, #)-?(/*', then the counterterm is

transformation

mass \i,

m,

C(ri)=-pole(ri)

Next

and gnew counterterm changes to we

use

\i



(n/v')3 ̄d/2g

instead of \i and g. The

Notice that we define CfTJ to be the negative of the pole part of Yx, with the ^-dependence of #new ignored. That is, we consider the function Ti ri(P>0new>m>AO and extract its pole at d 6 with gncw fixed. This =



prescription ensures that we may later replace gncw by its value at d without changing the renormalized value of the graph. Since both C^) and C'iFJ cancel the divergence, their difference



finite. We may therefore obtain the

same



is

value for the graph plus

174

Renormalization group

by letting the counterterm coefficients and b'm2, each be a sum of two terms: counterterm

in

(7.1.14), namely b'Z

(7.1.18)

3S4n3(d-6)

384tt

64w3(d-6)

64tt3



(7.1.19)

d-6

The first term in each equation is the minimal subtraction counterterm for the new coupling, while the second term is finite as d -? 6. Using the formula (7.1.12) for the Lagrangian,

C2





we

may regroup these terms to

/new

l +

384tt3

d-6

92

1 H +

384rc3



64k3

Hence

(7.1.20)

d-6



(7.1.21)

d-6

obtain the value of m':

we

m'2

1 +

m2



1 +

?m

Also at d

give





we

3|_

Sg2 384tt:



d-6

■InMO



Ofo*)

0te4) as

d->6.

(7.1.22)

have C(d

We can

Sg2

384tt

apply the



6)

same

value is the factor in curly

l +



768w:

?In

(fi/fi') + 0(g4).

(7.1.23)

procedure to the vertex graph, Fig. 3.6.1, whose brackets in (5.3.5). The counterterm b'g is written

as 3

.3

?'3"d/2

?newr



64n3(d



6)

,/3-d/2 p'3-d/2r(^)fe-d-i 64jr rtrc3 d-6

onew

|_

with the result that

C3g'n'3-d/2



2 ?new

g^n'3-dl2U + 64jr3 r^2 d-6

(7.1.24)

7.1

Change of renormalization prescription

175

It follows that 9'



9(n/fi')r\3-d/2 1 + L

Our

6-d"

■{Ml111

l1+2567r3L

3^2

0(g4)\

M(p/p') +

2567T3



d-6

O(04) (7.1.26)

asd->6.



strategy for understanding the effect of a change in renormalization is to absorb the difference into

prescription

counterterm will itself generate



finite counterterm. The

divergent bigger graph. Finally we reorganize the Lagrangian by putting all the finite counterterms into the basic Lagrangian. Then we see that the change in renormalization prescription is exactly compensated by a change in the counterterms when we insert it

into a

parameters of the theory. What happens when we go to

higher order? An example is given by the two-loop self-energy graph of Fig. 7.1.1(a). Graphs (ft), (c), and (d) renormal- ize its subdivergences and its overall divergence. We write the renormalized graph with the original value of p as

£i(p,flf,w,/i)

Za



(We thatZfc Zc.) we replace p by p and g by #new. exactly as at (7.1.24), so that: used the fact

Zb(p, g, w, p)







2Z6



(7.1.27)

Zd.

The unrenormalized

Zfc(p, #new, m, p')

graph

is unchanged if



But the vertex counterterm is treated

[Zfc(p, g, m, p)

S6(p, #new, m, p')\



(7.1.28) The first term contains the counterterm for

by compensates

minimal subtraction with unit of

one

of the

mass p.

the difference. It has the counterterm

(a)

subgraphs, computed exactly replaced by the finite

The second term

(W

-X- (d) (c) Fig. 7.1.1. Self-energy graph with counterterm graphs.

176

Renormalization group

part in (7.1.24), viz. Ag

We

now

£1



S?(P?

^neWJ ?>

64tt3

/O

j2ri

we wrote out

Zj with

a3 u'3-di2 y^^

(7.1.29)

]'

d-6

write



Here



unit of

2Zb(p, #new, m, //)





Zd(p, #new, m, //)

^r^j+ Zd(A^w,^)-Zd(p,^new,w,^)>. (7.1.30) the first three terms

mass \jl

2rlA#/(#new//3 ̄d/2)



as

the minimal renormalization of

Z1 is finite. The term self-energy graph with one of its divergence which can be cancelled by a

The remainder is finite, since

is the one-loop

couplings replaced by Ag.

It has a

minimal counterterm: C Hence the term in

2riAg g^ ̄dl2

-i/>2). i^=Hm2 6)



647r3(rf

(7.1.31)



curly brackets is + 2C



[Sd(p,g, m, /i)



Sd(p, gnew, m, /i')



2C]. (7.1.32)

The second term is finite, since there

are no

remaining divergences.

It is of

the form

i( and

so

gives



Am2



AZp2),

rise to another finite contribution to m'2 and to £.

7.2 Proof of RG invariance To show to all orders of perturbation theory that g\ m' and ( can be chosen so that (7.1.9) holds, we generalize from our treatment of the examples. We write

(3g'n'3-<"2=gn3-<"2 + Ag, C2 1 + AC2, C2m'2 m2 + Am2. =



(7.2.1)

The original version of the theory has counterterms computed with unit of mass fi: &

We

now

change



&bisic

to unit of mass

\i'



&ct(n)-

and wish to show that

(7-2-2) identically

the

Proof of RG

7.2

same

Green's functions and total

changes of the form (7.2.1). The Lagrangian is written cp

with the basic

Lagrangian

Phasic

(As before,



define #new of the form

Lagrangian'

^c



cp

**^

the

are

obtained if

make

we

basic

(7.2.3)

-L.cp-L.cpf ' ""^ ' °^ c

same as

cv

before, but written

as

(7.2.4) m2?/?2/2 -gnev,n'3-"^/6. (/i///)3 ̄d/2#.) The term ifc is a 'compensating

(#)72

we

Lagrangian

177

in the form



°^

invariance





A(2d(f)2/2



Am202/2



(7.2.5)

Ag<j)3/6.

Lagrangian £f'ct in (7.2.3) is computed using minimal mass \i'. We may later reorganize (7.2.3) so that the basic Lagrangian is taken as if basic + J2?c. We may drop the ^-dependence of the finite counterterms Ag, AC2, and Am2, since renormalized Green's functions are finite functions of The counterterm

subtraction

with unit of

renormalized

quantities. Finally we may rescale the fields to give (7.1.10). proof is most easily given with the form (7.2.3). Each of the finite counterterms is computed as a sum of terms, one for each IPI graph contributing to the relevant Green's function: But



A0=£Ar(0),etc.

(7.2.6)



Particular

cases are

given by

^.iC2



m2



A7.i.i(C2P2 where the label

on

m2)

examples

c/new

-(n'lnf-<-\ ̄ d-6

?20new V//x)6-"-i 64n3



in Section 7.1. Thus

384tt3

&3.6.l9=H -

the

d-6

'3 -J/2 ?new



(n'/n)6-"-i

64n2

d-6

i[I,d(p,g,m,n)



Zd(p, gntw, m, ft')

A indicates the number of the



2C],

figure depicting the

basic

graph. The

graph -

general proof

is by induction

on

the size of

G contributing to some Green's function.

gn3 ̄dl2(j>3/6 we renormalize it

with unit of



graph.

Consider



Using the basic interaction

mass ft

using

the method of

Section 5.11. The renormalized value of G is then R(G)



G+

£ Cy(G).

(7.2.7)

Renormalization group

178 Here the

sum

is

over

subgraphs of G that consist of one or more disjoint

1PI

graphs, and Cy(G) denotes the replacement of each of these 1PI graphs by a counterterm with unit of mass \i. Similarly we can renormalize G with a different

unit

of

mass

but

\i

with

the

basic

same

interaction

-^^-"'Wtoget R'(G) Here,

we use

the

prime



£ C'y(G).

G+

(7.2.8)

to denote use of the unit of mass \i

counterterms. These counterterms will be used to

compensating Lagrangian j^c in (7.2.3). The and

we

will arrange them

so

new

the original value R(G). The aim will be every 1PI graph y that is

Ay's

so

graphs

that when they

a vertex or

are

instead of \i.

graphs containing finite

We will now derive a series of new basic

generate the

will need renormalization, added to

R'(G\ we get back

to have a finite counterterm

self-energy graph.

Ay for

We will arrange the

that

R(G)



R'

G+X AV(G) + A(G)

(7.2.9)

>' £ G

Here the sum over y is over products of 1 PI graphs. The overall counterterm for a graph is computed using minimal subtraction, but with the d- dependence of #new and of the finite counterterms Ay ignored (see our remarks below (7.1.17)). That is, the counterterms in R'

at d



6 with their coefficients

vertexsubgraph,

are a

power series in #new. In the case of a there is also the usual factor jx'2 ̄ d/2. The finite subtraction

A(G) for the complete graph in (7.2.9) is only non-zero if G is a

series of poles

being

a 1 PI vertex or

self-energy graph. We will prove the

following

relation between counterterms for



1PI

graph C(y)



C'(y+ X AJLy))

The above equations (7.2.9) and (7.2.10)



(7.2.10)



trivially true for tree graphs, A(tree graph) 0. So let us assume they are true for all graphs smaller than a given graph G, with all the Ay's finite. There are two cases: (a) G is not overall divergent; then we must prove (7.2.9) with no counterterm A(G). (b) G is 1PI and overall divergent; then we use (7.2.10) with y replaced by G to define A(G). We must prove A(G) finite and prove (7.2.9). We must also prove A(G) is polynomial in the external momenta of G, with degree equal to the degree of divergence

where

no counterterms are

needed, for

are

we can set



7.2

of G;

we

will

assume

Proof of RG

inductively

invariance

179

that this is true with G

replaced by

any

graph. Consider the

smaller

terms in (7.2.7). For each we identify a contribution in G is the same as G in R\G). Next let y be a 1PI subgraph of The term (7.2.9). G (G itself being excluded). Decompose Cy(G) by (7.2.10):

(1) (2) (3)

The

C\y)

The term

The term

term occurs in

A(y)

occurs as

Cy(A#(y))

R\G)

as

Ay(G).

occurs

C'y{G).

as a counterterm

in the renormalization

R'(AS(G)). divergent, these exhaust all of the terms in R(G) and on of side right-hand (7.2.9). But if G is 1 PI and overall divergent then there

If G is not overall the

remains C(G) in (7.2.7) and the terms C

G+

Z A,(G)



A(G)

6¥G

in (7.2.9). We are therefore forced to define A(G) prove

A(G) finite. This is

now

easy, since

R(G\ R'(G\ and R'\ are

all finite, while

we

by (7.2.10), and it remains to

£ A,(G)

(7.2.11)

have K'(A(G))



A(G).

Moreover all the terms in (7.2.10), except possibly A(G), are ordinary minimal-subtraction counterterms. So they are polynomial in the external momenta of G, with degree equal to the degree of divergence of G. So A(G) is

polynomial, of the same degree. (Note that the replacement of a subgraph y by Ay does not change the overall degree of divergence of any graph T satisfying ygrgG. This is because of our inductive assumption on the polynomial degree of Ay.) The theorem is also true if R and R' stand not for renormalization with different units of mass, but for any two renormalization

important

that the

^-dependence

prescriptions.

It is

of Ay is taken outside of the extraction of

pole parts when computing a counterterm like Cr(Ay). This can be seen from the example of Fig. 7.1.1, at (7.1.31). The primed counterterms must be

particular function of #new, //, m', and d. The fact that g' itself is a function of other variables is ignored. Equation (7.2.10) expresses the counterterm C(y) for a graph y (with unit a

of mass pt) in and



terjns of renormalization counterterms with unit of mass \i

finite counterterm

A(y). We can use the finite counterterms to generate

180

Renormalization group

the compensating

Lagrangian i?c in (7.2.3), and the primed counterterms to new Lagrangian is the same as the original one (with unit of mass p\ considered as a function of 0 and 3</>. If we set (j)0 Z1/2</>, with Z Z(g9 m, p), then we can deduce that the bare parameters m0 and g0 are renormalization-group invariant:

generate

<g'cV

It should be evident that the





m0(g9m9p) Qo(9> m> I*) 7.3





m0(g'9m\p')9 9o(9'> m'> V'\

(7.2.12)

Renormalization-group equation

change in the unit of mass \i mass does not change the accompanied the while Green's functions satisfy theory, We saw in Sections 7.1 and 7.2 that



by suitable changes in coupling and

GN(x19...9xN'9g9m9p)



C ̄NGN(xl9...,xN9g'9m'9p).

(7.3.1)

m29 p9 and p\ If the ratio it not is is sufficient to lowest-order use large, p/p perturbation theory, since, for example, in (7.1.26) the coefficient of g2 may be large. An important device is to consider a large change in p as being made up of a sequence of very small changes, so that g\ m'2, and ( are obtained as We wish to compute g'9 m'2 and ( as functions of g,

solutions of differential equations. This is the subject to which we now turn. The consequence of our work in Sections 7.1 and 7.2 is that for a given

physical theory, we have for each value of p a definite value of the coupling mass m(p). These are called the effective (or running) coupling and

g(p) and

will derive differential equations for g(p) and m(/i). The easiest way to derive the results is to look at the Green's functions and the Lagrangian expressed in terms of the bare field 0O. The important

mass. We

point is that the Green's functions of 0O are invariant under our change of parametrization (fi9g9m)-+(n'9g'9m'). (This is because the mass and coupling in the bare Lagrangian are invariant.) 7.3.1

Renormalization-group coefficients

Physical quantities like the S-matrix are invariant under the change of variable (/*,#(/*), m(/^))^ (//,#(//), m(//)). This invariance is conveniently expressed by considering a small change in p9 accompanied by the corresponding changes in g and m. We write the result as

/idS/d/i The total derivative with respect to p

/jd/dfi





(7.3.2)

0.

can

pd/dp + pd/dg

be written



as

ymm2d/dm2.

(7.3.3)

7.3

Renormalization-group equation

181

right-hand side the partial derivatives with respect to n.gorm are fixed, and the coefficients p and ym give the variations of g{fx) and m2(/i) when /* is varied. The sign of the m2d/dm2 term is the usual convention (Weinberg (1973)). We have On the

taken with the other two





l*dg(/i)/dfr

(7.3 Aa)

ym=-m2tidm2<ji)/dii.

(7.3.46)

ft and ym are called renormalization-group coefficients. they are easy to calculate in terms of the counterterms, as

The coefficients

As

we

will

see

functions of g, m, \x, and d. If we use minimal subtraction, they have no mass dependence. This means that (7.3.4) can be readily solved as differential

equations

for

and

g(fi)

m(/i).

Indeed this is the easiest way in

compute the effective coupling and

practice

to

mass.

To compute /? and ym it is convenient to consider the Lagrangian expressed in terms of the bare field </>0 see (7.1.1a). We saw in Section 7.2 -

that

ml

and g0

are

jid0o/d/i

H&ml/dix Suppose

we

be solved to An

invariant:

renormalization-group =

O,



0.

have computed g0 and m0 to give p and ym.

example of such



calculation

(7.3.5)

some

comes

order in g. Then (7.3.5)

from

our

results

on

can

</>3 theory,

where

ml

m:

?}

^-"2i



3<72

1+2*^+<**} I+384^+0,*4>l-

9o=

"'3'6)

Thus 0



■%-*-*{?-<*-■£>

n^ n3-'?2\g(3-d/2)-^r-3









O^) 9a2

256k (a



6)

so

P(g)



(d/2-3)g--^ 3a3 y 256n



0(g5)

+0(g5)atd



6.

(7.3.7)

182

Renormalization group

Similarly 0

^ d/z



-m2ym[l



so



0(g2)-\



5g + 0(g3) 192n3(d-6)

m2p(g)

that

Observe that the (d/2 3)g term in /? is important in these derivations, even though the term disappears at d 6. Observe also that, even though the —



coefficients in g0 and m0 diverge at d 6, it is crucial to expand strictly in powers of g. A phenomenon true to all orders is that p and ym are =

independent

of



and \i,

provided

that

we

renormalize by minimal

subtraction. The general calculation of p and ym can be organized with the aid of (7.1.4) for the counterterms. Since we use minimal subtraction, the m- and \i-

dependence of

the bare parameters is 9o

ml





simple:

^ ̄dl29o(9>^ m2Zm{g,d).

(7.3.9)

Then

p(g,d)



ym(g) The

expressions (7.3.9)

are to



(d/2-3)g0 P(g,d)

Sg0 dg' (7.3.10)

dg

be expanded in powers of g with the aid of

(7.1.4). Now g0





(ti3-d/2g + ¥g)Z-3/2 axig)-lc,{g) /i3-d/2 g+- 6-d

Zm



(1



+ higher poles



8m2/m2)Z-

1+^4^ +higher poles' 6

where

we

h^ve picked



out the

counterterms (7.1.4). (These

(7.3.11)



single poles

are

in the series

expansion

of the

all that will be relevant.) Then (7.3.10)

7.3

183

Renormalization-group equation

becomes

r P(g,d)





y?(?)

f>



(rf/2



a 9 +



6--d

[i+a-\{g)- 6- 3)0

+ i



te)

g—



ifo) +

5g



higher poles higher poles

[fcto)



fll(ff)]



poles

(d/2-3)ff + /KffX





(7.3.12a)

"fci(g) ̄ci(g) +

[(d/2-3) + j5^)]-

6-d

higher poles

^[^1to)-*1to)].

These manipulations

are

made

(7.3.126) in powers of g. used have the fact that although

by expanding

In the last line of each of (7.3.12)

pole principle present, they must cancel in order that p and ym be 6. Notice that p and ym are independent of m and /*, and that the finite as d only ^-dependence is the (d/2 3)g term in p. Only the single-pole terms are we

terms are in -?



needed for the calculation. There is a series of relations between these and

higher poles that investigate these later. the

ensures

that the

poles

cancel in

(7.3.12);

we

will

7.3.2 RG equation

computed from two out of three equations (7.3.4) then be computed. To complete the calculation we

The RG coefficients p and ym

are

combinations of the counterterms. The differential

enable g and m' in (7.3.1) to

need £. This is related to the wave-function renormalization. It is easiest to obtain by observing that the bare Green's functions G(§) ZN/2GN are =

renormalization-group invariant,

so

that



Let

us



y^dn Z)GN.

(7.3.13)

d7lnZ;

(7.3.14)

define the finite coefficient ^



184

Renormalization group

GN satisfies the following renormalization Weinberg (1973)):

then

[^+r']G?=[^





38471-



""'"

dg

[(d/2-3)g + p(g)]

"2



Pir-ymm2 8m2

In the minimal subtraction scheme for y

group

dg

<p3

we

6-d



equation (e.g.

' ̄k=o.

(7.3.15)

find

higher poles

+0(g*).

(7.3.16)

7.3.3 Solution

equations to find g(n'), m(fi') and C,(n',n) in (7.3.1), The RG equation tells us that m(fi)

We wish to solve the RG

given

that

g{n)



g and



m.

"'d7ln c=^'d7ln tG?M/GN<M)l =

So

we

must

-M0&O].

solve this equation together with (7.3.4) for g and

boundary conditions



The

m,

(7.3.18)

C(/i,/i)=L can

m.

are

m(fj) Explicit solutions

(7-3-17)

be written:

ln(Ai7Ai)





m2([i')



m2([i)exp

dji

W) =

C(M'^)



m2(/i)exp<^



(a')



^pi-i\fljy(^))

H-Cvm[

(7'3I9)

Approximations can be made by taking a finite number perturbation series for /?, ym and y. For example: <

ln(Ai7rt=-



ruin')

dg

J g(n)

[1





128tt3

185

of Greens functions

7.4 Large-momentum behavior

of terms in the

O(02)] + 0

In

(7.3.20)

0(/*)

This is

accurate if g(fi') and #(/*) are small. Notice that g(ti')->0 is the property called asymptotic freedom. It is determined This \i the negative sign of the first term in ft at d 6. -?

oo.

as

by



The full solution to the RG

equation

is

N f dp, f  ̄—\  ̄| —y(g(fi))\GN(x\g(/i,)9nifi%fif). "

GN(x;g,m^)



exp\

(7.3.21)

7.4 Large-momentum behavior of Green's functions The most important application of the renormalization group is to compute large-momentum behavior. In this section we treat the simplest case, that of a Green's function

momenta

are

made

GN(pu...,pN) all of whose external large. (Notice that we have used our standard notation,

where the tilde indicates Fourier transformation into momentum space.) Let

suppose initially that all the Lorentz invariants formed from the momenta are non-zero. Then we scale all the momenta by a factor k: Pi

-+

by

us

KPn and let k get large. Thus all the Lorentz invariants pi-pj are scaled factor k2 and become large. Under these conditions, Weinberg's



theorem tells

us

that at least in

graphs for GN carry

asymptotic behavior



renormalizable theory all internal lines of and that graphs for GN have the

large momenta,

/cdim ^'(logarithms

of k).

(7.4.1)

Since all propagator denominators are large, we should be able to neglect masses and make only an error a power of k smaller than the leading

behavior (7.4.1). For example consider the propagator in </>3 theory at d 6. The at like is the while correction graph goes i/p2 large p2, one-loop =



128tt3

&>-!)+

dxx(l-x)ln

°bln

p2x(l



x)

47T/i2



(We used (3.6.10) for the self-energy graph

tree

(7.4.2) to derive this

equation.)

186

Renormalization group

Now there is a term proportional to ln(- p2/p2)/p2 tnat 8ets large relative to the tree graph if p2 is large enough. Thus the perturbation series has large coefficients and is not directly useful. However we may use the

renormalization group to set p2 0(p2). This makes the coefficient small again. So we use the following strategy to compute GN(icp) at large k: =

(1) Set p!



Kp, and use the solution of the RG

GN(Kp, g, m, p) (2) Neglect analysis



to

to write

C(/c/i, p)-N,2GN(Kp, g{Kp\ m(Kp\ Kp).

(if m(Kp) does give



equation

not get too

large). Then

use

(7.4.3)

dimensional

GN(Kpy g,myp) ̄( ̄ ",2GN(Kp, g{Kp\ 0, Kp)

^^a^py^'G^g^Apy (7.4.4) (3) Large /c-dependent coefficients, as in (7.4.2), are now removed, so if g(Kp) =

is small,



low-order calculation suffices.

This procedure makes it evident that the coupling that is relevant is the effective coupling at the scale of the momenta involved. we have related the large-momentum behavior of GN to the finite-momentum behavior of the zero-mass theory. It is therefore crucial that the zero-mass limit exists. However this limit does not always

It should be noted that

exist: if

we use

mass-shell

or zero-momentum

from, for example, (3.4.7), that the

subtractions, then

we see

considered in (7.4.2)

self-energy Weinberg's theorem tells us that, in the dominant momentum region for a graph without counterterms, all lines are far off- shell; hence masses can be neglected. So the problem must be that with mass-shell or zero-momentum renormalization prescriptions, the counter- terms diverge as m -? 0. This is easily checked from our explicit calculations diverges

as

same

as

m->0. Now

(see (5.10.2)). We can now

see the practical importance of the theorem whose proof was summarized at the end of Section 5.8, that the counterterms may be chosen

polynomial in masses. It ensures that the zero-mass limit may be taken directly and used to compute large momentum behavior. The minimal subtraction scheme is one way of ensuring that counterterms are polynomial in mass. If one uses, say, zero-momentum subtractions, large-momentum behavior may be computed by changing renormalization prescription to, say, minimal subtraction. Another approach is to observe that the logarithms of p2 break a,possible symmetry of the theory under scaling transformations. The consequences of this point of view were worked out by Callan (1970)

7.5

Varieties

of high- and low-energy behavior

187

and Symanzik (1970b). They derived the Ward identity for scaling transformations. It is called the Callan-Symanzik equation and looks similar to the RG equation. This equation may also be used to discuss high- energy behavior. 7.4.1

Generalizations

GN when all momenta are scaled by a large factor k is not normally experimentally relevant, for all the external momenta are then far off-shell. In coordinate space the corresponding region is the short-distance limit of GN(xl9...xN)9 where every Xj xk^> xk is made small: xj (Xj Xk)/K. The behavior of







This

means

that

we

should be able to

renormalization of the

use

RG methods to discuss the

theory, for renormalization is



purely

short-

distance phenomenon. We will work out the details in Section 7.10. On the other hand, physical experiments involve long distances. To get results for high-energy experiments we need the so-called factorization theorems. The simplest of these is the operator-product expansion which we will treat in Chapter 10. These theorems typically give a cross-section as a product of a factor which can be computed by pure short-distance methods and of

simple factors particles.

related to wave-functions of the

We could also

use

incoming and/or outgoing

RG methods to discuss the infra-red limit

k -?

0. This is

only useful if masses can be neglected. Certainly this is true in a purely massless theory, as we will see in Section 7.5.4, and then IR behavior is computable if and only if the theory is not asymptotically free. But in a massive theory, it is not useful to take \i much less than one obtains logarithms of m//i, and these prevent

perturbative



typical mass, for simple use of



methods when \x^m. 7.5 Varieties of high- and low-energy behavior 7.5.1 Asymptotic freedom

solving the RG equation to obtain high-energy behavior, we find two according to whether fi is positive or negative. In this section we discuss the asymptotically free case, when p is negative. Suppose the effective coupling g(n) is small for one value of \i, so that p is well In

cases

approximated by its first term. Then the evolution equation (7.3.4a) shows even smaller at larger values of fj, and in fact goes to zero as

that g becomes /^-?oo.

Thus perturbation theory is reliable for

computing high-energy

188

Renormalization group

behavior. From

(7.3.7)

03

that

we see

free.

in six dimensions is

asymptotically

It is instructive to compute the behavior of g from the first two terms in Let

us

/?.

define

P=-Aig3-A2g5 + 0{g7). We have the

(7.5.1)

equation for the evolution of the effective coupling:





(7.5.2)

P(g)>

Equations of the same form as (7.5.1) and (7.5.2) hold in any renormalizable theory, for example in QCD, though A { and A2 are not necessarily positive in the general case. The solution of (7.5.2) is In /i



constant +



constant +



constant +

dg'/f}(g') fg{n)

The constant

can

is conventional



0(g2).

(7.5.3)

(Buras, Floratos, Ross & Sachrajda (1977))

mass.

\A 2A, ̄



to write the

In (A,), where A is a parameter with the

Then

ln(|z7A') ln(M2/A2) so

^|ln [^)]

be computed from a knowledge ofg(n) at one value of \i. It

constant in the form In A +

dimensions of

^. +^+°(9': A.g'3 A\g 2 +



.,-,],

-1









——-J + —

Alg(nY

\n[Ald(fi)2 ̄\ + 0(g(tf), -I A]

(7.5.4)

that 9W

AMn'/A2)

A]\n2(n2/A2)



U\

ln3(,i/A)



(0^

specification specification of A.

of g(n) at one value of n is exactly equivalent to a The precise choice of the scale of A is that in (7.5.5) the

omitted terms

of the order shown rather than of order



are

expansion (7.5.5) is much used in QCD. A higher-order be made from the following form of the solution ln

(^/A2)





t-1732 +

At'gW + 2

I?

l/ln2(^/A). The

calculation of g can

A2

lA i^)2] -rlln A2'

,rj_



l0 9\_W)+^7q A.g'3

A\g\

(7.5.6)

7.5

Varieties

of high- and low-energy behavior

It is necessary to go to

two-loop order to

of (7.5.6) that

as

diverge

obtain both the terms

189 on

the

right

#-?(). The values of m(fi) and £ may be similarly calculated. For example, if y=Cl92

then

(7.5.7)

+ -,

L2Jgw



At



"""it. [l oc

7.5.2 Maximum



[InOi'/A*)] "Cl/4i4S

accuracy in

an

0(^2)] as

//

- oo.

(7.5.8)

asymptotically free theory

The results above enable calculations of Green's functions to be made at

high energy. By taking more and more terms in the series, we may improve predictions. However, in general, perturbation series are asymptotic,

the

not

convergent. A trivial example is the ordinary integral /*00

I(g,m) This

can

theory at J





m(2n)-112



dzexp(- m2z2/2-gz4/4\) —

00

a functional integral in Euclidean 04 field space-time dimension with the normalization chosen to give 0. The perturbation expansion is

be considered to be zero

1 when g



I ̄m(2n)-U2













oo

■-gz*Y± dzexp(-mV/2)(^-) JV! V ^

-gy__w2r(2Af r(jv



w?oU?!V t (zxYh-











r J





l/2) i) (7-5-9)



Now

IN



NN(^\ (nNyl/2[l



0(l/N)l

asN^oc,

(7.5.10)

divergent. The divergence is associated with the fact that the defining integral diverges if g is negative, so that J is not analytic at g 0. The corresponding property of 04 theory is that the Hamiltonian is unbounded below (i.e., the vacuum does not exist) if g is negative. so

the series is



190

Renormalization group

What

say is that

we can



by truncating



the series:

(7.5.11)





error

approximate

I (#T^ + 0[(#Y 1].

/= The

we can



is estimated by the first term omitted, i.e.,

These results

are

standard in the

theory of asymptotic expansions for simple

integrals. All experience, together with rigorous theorems for quantum mechanics and super-renormalizable field theories (Glimm & Jaffa (1981) and references therein), indicate that this behavior is in a non-trivial dimension (i.e., d > 0).

typical

for functional

integrals

Next, let

suppose we wish to compute some quantity in free theory by truncating its perturbation expansion

us

asymptotically

an

I 9(H)2NIN-



We assume that the set

jj,

the unit of mass







quantity depends

on some momentum p,

and that

we

be of order p. A case would be the propagator with that the coefficients in the expansion behave like

\i to

0(\p\). Suppose

IN



NNbNNad[l



0(1/N)],

(7.5.12)

for large N. What is the best accuracy with which we can calculate the quantity ? This is given by the minimum error, i.e., the minimum of INg2N as N

varies. The result is that the minimum possible

error in a

perturbative

calculation is of order

constant|p|-2^/e*(ln|p|)fl. that beyond a certain level, power-law corrections to the behavior computed in perturbation theory are meaningless since they are smaller than the irreducible error in using perturbation theory. Power-law corrections are those that are a power of p2 smaller than

This

means

asymptotic

the leading term.

7.5.3 Fixed point theories In four dimensions, the only theories that are asymptotically free are non- abelian gauge theories with a small enough number of matter fields see Coleman & Gross (1973) and Gross (1976). Other theories, like 04 and -

QED, have an effective coupling that increases with energy. Thus, in such theories it is impossible to compute the true high-energy behavior by perturbation theory. (Note however that the coupling in QED is a/n

of high- and low-energy behavior

7.5 Varieties

191

9* Fig. 7.5.1. fi(g) in



1/430.



non-asymptotically free theory with

This is very small,

not occur until very many



fixed point at g



g*.

non-perturbative region in QED does orders of magnitude beyond experimentally so

the

accessible energies.) An interesting possibility is that P(g) has the form shown in Fig. 7.5.1, with a zero at g g*. Then #(/*) approaches the fixed point' g* as \i -? oo. At =

large momentum Green's functions behave like

GN(Kply..., KpN9g9 m9fi)



const

Kdim^+iV^)/2GiV(p1,.. .,pN,g*Afi). (7.5.13)

This behavior is

Consequently, 0.

as

if

<j>

the function

had

an

extra

term

y(g)/2 is called

y(g*)/2

in its dimension.

the anomalous dimension of the

field

7.5.4 Low-energy behavior

If m



0, then the renormalization group

ofmassless theory can

be used to compute infra-red

calculability is the opposite of that for the UV behavior. Consider first an asymptotically free theory. There, the effective coupling g{fi) goes to zero when \i goes to infinity, so that short-distance behavior is computable perturbatively, as we saw in Section 7.5.1. But, when \i is small, g{jx) is large, so the infra-red behavior cannot be computed reliably by perturbation theory. (This is the case for strong interactions, according to QCD.) Let us now consider a non-asymptotically free theory. For large /*, the behavior. The

effective coupling is large,

so

the short-distance behavior is not

perturbatively computable. (For example, a perturbative calculation in low order of the position of the fixed point, g*9 in Fig. 7.5.1 and of the value of /?(#*) is subject to large errors from higher-order uncalculated corrections.) But when /* goes to zero, so does the effective coupling. We are assuming the absence of a mass term for the field, so there are no large logarithms of mj\i as fx goes to zero. Hence, we can compute IR behavior in such a theory, just

192

Renormalization group

as we

computed

U V behavior in

an

asymptotically free theory.

We will

now

do this. Now, for almost any

graph

there

large logarithms of p2/fi2

are

as

p2 ->0

massless theory, just as in the ultra-violet. In the case of the propagator, these logarithms mean that the propagator's singularity is not a pole, at in



least

To

order-by-order. investigate this singularity y{g)



let

Cl929

us

again write

(7.5.14)

P(g)=-Aig\

using the same notation as before, but now with A x



below the first

one.

non-zero

fixed point #*, if there is

0. We

that g is The propagator is assume

rg(Kfi)

G2(Kp;g(fi\fi) =



i/(/cV)[l

K ̄2G2{p;g(Kii\fi)Qxp +

J gin)

dgy(g)/fi(g)

O(0M2)]expj- !dg(CJAl9)[l



O^2)] (7.5.15)

-i/(K2p2)-constant[ln(l//c)]Cl/2yil, as

k:->0. Hence if

nonzero

limit

as p

Cx

is non-zero, then

p2G2(p) does

-?0; the singularity of G2(p2)

p2

not have



finite

0 does not correspond The massless particle that gives rise to the at



simple single-particle pole. singularity cannot be treated as an ordinary particle, because its long-range interactions are too strong. Positivity of the metric of the state vectors

to



Cx to be positive, and we assumed a theory with At negative, so K2 times the right-hand side of (7.5.15) goes to zero. (The positivity

constrains

argument is the

one

given in the the pole in

textbooks (e.g.,

Bjorken

& Drell (1966))

a propagator is less than unity if the propagator is of the canonical field. Application of this argument in the theory with an ultra-violet cut-off shows that the divergence of the self-

that the residue of

energy must be such that Cx is positive.) Notice that if Cx 0, then the propagator does have =







finite residue at

0:

-iK2p2G2(KP,g(ii\ii)^exp\

Fig. 7.5.2. Lowest-order self-energy graph in 04 theory.



dgy(g)/my

(7-5.16)

self-energy Fig. 7.5.3. Lowest-order to the contributes that graph anomalous dimension in (p4 theory.

Leading logarithms,

7.6

etc.

193

Observe that only the first term in y is relevant to the finiteness of the limit, and that the limit does not exist order-by-order. A case where Cx 0 is the =

04 theory,

because the one-loop graph

two-loop graph dimension.

Fig.

7.5.3

Fig.

7.5.2 is

independent

of p. The

provides the lowest-order term in the anomalous

7.6

Leading logarithms,

etc.

Renormalization-group logarithms

7.6.1

We saw in Section 7.4 how to compute the

large momentum behavior by approximating it by a Green's function with m 0. Then we used the renormalization group to reorganize the perturbation series into a form with small coefficients. It is of interest to examine how the complete result can be obtained by a systematic resummation of the perturbation expansion. For concreteness, let us examine the propagator G2(p, g, p) in massless 03 theory in six dimensions. We write its perturbation expansion as of a Green's function =

G2



(i/P2)f

n = 0

where the lowest coefficient is

T0

(7.6.1)

92nTn(-p2ln\



1. We will prove that each

Tn is



polynomial in In \i (and hence in In ( p2l\i2)) of degree at most n, with n being the number of loops. To do this we will regard the RG equation (7.3.15) not as an equation to give the variation of G2 when \i is changed with g set equal to the effective coupling g(p) (thus keeping the theory fixed), but as an equation for the /^-dependence of G2 with g fixed. Picking out the order g2n term gives —



T din fi'



-y(g)-P(9)f]ni Tn,g2n] C9 Jn'





T?





coefficient of g2n.

0(g2) and ft is 0(#3), this equation determines Tn,'s:

Since y is order



(7.6.2)

Tn in terms of lower-

constant

I -\y{G)^mP\g2n'[^^^TA-P2/ii'2)] {ln'<n L J 09 J

JO

coefficient of 32".

(7.6.3) Iteration of this procedure another n constants

of integration.



gives Tn in terms of T0 and n polynomial of degree n in ln/^ as

1 times

Evidently Tn is



194

Renormalization group

claimed:



T?=

Tnt?_l[ln{-P2/H2)y-

/=0

All but the constant term

are

(7-6.4)

determined in terms of lower-order

coefficients. A convenient way of organizing the series is to define for each term L=



This

number of

logarithms

g2

+ number of powers of



-l +

is

n.

(7.6.5)

non-negative. The sum of the terms with L 0 (viz., Tn0[ln(- p2/p2)Y) gives what is called the leading logarithm approximation to G2. Application of pd/dp, pd/dg or y to G2 (with one-loop values for p and y) increases L by 1. All the non-leading logarithms give even =

higher values of L. So the leading logarithm series exactly satisfies the one- loop approximation to the RG equation. Hence we may sum the leading logarithm series by solving this approximation to the RG equation

G2(p2^p)L=0=Up\ T I



Here

[1



%y{g{p'))\{p2-g{J-p2)J-p2)} ^

LJm

used the

we



A{g2\n(





notation

same

Jl=o

(7.6.6)

p2/p2)]Cli2Ali/P2-

in Section 7.5.1. This equation

as

reproduces the

approximation derived at (7.5.8). Another way of treating both the leading and non-leading

to

use

the RG equation

*Z







(7.6.2)

to

give

[2^1(?-l)-C1]E1T?-uDn(-pV)rl"i L







\2A2{n-2)-C2-}\

1=0

The



logarithms is TnL's:

recursion relation for the

2(L-n)Tn,L[\n(-p2/p2)f-L-i =

where y



Cxg2 + C2g4 H

leading logarithm -

Tn_2>L[ln(-p2//i2)]n"2-L+-", (7.6.7)

and /?= -Axg2 A2g5 + part of this equation is

2nTnt0







[2AM



1)



CJ T?_

1>0.

??■.

(7.6.8)

This equation determines the leading logarithm series in terms of T00 and of At and Cx; this series sums to (7.6.6). Equation (7.6.7) also determines the non-leading logarithms. example the next-to-leading terms are

-2{n-\)TnA= [2AM- ^-C^Tn_1A





1,

For

[2A2(n -2) -C2]Tn_2,0. (7.6.9)

7.6 n-L



Leading logarithms,

195

etc.

#logs





# loops

Fig. 7.6.1. Illustrating the leading logarithms and non-leading logarithms of



Green's function.

Again a convergent series results. Its sum is equally accurate as the result using the two-loop approximation to p and y in the solution (7.3.21). There \i is set equal to (— p2)1/2 and we take the one-loop approximation to

of

G2(p\g(y/-p2),

0, J ̄p2\ i.e., (i/p2) (1 +^2Tn). The series for larger L may be similarly determined. In Fig. 7.6.1 m



we

illustrate the structure of the calculations. The diagonal lines are lines of constant L, and the recursion relation (7.6.7) determines a coefficient TnL in

terms of lower-order terms

its

and

the

higher diagonals. Suppose we have computed perturbation theory to n 1 loops for G2 and wish to compute the rc-loop term. In this term the coefficients of all but ln( —p2//i2) and the constant are fixed by the lower-order calculations. on

diagonal

on



information is in the nth order coefficient Cn for y and in the terms with one and no logarithms, i.e., in Tnn_ x and Tnn. The (In)0 term in Thus the

new

(7.6.7) is



2Tn,?_x





Cn



V l2Am_jj



i= i

CB_,] Tjj.

(7.6.10)

knowing the coefficient, Tnn_li of the singly logarithmic is Tn equivalent to knowing the rc-loop coefficient Cn in y(g). Exactly the same procedure may be applied to any Green's function GN. The only difference is that there are several external momenta. It is enough to consider the connected graphs. Let us write This shows that

term in

^(connected)



(p2)dta'G>/25"-2

f £

G^ln""^- pV/'V,

n=0L=0

(7.6.11) where p is

one

of the external momenta and

which is the power of g

appearing

we

have factored out

in the tree

approximation.

gN ̄2, The

196

Renormalization group

coefficients

G{£1

are

now

functions of the dimensionless ratios of the

Lorentz invariants formed from the external momenta. The leading logarithm series and all the non-leading series

convergent,

so

they

appears when

we

be summed. The n\ behavior of

can

consider the

are

large orders only

single log and constant terms, and thus in the

sum over L.

7.6.2

Non-renormalization-group logarithms

In all of the above cases there was

loop. There

are more

one logarithm of the large momentum per complicated situations where not all invariants get

large. A simple standard example is the form-factor of the electron in QED with a massive photon (Fig. 7.6.2). Here q2 (px p2)2 gets large but pi =

and

p\ are fixed. It turns out

two

logarithms

per

(Sudakov (1956), Jackiw (1968)) that there are must be in the coefficients G^l in (7.6.11),

loop. These

since the power of the logarithms is too logarithms in (7.6.11).

Fig.



high

for them to be the

explicit

7.6.2. The electron's form factor in QED.

large-momentum behavior in such situations. sum the leading logarithms, which are often to relatively easy compute. For the on-shell form-factor this gives a which series sums to (Jackiw (1968)) convergent One would like to find the

A much-used

technique



is to



exp





(e2/l6n2)\n2q2l

(7.6.12)

(See

also Mueller (1981). For the simple cases that we considered earlier, with all momenta large, the leading logarithm approximation is justified by renormalization-group methods, as we have seen. For cases like the present one

of the form-factor, it may be



bad approximation

(1981,1982b)). behavior in some

However methods are available to obtain

(Collins & Soper large-momentum

of these situations. See Mueller (1979, 1981) and Collins (1980) for the electron form-factor and Collins & Soper (1981) for cases in strong interactions.

7.6.3 Landau The

ghost

leading logarithmic approximation (7.6.6)

for the propagator has



7J Other theories

197

singularity when (7.6.13)

p2=-Ai2exp[-l/M^2)].

This singularity, if present in the true propagator, would signal a state of squared. Since the residue has the opposite sign to that for a normal propagator pole, this would be a state with unphysical properties.

this value of mass

It is called the Landau ghost (Landau & Pomeranchuk (1955)). In

a non-

asymptotically free theory like asymptotically free theory

an

case it

QED, it occurs at very large energies and in like QCD, it occurs at low energies. In either

where perturbation theory is

occurs

leading logarithmic approximation is



inapplicable

and

so

where the

bad approximation.

7.7 Other theories We restricted

our

attention in

theory with one field,

to a

one

deriving the renormalization group equation coupling, and one mass parameter. However

by exactly the same method, a theory with several fields </>!,...,(j)A (each may be Bose or Fermi), several couplings gl9...9gB, and several masses. A change in the unit of mass \i is compensated by a change in each of the parameters and in the scale of the fields. The main problem is a we

may treat,

proliferation of indices. It is easiest to treat couplings and masses on the footing. So we have a collection gl9...9gc of renormalized parameters, with C being the total number of couplings and masses. Then we must write same

each function

fi} being, a priori, a function of all the parameters. For the case

theory with a single coupling, and a single mass, would (g, m2), and px f}(g) and p2= ym(g)m2. Given a function / of the renormalized parameters and of





The RG coefficients gj{0)

are

have

(gl9g2)





*h"**-(%+¥4ry can

of //,

we

have

,772)

be determined by noting that the bare couplings

invariant:

These form C equations for G unknowns. The RG equations for Green's functions are complicated by the possibility that fields of the same quantum numbers may mix under

198

Renormalization group

renormalization.

Writing <t>(0)i

we



IJ£ij(&d)<t)p j

find, for example, that 'kinetic energy' i

terms in ¥£ are of the form



hi

JJ

Hence, we have a matrix counterterm for the field-strength renormalization

Zi/g,d)=EC?Cy



(trOy,

(7.7.4)



denotes transpose. (Note that ( has a different meaning here than in Sections 7.1 and 7.2.) If we define a matrix anomalous dimension by

where

then invariance of the bare fields gives °



^^(0).



(^^-Ky)^

(7.7.6a)

i.e., d

*'-"s?- or

^Z=|{Z,y}.

(7.7.66)

If Z is diagonal (as is the case in most theories we consider): Ztj (7.7.6) reduces to

an



btiZh then

anomalous dimension for each field

^4>i=-hi4>i, dp

(7-7.7)

d In the case of a diagonal Z, the renormalization group equation for an Appoint Green's function is '









E KteXV

dl

(7-7.8)

7.7 Other theories

199

Here

yia is the anomalous dimension function for the ath external field of If the renormalization matrix Z is not diagonal, then we have a similar,

GN. but

complicated, set of equations for the Green's functions. equations for the evolution of the couplings are coupled, so their solution is in general complicated. Considerable simplification can be achieved by using our knowledge of the dependence of the counterterms on massive couplings. Let the couplings gl9...,gAbe dimensionless and let the more

The

corresponding

bare

couplings

be

where the g(0)i depend only on the renormalized dimensionless couplings and on the UV cut-off. For the sake of definiteness, we assume that the

physical

dimension of space-time in the theory is d 4. The wave-function also on the dimensionless couplings and only depend Zf =

renormalizations

d. Let the other parameters (masses and super-renormalizable couplings) be denoted by /s, and let the dimension of fs be (4 d)zs + as. If fs is on



the

mass

of a fermion, then os



1, ts



0. If it is



boson

mass

squared, then

2 and ts 0, while for a super-renormalizable coupling ts ^ 0 and <js > 0. Then (by Section 5.8) the bare quantity corresponding to fs has the cs form =



/(o,s= I Hi4-d)t*XFsX(g,d). dx

Here X is any





4) equal

Requiring

product of

(7.7.9)

-as

the dimensional

to the dimension os of

couplings with dimension (at

fs.

invariance of the bare couplings gives 0



yj= 0



(4



d)Pigm + X Pj7r-gm, j=i G9j

I^lnZ,

(7-7-10?) (7.7.10b)

(4-dXXXFsx+ZXPj4-FsX X

t,X

X, J

"Vj

Ch

triangular structure to the evolution equation: the evolution of a coupling depends only on couplings of the same and lower dimensions.

There is then



In these equations, the index; runs over the values 1 to A9 i.e., over those values that correspond to the dimensionless couplings, while the indices 5

and t

run over

the labels for the dimensionful

couplings.

Renormalization group

200 If we

minimal subtraction, the calculation of the coefficients is rather

use

easy. Let

Gt(g\ Zt(g\ and PsX(g) be the coefficients of single poles in g{0)i, Zt

and

Then

FsX.

we

have

Pj(g,d)=(d-4)Pjgj + pj(g)9 fc(g, t9d) (d 4)tJs + &(g, f), =

with

yj9

and

pj

/?s satisfying A









{(x." Z^)(^) ^I P^tf}. -

7.8 Other renormalization It was

only for

(7.7.11)



the sake of simplicity that

(7-7.12)

prescriptions

we

restricted

our

attention to the

minimal subtraction procedure. The proof in Section 7.2 in fact shows that any change in renormalization prescription can be compensated by a

change in renormalized parameters and a change in the scale of the us examine what happens in more general schemes. It is sufficient to restrict our attention to a theory with a single coupling and

renormalized field. Let

mass, like

03 theory

If we choose

be





in six dimensions.

renormalization scheme with

renormalization

point,

still be defined and computed

general

an extra mass \i,

which might

coefficients

then

can renormalization-group by (7.3.4), (7.3.5), and (7.3.14). What we lose in

are:

(1) the simple formulae (7.3.12), (2) the lack of dependence of /?,

ym and y on the masses.

In order to discuss UV limits, it is sensible to choose a scheme in which the limit m->0 exists. This means that

/?,

ym9 y are finite,

order-by-order,

as

m->0.

related by finite renormali- may relate the RG coefficients in different

Now, different renormalization schemes

zations of the parameters. So we schemes by looking at the theory

then computing /id/d/i in

one

are

in the physical space-time dimension and of/id/d/i in another scheme

scheme in terms

with the aid of the chain rule.

Suppose

we

have



second scheme in which the

new mass,

coupling

and

7.8 Other renormalization prescriptions field

are

m', g and 0': g' m'2





g'(g,m2/n2),

m2zm(g,m2/n2),\

Then the Green's functions in the

new

G'N(p;g\m'2yp) The

201



scheme

are

(7.8.2)

t:NGN(p;gim\p).

coefficients in the

renormalization-group

p'(g',m'2/n2)



(7.8.1)

new

scheme

pl ̄ ymm2 A V^> n^g' (n ̄ \ dp dp' +



are:



dm2

og

V), (7.8.3)

ym(9'^'2/p2)^-^j-m' =

Jm



+ L-^ {.zm{g,m21 p2)\ Py ymm2^)ln eg J \ cp  ̄

cm

(7.8.4)

Our definition of the total derivative d/dp is as the derivative with respect to p when the bare coupling g0 and bare mass m0 are held fixed. Therefore, it is the same in both schemes. Notice that there are two steps in computing /?' or

y'm\ First, compute the right-hand side expressed in terms second, change variables to the new coupling and mass. The anomalous dimension of 0' is obtained as

of g9 m, and p\

y'(g\mf2/p2)=-^p^\nG'N -2p—InC dp =*

A considerable



+ y

2(/V% y^to?)ln ^ m2'^  ̄

simplification

schemes. Then g\ ( and

zm

occurs

are

ym(g')

/(0')







relating mass-independent

functions of g alone,

renormalization-group coefficients in the

p'(g')

in

(7-8-5)

new

scheme

so

that the

satisfy

l}(g)Tg'(9l ym-P—^zm(g), 7-2/?-lnC(0). ?g

(7.8.6)

202

Renormalization group

In this case, let

g\

zm and £ have

perturbation expansions

fli03+fl205 + "O zm=l+M2 + fc204 + "-A Q'



0 +

and let the expansions of /?,

ym(g) yte)





The expansions of /?',

primed. Then example: =



ym, and y be



Big2 + B2g* + ---y c^2 + c2^4 +

y^,

and y'

are

gs(A'2



3axA\)



(7.8.8)



---

written similarly with all quantities

by using (7.8.7).

express them in terms of g

we can

A[g3



(7-8.7)

07(^

5^2flj



3^;a2



+ 3,4

X) +

For





?.

(7.8.9) This

must agree with the

perturbation expansion of the right-hand side of

(7.8.6) pdg'/dg=-Alg3-g5(A2



3alAl)-g7(A3

3A2al





5A1a2) +

' ̄.

(7.8.10) From these

equations

we see

that the first two coefficients in ft do not

prescription is changed, i.e., the above 2. By generalizing equations to all orders we also j A\, A2 see that, by adjusting the terms in the expansion of g\g\ we may choose the when

change A



the

renormalization





terms beyond the second in /?' to be whatever we want. In similar fashion that only the 0(g2) terms in ym and y are invariant.

we

see

one-loop term in y or ym is zero then the whole of y (or ym respectively) may be made zero by a choice of renormalization prescription. This privilege does not extend to /?: if the first non-vanishing term in p is at rc-loop order (n > 1) then that term is RG invariant (but not the (n + l)-loop term). In a theory with more than one dimensionless coupling we may try to Note that if the

apply the

same

methods. This is left

as an

exercise. It will be found that only

the first term in each p is now invariant, except in the ^-function does not mix the different couplings.

The

invariance

of

these

case

applies

coefficients

only

independent renormalization prescriptions.

If one

are

zero.

Then

we

have

f}







one-loop

within

were to use, say,

subtractions, then the parameter \i would not appear, respect to \i

that the

ym

so



mass-

on-shell

all derivatives with

0.

(The asymptotic

7.9 Dimensional transmutation

203

behavior that we extract by varying \i can no longer be computed by renormalization-group methods, if we stay within this renormalization

prescription. In this instead



case the Callan-Symanzik equation must be used Callan (1970) and Symanzik (1970b).)

see

7.9 Dimensional transmutation

theory with one dimensionless coupling g and no masses. A physically important case is QCD with several flavors of massless quark; with two or three flavors this is an approximation to actual Consider



renormalizable field

strong interactions. Since the basic theory has

no masses we must use a

renormalization

an arbitrary renormalization mass /*. Although the theory apparently has two parameters, g and /*, we saw that this is not so: a change in \i can be compensated by a change in g. In fact, as Coleman & Weinberg (1973) pointed out, the theory really has no parameters at all. The point is simple but somewhat elusive, so we explain it at length. A physically measurable quantity must be renormalization-group invariant. For example, let M(#,/i) be a particle mass. By dimensional

prescription with

analysis, it is

\i times a function of g alone. So 0





n—M d/i





(7.9.1)

p— M. dg

M +

Hence M



/r constant exp

t-r







P(9l.

A2,

['A ,{.A\g" + ^g')(A,-A2g'2)-]

nCexpl-——-—— \n(g)-\ dg 2Al92 A\ J 0 J0 ,





iiC exp





^—2



2Al92

Here C is a constant and

A2g'3p(g')

In (g) + 0(g2)}. -f A2 we

have written

f}(g)

(7.9.2) =



A{g3



A2g5 H



as

usual. Note that the Green's functions invariant :

to measure a

This definition has The formula

Green's function,

an

(7.9.2)

are

not

renormalization-group

one must

define the field operators.

arbitrariness, which is the freedom has



to vary its scale.

number of consequences:

particle masses cannot be computed in ordinary perturbation theory (in a theory with no mass in the Lagrangian). For to avoid large logarithms one must set \i to be of order M, where M is the particle mass

(1) Non-zero

204

Renormalization group

being computed. is

(2)



Then (7.9.2) tells

us

that g(M) is not

free parameter; it



number of order

In a

unity. non-asymptotically free theory (A,



0), suppose

we

have



small

value for g(p). Then jx<M, where M is the value of the mass of any given massive particle. Perturbation theory is therefore only useful for

Green's functions when the external momenta

(3)

threshold for

massive

In



producing any of the an asymptotically free theory (A x

much below the

are

particles.

0), we have p, > M whenever g(p) a theory only when

is small. Perturbation theory is useful in such momenta are much bigger than particle masses.

(4) Since the ^-dependence of (7.9.2) is universal, i.e., the same for all particles, ratios of particle masses are pure numbers independent of g and p. Let us emphasize once more that these results are true when there are no explicit mass terms in the Lagrangian. The observation of Coleman & Weinberg (1973) comes from asking what

be measured in the theory. Suppose we start with p px and g gx and ask how the theory changes when we work with the theory with a different

can





value

ofg,g g2. (We suppose gx and g2 are between g first fixed point of /?.) Each version of the theory has an satisfying =

Now evolve

0 and g effective



g(p)

g*, the coupling =

in the second version to the value of p where

0vers2(/*2)



01-

theory is just the first version with all p2jpv For example let abea cross-section

Then the second version of the

factor

by depending on momenta analysis give us momenta scaled



p{y...ypN. Then RG invariance and dimensional

g(Pi>--->Pn'>92>Hi)





°(Pi>' ̄>Pn'>9i>I*2)

(?l\ \/*l/

J tkp \A*2

V±pN;gi^AJ 1*2

(7.9.3)

The last factor is the cross-section in version 1 of the theory, with its momenta scaled. We

see

that

changing the dimensionless coupling

does not basically change the theory, but only its dimensional transmutation. There

are many ways

might give

the proton

in

mass



massless theory

scale. This is called

of specifying the scale of the theory: in QCD one For perturbative purposes it is better to use

mass.

7.9 Dimensional transmutation

205

something that can

be directly used in perturbation theory, for example the value of /i at which g(ji) has some given value (e.g., 0.1) in one's chosen renormalization prescription. One standard way is to notice that for large /*,

g([i)

has its

asymptotic 2/

)=

behavior

given by



A2ln(lnfcVflo)) +

constant

rin^On,l(/V/*o)) L ln3(AW7*o) x

^,

Here fi0 is a reference value of /*. If /^0 is changed then the series is reorganized; only the first two terms are unchanged. As is conventional

(Buras, Floratos, Ross & Sachrajda (1977)), we define the scale A of strong as the value of jx0 for which the l/ln2(/i2//^) term is zero. This

interactions

gives (7.5.5). If we change from,

say, minimal subtraction to momentum-space

subtraction, then the theory is

unchanged provided

the

coupling

is

adjusted. This may be done in perturbation theory. For example, we might find that g in the MS scheme and in the momentum-space subtraction scheme are related by

given by (7.5.5) with A AMS, and let gmom be given by (7.5.5) with A Amom. (We already know that Ax and A2 are the same in both schemes.) Substituting these expansions into (7.5.5) and requiring consistency gives

Now let gMS be





AMs



Amoinexp(a1/^i).

(7-9.6)

Notice that both A{ and ax are obtained from one-loop calculations and that there are no higher-order corrections whatever (Celmaster & Gonsalves (1979)). An amusing consequence is (7.9.2). Since M is independent

by substituting (7.5.5) for g in may let /*-? oo. The higher-order

obtained

of fi

we

terms all go away and leave M



CA(Al)A*<2A*.

(7.9.7)

equation is not very useful for performing perturbative calculations. If the theory is a complete theory of physics, then measurements of a and the p's in (7.9.3) will be in terms of a standard of mass. This we may take to

This

particle (say, the proton). Let us now change the theory by changing the coupling from gl to g2 Just as we did earlier. Then the standard of mass is multiplied by /^2/Vi- So if we do experiments in be the

mass M

of

some

206

Renormalization group

equal to those in version 1, the actually increased by a factor /i2/Vi Therefore, (7.9.3) tells us

version 2 with numerical values of momenta momenta are

that



gets multiplied by

increases sense



by



factor

the same factor,

so

(/^2/Mi)dimr7-

But its unit of measurement

unchanged. In this coupling (or different

the numerical value is

massless theories with different values of the

values of A)

are

of Coleman &

indistinguishable. This is perhaps the most important result Weinberg (1973).

However, there are many experiments that claim to measure A. There are give (without qualification) a single measured value of g.

even some that

How

can

this be? The second problem is easy to dispose of. What is being

measured is the effective coupling g in some renormalization scheme with \i set to a value of the order of the energy of the experiment (typically in e+e ̄- annihilation). Strictly one should specify not only the value of g but also the scheme and the value of \x. Now the experiments are at around 10 to 30GeV, and A is at most



few 100 MeV. The variation of g

over

and the variations between the usual renormalization schemes

this range

are

often

no

than the size of experimental errors. So it is possible to talk loosely. However, we just asserted that massless QCD with different values of A is the same theory. The sense of a measurement of A is that we measure the numerical value of the ratio of A (defined by (7.5.5)) to a standard of mass. For the purposes of the argument, we may regard the standard as being that the nucleon mass is 939 MeV. In terms of dimensionless quantities the measurement is of the constant C in (7.9.7) when M is the nucleon mass. (In the MS scheme, we find that C is between about 5 and 20.) The non-zero masses of the quarks make a relatively small perturbation of the above more

argument. Notice that if we

standard

mass

unchanged. In QED

play God and double the size of A, then the size of the

also doubles,

so

that numerical results of experiments

the situation is different. The electron has



mass,

are

and its

Coulomb field is classical at large distances. A mass-shell renormalization scheme is natural. Since there is a very important mass-scale, an unqualified statement of a measurement of the

QED with

makes good unlike QCD in the absence of sense.

7.10

It is very convenient to cutoff in



QED coupling, viz.,

different value of

quark

is









(47r/137)1/2,

different theory,

masses.

Choice of cut-off procedure

use

dimensional continuation

perturbation theory. However,

there is

no

as an

ultra-violet

known construction of



7.10 Choice

complete theory in

207

of cut-offprocedure

arbitrary complex dimension, so one must beware of significance to use of dimensional continuation. This is especially true when we use minimal subtraction, which is a procedure that exploits the form of the cut-off dependence of the theory. However, the renormalized theory with the cut-off removed does not depend on the form of the cut-off. We saw this in our one-loop calculations. an

assigning too much physical

general the fact is easiest to see by using BPHZ renormalization, in which integrand is constructed that gives a manifestly convergent integral. The only freedom left is a change of renormalization prescription, otherwise known as a change of parametrization. In this section we will examine the renormalization-group properties when a different UV cut-off is used. For definiteness we cut off the theory by using a lattice, with spacing a. We consider any theory with a single dimensionless coupling g and a single mass m. It is, of course, possible to generalize to any cut-off procedure and to any theory. In general we will In

an

need



renormalization

mass \i,

The bare coupling g0, bare

in order that

mass

we can

take the massless limit.

m0, and the fie Id-strength renormalization m and \i, and of the

Z are written as functions of the finite parameters g,

cutoff

a.

Then the renormalized Green's functions

are

written in terms of the

bare Green's functions

and for them the limit

0 exists.

a -?

renormalization-group structure is essentially unchanged. Let us choose a mass-independent renormalization prescription, so that again #0, The

Z, and m0 have the forms :



mo





The massless theory has



Z{g^a\

(7.10.2)

rn2Zm(g,fia) + a ̄2Y{g9fia)J m



0, and,

as

before, g0 and

Z are

independent

of

the cut-off parameter is dimensional, so g0 and Z have explicit dependence on fj, as shown. But the dimension of g0 is fixed at zero,

mass. But now

so

the The

longer

^-dependent

power of fj, is not used.

m2-dependence of the bare mass squared true that m0



0 when





0. In the

is

again linear.

case

But it is

no

of dimensional reg-

dimensional parameter is \i, and it is not possible (Collins (1974)) to generate by minimal subtraction a counterterm 2 fi2 Y(gy d). But with a lattice cut-off a term a Y is both possible and needed, ularization the

only remaining



as we

will

now

verify by computing



low-order

graph.

208

Renormalization group 7.10.1 Example:

</>4 self-energy

simplest example that shows the existence of the Y-term in (7.10.2) is the self-energy graph of Fig. 7.10.1, not in the 03 theory that we have been using, but in the 04 theory in four space-time dimensions (with m 0). The The



Lagrangian is (2.3.1). With dimensional regularization the value of the graph is

^O(2n)-d[d'lk/k2^0,

Fig.

but with



7.10.1. Lowest-order

lattice cut-off

we

self-energy graph

in

<j>4 theory.

find

-ig0(32nYl



d3kdcoD(co,k;a).

(7.10.3)

|fc**|<7t/a

Here the Euclidean lattice propagator is

l/(a>2



k2) if co

and

(7.10.3)

is

|fe|

are

much

smaller than \ja. For general values of &", it is

a2\U

| sin2(fc"a/2)|,

which is positive definite, so that the integral diverges to a number of order \ja2 as a->0. A similar

divergence

occurs

in the

self-energy

of



non-zero

and

scalar field in any

theory.

7.10.2 RG coefficients continue our general discussion of the renormalization group lattice cut-off is used. As in the treatment using dimensional regularization we define a renormalization-group operator We

now

when



%=%+Kg> ?a)ig



U9>

?a)m2ib-

(7-104)

changed our notation slightly, and used an overbar to indicate renormalization-group coefficients in the cut-off theory. These coefficients p and ym have finite limits, fl(g) and ym{g\ as a -? 0. In our later work it will be rather important to distinguish the coefficients before and after the

We have

7.10 Choice

209

of cut-offprocedure

cut-off is removed. The coefficients

can

be computed from

>Vo=0;

dp.

with (7.10.4) used for fid/dfi. We also /id In Z/d/i, just as with dimensional y =

have the anomalous dimension

regularization. This all results in



}d{?aY

-[ -[



-d'

5(H

<3#

In

t>5



(7.10.5)

Zm{g,pa\

\nZ(g,pa).

In addition there is the constraint





^(a-

The information

on

the

/?, ym and y. If desired,

g0



/?



be finite set

-I

pa

g(/ifl)

d#_

g3Gl x

(7.10.6)

Y{g,pa).

divergences is all contained in the finite functions

we can use

minimal subtraction with the form

g5G2l

+ -'y

function of g alone. In order that of logarithms ap must cancel in p. This implies of relations for the counterterms, the first of which is 2G22

so that

ft

Y)

g3Gll\n(ap) + g5\_G22\n2(ap)+G21\n(ap)]

g +



2V\- =

3G\V













is



a-+0 all the

as



An

analogous regularization, as can

set of relations occurs when we use dimensional

from (7.3.12). These we will discuss further in Section 7.11. Note that for minimal subtraction with the lattice cut-off we

havej8



j5,

ym



be

seen

ymi etc.

7.10.3 Computation

If

we

were

to

approximation,

of g0 and Z; asymptotically free

compute the exact theory, rather than we

would need to know how

case

perturbative g0(g, \id) depends on a as a -? 0 a

with g and p fixed. A low-order calculation is not sufficient, for g0 has logarithms in its perturbative expansion. Provided the theory is

large

asymptotically free, we can remedy this by using the renormalization group to improve the calculation, just as we did for the large-momentum behavior of

210

Renormalization group

Green's functions. The

starting point is

the equation

4^40o+^°=0' which

is in effect the

it

as

(7-10-7)

renormalization-group equation for g0.

We may solve

for the Green's functions.

just Ultimately, we will let a approach zero while holding g and \i fixed. But first let us keep a non-zero. Then we can define an effective coupling g(p) by ti'dg(?')/d?'



(7.10.M

p(g(?n

with the boundary condition #/i)

(7.10.86)

0.



We will also need the effective coupling at denote it by the symbol g(ji'). It satisfies

/I'd&O/d/i' 9iv)



Implicitly

there is



dependence g



jJtfOi'))



Now, when



us

0),

0-

of g

on \i

g(\i\\i,g\

9o(9> W)



0. For the moment, let

Mv'Y^'a



and g, and of g





on /*, g

and

a:

g(\i\\i,a,g\

Of course, #(//)-? #(/*') as a->0. We can solve the renormalization-group 0o





equation (7.10.7)

for g0 to find

9o(9(Va)> !)?

might appear that we can replace g(l/a) by g(l/a\ approximated by the first term in its perturbation expansion (since g(l/a) is small). That is, and that



is small, it

g0 is well

g0



g(l/a) + negligible

error.

These suppositions are actually false, for two reasons. First, fi(g,an) in general depends on a\i, so we cannot just replace g at \i \ja by g(l/a) computed in the a 0 theory. Secondly, we cannot simply drop the higher- =



order terms in #0, since the dependence of renormalized Green's functions on g0 is singular. Thus small errors in g0 may give rise to large errors in a

Green's function computed

as a

To derive the correct formula more

closely. So 9o



we

quantities. a-dependence of ft expansion of g0 in the form:

function of bare we must

examine the

write the perturbation

9^93lGll\n(a^)^Gi0 + Gl(a^)] +

6f5[G22ln2M + G21lnM+G20+G2M]+---

(7-10.9)

7.10 Choice

of cut-offprocedure

211

have not specified the renormalization prescription, so in addition logarithms we need finite functions Gx(a^\ etc. We have explicit constant terms Gl0, so we define Gt(afi) to be zero at afi 0. Once and from been have subtracted divergences Feynman subdivergences Here

we

to the



graphs, the remainders converge with power-law convergence in momentum. This is a consequence of our treatment of Weinberg's theorem, and is further treated in

Weinberg (1960). Therefore

say

we can

Gi(aii)=0((afir) as a/i->0, for some positive number logarithms of a\i, so we can safely set First we compute the ^-function:

ct



general, 1/2.

G-

equals

a\i times

l8go





c(. In



{<■■



d In

(afi)

Gi

2G22ln(^)



G21

(l+302[G11lnM + Glo+G1]



d\n(an)G\







d(fxa) -r<G2l-3Gl0Gll



dln(fia)

[G2-iG2l-3GllG1\n(^) -3G10e,]V +

The relation limit a->0

2G22 gives

P(g) so

-?

ft is

as

a-+0. The

P(g,0)=-g3Gll-g5(G21-3GloGli)+-,

(7.10.11)



must hold in order that

our





0, their logarithmic derivatives a^dG}/d(a^,) also go

The first

of g0

finite

usual notation Ax Gn and A2 G21 3GnG10. Since and G2(a/i) go to zero like a power of a\i (times logarithms) when a\i

that with

G^api)



3G\X

(7.10.10)

-.

step in

implies

our



to zero like a power.

calculation of g0 is to observe that the RG in variance

that

go(9>l*a)=g0(g(ji'),ii'a) =

The next step is to examine the size of the

g(l/a) by

the

(7.10.12)

0O(0(1?,1). error

in the

that is made in

replacing Finally, we will

effeptive coupling g{\/a) g0(g(\/a)y 1) must be computed in order to obtain

find the accuracy to which





0 theory.

212

Renormalization group

the correct renormalized Green's functions at The difference between the two effective

g(l/a)3

turn out to be of order

when





0.

is small. So let



and g(l/a) will define the fractional

couplings g(l/a) us

g2A by

error

g(^;^a,g)



/*' l/a and a->0. (7.10.13) and the definitions of p and p we find that

We will now show that A is finite when

From

/i'^7 AM



r3 [j3((l -





A(//,/i, a,#) is finite

AQi') We

now

as







3g2A)p(g)]

0(£ V*)1/2) + 0(54A).

/i'->oc,

when a->0 and /i =



g2A)g, pta)- (1

/i'fl^-5,

g2(fi') ̄ \/A1ln(fi'/A)

Now

(7.10.13)

g(fi';^g)[l+g(^;^g)2A(^';^aig)Ji.



so \i

(7.10.14)

this equation tells



l/a.

G^ci/i') + 0(l/ln(l/fl)).

compute g0. It is convenient

us

implies

In fact it

that

that

(7.10.15)

to write a formula for its square:

g0(g^a)2=g0(g(l/a),l)2 =

{g(l/a)(l



g2A) + g\\

02A)3[G1O + ^(1)]



0(g5)}2

^2ln[ln(l/a2A2)]

1 _ "

^fln^l/^A2"")]

[^t ln(l/fl2A2)] +



+ 0{ln2[ln(l/a)]/ln3(l/a)}. r72l^7T7^T27T [^2ln2(l/a2A2)]

(7.10.16)

#(/i) in terms of \i and A (7.5.5). The formula for g0 in terms of l/a and A is the same as (7.5.5) except for an additional 1/ln2 term. Observe that it was essential to keep the a-dependence in /%,a/i); the G^l) term in (7.10.15) canceled the G^l) in the two-loop Here

we

used the formula for





coefficient in g0.

Finally 9oW>W

terms of g, a and \i\

,4 2 In (In (l/a V))



 ̄Alln(l/a2fi2) +

where

(7.10.16) in

we express



,43ln2(l/aV)

2G10 ̄\-^\n(Aig2)-2AJ(g)

9 ,42ln2(l/aV)[ + 0{ln2[ln(l/a,<)]/ln3(l/a/z)},

fig)-

d?'[l//J(ff') 0



Ai

l/M,ff'3)-^2/M?ff')].

(7.10.17)

(7-10.18)

7.10 Choice of cut-offprocedure

213

Similar formulae hold for Z, for Zm, and for Y. Thus Z

[Alg2\n(l/a2n2)yCll2Al





y(g')

<fo'

exp



cx

1 + 0

 ̄ln(ln(fl/x))' J

\n(afi)

(7.10.19) where y



Cxg2



0(g4). 7.10.4 Accuracy needed for g0

Let

us now

suppose

we

compute the renormalized Green's functions:

GN(x19...9xN;g9m9fi;a) We must

now

let

approach



Z ̄Nf2G{0)N(xu...9x^r9g09m0;a).



zero, and ask how

accurately

(7.10.20)

we

need to

compute g0 and Z. In (7.10.17) and (7.10.19) we gave formulae for g0 and Z, with explicit estimates for the errors coming from uncalculated corrections.

These equations tell

us

a\i) when we let a -? 0 while keeping singular dependence on might affect the values of the renormalized

the value of g0(g,

g and [i fixed. Since the bare Green's functions have

g0, the uncalculated corrections Green's functions. In fact these terms do not affect the renormalized Green's

functions in the continuum limit, The

as we

will

now

key observation is that the renormalized

show.

Green's functions

are

finite

functions of the renormalized parameters. Thus we do not need to hold the renormalized coupling and mass fixed while taking the continuum limit a -?

0. We may in fact let them vary continuously, provided only that their a 0 are the same as before. Now examine (7.10.17). It is evident

values at

that

we



may absorb the whole of the correction term into

variation

of

In

g.

ln2(ln(l/fl))/ln(l/fl)

92o

as

fact

a->0. So

1 =

^lnO/aV) +

A\

the

necessary we

change

in



may choose the bare

just

is

such

of

coupling



order to be

^2ln[ln(l/aV)] A\\n2{\la2n2

IootK-?-^^-2^ (7.10.21a)

Hence in (7.10.16), So we have the

we can

following

also drop the

0{ln2[ln(l/a2)]/ln3(l/a)}

formula for g0 in terms of



^2ln[ln(l/a2A2)]

^1ln(l/a2A2)

A\\n2(\/a2A2)





2G10 ^2ln2(l/a2A2)



terms.

A alone:

(7.10.216)

214

Renormalization group

In the

case

corrections

of the wave-function renormalization Z, the uncalculated

can

be absorbed into



£iV multiplying

factor

the Green's

function GN. This factor must approach unity in the continuum limit. Hence we may use



A^ln



-CulAy



d0'

exp

a2M2

y(g')

P(g')





ci

ai9'] (7.10.22)

where M is

arbitrary coupling we had a form (7.10.21 b) that had dependence on A, but not on \i or on g. This was because g0 is renormalization-group invariant: we may take \i arbitrarily large without affecting g09 provided that we also set g equal to the effective coupling at /*. When /* is very big, g is very small, and the higher-order corrections contained in f(g) go to zero. But Z is not invariant; it must depend on g. What we can say is that any dependence on a an

mass

that is irrelevant when

a -?

0. Notice that for

the

of the form Z



finite-[ln(l/fl)]-Cl/2Al

will produce finite Green's functions. Notice that if the one-loop divergence in Z vanishes, then

we

may let Z be

finite: Z

There will in

orders. What



exp

dg'y(g'W(g')

general be divergences

in the self-energy

have proved is that

must sum to

we

In the case of g0i any

gl



l/Ax ln(l/fl2)



they ^-dependence of the

graphs in higher something finite.

form

A2\n[\n(l/a)]/[A3l\n2(l/a2)]



finite/In2^

give finite renormalized Green's functions. Only knowledge

of A2 and and two-loop Ax is necessary for this. They are obtained from calculations. The coefficient of the l/ln2a determines the value of g.

will

one-

The formula (7.10.216) shows the fundamental significance of the A- parameter. In a renormalizable field theory, there are divergences, so one

single number

cannot

simply specify

Rather,

one must construct



the theory

as

as

the bare

coupling

the continuum limit of

lattice theory, with g0 depending on the lattice spacing, (7.10.216) gives g0 as a definite numerical function of a.

Unfortunately, there is

certain arbitrariness in

a.

precisely

some

Equation how

one

lattice approximation to a continuum theory. This is physically irrelevant (although some particular approximation may

constructs £ arbitrariness



constant.

7.10 Choice

be superior when

it is used for



215

of cut-offprocedure

numerical calculation). So

(7.10.216) is also

important because it expresses the bare coupling in terms of quantities (A, Al9 and A2), which have direct meaning in the continuum theory, and in G10, which depends on the lattice approximation, but which can actually be computed analytically (Hasenfratz & Hasenfratz (1980)). The result of such a lattice calculation is normally given as the ratio

terms of one number

of a

Ajattice

to the value of A in some standard continuum renormalization

scheme. The definition of





Alattice is

the value for which

l/[>?1ln(l/fl2A12Bttlce)] -^alnpna/^A^JM^^^l/^A,2.^.)]

gives the same continuum limit as (7.10.21). Alattice



It is

(7.10.23)

easily checked that this

is

(7.10.24)

Aexp(G10/^1).

Despite the fundamental significance of A, there is a convention dependence in its definition. In specifying a theory by its value of A, one must specify these conventions. This is analogous in its effect to the need for specifying a system of units in electromagnetism. The main convention is that of the renormalization prescription. The other convention is the

one

implicit in the choice of the constant in (7.5.3). It is sensible to follow the usual convention, to avoid confusion. We have

seen

that

higher-order

corrections

(beyond

two

loops) do not (7.10.21a),

enter into our formula for g0 in terms of A. This is in contrast to

which

expresses g0 in terms of g and \i. So it is sensible to treat A as a fundamental parameter of the theory-say in strong interactions. But measure A. A typical which a useful for measuring quantity e ̄ perturbation expansion exists (for example, a jet cross-section in e+ annihilation). The experiment therefore measures the effective coupling g(ji) at some value of pi which is of the order of the energy of the experiment. There are errors in this measurement caused by uncalculated higher-order

practical

considerations intervene if

measurement consists of

one

tries to





terms in the theoretical calculation of the

perturbative corrections. errors

We

due to corrections in

Since

g is more

directly

can

cross-section, not to mention non-

then deduce A from

(7.5.6), with further

p. related to the size of the cross-section, it is

perhaps correct to argue that experiments should quote their results as a value of g. But to give the value of A is equally valid. However, a small fractional error in g corresponds to a much can

be

seen

from (7.5.6). If

\dA/A\



we

change

larger fractional error in A. This holding \i fixed, then

g and A while

\dg/g\[2Aig2yi[l



0(g2)l

216 If

Renormalization group

one

could do

then it is the mass.

As

we saw

transmutation, the

mass

mass of, say, the proton in QCD, in the value of A that would determine the error in the

real calculation of the



error

in Section 7.9, when

is

proportional

discussed dimensional

we

to A.

7.10.5

m20

Unfortunately ml has a 1/a2 term, but the variation of m0 with m2 depends [\n(l/a)] ̄hll/Al term. So we need the coefficient of 1/a2 to very high

on a

Any slight error (say of order 1/a) will be equivalent to making the renormalized mass diverge like 1/a as a-+0. The resulting need to be very accurate in m0 leads many people to consider scalar field theories accuracy.

unnatural. In the case of fermion theories there is a symmetry under m



0,

the Y term is absent and

so

m0



mZm

7.10.6



we

\jj

-?

y5\j/

when

have

[ln(l/a)]  ̄Bl/2Al.

m-constant

Non-asymptotically free

case

The values of g0, etc., as a -? 0 are not perturbatively computable unless the theory is asymptotically free. However, if we suppose that p in a non-

asymptotically

free theory has



fixed point, then

we

may write

(7.10.25)

^o(0*,l)as^O. Note that

the

same

g0(g(l/a), 1) is

which the limit is An



finite function of g, so the limit exists. However a 0 for any value of g{fi). So the way in

value is obtained for g0 at

example

approached



determines the value of g.

is easily constructed. Suppose

p(g)



we

have



sm2(g2)/(2g),

theory

in which

(7.10.26)

and

g0(sA) Then the effective coupling has 3

There is

fixed point g* a->0 must be a

9,o(9, W)









g.

(7-10.27)

the form

[arctan(l/ln(A/M))]1''2.

n1/2. We therefore find that the bare coupling

as

[arctan (1/ln (aA))]l'2

^l2-2^W)+°il/ln2{aA))-

(?-ia28)

217

Computing renormalization factors

7.11

It is necessary to know how

g0(g,an) approaches

its limit

g0(g*, 1)

in order

to determine the value of A.

7.11

Computing

renormalization factors using dimensional regularization

In the previous section, Section 7.10, we computed how the bare coupling g0 should behave as



function of the lattice spacing

a.

In this section

present the corresponding argument using dimensional continuation

we

as the

We do this by treating the defining equation (7.3.10) of p as a differential equation to compute g0. Our argument will be valid in any asymptotically free theory, like </>3 theory in six dimensions or QCD in four cut-off.

dimensions. If

let

we

regularize by going

d0

be the physical

space-time dimension, d0 s.

to a lower dimension d



then

we



First we compute the relations between lower and higher poles in the renormalization. Now we write

0o



^/2

£ dj(g)e-j\J

g+

(7.11.1)

j=i

and

we

 ̄p\

have the definition of

eg0/2 + p(g,d)dg0/dg Let us



0.

(7.11.2)

expand (7.11.2) in powers of e. The terms proportional to e and e° give

us:

p=- eg/2 We have

limit of

j5 ft

p{g)



changed notation from

the definitions that defined



we



eg/2

our



%gd/dg

original

definitions to correspond to

while

defined /?

as

we

the

pole e ̄j in (7.11.2) is

i(l -gd/8g)dj+l(g) + p{g)ddj/dg This is

we

the cut-off is removed.

Now, the coefficient of the



\)dx{g\

used in Section 7.10 for the lattice cut-off. There

to be RG coefficient in the cut-off theory, as





0.

(7.11.3)

differential equation which, when solved using the boundary

condition d.(0)



0, gives all the higher coefficients d.(g) in

single pole dx(g). Similar relations ft Hooft (1973)) hold

terms of the

for all renormalization counter-

logarithm expansion. They theory the only new information in the counterterm in a given order of perturbation theory is in the single pole. A convenient way of solving these relations is to work out the solution of terms. The structure is similar to the leading show that in each order of perturbation

the differential

equation (7.11.2).

ln[0o(0>/^ ̄£/2]=

This gives

fW[lg -2P(g)/e -VlJ Jo *



,W





faft

f7-11-4)

218

Renormalization group

i.e.,

g0/g^>n£'2 as #-?0 has been used. We now ask how g0 must behave as e d0 d 0, with g (and fx) fixed. If A xg3 + the theory is not asymptotically free (so that in P(g) *, A x is the then has a at negative), integrand pole The boundary condition





-?









g'2=-(d0-d)/2Ai+O(d-d0)2. (7.11.4) only unambiguously exists if g2 is less than this value, when d d0. To get to the d d0 theory with g non-zero we must continue g0 so that the integration avoids the pole. The result is that g0 has an imaginary part. This, among other things, suggests that the theory is unphysical (see Wilson (1973), Gross (1976)). Recently, evidence has The solution which is

zero





accumulated that the lattice

continuum limit



see

</>4 theory

Symanzik (1982) for

does not have a



non-trivial

review.

If the theory is asymptotically free then we may continue (7.11.4) to d d0, i.e., e 0. The integrand becomes singular when e 0, and we =





examine the singularity by expanding in powers of g'\

In0o



ln0

4-^ln/*

/%')] -|I>'{f^ A2g'5



l-eg'/l-A.g'^



\_-eg'/2-Aig'3y

"^lE-^-^^^-^-l^pv}. In the first

ln0o=^ln +

integral

we may set e



0 and have

errors

that

are

(7.11-6)

o(l). So

2A[

lh-^-2>(^hVw



o(s),

(7.11.7)

where/(g) is defined by (7.10.18). Thus

??-A-fe){i+#-in0)]+4

(7118)

7.12 Renormalization group

219

for composite operators

7.12 Renormalization group for composite operators We have

how

prescription for the interactions of a theory can be compensated by a change in the values of the seen



change

renormalized parameters. The

operators

we

in renormalization

same

property holds for the composite

defined in Chapter 6.

For example, consider the renormalized [02] operator in 03 theory at d 6. In Section 6.2 we calculated it in the one-loop approximation: =

<O|T0(x)0(y)[02](z)/2|O> =

tree

graph





{one-loop graph + counterterm graphs}

(7.12.1)

H—

A change in renormalization prescription amounts to a finite change in the graphs. Since the counterterms are of the form

counterterm

il>2] we





\&Za(f)2 + dZbm2(j)





3ZC □ 0



higher order, (7.12.2)

have

il>2]?ewRP



il>2]o.dRP + ??[>2]



b<j,



cU4>.

(7.12.3)

Here a, ft, and c are finite quantities of the same order in the coupling as the one-loop counterterms. The equation (7.12.3) is, so far, only derived at the one-loop order-so the finite counterterms are to be used with their

operators inserted in tree graphs. Let us examine the situation we expect to all orders. We will use minimal subtraction. Then the renormalization in the notation of

(6.2.12)

is

(7.12.4) Now the bare field

is







independent

of n,

so we may

write

^2-3ZbZ-l'2m2^0ti^-\n(fid'2-3ZbZ-1'2m2) ndl2-3zcz-ll2n<Pon-fln&'"2 ̄3zcz ̄112)

l[^2]^ln(Za/Z) +

/z"'2-3Z6mV^ln(//2-3Z6Z1/2Zfl-1m2) d^



/'2-3ZfD<^ln(//2-3ZcZ1'2Z;1), d/z

(7.12.5)

220

Renormalization group

which has the form

^i[*2]= v&W We

can

(6=|

formulate this

fi|>2]| <£



as a



yt>m2S'2 ̄3<t> + yy/2"3 □ *■

matrix

equation:

/z^-'m^Z-'-V'2-3 Z,Z_1/V/2"3| 0

D<M

<7-12-6)



Z"1'2





200

C0o/

Z-1'2

(7.12.7) dM



^—0 d/x



dp

The coefficients ya,y6, and yc Section 6.2, we have

where X





/64k



.,

1<b

finite at d

are

V 384rc3





(7.12.8)

y<p<b. 6. From





64^

?c



384tt3"6X/^



(7.12.9)

AM

.1



Thus

0 0



AbiV'2-3 y/2'3 a 3 6g 0



higher order.

arguments that

we

used to prove the

From the RG equation (7.12.7)

equations for Green's functions

% pT9 ̄ +

we

of the

same

m.

This follows from

property for ym and y.

prove renormalization-group composite operators. For example: we

ymW2i)<01 T<t>(x)<fi(y)K4>2W\o> =

where

(7.12.10)

i9

Observe that ya, yb, and yc are all independent of/z and same

calculations in







the

our

<*' 6'

ya

,_5 ?*=!

/z-— M

have used

(ya-y)<o\T<t>4>K<P2]\oy + (ybm2 + ycnK0\T<l>(x)4>(y)<P(z)\0>,

^.dtp/d/x

 ̄3 factors. eliminating the //'2





y<j>/2,

and

we

have

set

(7.12.11) d



6, thus

7.12 Renormalization group We must prove

generalizations

automatically

(7.12.8) both to

221

[02] operator and in its bare operators are Since operators.

all orders for the

to deal with any

RG invariant, the

for composite operators

only question

is whether the anomalous

finite. This is handled by a simple generalization of the proof given in Section 7.2 for the ordinary Green's functions. We will not spell out dimensions

the details

are



for that is

just



mathematical exercise.



Large-mass expansion situation in physics is that in investigating phenomena on a certain distance scale, one sees no hint of those phenomena that happen at A

common

much shorter distance scales. In a classical situation this observation seems evident. For

example, one can treat fluid dynamics without any knowledge physics that generates the actual properties of the fluids.

of the atomic

quantum field theory this decoupling of short-distance phenomena from long-distance phenomena is not self-evident at all. However, in



Consider an e+



e ̄ annihilation

experiment at a center-of-mass energy

GeV, the threshold for making hadrons containing the fe- There is, for practical (or experimental) purposes, no trace of the quark. existence of this quark. However, the quark is present in Feynman graphs well below 10

virtual particle, and can have an apparently significant effect on cross- sections. Our task in this chapter is therefore to prove what is known as the decoupling theorem. This states that a Feynman graph containing a as a

propagator for



momenta of the

graph is

field whose

mass

is much greater than the external by a power of the heavy mass.

in fact suppressed

physics at low energy is described by an effective low-energy theory that is obtained by deleting all heavy fields from the original theory. The decoupling of heavy particles is not absolutely universal. One important and typical exception is that of weak interactions. Let us

The

consider the interactions of hadrons at energies of a few GeV. The effective low-energy theory, in the sense just described, consists of strong and

electromagnetic interactions alone, without weak interactions. So weak interactions should be ignorable at low energies. However, it is well known that there are in fact many observed processes, particularly decays, that are due entirely to weak interactions. The point is that, in the absence of weak interactions, these processes are exactly forbidden by symmetries, such as parity, charge-conjugation, and strangeness conservation. Weak- interaction amplitudes for the processes in question would be power-law corrections suppressed by a power of energy divided by the mass of the W-boson were it not that they are corrections to zero. The consequence of this particular situation is that, at low energies, weak interactions are -



222

223

Large-mass expansion described by

raw-renormalizable theory, viz., the four-fermion



interaction. Efforts to find a

prediction of

a renormalizable theory led to gauge theories, and the W- and Z-bosons from phenomena at energies much

lower than their

clues In

low-energy phenomena have indeed provided as to what might happen at much higher energy. this chapter, we will treat the cases where decoupling occurs. The masses:

theorem that tells

expect decoupling to

us to

occur

in many theories

was

formalized by Appelquist & Carazzone (1975) and Symanzik (1973). They work with a renormalizable theory in which some fields have masses very

large compared with the others. They then consider Green's functions of the low-mass fields at momenta much less than the large masses. The theorem is that the Green's functions are the same as those in an effective low-energy

heavy fields. Corrections are smaller by a heavy mass. The sole effect of loops of heavy particles is that the couplings of the low-energy theory can have different values from those in the complete theory. Since the renormalized couplings have no particular a priori value, the theory by



obtained by

deleting all

of the

power of momentum divided

heavy particles

are

unobservable until close to threshold. The practical

importance of the theorem is that

without having



one can

understand low-energy

physics

complete Lagrangian for all phenomena.

applied in the the low-energy

We will also show how the renormalization group can be

computation of

the relation between the

couplings

of

effective theory and those of the full theory. There are many ramifications of the decoupling theorem, but

we

will not

treat these. One of these is the detailed application of the decoupling theorem to gauge theories (see, for example, Kazama & Yao (1982)). Another is the large-mass expansion of Witten (1976)-where Green's

heavy fields inelastic scattering.

functions of the

We will also not treat the

are

computed; this expansion is used is deep-

exceptions

to the

decoupling

theorem. These

techniques those used to prove the decoupling theorem itself. We have already mentioned weak interactions as one of the can

be treated by the

same

typical exceptions. Let

(1)

us

as

just

note two other main classes of

In theories with spontaneous symmetry

breaking,

a mass

exception:

is often made

dimensionless coupling (Veltman (1977) and Toussaint (1978)). The decoupling theorem assumes that a mass is made large by increasing dimensional parameters.

large by increasing



power-counting violate renormalizability of the low-energy theory (see Collins, Wilczek & Zee

(2) Some dimensionless couplings needed by (1978)).

224

Large-mass expansion

In any event the effective low-energy theory is non-renormalizable. It might be supposed that since General Relativity is non-renormalizable in

it contains some clues to phenomena at very high for energies (see, example, Hawking & Israel (1979)).

perturbation theory,

8.1 A model We will restrict

two fields in six if



simple model. space-time dimensions: attention to

our



very

(3^)2/2 + (dcf>h)2/2 m2ti/2 M20h2/2 /*3"d/2[0i</>i3/6 + 02</>i</>h/2] /2" V>i -

It is



03 theory with







+ counterterms.

(8.1.1)

Symmetry under 0h 0h has been imposed to cut down the number of Then possible couplings. (8.1.1) contains all couplings necessary for We assume that the renormalized mass, M, of the heavy renormalizability. field is made large while all other parameters are held finite. The factors of -?

the unit of



needed with dimensional

mass fj,

regularization

are

explicitly

indicated. All

techniques (realistic) theories. our

can

be

readily

As usual we have introduced

tadpole graphs. This



extended to treat

linear term in the

is determined

by

more

complicated

Lagrangian to cancel the

the renormalization condition that

<O|01|O>=O. The

remaining

counterterms can be put in the form

j5fc,=(Zl-l)^,2/2 + (Zh-l)a^/2 \m\Zm 1) + M2ZmM/2 \M\ZM 1) + m2ZMJ<p2J2 H3-dl2ttgiB 9i)<t>?/6 + (02B 02)Mh/2] (8-1-2) -/'2"3(/b-/)<£.? As usual,







and the

we

#B's)









may choose the dimensionless renormalizations to be

independent

(viz., the

of the dimensional parameters

Z's

m2, M2,

and /. The

decoupling theorem

on energy scales much effective low-energy theory whose

asserts that phenomena

less than M are described

by

an

Lagrangian has the form Se^



zd(f)i/2



m*2z4>\l2



ii2 ̄dl2g*z*l2<¥>\/*>



?d/2 ̄3z1/2f*<t>i

+ counterterms =

c0*2/2



m*20*2/2

+ counterterms.



^  ̄d/V0*3/6



\idl2 ̄ 7*</>* (8.1.3)

225

8.2 Power-counting Here

we

have defined

and the coefficient



z can

scaled field

be chosen

z1/2</>,. We will prove that g*, m* that Green's functions of </>, obtained

0*

so



Lagrangian (8.1.1) by by M. field </>* which has unit

from J5frff differ from those obtained from the full

terms which are of the order of a power of external momenta divided

It is usually convenient coefficient for its kinetic

functions in the full energy

to work with the scaled term in the basic

theory

Lagrangian.

Then Green's

related to Green's functions in the low-

theory by

GN(pl,...,PN;g{,gh,m,M^) =



as

are

M —?

oo

(0\Tfa(p1)... 0l(PN)|O>full theory



z-^2G*(Pl,...,pN;^,m*,/i)[H-0(l/Ma)] z-N!\0\Tr(Pl)...¥*(pN)\0)[l + O(l/Ma)l (8.1.4)

with p1,..., pN fixed. The fractional

errors

go to

zero as a power

of M times logarithms; the power is typically M ̄2. We can therefore use 1/Ma, with a slightly less than two, to bound the error. As is our convention,

the tilde signs

over

the fields and Green's functions indicate



Fourier

transform into momentum space.

8.2

Power-counting

finding tjie leading power of M a simple generalization of graph Weinberg's theorem, and will involve us in understanding which regions of momentum space are important. We will mostly be interested in graphs for In this section,

we

in the value of



will establish the rules for as M -> oo.

These form

the Green's functions of the light field </>,. Our aim will be to find those graphs that contain lines for the heavy field and that do not vanish as M goes to

infinity. 8.2.1 Tree

Because

we

containing

choose to

graphs

impose the symmetry <£h </>h, the only tree graphs heavy field have heavy external lines. An example is -?



lines for the

Fig. 8.2.1. Since all momenta

on

the lines

are

fixed, and since the free 0h-

propagator is i/(p2-M2) ̄ -i/M2,

-< Fig. 8.2.1. A

tree

graph with



heavy line.

226

Large-mass expansion

the behavior of any

given

tree

graph

as M -?

oo

is

M ̄2H,

(8.2.1)

where H is the number of

heavy lines. (We use the natural terminology of calling Feynman graph heavy or light according to whether its free propagator is for c/>h or c/>, respectively. Our graphical notation is that a

line of

heavy lines

are



thicker than light lines.)

8.2.2 Finite graphs with heavy loops

graph that has one or more loops but no ultra-violet divergences or subdivergences, and that has some heavy internal lines. The lowest-order example is Fig. 8.2.2(a), for the four-point function. Its external momenta (if small) may evidently be neglected on the lines of the Consider

now a

loop, whose value is then r.4 r, 2

(We

= -

V-M2)4 ̄3847r3M2'

(2n)6

label the symbol T by the





figure number.) 102 384ti3M2



(a)



Fig. 8.2.2. Large-mass behavior of

(b)

graph without

an

ultra-violet divergence.

oo. The precise power of M can be obtained the considering by possible regions of momentum space (after Wick as follows. rotation), Any region of k that is finite as M->oo gives a

The

graph

vanishes

as

M ->

graph is UV finite, the only other 0{M). Simple power-counting gives M ̄2, as found in

contribution of order M ̄6. Since the

possibility

is k



(8.2.2). This power-counting is the same as for the UV degree of divergence. The graph is negligible (by a power of M2) compared to graphs with no

heavy

wanted its leading contribution, then it the local four-point vertex symbolized in Fig. 8.2.2(b).

lines. If, nevertheless,

would be

we

effectively The non-renormalizability of this coupling (when the space-time dimension is six) is tied to the negative power of M2.

8.2.3 Divergent one-loop graphs Consider the logarithmically divergent vertex

graph

in Fig. 8.2.3(a). After

8.2 Power-counting

(?)

(*)

Fig. 8.2.3. Large-mass behavior of graph with minimal subtraction the

ultra-violet divergence.

dx

dyx

M2-(p2x + p22y)(l

■x-y)-

pixy

-}}

In

an

loop gives

+ 64n3 b



227

(8.2.3) The

same power-counting as for Fig. 8.2.2 confirms the power M° for the leading behavior as M -? oo. There is also a logarithm. This occurs because two regions contribute to the leading behavior: the first is where the loop momentum k is of order M. The second region is the UV region where

k ->

After subtraction of the ultra-violet divergence remains. oc.



finite contribution

Evidently the graph gives a contribution that increases with M. Fortunately the non-vanishing part of the loop is independent of the external momenta. So for large M, the loop is effectively a three-point vertex, as shown in Fig. 8.2.3(b). A proof which generalizes to higher order is to differentiate with respect to any external momentum. Since the differentiated

graph

is finite, it vanishes when M -> oo, like a power of M. decoupling theorem, that at low energies, we

Recall our statement of the

could calculate Green's functions from the effective low-energy Lagrangian (8.1.3). The result of our calculation of Fig. 8.2.3 is that the graph generates an extra

piece

in the

</>,3 coupling of the low-energy theory.

Let

us

therefore

write

z3/V



0i-

128ti3

y + ln

M2

4njn2



0(g5).

(8.2.4)

We may drop the graph Fig. 8.2.3 and replace it by the order g3 term on the right of (8.2.4). The loop has been replaced by a local vertex where all the

lines

come to a

single point.

line is far off-shell and

can

This corresponds to the fact that the internal only exist for a short time.

228

Large-mass expansion

Fig. 8.2.4. Large-mass behavior of graph with

The self-energy graph, Fig. 8.2.4, gives value of the loop is





quadratic ultra-violet divergence.

leading

term of order

M2. The

102

]}

f-"W1-*)]lnL A& ̄ i dx[M2-p2x(l -&{4"+"(&)]-fr+,-(£)] +

+ 0

(8.2.5)

M2

Again the loop momentum so

there will be at most

k can be either UV or of order M to



single logarithm of M2/fi2. Since

differentiate three times with respect to p^ before obtaining

graph, the non-vanishing terms, effective

Lagrangian (8.1.3),

ztn*2 We



m2

we see

that the

9l

convergent

quadratic in p. From the graph may be replaced by a

i[(z





768n3



have to

as M-> oo, are

contribution to the basic self-energy vertex the low-energy theory, with z=l-



contribute, we

g\M2 1287T3

i)p2



(m*2z



0(g%

m2)]

in

(8.2.6)



0(g%

(8.2.7)



O(05),

(8.2.8)

compute g* and m*:

can now

g* m



gl(g2

9i-:





*2

y +

Notice that there is



ln(^?III>



O(04)-

(8-Z9)

contribution of order M2 to the self-energy and

8.2 hence to m*2. In order to

m* finite

229

Power-counting

keep the physical mass of </>j finite and hence keep let m2 have a term proportional to g\M2 (with

as M -* oo, we must

higher-order corrections): m2 On

finite +



#2

expanding m*2 "*2 m*



M2





'128tc3L#

M2

1 + In  ̄

in powers of





"W

coupling,

we

higher order.

(8.2.10)

find

finite term in m2 '



Since



is the

unnatural

y +In

6Vl287tV mass

M2

4^

4-

higher

order.

(8.2.11)

parameter for the light field, it is generally considered

to have to fine-tune it within a fractional accuracy of

is required

m*2/M2, as

oo. In the by (8.2.10), context of grand unified theories this is called the gauge hierarchy problem (Weinberg (1974,1976), Gildener & Weinberg (1976)). It is hoped to solve it by finding a phenomenologically sensible theory with no need for fine- tuning. to obtain a finite value of m* when M-?

8.2.4 More than

one

loop

divergent one-loop graphs occurring inside a larger superficially convergent graph. A typical example is Fig. 8.2.5. When We may have

one

of the

Mg*z

\r- Fig. 8.2.5. Large-mass behavior of two-loop graph with M->co with the external momenta momenta that

gives



non-zero

an

ultra-violet divergence.

fixed, the only region of the loop

contribution is where the outer-loop

momentum / is finite and the

larger. So the heavy loop can

inner-loop momentum k is of order M or be replaced by its effective low-energy vertex

computed at (8.2.3). This procedure does not change the overall degree of divergence. The situation at subtle a

as we

with overall-divergent graphs is more The graph of Fig. 8.2.6 is typical. Now, it contains

higher

will now

see.

order

or

subgraph, consisting of the heavy loop, which we have already considered.

230

Large-mass expansion

?P03-O] Fig. 8.2.6. Large-mass behavior of another two-loop graph with

an

ultra-violet

divergence.

Therefore, the low-energy theory contains a graph where the heavy loop is replaced by a vertex using (8.2.4) for g*z3f2 g. This graph exactly reproduces the region where k is finite and / is large for Fig. 8.2.6. We add and subtract this graph from the original graph as indicated in the figure. The subtracted term (in square brackets) has a vanishing contribution from finite k (as M-> oo). So we replace it by an effective vertex A6. The same -

arguments

used for one-loop self-energy, Fig. 8.2.4, show that it has proportional to p2, m2, and M2, with coefficients polynomial in

as we

three terms,

ln(M2//x2).

graphs there are UV divergences for the whole graph and for subgraphs. Implicitly, the counterterm graphs are to be included. Provided we use mass-independent renormalization we are guaranteed that the counterterm graphs satisfy the same power-counting as the original graphs. In particular they are polynomial in the light masses. Thus the counterterm graphs do not change the power-counting and differentiation In this and in other

arguments that

are

crucial to

our

work.

8.3 General ideas Structurally, the arguments in the last section appear similar to those used in

Chapter

7 to show that

we

renormalization-prescription dependence

by finite counterterms. In fact, as we will see in the next a proof of the decoupling theorem can be constructed exactly by changing the renormalization prescription. We will show that a can

be compensated

section, Section 8.4,

renormalization prescription

properties,

constructed

couplings

can

important simply by deleting

and

masses

decouplings and subsection 8.3.1.

we

be chosen to have



number of convenient

of which is that the low-energy theory is

the most

all

heavy fields without changing the

of the light fields. This property is called manifest explain it with the aid of an example in

will

231

8.3 General ideas Our

follows the method given by

approach

Appelquist

& Carazzone

(1976). This approach generates the effective theory as a series of subtractions. The simplest non-trivial case is given in Fig. 8.2.6. There is another approach due to Weinberg (1980) in which the decoupling is considered by first integrating over the heavy fields in the (1975) and Witten

functional integral. (See also Ovrut & Schnitzer (1980).) This method is less convenient for

treating graphs like Fig. 8.2.6,

so we

do not

use

it.

8.3.1 Renormalization prescriptions with manifest decoupling Suppose we used BPH(Z) renormalization instead of minimal subtraction. Then the renormalization condition is that the terms up to pd(f) are zero in the Taylor expansion of a graph T about zero external momentum. Here

5(F) is the degree of divergence. For a graph with a single loop, consisting of a heavy line, these terms are precisely those that are non-vanishing as M-> oo. Examples are given by the graphs of Figs. 8.2.3 and 8.2.4. In fact, for a general graph, the effective low-energy theory in this renormalization presciption is obtained merely by deleting all graphs containing heavy lines, together with all their counterterm graphs. The values of the couplings and masses are not changed. Therefore the BPH(Z) prescription has the property we called 'manifest decoupling'. It might appear sensible always to use a renormalization prescription that has this property. However, for many purposes it is useful to use other renormalization prescriptions, e.g. minimal subtraction and its relatives. Particular cases

are

theories

containing massless fields, especially

non-

abelian gauge theories, and theories with spontaneous symmetry breaking. In any case, it is good to have a direct method of proof of decoupling that can

work with any

subtraction is

prescription. Furthermore, convenient if



prescription

like minimal

high-energy behavior (Section 7.4) with the aid of the renormalization group. In fact, the method we will use will start from a mass-independent renormalization prescription defined for both the full theory and for the effective low-energy theory. Then the renormalization of the low-energy theory is extended to a renormalization prescription of the full theory in such a way as to satisfy manifest decoupling. This method was first stated by Collins, Wilczek & more

one

also wishes to compute

Zee (1978). One renormalization prescription that gives manifest decoupling at low energies and that allows the use of renormalization-group methods at high energies is due qriginally to Gell-Mann & Low (1954). In this scheme,

makes subtractions at

some

arbitrarily

one

chosen value of momentum. This

232

Large-mass expansion

applied to the large-mass problem by Georgi & Politzer (1976). The disadvantage of this scheme, compared with the scheme that we will actually use, is that renormalization-group coefficients are explicitly functions of M/p, and of mj\x\ scheme was

Pi=Pi(9i>92'>M/iJL,m/ii). This

makes calculations

some

complicated. Furthermore,

this scheme obscures

symmetries. 8.3.2 Dominant regions

a proof of the decoupling theorem, let us give a that give unsuppressed contributions (i.e. regions precise not suppressed by a power of M2). We consider each graph in the full theory together with the set of subtraction graphs needed to cancel its divergences and subdivergences. We do not consider the subtraction graphs separately.

Before actually constructing statement of the

First of all, any

graph

with

no

heavy

lines at all contributes without

suppression. A graph with one or more heavy lines cannot give a contribution unless at least one loop momentum is of order M. The contribution to a graph when M is large can be considered

possible regions of momentum

as

the

sum

space. The

of contributions from various

regions

can

be

specified by

the

sizes of the loop momenta. For our purposes, it is enough to classify a momentum as either finite or large. 'Large' we define to mean 'of order M or

power-counting for each region in the obvious way. For the loops carrying large momenta, counting powers of M is the same as for the ultra-violet degree of divergence. This gives a factor Md, where S is the bigger'.

We

can

do

ultra-violet degree of the lines carrying large momenta. A heavy line carrying finite momentum counts as M ̄2. A light line carrying finite momentum counts as M°.

power of M for a graph is obtained as the maximum of the possible regions. Let us define SM(T) to be this highest power. In general there will be logarithmic enhancements. But Weinberg's theorem The

leading

powers for the

guarantees that the power SM(T) is correctly given by considering only the regions we have listed. The graphs treated in Section 8.2 provide examples of this procedure. (The subscript 'M' is to distinguish 5M(T) from the ultraviolet degree of divergence.) A region contributing to a leading power that is M° or bigger is symbolized by contracting to a point both the heavy lines and the lines carrying large momentum. The points represent vertices in the effective low-energy theory; we have already used this notation in Section 8.2, in the

8.4 Proof of decoupling

Fig. 8.3.1. Graph with two contracted subgraphs,

233

Fig. 8.3.2. One-particle reducible sub- graphs may have to be contracted.

figures. In general (see Fig. 8.3.1) the contractions will result in several vertices. We include in our definition the restriction that a subgraph is only contracted to a point if it contains at least one heavy line. A contracted

subgraph is 1PI in the light cutting a light line then that

lines; for if it

can

be split into two parts

line is not carrying



large loop

by

momentum.

However, the contracted graph may be 1PR in the heavy lines. For example, in a theory where the symmetry </>h </>h is not valid, a graph -?



power M°; the self-energy gives a power M2 in the propagator. A subgraph that is contracted to a single vertex gives the same power of M as its UV power-counting. So

like

Fig.

8.3.2

gives

which cancels the



leading

1/M2

dim (subgraph)



power of M + dim (couplings).

Hence in a renormalizable theory (where couplings have non-negative dimension) the only contracted graphs that have a non-vanishing value as M -? oo correspond to vertices whose couplings have non-negative

dimension. These vertices

give

the difference between

g*z3/2

and g, etc. Thus the

low-energy theory satisfy the dimensional renormalizability. In a scalar theory, this implies actual renormalizability, provided all the couplings are used that have non- negative dimension and that obey the symmetries of the full theory.

couplings

in the effective

criterion for

8.4 Proof of

decoupling

8.4.1 Renormalization prescription R* with

manifest decoupling

us work with the theory defined by (8.1.1). We choose to renormalize according to a mass-independent prescription, which we will denote by a symbol R. For definiteness we choose this to be minimal subtraction. By deleting all heavy fields from (8.1.1) and by changing the values of the couplings we obtain the form of expected low-energy theory (8.1.3). We choose to renormalize the low-energy theory by a mass-independent prescription R*, which we also take to be minimal subtraction. Our proof will consist of extending R* to a renormalization of the full

Let

it

234

Large-mass expansion

theory. The

satisfy manifest decoupling. Since different prescriptions differ only by a reparametrization, the statement (8.1.4) of the decoupling theorem will hold. The structure of/?* in the full theory will give a 'mass-independent' form for g*, m*2 and z: extension will

renormalization



0*=0*(01>02>MM z



m*2





z(gl9g29M/ri9

m2zJgvg2,M/ii)



(8.4.1)

M2zmM(gl9g29M/fi).)

independence independence of the light mass. As before, to complication we choose to renormalize the linear coupling 0. We then ignore both the linear f(f>} by the prescription that <0|</>,|0> coupling and the tadpole graphs. The reason we use mass-independent renormalization prescriptions for all the couplings other than the term linear in 0 is that we can thereby make very clear the decoupling of phenomena at small mass scales from large mass scales. In addition, the renormalization-group equations for (8.4.1) are much simpler to work with than they would otherwise be. Mass

save

means

notational



It is convenient to define two concepts: (1) A heavy graph is heavy field </>h).

(2)



light graph is

one

containing at least one heavy line (i.e. a line for the

one

that contains

no

heavy

lines.

theory we have a series of counterterm graphs that are used to cancel its divergences. If T is a heavy graph, then we also consider its counterterm graphs to be heavy graphs, even though they may contain no explicit heavy lines. We have chosen a renormalization prescription R* for the low-energy theory. This defines the renormalized value of any graph in the low-energy theory, and therefore of any light graph in the full theory. We now wish to For each basic

graph

T in the full

extend this prescription to heavy graphs, in such a way that it satisfies manifest decoupling. That is, the renormalized value R*(T) of a heavy graph goes to zero

as M -> oo.

The basic idea is to subtract such

momentum. That this is a sensible

procedure is easily

graphs from Section 8.2. example, we saw that Fig.

seen

graphs at zero by examining a

few of the For

M-?oc,

if

subtraction

we



we use

8.2.3 diverges

logarithmically

when

minimal subtraction. But with zero-momentum

have

0(p2/M2)

as M -> oo.

(8.4.2)

235

8.4 Proof of decoupling

Clearly, the difference between the difference given by (8.2.4) for g*z3/2 minimal subtraction

two -

renormalizations is

high

g. At

just the

energy we would use

be neglected compared with momenta would use the R* prescription so that we can

so M can



but at low energy we simplify calculations by





dropping heavy graphs.

Fig.

8.4.1.

Fig. 8.4.1 for the self-energy of the heavy field. This graph light and heavy lines. It would behave like M2 In (M) for large

Consider next

contains both

M, if we used minimal K{1"'

subtraction. Instead, let

us

define the subtraction by

64n3(4nfi2Y'2-3

dx{[M2x + m2(l-x)-p2x{l-x)]d'2-2-(M2x)dl



2-2

-{d/2-2)(M2x)'"2-3[m2(l-x)-p2xil-x) ̄\}

dx\\_M2x I

—^773

^6)64n3J0 :ln =

m2(l-x)-p2x(l-x)]x

m2(l-x)

p2(l-x)

M2x

M2

1+-

0(1/M2)





m2{\ -x) + p2x(l



x)

(8.4.3)

asM->oo.

Here, we observed that when M graph is linear in m2 and

-> oo

p2.

So

the dependence of the unrenormalized we

about p

expanded





0 and



quadratic in m and p. This means that the polynomial in m, i.e., 'mass-independence' holds good.

subtracted the terms up to

counterterms

are

Normally subtractions at zero mass and momentum have infra-red divergences, but the presence of a heavy line prevents this here. As a final example let us examine the two-loop graph of Fig. 8.2.6. The unrenormalized

graph

cannot be

counterterm, because there

p-dependence



,,,

are two

expanded about m light lines. At m, p  ̄

(value of heavy loop at p fc 771 2irv + Ln: 21 (k2 m2)[(p k)2 m2] =



The







0 to

they give

give



m-

and



2\

of the form

d k- ?>m>0







0)

" ̄



(P







rn2)ln(p2 + m2). 2,,



(8.4.4) right-hand

side of this equation is schematic and

symbolizes

the

236

Large-mass expansion

o o W

(b)

Fig. 8.4.2. Counterterms for Fig. 8.2.6. maximum powers and logarithms of obtain

the

renormalized

value



and of p that

of the

graph

we

occur.

However,

to

first subtract

must

subdivergences by the counterterms Fig. 8.4.2, constructed by the R*- scheme. Now, the counterterm graph (b) has infra-red behavior exactly equal and opposite to that of (8.4.4), because the counterterm is minus the value of the heavy loop at p k 0. Thus the sum of the two graphs has the —



extra convergence we need. The overall counterterm is then linear in m2. There are no logarithms of m as M -> oo.

8.4.2

p2 and

Definition of R*

To define the renormalization

summarize and generalize what We define the renormalization

prescription we

R* in

general,

we

simply

have just done for particular graphs.

prescription R* in the full theory to be the for the low-energy theory whenever it acts

prescription purely light graph. For a heavy graph T, we assume inductively that we have defined the quantity R*{T) in the usual way to be the unrenormalized value of r plus counterterms in the K*-scheme to cancel its subdivergences. same as our chosen

on a

If T has degree of divergence S(T)



0, then its overall

counterterm is

by subtraction at m p 0. The renormalized value of T is R*(T) + C*(T), as usual. To define C*(T) precisely, we first expand R*(T) in a Taylor series about the point where its external momenta and the light mass m are zero. Pick defined

R*(T)







with dimension up to 5(T), and let the counterterm C*(T) be the negative of these terms. Our examples tell

out the terms where momenta and m2

us to

expect that with such

(1) the leading (2) there

M->

are no IR

oo

occur

a counterterm:

behavior is canceled,

singularities

in the counterterm.

general. The proof will generalize from Fig. 8.2.6. There, the UV divergent

We must prove these statements in

the simplest non-trivial

case,

unrenormalized graph is not polynomial in



and p, but after subtraction of

subdivergences by the fl*-scheme, it becomes polynomial. Then the R*- prescriptioi} can legitimately generate the overall counterterm. Moreover, after subtraction of the subdivergences, the leading large-M behavior is also

8.4 Proof of decoupling

237

polynomial in m and p with degree equal to the degree of divergence, so that it is cancelled by the overall counterterm. Even with the subtractions for

subgraphs, there are in general IR in the singularities Taylor expansion of a graph. For example, consider Fig. 8.4.1 and expand its integrand see (8.4.3) in powers of m2 and p2. All the terms beyond the second give divergences at x 0; it is only the terms needed to cancel the UV divergence that are non-singular. -





8.4.3 IRfinitenessofC*(r) Suppose r is a heavy graph, 1PI in its light lines. Potential infra-red divergences in C*(T) arise when m and the external momenta are made small. They come from regions where some or all of the loop momenta are of order

m.

The

simplest case is where all the internal momenta are of order

m.

If r

were a

light graph,

we

would obtain



contribution of order

where d(T) is the UV degree of divergence. So let us call canonical IR degree of divergence of T. If d(T) 0, this is a =

m3{r\

d(T) the logarithmic -

divergence. If (5(T)>0, then the graph is finite as m-?0. But to get the coefficients of the polynomial counterterms we differentiate up to S(T) times with respect to m and the external momenta. The highest terms in the polynomial are therefore always logarithmically IR divergent, for a light graph. However, r is actually a heavy graph. So at least one of its propagators counts

as

1/M2 instead of 1/m2. Thus all the counterterms

contribution from this

region,

where all its

This discussion is sufficient for all

graphs have small

have

loop one-loop graphs.

divergences coming momenta. For example, Fig.

IR finite

multi-loop loops divergence from the But

from regions where only

have IR

an

momenta are small.

8.2.6 has



seme

region where p and k are small, i.e., order m, and / is finite or large. This 2, and is given by (8.4.4). As we saw, the IR corresponds to IR degree the is canceled divergence by graph with a counterterm for the heavy loop. The general case is that some light lines carry momenta of order m and —

(a) Fig. 8.4.1, with k and p of order (b) Fig. 8.2.6, all loop

momenta

(c) Fig. 8.2.6, k and p of order





large



V?^









Fig. 8.4.3. Examples of reduced graphs.

^—

238

Large-mass expansion

the remainder of the lines either

are heavy or carry large momentum. Each region is symbolized by a reduced graph in which the subgraphs consisting of the lines with large momenta and of the heavy lines are contracted to points. Examples of reduced graphs are shown in Fig. 8.4.3.

such

Note: (1) Counterterm graphs can also have infra-red divergences. The counter- terms are inside the vertices of the reduced graphs. All lines of reduced

(2)

graphs are light, so at graph corresponds to a heavy subgraph.

least

one vertex

of

reduced



write the infra-red degree of divergence for the region corresponding to a particular reduced graph y as We

can

S1R(T;y)= ̄S(r)+

[*W +IR degree of K*(P)].



(8.4.5)

reduced vertices V

The

meaning of this equation

can

be

seen

from

an

example.

Consider

Fig. 8.2.6 when k and p are of order m. If the graph were purely light, we would have IR degree equal to 2, which is the negative of the UV degree. —

This would imply that the m2 and p2 terms in the expansion about m would be

divergent. However,

the

single reduced

vertex



as









illustrated in

Fig. 8.4.3(c) has a counterterm. This counterterm ensures that the vertex's value is of order m2/M2 instead of m°. The IR degree for the whole graph is -

thereby decreased by 2. The second term in (8.4.5), where the sum is over this 4; single vertex, indicates this reduction. The degree for the region is then therefore expand up to order m2 and p2 without an infra-red —

we can

divergence. The terms of order are not

m4, p4, etc., are infra-red divergent,

but they

needed for ultra-violet renormalization.

general case of (8.4.5), each reduced vertex V would contribute if it were light and all its internal lines had momenta of order m. But d(V) actually contributes what we will now prove is a smaller amount.

In the -

it

Remember that counterterm counterterm vertices

R*(V) is its power cases for V are:

graphs

also contribute, and

we assume

that

included inside reduced vertices. The IR degree of its external momenta are scaled like m. The possible

are

as

(1) If V is overall convergent and contains a heavy line, then its infra-red degree is greater than its ultra-violet degree. Fig. 8.4.3(a) has a vertex 2 and IR degree zero. with UV degree If and contains a heavy line then ordinarily we V is overall divergent (2) -

would expect it to behave

as

m° when

m ->

0 with fixed M. But

we

make

239

8.4 Proof of deco upling

subtractions by the R* scheme so that its behavior is actually m3{V) 2. By induction we may assume its subtractions have no IR divergence. +

Hence,

heavy graph T always has at least degree below d(T) and none to increase it.

in every region of momenta

one mechanism to reduce its IR





Thus the overall counterterm C*(T) is IR finite. It is crucial to

proof

that

we

inductive

our

first subtract subdivergences by the R* scheme.

8.4.4 Manifest decoupling for R* A purely light graph is



graph in both the full theory and in the low-energy

theory. heavy graphs vanish when

It survives unaltered when we let M -?

the



->

oo,

given

oo.

now prove that all renormalize them by

We will

that

we

the R* scheme. To do this, decompose each heavy graph into its skeleton, i.e., a series of graphs connected by lines that are not part of any loop. Since a heavy

1PI

line

loop vanishes as M-> oo, all heavy graphs vanish as graphs vanish. limit of a 1 PI graph can be related to an IR limit by scaling

that is outside



M-> oo, if the 1PI

The all



-> oo

masses

and momenta: M-*l,

p^p/M,

m-+m/M.

Then T(p, m, M)



(8.4.6)

Mdir)T(p/M, m/M, 1),

where d(V) is the dimension of T. So T vanishes as M -> oo infra-red behavior is less singular than m ̄d(r). But this is what

provided we

the

showed in

the proof of IR finiteness of the counterterms. (Note that the dimension of a

graph

is greater than

or

equal

to its UV

degree of divergence.)

8.4.5 Decoupling theorem We have constructed two renormalization

prescriptions, labelled R and R*,

for the theory under consideration. The Green's functions in the schemes R and R* are equal provided we make appropriate changes in the parameters:

0i-?0*, 0h-*0h> coefficient of dcp\/2

-?

z,

oid^l/l

-?

zh,

coefficient

m2^>m*2, M2-+M*2

(8.4.7)

240

Large-mass expansion

just a particular case of a renormalization-group transformation, and is proved by Section 7.2. When M may drop all heavy graphs in the R* scheme (so also g£, zh, M* drop out of consideration). This then gives (8.1.4), which is the decoupling theorem. Mass-independence is true because we have arranged all counterterms to be polynomials in the light mass of the appropriate degree. This is

-* oo

8.5

When

one

computes



we

Renormalization-group analysis graph containing lines for fields with widely different

finds, in general, that its value gets large as a power of the of the mass ratio. Such large coefficients are undesirable in a logarithm

masses,

one

perturbation expansion,

for

they

mean

that the reliability of

using



few

low-order terms is worsened. This situation arises in both strong- and weak-interaction

physics.

We will now show how to combine the

decoupling theorem and the renormalization group to do calculations without their being made unreliable by the large logarithms. A convenient method is to use a mass-independent scheme (specifically minimal subtraction) for high-momentum calculations, where one often wishes to neglect all masses, and to use the R* scheme, as defined in Section 8.4, at low momenta, where one wishes to neglect heavy graphs. An

advantage of this method is a simplification of many of the calculations high-energy and low-energy calculations. One needs only the pole parts of graphs and the values at zero external momentum. needed to match

We will explain how to use this scheme in the toy theory (8.1.1). First let us write the RG equations for the Green's functions. For a Green's function of Nx

light

and Nh

heavy fields,

we

have

{% +±2Niyi+iiVh7h) G?>=a

where



(M2yM +

m2ymM)^



(m2ym +

M2yMm)-^.

(8-5'1)

(8.5.2)

The RG coefficients are obtained from the renormalization counterterms

usual. Their lowest-order values

are

as

8.5 Renormalization-group analysis

Pl

1 =

128ti

3(502



Vi



yh





0192—z92i92) +

241



+ -.

02 mn3*2 1

In the effective

(8.5.3)

384rr

low-energy theory,

the RG equation is (8.5.4)

with d*











(8.5.5)

464tc3 *2

7*





384tc

r +

To compare the low-energy theory and the full theory, we extended the renormahzation scheme of the low-energy theory to a renormahzation scheme R* for the full

d*





theory.





?.

-(M*2y*M +

In this scheme the RG operator has the form

m*2y*mM)^i  ̄i^*2y* M*2y*Mm)-^2, +

(8-5-6)

y* and y£. In fact /?*, y*, and y* are low-energy theory (see (8.5.4) and (8.5.5)), while f}% y*M y^m y£ y£ 0. This is easily seen by examining the Green's functions which provide the normalization conditions for the and the anomalous dimensions of the fields

are

identical to those in the =









renormalizations.

heavy propagator when

For example, consider the inverse of the

and m2

are

both

p2

much less than M2:

1/G0i2= -i[p2-M*2



0(p\m2p\m4)]

(8.5.7)

which satisfies

^ ̄Go,i This is only consistent if y*



yti





ytGZ,l

y*M



0-

(8-5.8)

242

Large-mass expansion 8.5.1 Sample calculation

We wish to start with the full

theory renormalized by minimal subtraction. theory, we know the evolution of the couplings. Our aim is to compute Green's functions in the low-energy theory and the values of the mass and coupling. The low-energy effective couplings are In that version of the

?2/

128*3



3 In (p/A*) =

fixed,

m*2



constant

M*2



fixed.

qI

(8.5.9)

[In (/x/A*)] ̄1/9,

The effective couplings for the full theory with minimal subtraction are more complicated because they solve a coupled equation for two variables. To make the transition between the schemes we compute the lowest- order divergent graphs. We equate the self-energy for 0, in the two schemes,

with

use

of the Lagrangian (8.1.3) for the low-energy theory. This gives

Fig. 8.2.4 4- pole counterterm =

Fig. 8.2.4

4- zero-momentum counterterm 4-

We thus obtain



and not of order

and m*2

M2,

m2



as

i(z



l)p2

given by (8.2.6) and (8.2.7). replace m2 by

To



i(m*2z



m2).

keep m*2 finite,

we must

T^pM2^2[y-l +ln(M2/4^2)].

Notice the presence of logarithms invalidate the use of perturbation

(8.5.10)

of M/p. If they are large enough, they theory to compute g*, m*2, and z.

However, the equations

we write are valid at any value of p, so we may with the calculations p of order M. After computing g*, m*, and z in perform

terms of g{, g2, m, and M,

we can

evolve them to the value of p that

for calculations in the low-energy theory. point to do the matching is where g{ g*9 i.e., at where pi M2ey/4n. Then for a general value of p we have to

we

wish

use

A convenient



p2



pi,



^2(/z)=l/^(/,0) A similiar



(3/1287r3)ln^0)+--

(8-5,lla)

equation holds for m*: m*2(Ai)

The solution for





is

m2[(1287c3/3ff?) + lnOi/Ai0) + -]"l/9.

more

(*.5.Ub)

complicated since the renormalization-group prescriptions is needed.

equation for both renormalization

8.5

243

Renormalization-group analysis 8.5.2 Accuracy

We compute

g*(fi) in

low-energy theory by matching to the full theory evolving to an arbitrary renormalization mass \x from fi0. Given the accuracy in g*(fi) that we need for a particular calculation, we will find the order to which we must perform the matching the

at some fi0 of order M and then

and to which

The RG

we must

know /?.

equation for g*(fi) gives ln(fi/tt>)=

So if there

are

small

errors

fff(M) J

(8.5.12)

do'/Pig*).

0(Mo)

Ag(fi0) and A(l//?) in ^(^0) and 1//? then the error

in g(jj) is

Suppose we perform matching up to nm-loop order; then the error in g(fi0) is + of order g(fi0)2nni 3. Suppose ft is computed to rc^-loop order; then the error in

\/fi(g)

is of order

g2nfi ̄3.

These translate to

errors

in g(fi) of order

QitfgM2"'" and

g(fi)3\n[g(fi)/g(fi0)]\infi



l,

or +

0[g(fiV 2n>-\ For we

example,

if

we



wish to

O[ff0i)3^0io)2("'-1}] if ?, perform

need Ag to be much smaller than

>2.

reliable two-loop calculations, then This means that we need to do the

g3.

one loop and that the /?-function is needed to two loops. minimum This is the accuracy needed to correspond to a fractional error on A (defined in Chapter 7) which is much less than unity.

matching correct to

9 Global

In this chapter on

we

symmetries

consider the impact of global symmetries of a field theory an example consider the theory of a charged

its renormalization. As

scalar field : <£



dtfd<f>



m2^(f)



(9.0.1)

g(ft(t>)2/4.

e"iw</>. The quantum theory is also invariant. For this particular theory, the quantum invariance is not a very deep statement. However, symmetries do This classical Lagrangian is invariant under the transformation 0

not

always survive quantization,

as we

will

see

in

Chapter

-?

13. Thus it is

useful to examine the consequences of the symmetry in this theory. One consequence is that only invariant counterterms are needed; for example, we

do not need to

use

non-invariant counterterms proportional to

02 + 0t2

or

i((f>2



<f>u).

Other consequences are the Ward identities, which characterize the action of the symmetry at the level of Green's functions. The main step in proving the statements is to impose an ultra-violet

putting the theory on a lattice or by using dimensional regularization, the symmetry is preserved. The arguments given in Section cutoff. If this is done by 2.7

are

sufficient to prove Ward identities in the bare theory. From the

invariance of Green's functions follows invariance of the counterterms. As we

will

see

in Section 9.1

we can

then write renormalized Ward identities in

the renormalized theory, which therefore exhibits the symmetry. In

cases this simple procedure fails. is that the UV cut-off breaks the symmetry. For example, the theory on a lattice breaks Poincare invariance. Luckily, other

more

One

putting

general

case

regulators, like dimensional continuation, preserve this invariance, and the renormalized theory with no cut-off is Poincare invariant. Some symmetries cannot be preserved after quantization. It must be true that no regulator can preserve them. An example, to be treated in Chapter 13, is the chiral invariance of QCD. 244

9.1 Another case, which

symmetry breaking,

by the



we

245

Unbroken symmetry

will treat later in this

typified by the theory given

is of spontaneous by (9.0.1) with m2 replaced

chapter,

m2. This is called the Goldstone model. In this case the ground-state

vacuum

vacuum



is not invariant under the symmetry, and the field acquires

expectation





value:

<O|0|O>



\2\m2\lg]X12 + higher

order.

If we use an invariant regulator, like dimensional continuation, we will still be able to prove Ward identities. Hence, we will be able to prove that only symmetric counterterms are needed, so that the symmetry is preserved.

From the Ward identities follows Goldstone's theorem, that there is massless boson for each generator of





broken symmetry.

9.1 Unbroken symmetry a totally unbroken internal symmetry. The fields representation of the generators. Thus:

We first consider matrix

carry

S*<t>i=-i(Qij<t>j, in the notation of Section 2.6. The proof that the symmetry elementary. We in less trivial

spell

out the

steps

can so



(9.1.1)

be preserved under quantization is we can see what needs to be done

that

cases:

(1) Regulate in a way that preserves the symmetry. Lattice and dimensional regularization both do this since the symmetry commutes with all space-time transformations. Include in <£ all possible invariant counterterms up to the appropriate (2) dimension. Thus 8aJ? 0. For the model (9.0.1) we replace jS? by =

<£



Zdftdcf)



m2<t>i(f>



0B(</>W/4.

(9.1.2)

perturbation theory, let the free Lagrangian be invariant: SaJ?0 0. Then the interaction Lagrangian is also invariant. (4) At each order, choose the counterterms to cancel the divergences in 1 PI

(3)

To do



Green's functions. Since the free propagators and the interactions are divergences are symmetric and

all invariant under the symmetry, the

non-invariant counterterms

are not

needed.

(5) Remove the UV cut-off. The Green's functions

are

^<O|T0ji(>;1)...0J.N(yJv)|O>=O. an arrow

(9.1.3)

(9.0.1) the propagator for the charged field the direction of flow of charge. All vertices have indicating

In the case of the model

carries

symmetric:

246

Global symmetries

equal numbers of ingoing and outgoing lines. In and d(f)f i(f)\ so this equation is literally

(9.1.3) a



we

have

6(f>





ic/>

of charge

statement

conservation. The

current

for



symmetry is

defined

by

Noether's

theorem

(Section 2.6): &

In the

case

^<wz-



of the simple model (9.0.1) there is



(914) a

single

(9.1.5)

We derived the Ward identities of the bare theory

(9.0.1) these

(2.7.6). For the theory

are

^<0| Tf(x)<t>(yiy ■■<t>(yf,)4>\ziy =

current

iZ</>t^.





-0t(zw)|O>

i£ [6(x-yj)-S(x-zJfi(0\T<t>(yiy--<l>(yN)4?(ziy--<l>\zN)\0>. (9.1.6)

We showed in Section 6.6 that the current is in fact finite;

renormalization counterterms factor Z in (9.1.5). It is of interest to

see

are

needed beyond those

how the divergences that

the factor Z. For the two-point function of/ Fig. 9.1.1, up to order Z by 1 everywhere

g2.

Since Z

except



1 +

0(g2)

in the tree

no extra

implied by

the

are

present get cancelled by

we

have the IPI

in this theory,

we

graphs

may

graph (a). Graph (b)

of

replace

could be

logarithmically divergent by power-counting, but is in fact zero,

so no

is needed at order g. Graph (e) is also zero. Graphs (c) and (d) are finite after their subdivergences are cancelled by a counterterm; they also cancel each other. These cancellations arise since these graphs have a

counterterm

< <X (?)

(b)

<X <^K (c)

[d)

0CX<J <K (?)

(/)

(9)

Fig. 9.1.1. Graphs up to order g2 for the two-point function of y.

Spontaneously broken symmetry

9.2

subgraph

which is

for

graph



247

<0|T/*(x)4>tyO>)|0>. In momentum space this is of the form

that its divergence is

q*f{q2).

The Ward identity implies

zero:

^<O|T/(x)0VW|O>=i(5(x->;)<O|<5(0V)|O>=O, that

so

q2f{q2)



0.

Graphs (/) and (g) each have graph of the form (b), which is cancelled by using the order

9.2

which is cancelled by a Their overall divergence must be

subdivergence

zero.

Spontaneously

broken symmetry

symmetry it will be sufficient to consider the if





graph (a).

term in Z in

the renormalization of theories with

explain

To

g2



Zd^dty + m20V

case

spontaneously

g(ft(f))2/4 + 5m2^(f) 5g(^(f))2/4 + + (^t)2/2 (502)2/2 m2(cf>2 + 022)/2 g(<f>2 + 02)2/16 -





(9.2.1)

+ counterterms.

Here (f>



broken

of the Goldstone model:

have written the complex scalar field in terms of real fields: ((j)l-\- i(f>2)2 ̄1/2. The mass term is of the 'wrong sign'. This will result in we

spontaneous breaking of the symmetry under 0-x/>e"lw. The Noether current for this symmetry is / For small



couplings

iZ(pFl4>



Z((f)ld^(t)2



fad^).

(9.2.2)

potential

integral is dominated by in (9.2.1). This is at

\(t>\=2m/g^2.

(9.2.3)

the Euclidean functional

fields close to the minimum of the

The perturbation expansion amounts to a saddle point expansion about the minimum. It is set up by

making

cf>1 to



the substitution

<t>'l



(9.2.4)

2m/g1'2,

give <£



(dcf>[)2/2 + (d(f>2)2/2 ̄m2cf>'2 -g(<f>'2 + <I>22)2/16 -mgl'2<f>'l(<f>,2 + cf>2)/2 + ^ci,

(9.2.5a)

where ^?









Wt? + 4>l)2/H> Sgmg- il2(f>\(4>\2 + <t>22)/2 <P'2(3m25g/g Sm2)/2 4>22(m2Sg/g 5m2)/2 24>\mg- ll2(m2dg/g dm2) + (Z l)^;2 + d<t>22)/2. (9.2.5b) -











248

Global symmetries

making this perturbation expansion is that in the functional integral we impose a boundary condition that fixes the phase of the field at oo. By the symmetry we may make this phase real, without loss of In or three more generality. space-time dimensions, fields that have a different phase over a large region have an action so much larger that The idea of

quantum fluctuations cannot destroy the boundary condition. Then forced to have a real vacuum expectation value close to 2m/gl/2. In

setting

perturbation expansion we have tadpole generate a vacuum expectation value for <f>\

up the

Fig. 9.2.1. These

<o|0;|o>



<j)1

is

graphs like



expectation value 2mg ̄l/2 tadpoles graphs appear as subgraphs. It is possible to recast the Feynman rules by writing 0, 0J + 2mg"1/2 + 8v and requiring fy\ to have zero vacuum expectation value. A better practical approach is to impose dv 0 as a renormalization that starts at order

gl/2.

It

means

+ dv. There are then

that

(pl

has

vacuum

like Fig. 9.2.2, where the





condition

dm2.

on

Fig. 9.2.1. Graphs for <O|0i|O>.

Fig. 9.2.2. Graphs containing tadpoles If

we

start

as

subgraphs.

with the theory (9.0.1) without spontaneous symmetry

breaking and vary m2 until it is negative, then we should pass through a phase transition and thereby reach the Goldstone model (9.2.1). There must be an actual phase transition because <O|0|O> is exactly zero in the phase with unbroken symmetry. Since this expectation value is non-zero in the Goldstone phase there must be non-analyticity of the theory as a function ofm2. Now,

we

normalized functions of

must

renormalize the theory: the continuation in the

m2 is sensible only if the m2 and g in the two phases. It is sensible to use

mass

re-

counterterms are the same a

mass-

independent prescription, for then the dependence on m2 of the counterterms is the simplest possible. We will prove the following: renormalization

Spontaneously broken symmetry

9.2

249

phase is accomplished by using only Lagrangian (9.2.1).

(1) Renormalization of the Goldstone

symmetric

(2)

counterterms in the

The dimensionless counterterms Z to be the same

in the

as

phase



1, Sg, and

Sm2/m2 can

of unbroken symmetry

be chosen

(the so-called

Wigner phase). (3) The current given by (9.2.2) is finite just

as it is in the Wigner phase. Since the bare Lagrangian is invariant under 0->0e"1<o, Ward identities are valid and from them Goldstone's theorem follows, that the

physical

mass

of (j>2 is exactly

zero.

We must also discuss the choice of a

practical renormalization prescription.

9.2.1 Proof of invariance of counterterms We will do

perturbation theory by choosing JS?0



the free

Lagrangian

m2(f>[2,

(9.2.6)

mg^VM? + *D/2.

(9-27)

d(f>[2/2 + d<f>22/2



and the basic interaction

JS?b





9(<t>? + 0D/16



The counterterms are given the form (9.2.5b) and Sg, Sm2/m2, and Z are given the same values as in the unbroken theory with a mass-independent renormalization scheme. We will prove that these counterterms are

sufficient to make the broken-symmetry theory finite. Some of the interaction vertices are the same as in the Wigner phase. The others are obtained by substituting 2m/gl/2 for <j)v Therefore graphs

involving the extra vertices are obtained by erasing external (f>1 lines on symmetric graphs. Examples are shown in Fig. 9.2.3. The only complication is that mass terms generated from the basic interaction go into the free rather than the interaction Lagrangian. This is the sole source of complications

in

our

proof. + other terms

+ other terms





Fig. 9.2.3. Generation of graphs in theory with spontaneously broken symmetry from graphs in the symmetric theory.

250

Global symmetries

theory let

To relate the counterterms to those in the unbroken

the free propagators



as









2m2

p2



M2









-iM2

^''p2' p2 -M2+ (p1 Here M2 is







j__

write

i(2m2 M2)2 i(2m2 M2) + (p2 M2)2 (p2 M2)2(p2 2m2)' -

_  ̄

>2

us

follows:

iM4

-M2)1* (p2 -M2)^1'

arbitrary parameter.

(9.2.8)

We substitute (9.2.8) for every line in



graph. Suppose we substitute the first term on the right of (9.2.8) for every line of a basic graph which has only four-point basic vertices. Then we obtain a graph in the symmetric theory with mass M. The difference between these symmetric graphs and the true theory is given by: (1) graphs with one or more three-point vertices, (2) graphs with the second or third term on the right of (9.2.8) substituted for one or more propagators. In either

the degree of divergence is reduced. Now the maximum divergence is two. So substitution of the third term in (9.2.8)

case

degree of

always makes a graph overall convergent. We substitution of the second term.

Let

us

counterterm

Zm

one

graphs with fewer than N loops are symmetric counterterms. We will prove graphs are renormalized. The induction starts

in ¥£ct

renormalization. We decompose the

^'2[_3m2Sg/g + (Zm 1)( h<t>l\jn25glg + (Zm 1)( -



mass

as



Here

allowed at most

suppose that all renormalized by our

now

successfully inductively that all AMoop because tree graphs need no -

are



(m2 + Sm2)/m2

is the





mass

M2) + (Zm 1)M2] m2 M2) + (Zm 1)M2].

m2









(9.2.9)

renormalization factor.

After substitution of (9.2.8) for each propagator in a basic 1PI graph with AMoop all subdivergences are cancelled by counterterms of lower order,

according to the inductive hypothesis.

We

are

left with the following overall

divergences: (1)

Logarithmically divergent graphs

for the

four-point

function with all

M2) and with only four-point vertices. Such propagators set to i/(p2 graphs have an overall divergence independent of M which is removed -

by counterterms in 8g for the symmetric theory. No other 1PI graph for the four-point function has an overall divergence.

Spontaneously broken symmetry

9.2

(2) Self-energy graphs with four-point vertices only

i/(p2



M2).

Field renormalization and the

renormalize these, again exactly as in (3) Self-energy graphs as in (2) but with

the

and with propagators

(Zm 1)M2 terms symmetric theory.

one



2m2

-M2

M2 =

The terms



m2

3m2

in

(9.2.9)

propagator replaced by

second term in (9.2.8). In the numerators of (9.2.8) -

251

we



write

(m2 + M2) (2m/gl/2)2{3g/4) (m2 + M2\ m2- (m2 + M2) (2m/gi/2)2{g/4) (m2 + M2). (9.2.10) =









with

M2) parts

—(m2

M2)



of the



are

renormalized

by the

(Zm—1)

They correspond to the of the effect differentiating self-energy graphs with respect to M2. The other terms in (9.2.10) we will regard as an insertion of a four-point vertex on a line when two <j>l fields are replaced by 2m/g1/2. These terms are considered under (4). (4) Graphs of classes (1) and (2) in which one or more external (f>1 fields are deleted and replaced by 2m/gl/2. Examples are Fig. 9.2.3(b) and Fig. 9.2.4. The same replacement generates the counterterm Lagrangian (9.2.5b) from the symmetric theory, so we have counter- —



mass counterterms.

terms for them. This

Fig.

completes

the

proof.

9.2.4. Generation of graphs with loops in theory with spontaneously broken symmetry from graphs in the symmetric theory.

9.2.2 Renormalization The same

procedure shows /



of the

current

that the current

Z((j),1d^2



4>2d^\) + 2Zmg ̄l/2dtl(t)2

2 looP



Fig. 9.2.5. Renormalization of

(9.2.11)

current in

spontaneously broken theory.

252

Global symmetries

has finite Green's functions. Note that the term 2Zmg ̄ graphs like Fig. 9.2.5.

l'2dfi(j)2 contains the

counterterms that renormalize The Ward identities

are

then true and involve finite quantities. A typical

is

case

^<O|T/(x)02(y)|O>=i<O|<502(y)|O> -i<O|01(y)|O> =



-i(2m/gl/2



dv).

(9.2.12)

By multiplying by the inverse propagator for 02 and going

to momentum

space, we find

P^UP) where

r£2

is the set of

is illustrated in

has

a zero at

Fig.

p2





graphs

for

9.2.6. Since F1

<O|T/02|O> that are 1PI in 02. This p" as p2 0, (9.2.13) implies that G22l

oc

->

0, in other words that 02 is massless

perturbation theory.

(9.2.13)

(2m/9l/2 + Sv)(i/G22(p2)),

to all orders of

This is the Goldstone theorem (Goldstone, Salam &

Weinberg (1962)).

Fig. 9.2.6. The Ward identity that implies Goldstone's theorem.

9.2.3 Infra-red divergences

self-energy insertion on a </>2 line have infra-red graph is illustrated in Fig. 9.2.7, and the divergence region where the momentum k on the 02 line is close to

Individual graphs with

divergences. Such comes

from the





zero:

If uncancelled, this to a

divergence indicates

value other than

energy is

zero

at

zero.

Jc



that the self-energy shifts the

But the Goldstone theorem tells

0. So the infra-red

divergences in other graphs of the

same

us

mass

that the self-

divergence cancels against

order.

Fig. 9.2.7. Graph with infra-red divergence.

9.3 Renormalization methods

253

9.3 Renormalization methods One of the practical problems that arises in making calculations in a theory with spontaneous symmetry breaking is to find the most convenient renormalization prescription. Fundamentally, there is no problem, for all renormalization prescriptions are related by renormalization-group transformations, and are therefore equally good. But, in practice, choice of one prescription over another can save some labor. The problems become particularly acute in gauge theories of weak interactions (Beg & Sirlin (1982)). Among the issues to be considered in choosing a renormalization prescription are:

(1) If we ignore higher-order corrections, then some parameters are equal to quantities, like particle masses, that are easily measurable. It is often convenient to impose exact equality as a renormalization condition. (2) One must treat tadpole graphs. Their effect is to provide an additional shift Sv in the vacuum expectation value of the field. Leaving these

graphs

as

they

contributing to

considerably increases the given Green's function. Shifting

are a

number of graphs the field

many extra terms in the formulae for the coefficients in

(9.2.5b). One

can

impose

Sv





by Sv gives (9.2.5a) and

renormalization condition, at the connection between the phases of

as a

expense of removing the simple broken and unbroken symmetry (as

was exploited in Section 9.2). to is relate It calculations done by different people. Direct necessary (3) can if be made the same renormalizations are used. It only comparisons to a is evidently useful standard. agree on (4) If the coupling is not very small or if there occur very large ratios of masses

and momenta, then

prescription

with the

ability

one

must

choose

to remove the



renormalization

large logarithms.

One approach is to use dimensional regularization with minimal Graphs can be renormalized by the forest formula. At one-loop order this amounts to subtraction of the pole part from each 1PI graph. We can do this without regard to the symmetry relations between counterterms for different Green's functions. Since the counterterms have the pure-pole form, these relations are automatically satisfied. subtraction.

approach is to compute Z, Sg, and Sm2 by three renormalization conditions imposed on some of the 1PI Green's functions in the broken- symmetry phase. Then the values of counterterms for other Green's functions are c6mputed from (9.2.5b). It is convenient to determine m2Sg/g -Sm2 by requiring Sv 0, i.e., <O|0j|O> 2mg ̄112 exactly. Then Another





254

Global symmetries 1



or 2



















+ 2







Fig. 9.3.1. Renormalization of self-energy

the

1 or 2

/-^

at

one-loop order.

for

02 forces the IPI self-energy of 02 to be exactly one-loop graphs are shown in Fig. 9.3.1. Both of these approaches require explicit computation of the values of Z, mass counterterm

zero

when k2



0. The

8g, and dm2 to find the values of counterterms for the various Green's functions. It is also possible (Symanzik (1970a)) to use the Ward identities to generate renormalization conditions for all divergent IPI Green's functions from the three basic conditions. These conditions are simple if the three basic conditions

9.3.1

Generation

are

imposed

at zero external momentum.

of renormalization conditions by Ward identities

The general Ward identity is



-i

n ^-^xoi^^wn^^io). i

J=l

(9.3.i)

+j

the Goldstone model, the labels nt take the values 1' or 2, and we have 0X (</>[ + 2m/g1/2). For simplicity we impose (50i 02 and 8(f>2

In





the condition



<O|0i|O>



conditions valid up to /





0.

Suppose we have obtained renormalization loops. We will now find the appropriate





conditions for /-loop graphs. 1 was given in (9.2.12)and (9.2.13), andin Fig. The case JV =

9.2.6. Wesaw

self-energy of 02 is that it is zero 0. Another condition on the derivative can be chosen arbitrarily, at p2 corresponding to the freedom to multiply Z by a finite factor. Suppose we choose to make the residue of the Goldstone pole equal to unity, and we

that

one

renormalization condition

on

the



choose to make dv on



condition is that it is is

0. Then

the Green's function

equivalent

to

we

also obtain the renormalization condition

of the current with 02. The 0. This condition to its lowest-order value at p

<O|T/(x)02(^)|O>

equal making the



counterterm

equal

2m0-1/2(Z-l)d"02,

to

9.3 Renormalization methods as

required if/ is to be the Noether Similarly we may treat the case N

255

current. =

2:

d,<0\Tr(x)4>[(y)cf>2(z)\0y -id(x-yK0\Tct>2(y)(f)2(z)\0> + i<5(x-z)<O|T0;(>;)0i(z)|O>, (9.3.2) where we used <O|0i |0> 0. In terms of IPI graphs in momentum space =



this gives Fig. 9.3.2. After

use

of

Fig.

9.2.6

.2

find

we

Fig.

9.3.3.

+ i

^ Fig.

two-point

9.3.2. Ward identity for



Sj



2m =1

gm2



-<T^

+ i

now set p + q











Fig. 9.3.3. Result of multiplying Fig. 9.3.2 by We

function of/.

propagators.

0 to eliminate the left-hand side. This

^ir212(0,p,-p)-[p2-2m2-i:1(p2)]









two inverse

%ir212(0,p,-p)

+ 2m2



^(p2)



gives

[p2-Z2(p2)]

Z2(p2).

(9.3.3)

Zx and Z2 are self-energies and T212 is the IPI Green's function for 02 fields and one ¥[. We choose a mass renormalization condition for 0. Since we already know that I2(0) I x, say I x (0) 0, this tells us that the renormalization condition on T212 is Here

two





r212(0,0,0) Since graphs for T212 r212(p,4,r) completely





img-1/2



lowest-order value.

(9.3.4)

logarithmically divergent we know at this order. From (9.3.3) we can now determine Zj(p2). But the calculation of Zx(p2) from its graphs is already fixed except for



are at worst

renormalization condition that determines the value of the field-

256

Global symmetries

strength

counterterm. So

that the counterterm is

(9.3.3) gives (Z l)p2.



Similar arguments may be

all the remaining are

us

the renormahzation in such

a way



applied to give renormahzation conditions for divergences. They

1 PI Green's functions that have overall

easiest to express in terms of Ward identities for 1PI Green's functions.

(See

Lee (1976), and references

therein.)

The structure of these arguments generalizes what unbroken phase. For example, (9.3.2) integrated over

would be done in the

would give Zx Z2 in this phase. This condition would say that the counterterms for Zx and Z2 are equal. But in the Goldstone phase this is not so. Integrating over x is

equivalent

to

setting the





momentum at the vertex for the current to zero.

gives a factor of this momentum, but since there is a pole l/p2 the right-hand side of (9.3.2) is not zero. The argument that we had to use is more complicated. The derivative with

respect

to x

10

Operator-product expansion In this

chapter

will

we

(f>4 theory in d

investigate

two

closely

related

problems.

We work

time-ordered product of two fields, T0(x)0(O), together with its Fourier transform with



4 dimensions and consider

70(4)0(0)=



d4xe^*70(x)0(O).

(It is easiest

to work with time-ordered products. The methods work with of any pair operators TA(x) B(0) in any theory.) The first problem is to ask how T0(x)0(O) behaves as xM -> 0. If the theory

totally finite

were

then the result would

just

be

02(O). However, there

are

ultra-violet divergences that prevent the product from existing, so the limit does not exist. It was the idea of Wilson (1969) that 0(x)0(O) should behave like The

singular function of x times the renormalized [02] operator, as full result is that we have an expansion of the form



T0(x)0(O) as x -?

0. Here the sum is

C<p(x)'s [0] generalizations, is called and the

coefficients C6

are

are

smaller by a power The second problem

\q2\-+

More

X C,(x*)[0(O)]

0.

(10.0.1)

over a set of local renormalized composite fields onumber functions. This formula, or one of its an

operator product expansion (OPE), and the

often called Wilson coefficients. Corrections to (10.0.1)

are

oc.



x -?

of x2 than the terms we

precisely

given.

wish to treat is the behavior of

we

T0(g)0(O)

as

will consider the momentum-space Green's

function

Sw+2



<O|T0te)#>)0(p1)...0(pw)|O>,

(10.0.2)

when q11 along a fixed direction with pl9...,pN fixed. In other words we scale the invariants q2^>K2q2, Pt'q^Kp^q. There is an operator product -?

oo

expansion

6N 2-EC^)<0l^(°)^i)---^iv)|0>. +

(10.0.3)



The relation between the

coordinate-space 257

and momentum-space expan-

258

Operator-product expansion

sions is elementary. Let

us

take the Fourier transform from

x to

q of a

Then the large-g behavior is dominated by the singularities in x-space. The only relevant 0. So C^q) is the large-4 part of the Fourier transform of singularity is at x if Q,(x). Conversely one Fourier transforms GN + 2 to get

momentum-space Green's function of

T(f)(x)(f)(0).



/?

<O|T0(x)0(O)^(pJ...^(pw)|O>



14

J^e-i?-*ejv

+ 2,

(10.0.4)

then the limit x -> 0 fails to exist if GN + 2 falls only as 1/q* or slower as q ->

00.

Thus knowing the large-g behavior is equivalent to knowing the singular small-x behavior, but the coordinate-space expansion also includes the leading non-singular part of the small-x region. expansions have a number of uses, particularly in an

information

These

on

asymptotically free theory. There the perturbation theory when improved by the renormalization group gives an effective method of computing the Wilson coefficients. Among the

uses are

the following:

(1) The expansion (10.0.1) in coordinate space provides a definition of renormalized composite operators that does not involve any reg- ularization (Brandt (1967)). (2) Although there is no physically important process which directly uses the limit taken in the momentum-space expansion (10.0.3), it is used

indirectly for deep-inelastic scattering of involves a matrix element of the form



lepton

on a

hadron. This

<p|M/(0)|p>. Here

q2

-?



00, but with the ratio

A dispersion relation relates this OPE is used indirectly,

as we

q2/q-p fixed instead

case to

will

see

of q2/qp2 fixed.

the limit used in

in

Chapter

(10.0.3), so

the

14.

(3) The form and the method of proof of short-distance operator-product expansion can be generalized to handle many interesting high-energy

scattering processes. (See Buras (1981) and Mueller (1981) for a review.) The results in the present chapter form a prototype for these other

results.

10.1 Examples 10.1.1 Cases with We will

no

divergences

mainly restrict our attention to Green's functions of </>(x)</>(0) in 0(x) and (f)(0) are connected to other external lines. This is the

which both

10.1



Ox



/\.P2

Pi

259

Examples

Pi

yC

>Pi/\p2

p2

(fl)

(b)

Fig. 10.1.1. Lowest-order graphs for operator-product expansion of 0(x)0(O). case

that is relevant to most &

First consider

d(f>2l2





applications.

m2(f)2/2



Our

#4/24

graphs Fig. 10.1.1(a)

the tree

theory

will be

04 theory (10.1.1)

+ counterterms.

for

<O|T0(x)0(O)^(Pl)0(p2)|O>.

(10.1.2)

These give p2

Expansion

in

_



m2



_

m2

CeXP(

power series about





{Pl'X) + eXP(









(10.1.3)

iP2"^)]-

gives

—i[2-i(p1+p2)-x-(prx2 + p2-x2)/2+--]-

p2-m2p2-m2 This is equivalent



to the

T(f>(x)(f>(0)





"-





(10.1.4)

replacement

(f>2(0) + 12xfld,<t>2



:b^xv(K3v0 + *"'

(10-L5)

Fig. 10.1.1(b). This equation has the form of the operator- product expansion (10.0.1). Thus the operator-product expansion in this case (free-field theory) is 0. The power of x in each term is really a Taylor expansion of 0(x) about x as

illustrated in



just such that

Ce(x)



no

dimensional coefficients

constant x

are

needed:

|x|a, with a ==dim(0)-dim[0(x)0(O)].

(10.1.6)

This result also correctly gives the power-law behavior in the presence of renormalizable interactions, as we will see. But there will also be

logarithmic

corrections.

A feature which does not appear to survive inclusion of interactions is on the right of (10.1.4) is convergent and sums to give

that the series

T0(x)0(O). Consider next the graphs of Fig. 10.1.2 for the four-point function of T<f>(x)<f>(0). The important factor comes from the lines carrying momentum

fd^

{exP(-i,-x) + exp[i(g-p,-Pi,)-x]} m2)HpA-q)2-m2X(q-pA-PB)2-m2J

260

Operator-product expansion

¥Pa

Pb

Fig. 10.1.2. Higher-order graphs for operator-product expansion of </>(x)</>(0). We

now

expand the integrand

in powers of

'4

(10.1.7)'

)A (q2



m2)[(pA



x to

obtain

2-i(pA + pB)-x q)2 m2] [(a -pA- pB)2  ̄



2T + -".

m2]

(10.1.8) These first two terms

higher

are

just those we would

expect from (10.1.5). But the

terms have at least two extra powers of q in the numerator and are

divergent. The divergences are those of the Green's function of the composite operators. They indicate that modification of the higher terms of the expansion is needed. For example, the behavior of the coefficient of c/>dMdv0 is modified by a logarithm of x. therefore ultra-violet

Similar modifications will be needed for the coefficient of

<f>2,

when

we

consider

higher-order corrections. Therefore we will find it convenient just to restrict our attention to the leading-power behavior, corresponding to the (f)2 term in (10.1.5). 10.1.2 Divergent example

Aside from trivial propagator corrections the contribution of order g to the two-point function of 7c/>(x)0(O) is given by Fig. 10.1.3(a), which gives

I (pl-m When



fd^_

l)iP22-m2)g)(2n)\q2- m2)[(q-pl ̄p2)2-^2]

(10.1.9)

integral diverges logarithmically. This is a symptom of the fact that there are two important regions of q that contribute. The first is x -?

0 the

X/ v0 -uv

+ UV

(c) Fig. 10.1.3. Gr,aph for operator-product expansion of 0(x)0(O) with divergences for the operator.

10 A

where

q is finite as

T0(x)0(O) by

261

Examples

x-?0; the contribution is correctly given by replacing

02(O)

in the

large, up to 0(l/x)

as

graph.

The second region is where q becomes

x->0; in this region the interaction

vertex in

coordinate space is close to x and 0. In the second

space. From the

region the loop is confined to a small region in coordinate point of view of px and p2 the loop is a point. So we should

be able to represent the contribution of this region by Wilson coefficient of 02:

an extra term

70(x)0(O)^Q2(x)[02], C+2 Let

us now

calculate

(10.1.10)



l+(g/16n2)cdx2)1 ,



in the

c2(x2).

Now the contribution of the first region is given by replacing T</>(x)</>(0) by c/>2(0). However, this operator has an ultra-violet divergence. So let us

add and subtract the renormalized Green's function of

[02], i.e.,

<O|T[02(O)]^(p1)^(p2)|O>,

(10.1.11)

give the equation depicted by Fig. 10.1.3. The contribution of order 1 region is entirely contained in the term (b) representing (10.1.11): to

from the first

(p2-m2)(p2-m2)16^2|J0dXlnL

J T

4V

(10.1.12) Here

we

have used minimal subtraction. The remainder, term (c), is 1

-ig

{p\-m2)(p22-m2){2nf

fd4?rr- to2



-'



)l(q-Pl-p2)2-m^

UVdivergencei.

"—

*—J

(10.1.13)

When x->0 the contribution from finite q is of order |x|. But there is a contribution of order 1 from large q: this is the contribution to the original graph minus whatever is taken care of by graph (b). the x->0 behavior of the curly bracket factor of (10.1.13) (aside from a normalization factor), since to We

can

identify cx(x2)

in

(10.1.10)

as

lowest order

<0|7>2#(pt)#(p2)|0>=- (P?-m2)(p22-m2)' Now the

leading-power behavior of the curly-bracket factor is independent yn. This is easily seen by differentiating with respect to any of

°f PmPi and

262

Operator-product expansion

these variables. The result is power of x when x

-?

0. So



we

convergent integral which goes may define

cx(x) by setting m

to zero like a



c^-M^h^+M

pl



p2



0:

,iou4)

The integral is easily done by using

-j;pdzze-2<-?2)

l/(q2)2



form, with

to turn it into a Gaussian

*i(*)



the result

i[y + ln(





V*2)].

(10.1.15)

10.1.3 Momentum space We

now

Fourier transform Fig. 10.1.3 to obtain the 0(g) contribution to

<o|r^(?)^(0)^(p1)^(p2)|o>. As

q2

->■ oo

we

find

Fig. 10.1.3(a) This

gives a (10.0.3) with



—2

=--!, + 0[l/(o2)3]. (10.1.16)

j—

(p2-m2)(/>22-'?W)2

contribution to the term in the 0



02.

operator-product expansion

The coefficient is

C+*(<l2)







(10.1.17)

0(g2l

which is just the Fourier transform of the order g term in the coordinate- space expansion, ^c1(x)/(167i2). The g° term in the coordinate-space

coefficient is independent and hence

nothing

at

of x,

so

that it

gives



d(A\q) in momentum space,

large q2.

10.1.4 Fig.

10.1.3 inside bigger

graph

(10.0.3) indicates that the same asymptotic oo) is obtained independently of the Green's function considered. This happens because graphs like Fig. 10.1.3 can occur The

expansion (10.0.1)

behavior

as

x->0 (or

or

as q-+

167T

Fig. 10.1.4. Even higher-order graph for operator-product expansion of 0(x)0(O).

10.2 Strategy

subgraphs Fig. 10.1.4.

as

of

graphs

with

more

10.2

First

we

will make

expansion applies.

If

precise we are

263

ofproof

external lines. An

Strategy

of

is given in

example

proof

the limits in which the operator-product in Euclidean space (i.e., with imaginary time

and energy) then there is really only

one

way in which

we can

take xM to

zero

infinity. However, in Minkowski space we can let x2 0 without each component going to zero, and we can let components of q*1 go to infinity without g2->oo. These cases are interesting physically. For example, the q^^co limit with finite q2 is the case of high-energy scattering. Much is known about these limits, but they are beyond the scope of this cf

or

to

-?

book. We will prove the coordinate-space expansion (10.0.1) in the case that all components of xM go to zero with their ratios fixed. The corresponding

momentum-space expansion (10.0.3)

we will prove in the limit that all components of q* go to infinity with a fixed ratio, and with q* not light-like, so that q2 -?oo. These limits are essentially Euclidean.

Our proof will be

in

perturbation theory.

The first step is to

regions of loop-momentum space that give the

the

x -?



or q -?

oo

limits. Then

section, which applied

to

we

leading-power

identify

the

behavior in

generalize the arguments of the previous

specific graphs.

The region of large q which we investigate in the momentum-space expansion (10.0.3) is precisely the one to which Weinberg's (1960) theorem

applies. The theorem tells us to consider all subgraphs connected to the (f>(q) and c/>(0) in (10.0.2). For each such subgraph we let all its loop momentum be of order q, and count powers just as we did for UV divergences. The subgraph(s) with the largest power of q2 correspond to the dominant regions of momentum space. Then as ^2->oo, the complete graph is proportional to this power of q2 times possible logarithms of q2. smaller by a power of q2. Although Weinberg's result also the highest power of ln(^2) that appears, it is easier to determine this

Corrections tells

us

are

by first constructing the operator-product expansion and then applying renormalization-group methods (as in Section 10.5 below) to the coefficients. Essentially the same method can be applied to obtain the short-distance behavior, (10.0.1). For example, in Fig. 10.1.4 we have leading contributions with q finite or with q large (of order 1/x), but always with the lower loop momentum, /c, finite. The leading power is (x2)0. The logarithm of x in

264

Operator-product expansion

(10.1.15)

for the corresponding Wilson coefficient

over momenta

Suppose

we

comes

from integrating

intermediate between these two regions. have large momentum confined to a subgraph T. Then the

(for the momentum-space expansion) is exactly the dimension of our since T, theory is renormalizable, with dimensionless couplings. The leading power of q2 comes from subgraphs with the largest dimension, i.e., power of q

subgraphs with the smallest possible number of external lines. This number is two (beyond 4>(q) and (f)(0)), so that the leading power is l/(q2)2. The subgraphs have the form of the subgraph U in Fig. 10.2.1. In the ultra-violet

subgraph U, all lines carry momentum of order q. This subgraph is 1 PI in its lower two lines. All momenta in the infra-red subgraph / are finite.

Fig. 10.2.1. General

structure of lead-

Fig. 10.2.2. Factorization

of Fig. 10.2.1.

U is

independent

ing regions of momentum space for Appoint function of </>(x)</>(0). Now, to the

leading power of q2, the

ultra-violet

graph

of the external momenta k and / flowing into it. Thus we may replace U by 0 (and we may set the mass m l its value when k 0). We may also the infra-red subgraph / by an insertion of a vertex for c/>2/2 in an =

replace Appoint Green's structure as





function. This is illustrated in Fig. 10.2.2. This has the same operator-product expansion. But it should be emphasized

the

supposing that loop momenta are restricted to certain regions. These regions are not defined very precisely, and it is one of the tasks of the proof to remedy the impreciseness. Schematically we have that

we are

X Gn 2(4>Pi>' ̄>Pn) +

graphs

£!/(?,*



0, J



d*kd*ld(k +1

0)£ /

ZT,



£p,)/(Upi,...,Pn)- (10.2.1)

generalize the technique that we applied to Fig. graph U that could appear in Fig. 10.2.1, but not its momenta. restrict It can occur as a subgraph of some graph for we do the complete Green's function. If all momenta in U are of order q and if all To construct the

expansion

we

10.1.3. We consider each

265

10.2 Strategy of proof

large

momenta

small momenta

Fig. 10.2.3. A possible leading region for momenta outside are



subgraph of the form of

U in Fig. 10.2.1.

small, then we get a leading contribution to the cross-

leading contribution where the large momenta proper subgraph of t/-as in Fig. 10.2.3. Suppose we subtract off all of these contributions. Then we integrate over all loop section. We can also have a occur

inside



momenta of U and find that the result only gives a leading contribution when all its momenta are large. We therefore define the contribution of U to the Wilson coefficient

C(U) all evaluated



as

U- subtractions for

at k











regions of form Fig. 10.2.3 (10.2.2)

0.

resulting formula for C(U) is very similar to that of the formula for renormalizing the ultra-violet divergences of a graph. In fact, as Zimmermann (1970, 1973b) explains, a good way to prove the operator- product expansion is to treat it exactly as a problem in renormalization. His method, used in the next section, is not to compute directly the Wilson coefficients but to define first a quantity which is a Green's function minus The

the

leading

terms in its

operator-product expansion: C

This is constructed

as a sum over

graphs

T for

GN + 2-

Each graph has

subtracted from it not only counterterms to remove ultra-violet divergences but also counterterms to cancel the large-Q (or small-x) behavior. The result we

call

Now,

RW(T). the subtractions that

oversubtraction. So

is

remove the

large-Q behavior

of

are a sort

related to R(T) in the

style

of

simply RW{T) renormalization-group transformation. This transformation can then written as the Wilson expansion, as we will see. A disadvantage of Zimmermann's proof is that it uses BPHZ



be

re-

normalization. He takes advantage of certain short-cuts available through the use of zero-momentum subtractions. We will choose not to take these be

applied to ultra-violet divergences. Our method of proof will be essentially the same as the one used for problems with large masses (Chapter 8). The same techniques apply to the coordinate-space expansion. Here, the shortcuts

so

that minimal subtraction

can

266

Operator-product expansion

integrated over. When we do power-counting, large momenta are regarded as of order 1/x and small momenta as finite when x -> 0. The leading behavior again comes from graphs of the form of U in Fig. 10.2.1, and the power is (x2)0. There is a difference in the form of the possible graphs U that carry large momentum. All the graphs that we use in

momentum q is

the momentum-space case are also used in coordinate space. But we can also have the graph consisting of the vertices for 0(x) and (f)(0) and of

nothing else. 10.3 Proof We must

now

prove the operator

product expansion.

In

04 theory

we

consider the part of the Green's function, Gw + 2(x,p1,...,piv)



<O|T0(x)0(O)^(p1)--^(PiV)|O>.

in which each of 0(x) and (f)(0) is connected lines. We will now scale x

(10.3.1)

to some of the other external

by a factor k and construct a decomposition of the

form:

Gn 2{kx,Pi,.-',Pn) +



C(icV)<O|^[02](O)^(pJ--^(pw)|O> + rN

In every order of

perturbation theory,

the coefficient

(k2)0 times logarithms, when the scaling parameter remainder goes to zero like a power of k. Fourier transformation

on x





2(icx,p1,...,pAr). (10.3.2)

C(k2x2)

behaves like

goes to zero, while the

gives the result

GN 2(<i/K>Pi>-- >Pn) +



CtaVic2)<O|T¥[02^ (10.3.3)

4/jc-xx>, C(q2JK2) behaves like K*/q* times logarithms, while ^ + 2(^Piv) is smaller by a power. Our proof is given for a specific Green's function in a specific theory. However, it can easily be generalized. Features specific to a gauge theory will be pointed out in Chapter 12. The particular case of QCD with the application to deep-inelastic scattering will be treated in Chapter 14. When

10.3.1 Construction of remainder We consider the set of graphs

for GN

+ 2.

graph T, we will construct Each graph V we consider to be

For each

its contribution r(T) to the remainder rN +

2.

103 an

Proof

267

unrenormalized (but regulated) graph containing only basic interaction are derived from

vertices. These

^basic



o4>2/2



m2cf>2/2



/i4-d#4/24,

where g and m are the renormalized coupling and mass. As usual the renormalized value of the graph is R(Y), which is T, plus



series of counterterm graphs to cancel its ultra-violet divergences. For our proof we will use a renormalization prescription in which the theory is finite when the renormalized

mass m is set to zero. The renormalized value R(Y) is then the contribution of Y to the Green's function GN + 2. The remainder term r(Y) is equal to Y plus a somewhat different series of

counterterm graphs. These counterterms will be constructed so that they cancel not only the ultra-violet divergences but also the leading x->0 or 4

-?

oo

behavior of Y.

Now r(Y) is in effect a oversubtracted form of Y. The oversubtractions are of the form of an insertion of the operator [02] times a coefficient. Thus

r{Y) is the Wilson expansion, i.e., the first term on the right of (10.3.2) (10.3.3). The coefficient C(x2) (or C(q2)) depends on the coupling g and on the renormalization mass \i. It must be independent of all the momenta. In R(Y)



or

order to be able to neglect m in the ultra-violet limit x -> 0 or q -> oo, we must use a renormalization prescription in which the counterterms do not become infinite when use



->0

(with fixed regulator).

For concreteness

we

will

minimal subtraction in what follows.

In order to define r(Y\ let us recall the definition of the ordinary renormalization R(Y). This starts from the fact that the divergences of Y come from regions of loop momenta where all lines in some set of IPI subgraphs carry

a momentum

subgraph consisting

approaches infinity. We label each region by the all the lines carrying large momentum. Then

that of

R(H



r-j;qr).

(10.3.4)



is

subgraphs subgraph y replaced by its Cy(Y) to be non-zero only if y is The

sum

y!,...,7?.

over

all

y of T, and

Cy(Y)

is essentially Y with the

large-momentum divergence. We define a disjoint union of one or more IPI graphs

In that case each y. is

replaced by

a counterterm vertex

C{y^,

which is the divergent part of y.. To avoid double-counting, the sub- divergences are subtracted off first:

C(yt)





y,-



Ca(yA

(10.3.5)

268

Operator-product expansion

Here \^' denotes

'pole part

at d



4' and the sum is over all proper

subgraphs 8 of y-. As usual, if we use some other renormalization prescription than minimal subtraction the operator 0> is replaced by the operator appropriate to the renormalization prescription. The definition of the remainder r(T) is almost identical to the definition of R(T). Now, the leading short-distance (i.e., x -> 0) behavior of T comes from the following

regions:

(1)

where the momentum q is finite, (2) where q gets large and the momenta in Fig. 10.2.1 also get large.

Further leading contributions momenta get infinite in

contributions

come

some set

to

correspond

momentum-space expansion the region of finite q.

of



graph of

the form of U in

from regions where in addition divergent IPI graphs. These extra

the

ultra-violet

divergences.

In

the

(10.3.3) the same regions are leading except for

We define r(V) to be r with all ultra-violet divergences subtracted and then with all the leading small-x behavior subtracted: r(r)



R(r)-ZlLyUn d

Here the

graphs of the form of U in Fig. 10.2.1 and the subgraphs y of T that do not intersect 8. We use Lyu5 to subtraction operation defined below. It is used to extract the

sum over

sum over y

is

symbolize



(10.3.6)



over

8 is

over

all

all

comes from the region where the momenta in graph 8 are of order 1/x and the momenta in y go to infinity. The case of finite q in Fig. 10.2.1 is covered by the case that 8 consists of the vertices for 0(x) and

contribution that

0(0) only.

L^ud(T) to be zero unless y is a disjoint union of replaced by its overall counterterm C(yt) defined in (10.3.5) while the graph 8 is replaced by a quantity L(8). L(8) is to contain the leading behavior of 8 when all internal lines have large momenta. This is the same idea as that C(yt) is the overall divergence of y-. Now there are regions where a subgraph 8' of 8 carries momenta of order q and other lines in 8 carry small momenta. To avoid double-counting we We define the subtraction

IPI yl9..., yn. In that

case

subtract them first. So L(8)

each yt is

we



write: T





<5'<£ 5 ynS

Here T is to be

an

operator that picks

argument. Now this behavior is





out the

independent

(10.3.7)

L7U6.Q) leading

x ->

0 behavior of its

of m, and of the finite external

103

269

Proof

(We prove this by differentiating with respect to the variables m p?.) So we define T to set the values of m and the finite external

momenta.

and

momenta to We now

zero.

have



complete definition of r(r).

10.3.2 Absence In

Fig. 10.1.1(a)

of infra-red

the only

that consists of

and ultra-violet

possible graph

divergences

in

r(T).

of the form of U is the

alone. We call it d{. Its value is e"ipi

Q-1P2* jt kas no

0(x)0(O) subgraphs

so

L01) (e-,'?-* The remainder is therefore =





that



/-(Fig. 10.1.1)

one



Fig.

e-,"-*)Pi_P2=o 10.1.1





(10.3.8)

2.

(10.3.9)

Fig. 10.1.1(b).

Here we regarded L(<5 () as a [<£2] vertex. It is manifest that this remainder is exactly the graph minus its Wilson expansion. Now we turn to Fig. 10.1.3(a). It has Sx as a subgraph of form U and also the loop, which we call 32. Then by (10.3.7)

L(S2)



{S2-Ldi(d2)}m=pi=P2=0

\(2nY =



q(q2-m2)[(pl+p2-q)2-m2-\Sm=Pi=P2

^k<?>





(10.3,0)

Hence

Kr)=*(r)-Mr)-wn

W*-'\ f

i2

J'

(pi2_m2)(p2_m2) [2ny -\d"q



e^Z

|Jd V-?2)[(Pl+P2-?)2-?2]

(q2-m2)[(Pl+p2-q)2-m^

.■(e-1'"-!) ddq (10.3.11)

where The

we use T to

denote Fig. 10.1.3(a).

following properties

hold:

is infra-red convergent even though it has zero mass and zero external momentum. Although d2 has an infra-red divergence when m,

(1) Ld2(T)

pl9 and p2

approach

the divergence.

zero, the subtraction term

Ldl(82) exactly cancels

Operator-product expansion

270

(2) Ldi(T) is ultra-violet divergent, since in replacing the vertex St =iQ ̄'lq'x by 1 we remove the ultra-violet cut off. However, Ld2(T) contains all the large-4 behavior of T and of subtractions for subgraphs of 82. Thus Ld2(T) cancels the ultra-violet divergence of Ldi(T). (3) When m, pl9 p2 approach zero, we find that

r(d2)

= =

R(d2)-LSi(d2)-L(d2) 82-LSi(82)-L(82)

-?0. This is p2

The



the statement that

just

0, after subtractions of these

explanations

demonstrating that they

on

L(82) is the value subgraphs are made.

properties

are

of

82

at m



p1



convoluted, but with the aim of

in general. Refer now to the general definition of r(T), viz., (10.3.6), and refer to Fig. 10.2.1 instead of Fig. 10.1.3. Then the above

are true

properties get replaced by:

(1) L{8)

is infra-red convergent for any graph only regions that could give infra-red

of form

U in

Fig.

problems subtractions. (2) r(T) is ultra-violet and infra-red convergent if m, p\ and p\ The individual terms

are

10.2.1: the

cancelled

by

are non-zero.

Lyud(T) are IR finite. The subtractions remove all

ultra-violet behavior.

(3) r(8)



0 when





p{



10.3.3 R(T)

p2



0.



r(T)

is the Wilson

expansion

Although w(n = R(r)-r(r)

contains

ZL^



c/n)

(10.3.12)

0 behavior of R(T), it is not yet in the form of the expansion, which is

the leading

operator-product



x ->

W(D



Xs C(S)R(r/8).

(10.3.13)

(of form U) to the Wilson point, i.e., replaced by a <f>2 for the divergent subtractions pole-part

Here C(d) is the contribution of a subgraph 6 coefficient, while T/S is V with 8 contracted to a vertex.

R(T/d) will

now

include

(f>2. Summing(10.3.13) over T can be done by independently summing over 8

Green's functions of

and r/8. This gives the operator-product

expansion (first term of the right-

271

10.3 Proof hand side of (10.0.1)) with

Q2(x2)=IC(<5).

(10.3.14)



(We have used the as

same

for the

symbol

for the contribution from



complete Wilson coefficient C^2(x) particular graph.) Hence to prove the

expansion (10.0.1) or (10.0.3), we have to prove that W(V) as defined by (10.3.12) equals the right-hand side of (10.3.13). Any graph r for the Green's function GN + 2 can be written in the form of Fig. 10.2.1. In general there will be several possibilities for the upper

subgraph U. For the following argument we will choose U to be the biggest possible graph. Then, in (10.3.12) for W(T), all the subgraphs d are contained in U. So

we

have

W(T)

prove

divergent (10.3.13) when

(10.3.15)

W{U)?R(I\



since no

fd4/cd4/ wW{U)R(l) (2k)8



IPI

subgraphs include parts of both U and /. Suppose we r is replaced by U. Then from (10.3.15) it follows that py(D





^C((5)R((//(5)?^(/)

(10.3.16)

YC(8)R(m- s

The last line follows because (a) the

graphs for T/S are of the form U/d times divergent IPI subgraphs are entirely within U/d or

/, and (b) ultra-violet within /. Notice that the extra divergences for T/S as compared with T are due to the presence of the 02 vertex. These divergences are confined to

graphs

of the form of U/d.

(10.3.13) for the case T U, we use it C(U\ recursive in terms of C(d) for smaller d: Rather than prove

2C(U)=W(U)-



as a

definition of

X C(d)R(U/d).

(10.3.17)

multiplying C(U) is the lowest-order IPI Green's function of the operator <j>2. It is evidently much too good to be true that we have reduced what ought to be a deep proof to a mere definition. The important physics was at the previous step, (10.3.16). By implicit use of power- counting arguments, to restrict ultra-violet divergences to within U/d, we proved (10.3.13) given its truth for the case T=U. The Wilson coefficient The factor



C^2(x) as it occurs in the expansion (10.0.1) is then the same no matter what

Green's function.we take of 7</>(x)</>(0). This universality is feature of the Wilson expansion.

an

important

272

The

Operator-product expansion other

coefficient is



important feature of the expansion is that the Wilson purely ultra-violet object. This will enable renormalization-

techniques to be useful in its calculation. For this purpose we must C(U) is independent of m, k and /. The proof is done by differentiating the right-hand side of (10.3.17) with

group

now prove that

respect to the mass m or with respect to one of the external momenta k or /. Let A represent this operation. It is applied in turn to each propagator in

Inductively, we assume C(d) satisfies AC(d) 0. This is true for the lowest-order graph, 619 for which 0(5^ 1 follows from (10.3.8). In the general case AC(U) is therefore given by W(U) and in R(U/5).





2AC(U)

AW(U)-



£ C(d)AR(U/d),

(10.3.18)

6Z.U

Fig. 10.3.1. We have a series of terms in each of which one particular propagator of U is differentiated. The differentiation improves

as

illustrated in

the ultra-violet behavior,

so

the differentiated propagator cannot be

of a leading large-momentum region connected therefore factor out



maximal two-particle

to the vertices

graph <5',just

as we



part

0(x)</>(O).

We

factored out

U in (10.3.15).

Fig. 10.3.1. Differentiation of Wilson coefficient with respect to

a mass or an external

momentum.

Now we can use the result

d"

for AC(U). (This argument is analogous to (10.3.16).) The operator-product expansion is now proved.

to

give

zero

10.3.4 Formula for C(U) From the above results

C(U)

we





may deduce

R(U) ̄

£ C(8)R(U/8)

(10.3.19)

Heuristically this says that C(U) is that part of U arising when all its internal

10.4 General

273

case

lines have momenta of order 1/x, less contributions taken

care of by terms C(5) in the operator-product expansion with smaller graphs 8 contained in U. The term in square brackets is the unrenormalized U minus (a) all contributions involving momenta not all of order q, and (b) contributions with all momenta of order q taken care of by the subtractions under (a). This equation reproduces exactly the calculation performed in Section 10.1 of

the 0(g) Wilson coefficient, viz.,

C(S2)



{S2-QS1)R(S2/51)]m=t,i=P2=0. 10.4 General

case

The most fundamental form of the operator-product expansion is (10.0.1), proved by obtaining its matrix elements from the Green's functions

which is

(10.0.2). We proved the Green's function expansion restricted to connected graphs in </>4 theory, and restricted to the leading power of x2. The only that then contributes is (9 <f>2. Our proof generalizes. We may take the full Green's functions in any theory and include non-leading powers of x.

operator



general case the operator L extracts the appropriate number of Taylor expansion in the mass m2 and in the external momenta. The result is that each Wilson coefficient CG{x) behaves in each order of perturbation theory like a power of x2 times logarithms of (x2^2) times a polynomial in xM and m2 with dimensionless coefficients. If (9 is a tensor In the

terms in



operator then the coefficient is also order example (10.1.5). The leading operator for



tensor,

as

illustrated by the lowest-

04 theory is in fact the unit operator since it has

lowest dimension. We have

Tcf>(x)cf>(0)



Ct(x2)l



Q2(x2)[02] + 0(x?logs).

(10.4.1)

Operators linear in <f> are prohibited by the <f> <f> symmetry. To compute the coefficient, Cx(x2\ of the unit operator to lowest order, -?

we use

the

graph of Fig.



10.4.1. We will extract terms up to



Fig. 10.4.L Lowest-order

term for

O(x0), so that we

<O|7>(x)0(O)|O>.

274

Operator-product expansion

find:

C1(x2)





C1(x2)<0|l|0> {Fig.

10.4.1



C02(x2)<O|[>2(O)]|O>} expanded

in pow of m2 to order ,

ie'? ddq (2n)dq2-m2 + ie *



m2

Hd-*m2

+ k

]}.

q2

)expanded

J(2^1(C

1)U2+(^)2J







"4m2 8^ 871V-4)

m2

-^2^2+Y^[y + ln(-7iVx2)]

at J



(10.4.2)

4.

The Wilson coefficient is obtained by expanding in powers of mass and an appropriate degree, which is two here. We have no external momenta, so the expansion is in powers of m2. In the next section, when we apply the renormalization group to compute the Wilson coefficients we will find that each coefficient is given as a series in the effective coupling with the renormalization mass \i set to (— x2)1/2. The main application of these methods will be to asymptotically free theories. If we truncate the perturbation expansion, then the error will be of the order external momentum up to

of the first omitted term, and hence the fractional

error

is of order

l/[ln(-x2)]'+1, where p is the number of

loops

in the

highest-order graph. This

dominates any positive power of x. Consequently, it is difficult to use the

power-law corrections

error

to the

cannot Wilson coefficient. This would suggest leading power in properly use anything but the coefficient of the unit operator. However, in applications we will normally work with connected Green's functions. For these the leading coefficient is of the two-particle operator (j>2. a

we

10.5 Renormalization group To

make

calculations

for



Wilson

coefficient,

we

must

renormalization group to obtain maximum information from calculation in the way we will now explain. The coefficient has the functional

dependence

C{x2,g,fi).



use

the

low-order

10.5 Renormalization group

275

\x2 to be of order | l/x21 to avoid large logarithms, just as in calculating Green's functions when all their external momenta get large (Chapter 7). The RG equations are simple, since, to the leading power, there is no mass dependence. (If we use non-leading powers of x then we have polynomial dependence on masses.) The renormalization-group equation for the Wilson coefficient can be derived most easily from the To

this

use

we must set

renormalization-group equations fashion. In the

04 theory

we

for Green's functions in the

have the renormalization-group operator:

%-%+^-i-m'^ The

renormalization-group equations

^AG(jonn)

£GN([<t>2l? Here,

Gff+p

restricted



.)(conn)

_



{N/2

(ym



we

ml

We



are

(10.5.2)

l)yGl*?*)9

i2Ny)GN([>2],...)<conn>.

[</>2]

(10.5.3) 0-fields,

denotes the connected

operator with N 0-fields. To

used the fact that

m2[02] and



need

graphs. GiV([02],...)(conn)

Green's function of the renormalized

derive (10.5.3)



we

,i,m)

denotes the Green's function of N + 2 external

to connected

following



m\§\ 4- coefficient

times

1,

Zmm2.

apply

the operator

fid/dfi 4- (N/2 + l)y

to both sides of

(o\T<t>(x)cf>(0)4>(p1)...4>ipN)\oy =

C<f>2(xK0\T[ct>^4>(pl)...¥ipN)\0} + 0(x2\

(10.5.4)

to obtain

0=r^^*2' ^(y y?>c^"j<°l

^t*2]*^^

"■"

Since the

0(x2)

terms are order





*<p^)l°> -?- °c^2>- (10.5.5)

x2 independently of g,

iid/dp is applied. (As usual, 'order possible logarithms.) Immediately we obtain

remain of order x2 after x2 modulo



^<^2



(ym + y)C>

m, and \i,

x2'

means

they

order

(10-5.6)

276

Operator-product expansion

be solved

0 theory. The it is effectively in the m anomalous dimension of C^2 is the anomalous dimension of 0(x)0(O) minus This

equation

can

the anomalous dimension of





[</>2].

This is the

same

engineering dimensions. The signs in (10.5.6) arise from definitions of ym and y.

relation

as

for the

peculiarities

of our

11 Coordinate space

previous chapters we set up renormalization theory in momentum space. In this chapter, we will give a treatment in coordinate space. Now, the utility of a momentum-space description, such as we gave in the earlier In the

chapters, comes from the translation in variance of a problem. However, the momentum-space formulation rather obscures the fact that UV divergences arise from purely short-distance phenomena. Thus a coordinate-

space treatment is useful from a fundamental point of view. There are also number of situations, essentially external field problems, where

a a

coordinate-space practical appropriate point of view. A particular advantage is that the coordinate-space method from

treatment is the most

makes it easy to

see

that the counterterms

are

the

a more

same as

with

no

external

field. An important case, which

we

will treat in detail in this

chapter, is

that of

thermal field theory at temperature T (Fetter & Walecka (1971), Bernard (1974), and references therein). There one works with imaginary time using

periodic boundary conditions (period 1/T). It is first necessary to work out the short-distance

singularities of the free

propagator. This we will do in Section 11.1. A number of forms of the propagator will be given, whose usefulness will become apparent when we treat some

examples

in Section 11.2. The reader should

probably skip

to

this section first and refer back to Section 11.1 as the need for various

properties of the propagator arises. We will explicitly show that the particular counterterms that we compute are independent of temperature. Our arguments will be of a form that will readily generalize not only to higher orders but also to other situations. In Section 11.3 we will show to all orders of perturbation theory that counterterms at finite temperature are independent of T. We will do this by constructing counterterms in a theory in flat space-time in such a way as to make manifest the fact that only the short-distance singularities of the propagator affect the counterterms. Another case, which we will treat in Section

277

11.4, is flat space-time with,

278

Coordinate space

electromagnetic field. If the field is weak it can be treated as perturbation, i.e., as part of the interaction. But if the field is strong, one

say, an external a

put it in the free Lagrangian. Thus the free electron propagator satisfies

must

(i? + eA- M)SF(x,y;eA)



id{x



(11.0.1)

y),

which

cannot in general be solved by expanding in a series in the field (Schwinger (1951)). Another common external field problem is that of quantizing a quantum field theory in a curved space-time, where there is no remnant of a global Poincare symmetry. Momentum space is then an inappropriate tool. One

wishes to show that the UV counterterms We will not discuss this

space-time. the techniques given.

can

be kept the here. But

particular a fairly simple case

described should enable

11.1 Short-distance

singularities

same as an

in flat

extension of

treatment to be

of free propagator

11.1.1 Zero temperature

graph is ultimately determined by the short- distance singularities of the free propagators that make up the graph. So we The UV counterterm of a 1 PI

need to obtain these singularities; and

we

will do this in this section.

It is sufficient to consider the scalar propagator

Propagators for fields with spin can be obtained by respect to x. The propagator satisfies the equation

differentiating

(11.1.2)

(□ W)SF=-i<5(d>(x), together with appropriate boundary equation reduces to

conditions. When

(4zd2/dz2 + 2dd/dz where

z =



x2. Thus SF



Bessel equation of order determined and that





m2)SF



z1/2_c//4w(mz1/2), where =

d/2



1. The

with



is non-zero this

0, w

(11.1.3) satisfies the modified

particular solution

we

need is

by requiring that the (5-function in (11.1.2) be obtained at x2

SF-?0

as

x2->



oo.





279

singularities offree propagator

11.1 Short-distance

A standard method of solving (11.1.3) is

to

expand in

powers of

z:

00

n = 0 =

There a



independent

are two

*2)flK

("



singularity

of these solutions. The

is to observe that, if

SF(m



0)



0, i.e., it has l—d/2. SF is a linear

solutions. One is analytic at

0; the other solution is singular, with

combination



T(d/2





(11.1.4)

m2x2)nwn. a



quickest





way to

compute

the

0, then (11.1.1) gives

l)/[4nd/2(



(11.1.5)

x2f2 ̄ *].

This normalizes the coefficient of the singular solution. The normalization of the regular solution is obtained from the properties of Bessel functions, by

requiring

that SF

-?



as

x2

-?



oo.

Then

we

find that

f fm2x2\nV(d/2-l-n)



00



md-2(4n)-"12 £ (m2x2/4)T(l-n-d/2)/n\ n = 0



Here,

we

SFsing + SFana.

have used the series

(11.1.6)

expansion

coefficient of the regular solution computing, from (11.1.1), that SF



can

T(l

of the Bessel functions. The

quickly by explicitly

be obtained -

d/2)md-2/(4n)d/2

at x



0, if d is

less than 2. For the purposes of renormalization

we will need the singularities of SF. of two types: the singularities of SFsing as x->0, and the singularities of SFana as d approaches an even integer. The fact that most of

These

are

poles when d is an even integer d 2a>, reflects the fact that the ansatz (11.1.4) is then incorrect for the singular solution. However, the F-functions have



the correct result is obtained by expanding in powers of d letting d -> 2a>. The limits x -> 0 and d^>2a> are non-uniform.

graphs

like the

</>3 self-energy

need





2a> and then

Application to

± 0, but tadpole graphs (e.g.,

Fig. 11.1.1 in 04 theory) have a propagator with

Fig. 11.1,1. Graph with







0 for all values of d. So

tadpole.

280

Coordinate space

it is convenient to extract all the x





singular behavior of SF close to



2a> and

by writing

Op

>FSing



Jpsing + ^Fana



4n*,{

jyo-i

_



n = (o- 1

s JFana



\^oy ̄

W V"

\"





n!(

11/7//

y Z^ M-r^w (4?r ?^0V

as



(-l)n^V

77 I





(<//2





&))(?

(-1)"+V ̄2<°

+ 1





____^_—_|.

x|r(1_n_,/2)(_j finite limit



/2-a>

eo)! J j'



7 n!



{m2\il2 ̄a

Neither term is





(- nx2f



Trx2)""2""



T(d/2-l-n)









n=0







solution of the equation for SF. The analytic part 0. The no singular behavior at x

d^>2a>, with



(11,.7) now

has

singular

x^O. It contains the x->0 we then first take d < 2, then x -? 0, SF(x 0) 2a>. Then the only singular contribution is from the n

term also has a finite limit as d^2a> if

singularities.

But if

and finally d

-*

co-l term in

need

we





SFsing: (_l)^m2co-2^d-2co

w^oi^:/^:,^ (^/2-a))(47r)<>)-l)!

(in.?

?,.

This result will

be needed in evaluating

tadpole graphs.

11.1.2 Non-zero temperature Thermal Green's functions at inverse temperature p

Wick-rotating time conditions (Fetter SF(x y;p) satisfies



->



h and then



Walecka

1/7 are obtained by by imposing periodic boundary

(1971)).

Thus



the

free

propagator







d2/dr2



V2



m2)SF(T,x)

Sf(t,x)->0, Sf(t



P,x)





(5(t)(5(3)(x),

asx->oo,

Sf(t,x).

(11.1.9a) (11.1.96)

(11.1.9c)

integrals over space-time are restricted to the time range between 0 and The p. propagator is obtained from the zero-temperature propagator All

SF(x;oo) by writing

SF(*'J)=



S¥{T + n09i;cc).

(11.1.10)

of counterterms

11.2 Construction

(Note that in

apply the







The

SF(x

d2/dT2





V2

y;P) both

tx and zy

+ m2 to it the

are

in low-order

281

graphs

between 0 and p, so when we in (11.1.9a) arises from

only (5-function

0 term.)

only singularities

in

SF(x\p) are from the n



0 term when xM



0,

so

have

we

SF(x',P) where

SFsing

is the

same as



in (11.1.7), with

term is different. It differs from

A5F=

which, in the (D + m2)SF



(11.1.11)

S¥u]ag(x) + 3¥m(x;P), SFana(x; oo)



x2



at zero

x2. The analytic temperature by

t2



£ SF(T + nP,x;oo),

range 0 < t < p, is a solution of the homogeneous equation 0. Notice that it is not a function of x2 alone, i.e., it is not

Poincare invariant. Another, direct, way of deriving the form (11.1.11) is to expand SF(x; P) in

require the differential equation (11.1.9a) right-hand side ensures that SF SFsing(x) + SFana(x;p), where SFsing is the same singular solution as in (11.1.6). Then SFana is a solution of the homogeneous equation analytic at x 0, and for which SFsing + SFana satisfies periodic boundary conditions. We then add and subtract the poles 2Xd 2o^ to obtain (11.1.11). The coefficient of the pole in 5Fana is the same as at zero temperature, because it must cancel a

power series about xM



0, and then

to

to be satisfied. The (5-function on the







the pole in

SFsing,

which is

independent

of temperature.

11.2 Construction of counterterms in low-order graphs To explain how renormalization works in coordinate space we consider the graphs shown in Figs. 11.2.1 to 11.2.4 for 04 theory in four space-time dimensions. These

are

sufficient to show the various

complications

that

occur. Our treatment

The

works

as

simplest example





(a)

well at any temperature.

is the

one-loop

correction to the propagator,

* y

(b)

Fig. 11.2.1. Self-energy graph and counterterm.

282

Coordinate space

Fig. 11.2.1. Its contribution to the full propagator is: M,



-yA*4"" \ddzSF(x- z,d;l/T)SF(z

y,d;l/T)SF(0,d;l/T). (11.2.1)

SF(x z) and SF(y z) are singular singularities are integrable, so that they

The functions the







divergence. The factor

(11.1.11)

Mx

to write

Ml







SF(0,d;l/T)

at z



and at



y, but

is divergent,

so we use

(11.1.8) and

as

WZS^X



Z)S^Z



>^Fana(0, d



term

4; l/T)



0(^4)-

evidently exactly cancelled by

is

(1L2*2) the

renormalization counterterm, Fig. 11.2.1(b). The result for the lized propagator at order g is

! Jd4zSF(x





do not contribute any

||ddzSF(x- Z)SF(Z"y)ri^ The divergent





z)SF(z



y)SFana(0,4, l/T).

mass-

renorma-

(11.2.3)

Notice that the renormalization only involved the singular term in SF. Thus we have shown that the counterterm is the same at any temperature. The unrenormalized

M2



graph

U- w4"")2

of

Fig. 11.2.2 for

IdV*sF(*i



the

y)sF(*2

four-point function



y)SAy

SF(y



z)2

divergence at y



z; it is

M2



Since the divergence is

d approaches 4

as

logarithmic.

g2

*? (11.2.4)

xSF(z-x3)SF(z-x4). The only



is

Let

conies us

from the

singularity

of

write

L'yd'zgSrly -z)2f(y,z,{Xi}).

(11.2.5)

logarithmic, it is governed by the leading behavior of

(a)



Fig. 11.2.2. Vertex graph and counterterm.

11.2 Construction the

of (11.2.5)

integrand

of counterterms

as y-?z.

Thus

in low-order

we can

graphs

generate



283

counterterm,

illustrated in Fig. 11.2.2(6), M,

-J-Vy&Wf{y,yA*ii) [pole

singularity is entirely given by (11.1.7) of SFsing The

sF1 The

pole-part factor



d"zSF1 (y-*)2]. (11.2.6)

part of

the first term

r(d/2 -1 )/[47rd/2(



SF1

in the

expansion

x2y"2 -*].

in (11.2.6) is then

-polejj d^SF1(y-z)2|=-pole|r(^ ̄1)2J = -—-

d'z(-z>?-

dzz-

pole <

1 fiu 8n24-d*

In the first line

we

and the notation The factor

\id ̄*

j*

shifted z

_

z to z +

(11.2.7)

y. In the second line we Wick rotated

0dz indicates that only the region near z

in the last line is needed,



as

usual,



z°,

0 matters.

to preserve

explicit

dimensional correctness. We thus find that the renormalized M2 is

M2

\im[ddzddy[g2SF(y- z)2+8^QW)('-z)



(11.2.8)

xf(y,z;{Xi}).

4. Again The factor in square brackets is a well-defined distribution at d observe that we only used the singular part SFsing of SF in obtaining the =

counterterm, The

so

temperature-independent. logarithmic subdivergence as well

that the counterterm is

graph Fig.

11.2.3 has



(?)



>c<



(b)

(c)

Fig. 11.2.3. Two-loop vertex graph and counterterms.

as

an

284

Coordinate space

overall

M3

divergence. =

i(





g3

iddWddyddzSF(x1

igfi4-")3

SF(x4



Its unrenormalized value is



z)SF(w



y)SF(w

jd'wd'yd'zf <3>(w,

y, z,





w)SF(x2

z)SF(y

{xj )SF(w

w)SF(x3



y)



z)2







y)SF(w



z)SF(y



zf. (11.2.9)

If w is not close to y or z there is a logarithmic divergence when y approaches z. This is identical to the divergence in Fig. 11.2.2. There is also an overall divergence when all of the interaction points w, y, and z approach

each

other.

We first subtract the

R(M3)



subdivergence by

g3

the counterterm (11.2.7):

j}(3)(w,y,z, {xj})SF(w [sF(y- *+ Sn2(4-d) 8?\y-z) -



y)SF(w



z)





(11.2.10)

z. However, it is The distribution in square brackets is singular at y to 4, if it is it d that a when finite result is, goes integrable; produces z. function with The a at test function continuous y integrated =



/(3)(w, y, z)SF(w



y)SF(w

z) is continuous



at y



z, unless also w







z.

The remaining divergence in R(M3) comes from the region w y ̄ z, where z are both equal to if and not continuous. also is is a (There singularity y /(3) x3 or x4; however, this region is again integrable.) The divergence at w  ̄









is logarithmic,

singular

terms of the

C=



so

again it is determined entirely by the leading

propagators. The

counterterm is therefore

JdW/(V,w,w,{xJ)rx)lej fdVdzSF1(y)SF1(z)

[ s^-z)2+s?,f^^-) shown in

Fig. 11.2.3(c). singular but integrable at y as



SF(y)SF(z)



(11.2.11)

Recall that the factor in square brackets is z. So the only divergent behavior comes when

is singular, i.e., at y







0. It is

easily checked that the

counterterm C then reduces to 3?2<<-8

just



as

JdV(3)(w,w,w,{xi})('-^

)2



2T2-



(4-d)2

4 ̄d

(11.2.12)

in momentum space (Vladimirov, Kazakov & Tarasov (1979)).

11.2 Construction

of counter terms

in low-order

&* *¥■ (a)

285

graphs

(c)

(b)

Fig. 11.2.4. Two-loop self-energy graph and counterterms.

Our final

graph, Fig. 11.2.4,

is



little

more

difficult. Its unrenormalized

contribution to the propagator is M4



±(



xg^-d)2\&dy&dzS¥(x,



y)SF(y



zfSF(z



x2)

^^2|dVV(4)(y^;k-})sF(y-z)3.

(n.2.13)

The difficulty is that in momentum space there is not only the quadratic overall divergence, but also three logarithmic subdivergences. However (11.2.13) appears to have only an overall divergence, when y goes to z; this involves all three propagators. Furthermore, the counterterm, calculated in momentum space, for a subdivergence from two of the lines gives a graph of the form of Fig. 11.2.1 (a). There the third propagator is at y z, whereas the —

propagators in (11.2.13)

are

also used at y ^ z. case is to write

The correct way to handle this SF

where SF1 is the ^

Fsing

same



SF1+SF2

leading term

as



(H.2.14)

Srcm,

before, and SF2 is the second term

in

?'

SF2(*)



r(d/2-2)



16^

(-nx2)"12-2

d-4



d/2-2_

(11.2.15)

z) is finite as d -* 4 and/or y-*z. We now substitute (11.2.14) for each factor of SF(y z) in (11.2.13) and obtain The remainder

STem(y





M4



g2\ \d'yddzf?\y,z)SF1(y + 3\

/<4>(y,z)SF1Srem U f/(4)(y,z)SF1srem|

-zf + 3 f?\y,z)S2F1SF finite.



(11.2.16)

SF1's has the quadratic divergence, while the terms with two SF1's are logarithmically divergent. The other terms are finite. The term with SF1 Srem has a divergence coming from the two SF1's, and its factor 3 reflects the fact that there are three divergent subgraphs. The graph with the The term with all

286

Coordinate space

counterterm for the

subdivergences

is Fig. 11.2.4(b):

2,,4-d

M*?

8^(4





^Wto'*' W>Sf<°>

g^fr^ldV(4)(y,>';{x1})[sF2(0)





srem(0)]. (i 1.2.17)

Here we used the value

we calculated at (11.2.7) for the counterterm vertex. The Srem(0) term here cancels the corresponding divergence in the third term in (11.2.16), while the SF2 term combines with the second term in

(11.2.16)

to

produce

302



logarithmic divergence from

iV4)(y,y){ [

ddzSF1(y-z)2Sh2(y-z)n

d'

8-2d

,4-<J

Let

/(4)

us return to

;SF2(0?. Sn2(4-d)*¥2K"'f-

(11'118)

(11.2.16). The divergence in the

first term is sensitive to

and its second derivative at y

g2

Jd'z f

z:

d"ySFl(y-z)3[/<*>(z,z) + (y -zf(df^ldy%=z +





finite +

[2ig2nd'2/r(d/2)]

Uy



zY(y



zy(d2f^ldy"dy%=z-\



y2 d2fA)

^-■?[/<<>M>- =

finite





Note that the

Tddy^dy.

ddyddzf{4)(y,z)nS{d)(y

16n2 term with

/(4)(z,z) gives



pole



at d

z)/(8 =



2d).

(11.2.19)

2 but not at d



4.

11.3 Flat-space renormalization In the last section

Our method

we

was not

computed counterterms for some low-order graphs. a good method for computing the finite parts. But it

made very explicit the fact that the counterterms

depend only on the short-

distance structure of the free propagator. In particular it made it obvious that the counterterms are independent of temperature. We must now spell out how to

generalize

the results to

an

arbitrary graph

in

an

arbitrary

theory. We define the renormalization of a graph G by the same structure that we had in momentum space (i.e., by the recursion method or by the forest

11.3

formula).

Flat-space renormalization

287

A counterterm is needed for the overall

divergence of every 1PI subgraph is obtained by first counting powers of position as all of its vertices approach each other, and by then taking the negative of the result (since it is an x -> 0 divergence). The integrations over relative positions of the vertices are included in the power-counting. The result coincides with the usual momentum-space

subgraph. The overall degree of divergence of a

definition for

1PI

1 PI

subgraph. The value of G is obtained by integrating over positions divergence and subdivergences come from regions where some of these positions approach each other. To each region corresponds a subgraph consisting of the vertices that go to the same point, together with all the lines joining them. It might appear that these graphs should be in one-to-one correspondence with the 1 PI subgraphs that in momentum space are divergent when all their internal loop momenta are large. However this is not so, for the following reasons: a

of its interaction vertices, and the

(1) There are graphs overall divergent in coordinate space that are not 1PI. (2) There are divergent subgraphs in momentum space that are not obtained directly in coordinate space; these are when some but not all lines connecting vertices of a 1PI graph are in the subgraph.

example of the first case is Fig. 11.3.1, where there is a logarithmically divergent graph consisting of the vertices at w, y, z, and of the lines joining them. In momentum space the counterterms for the self-energy subgraph cancel all the divergences. But in coordinate space SF(w y) is singular, so we have to justify the momentum-space result. Other graphs are divergent in momentum space but do not occur directly as divergences in coordinate space. A typical example is given by the one-loop subgraphs of Fig. 11.2.4. An



The divergence

comes

the trick

to

propagator,

from the region y ->z, and it appears to involve all

just two propagators. As we saw in Section 11.2, this handling problem is to make a decomposition of the

three propagators, as

never

in (11.2.14).

 ̄/w Fig. 11.3.1. The









subgraph is one-particle-reducible and overall divergence in coordinate space.

w-y-z

apparently

has

an

Coordinate space

288 First, however, let

us

consider the

of divergent 1PR

problem

graphs

like

of Fig. 11.3.1. After subtraction of the self-energy counterterms, Fig. 11.3.1 has the form the (w,y,z)

subgraph

ddwddyddzf(w)SF(w

x[SF(y-z)3 Here f(w) is

y)





(11.3.1)

AS(y-z) + B\jS(y-z)]SF(z-x2).



function

non-singular at w y. The factor in square brackets is the self-energy plus counterterms. At d 4, we have SF(w y) 1/ (w y)2 and SF(y z)3 l/(y z)6, so it might appear that there is a logarithmic divergence when w and y both approach z. A counterterm for this divergence would be a four-point vertex and therefore allowed. a











However, the momentum-space result gives

important

no

directly

to derive the same result

We first integrate







such counterterm,

so

it is

in coordinate space.

over w, and see that

iddwf(w)SF(w-y) non-singular as a function of y. Then we can perform the y- integral. Although SF(y z) is singular at y z, it is integrated with a function with no singularity there, so the counterterms are sufficient to cancel the divergence. Then, finally, we integrate over z. The crucial point is is finite and





that

find an order of integration with the following property:-each kills the singularity on precisely one propagator without integral new introducing singularities. If we replaced Fig. 11.3.1 by Fig. 11.2.3, say, we would have a not-quite similar integral: we

ddwddyddzSF(w The

SF(w

integral —

over

w,

for



y)SF(w



example,

z)[SF(y

we

z)2



counterterm].

involves two propagators

z). The result becomes singular

So far



at y



SF(w



y),

z.

have chosen to define renormalization by the forest formula

or

that, after subtraction, only 1PI by graphs have divergences. Some subgraphs that need counterterms in the recursive method. We have

momentum space

(like Fig. 11.2.4)

in coordinate space, since

we

subgraphs

we

saw

7 as in

general and

do not appear

explicitly

as

divergences

defined the subgraph corresponding to

coordinate-space divergence to Even so,

seen

include all the lines

in Section 11.2 that

Fig. 11.2.4(b).

we



connecting its vertices.

have counterterms for such

To show how these counterterms arise in

to show that the counterterms are

independent of temperature,

11.3 Flat-space renormalization we

decompose SF

289

as

SF



SF1







(11.3.2)

SFN + Srem.

singular part SFsing of SF. We larger subgraph of the graph G that is being renormalized. Thus when SF is replaced by Srem, any (sub)graph y containing it becomes overall convergent. When we substitute (11.3.2) for every propagator in a graph, we obtain a sum of graphs, in each of which every propagator is replaced by one of We or renormalize each of these new graphs separately. A 5F1,..., SFN, Srem. subgraph needs a counterterm if its overall degree of divergence is a positive integer or zero, and we make counterterms for all 1PI subgraphs that are divergent. The rules for computing the degree of divergence are essentially obvious. Any propagator with SF replaced by SF1 contributes as in the original graph, and a replacement by SFN reduces the degree of divergence by IN 2. The only subtlety is that replacement of SF1 by Srem(x y) is considered as contributing 2N to the degree of divergence, even though it d + 2N does not behave as (x 0: Its analytic part behaves when x y)2 y as (x y)°. However, the analytic part never actually contributes to any overall divergence, and the singular part of Srem does go like (x y)2 ̄d+2N Here

SF1

choose N

the first N terms in the

to

SFN

so

that 2N is

are

than the

degree of divergence of any













->





relative to A typical

SF1. case

is the third term

on

of the subdivergence, this term is

if{4)(y,z)Srem(y SF1 (y z)2  ̄(y- zf logarithmic divergence at

Now a





2d



proportional -

z)SF1(y

as y -> z, =

z.

[/(4)(y,z)Srem(y z)[sF1(y -

the right of (11.2.16). Before subtraction



to

z)2ddyddz

while Srem



SFana(0) + 0. So there is

But after subtraction, -

z)2



we

have:

s*(4*_d)Hy *)]d dyddz. -

integrated with an analytic function like/(4)(y,z)SFana(y-z), the factor in square brackets gives a finite result. So any divergence comes from When

singularity in 5rem. But the singularity is of order (y z)6 ̄d, so no divergence occurs. We now generalize our treatment of Fig. 11.2.4. We examine structures of the form of Fig. 11.3.2(a). The subgraph A is 1PI and overall divergent. The lines /,,..., /Join vertices of A. A graph that consists of A and some or all of the /-'s is 1PI andjs overall divergent. In coordinate space, these divergences come from the region where all vertices of A approach the same point; it

the



290

Coordinate space

Fig. 11.3.2. General structure where subgraphs but appear to have

no

are divergent in momentum space, corresponding divergence in coordinate space.

lv..., la all participate in every one of the divergences. In order to preserve the correct counterterm structure we must include counterterms for graphs of form A in our definition of renormalization. appears that the lines

This

ensures that counterterms are given by terms in the action. For simplicity consider the case in 04 theory of one line / with two external lines pu p2. Schematically we have

ddwddxddyddzSF(x Here

SF(x





w)A(w, x, y, z)f(y, z).

(11.3.3)

w) is the propagator of line /, while f(y,z) is analytic at y z graph. A(w, x, y, z) is the value of graph A. We =

and represents the rest of the

subtract off all subdivergences of the graph Au{l}. Among these is the overall counterterm for A, which is of the form

CA(w,x,y9z;d9g,ri since



gives





C(d,g,^d\w

logarithmic divergence.



z)d^(x



z)d?(y

The result of



z),

replacing

(11.3.4) A

by its

counterterm is then

C(d9g9ri(ddzf(z,z)S¥(Q).

(11.3.5)

by the Feynman rules, since when four-point Green's function without additional loops it needs

This counterterm is necessarily generated A

occurs

an

in



overall counterterm. We wish to show how this counterterm is needed to

cancel certain parts of the divergences of (11.3.3) as y-*z. Let us now substitute each propagator by (11.3.2) with JV



2. The

/may be replaced by SF1, SF2, or Srem. The terms with SFx and SF2 both contribute to a divergence at y z, whether in the original or in The from (11.3.5) with SF SF1 or (11.3.5). divergence graph (11.3.3) SF2 is local and temperature-independent, so that it can be cancelled by overall wave-function and mass counterterms. The terms with Srem also give divergences,,but cancel in the sum of (11.3.3) and (11.3.5), just as in the simplest case of Fig. 11.2.4. propagator for line





291

11.4 External fields

The argument

presented above

is

considerably

more

cumbersome than

needed for the particular case considered. However it was presented so emphasize its form as a special case of a general argument. This

happens

only

to be the

as to case

present in Green's functions. However in the

one

bubbles (used to compute ground-state energies) and in the presence of composite operators the degree of divergence can be higher, with the consequence of needing a more general proof. vacuum

In any event, the moral is that if

we

divergent subgraphs including those

choose counterterms to cancel

of form A in Fig. 11.3.2, then the

counterterms need only depend on the SF1,SF2, In particular, the value of SFana is irrelevant. It is only SFana that knows the boundary conditions, so only it knows the temperature. Note that the SF/s are monomials in mass. Hence a

our counterterms are

totally different

polynomials in mass parameters, as we saw by

method in momentum space.

11.4 Externa] fields Consider

as

an

generated

by











example QED

classical

source

in the external electromagnetic field

J*. We have

(in covariant gauge)

iFl* + #(#+ **- M)\l* J?A^ i(Z3 l)Fjv + (Z2 l)0(tf+ eAty -







W(Z2M0



M)

-im)d-A2. (11.4.1) To separate the classical and quantum parts of the electromagnetic field, we let stf ^ be a c-number potential that satisfies the classical Maxwell equations with source J". Then we replace A^ in (11.4.1) by A + st ^ with the result &









iFl* i(Z3 (Z2







0(tf+ est- Mty + e¥Aty 1)F*V + (Z2 l)0(i*+ eA)iP (Z2M0 M)^

(V2OS'A2





\)\\iei#\\i





st^

(11.4.2)

+ total derivative.

We assume the gauge condition d -srf







with the classical c-number part

3-^ Q^







J".

electromagnetic field is satisfying

0. The total

(11.4.3)

¥t^ is large then we are not allowed to expand in powers of st^. Indeed, as,Schwinger (1951) pointed out, the electron propagator If the field

SF(x,y;ejtf) in the external field is not necessarily

analytic

at jtf



0.

292

Coordinate space

Moreover, he showed that it is the

production in an electric field. without assuming that esi\

non-analytic part that is relevant for pair perturbation theory in e

We will therefore do

is small. The lowest-order propagator (in

powers of e) therefore satisfies the equation: (iy?d/dx* +

Since SF is

no

efd^

longer simple

renormalizations

are

the



M)SF(x, y; est)



i(5(4)(x



y)\.

(11.4.4)

in momentum space, it is not clear that the with si\ 0. Let us work in coordinate

same as



(11.4.4). The solution is y, plus series analytic there. We only need explicitly terms, and to prove renormalizability we decompose SF as in (11.3.2). It is important not to use an infinite series for SF at that stage, because the series will not converge for general values of x, y and si. The singular series is obtained by treating both e and M in (11.4.4) as space. We construct

power series in



series singular at x the first few singular a





y to solve





perturbations and expanding in powers. This is similar to the way in which the singular part of the scalar propagator is an expansion in powers of m2. We obtain the leading power of x y by solving —

vfdldx?S¥1 =i(5(x



(11.4.5)

y)\.

Each non-leading term SFn is obtained in terms of earlier terms by solving at x±y the equation n-l

iy?d/dx?SFn



(M



est ^f)

(11.4.6)

£ SF,

SFn uniquely, we require it to be a singular power of x y times a function of y. The operation iy^d/dx* makes SFn more singular while multiplication by M or by ey^st^ leaves the degree of singularity unchanged. After this work we see that the renormalizations are correctly given by treating as a perturbation the term JM^ in the Lagrangian (11.4.1) before To obtain



the shift of the field to

A^ + si ^ Since A^ is the renormalized field, this term

cannot affect the divergences

with each of

one or more



it

simply tells us to integrate J^(x) separately an ordinary Green's

external photon fields of

function. We

see

that, after the shift

to

the divergences are correctly si and then taking the first few terms.

A^ + si^

treated by expanding in powers of The non-analyticity is entirely confined to the remainder term in SF; this contributes to important physics, but not to the divergences.

12 Renormalization of gauge theories important to show that renormalization of a gauge theory can be accomplished without violating its gauge invariance. Gauge invariance is physically important; among other things it is used (via the Ward identities) to show that the unphysical states decouple ('t Hooft (1971a)). In Chapter 9 we considered the case that the basic Lagrangian of a theory is invariant under a global symmetry, as opposed to a gauge symmetry, such as we will be investigating in this chapter. We showed that the counterterm Lagrangian is also invariant under the symmetry. Suppose now that the basic Lagrangian is invariant under a gauge symmetry. One might suppose that the counterterms are also invariant under the symmetry, just as for a global symmetry. This is not true, however, since the introduction of gauge fixing (as explained in Sections 2.12 and 2.13) destroys manifest gauge invariance of the Lagrangian. One might instead point out that the theory with gauge fixing is BRS invariant and deduce that the counterterms are BRS invariant. This deduction is false. To see this, we recall that an ordinary internal symmetry relates Green's functions with certain external fields to other Green's functions differing only by change of symmetry labels. However, BRS symmetry relates a field to a composite It is

field (2.13.1). This wrecks the proof of BRS invariance of counterterms except in an abelian theory, where the Faddeev-Popov ghost is a free field.

Before treating the non-abelian case, let QED. The

Lagrangian





us

examine

an

abelian theory,

is

i(Z3 \)Fl + (Z2 l)0(i0 + eRA- MW Z2{p(M0 M)¥ (l/2£)d-A2. -







(12.0.1)

Here eR and M are the renormalized coupling and mass, while A^\p and \j/ are the renormalized fields. We have chosen to include only counterterms

invariant under the gauge transformation

^->e-^, 293



(12.0.2)

Renormalization

294

of gauge

theories

special case of our later results will show that coefficients of the possible gauge non-invariant counterterms actually vanish. (We will write eR fi2 ̄d/2e if dimensional regularization is used, with e being the dimension-





less renormalized charge.) In terms of unrenormalized quantities Z\/2*?, <Ao Zj'V, e0 Z" *'2eR9 and £0 ^ {Z3, we have ^



i(dHA0v Mo/ + <Mtf + eo4o -



The counterterm for the

photon-electron (Z2

which has



A^











Mo¥o

vertex



i^i0)d'A2. (12.0.3)

is, from (12.0.1),

(12.0.4)

Ije^A^

coefficient

It is easy to

proportional to the wave-function renormalization. verify that to one-loop order this result is correct. Our later

results prove it to all orders. If the non-abelian

theory of (2.11.7)

and (2.11.8)

gauge-invariant counterterms, then the

same

were

renormalized by

relation of the

iJ/Aip

counter-

quark field strength renormalization Z2 would hold. It is easily verified by explicit calculation that this is false. As we pointed out above, the

term to the

reason

that the counterterms need not be BRS invariant is that the BRS

invariance is not



symmetry that relates elementary fields

to

elementary

such that the Lagrangian is BRS invariant after renormalization, but under a renormalized BRS symmetry,

fields. Even so, the counterterms

as we

will show in this

are

chapter.

One result will be that whereas the gauge transformation of a fermion field in QED is given by di// ieRa>il/, with eR being the renormalized charge, =

the

where X is

is finite. There



fermion field in QCD is dxjj

ig^fXxjjdX, renormalization factor. The divergent composite operator dip

BRS transformation of





number of strategies for proving renormalizability. Before let us remark that the aim is to show that a finite theory exists which has gauge-invariance properties. The gauge invariance is are a

explaining them,

exhibited by Ward identities. It is possible to choose counterterms in such a way that gauge invariance does not hold. For example, we could add a term to the QED Lagrangian. Since the coefficient of this term is dimensionless, we obtain a finite theory by adding appropriate extra infinite

(A^A^2

is not gauge invariant. This implies, for that negative metric states do not decouple from physical

counterterms. But this

example,

theory

processes, and the theory is unphysical.

The proofs state that it is possible to

choose counterterms so that gauge invariance holds. Several

approaches

can

be distinguished:

295

Renormalization of gauge theories

1. Invariant regulator. Use an ultra-violet regulator that does not break gauge symmetry, for example dimensional regularization. Then Ward identities are true when the Lagrangian is given by (2.11.7), (2.12.5), and (2.12.9). Allow all parameters and fields to get renormalized. The theorems

proved are that this is sufficient to make the theory finite when the regulator is removed. This is the traditional approach. The advantages center around the manifest preservation of gauge invariance. The to be

that chiral symmetries cannot be regulated gauge invariantly; this symptomizes the fact that not all chiral symmetries can be preserved after

disadvantages

are

quantization



see

Chapter

13.

2. Gauge invariant regulator + MS. Let

us

again

use

dimensional

regularization (or another gauge-invariant cut-off). But now let renormalize each separate graph by the forest formula or by

us

choose to

the recursive method (as given in Chapter 5). We do not explicitly constrain the counterterms to satisfy gauge invariance; so in general we have violated gauge invariance. But if we use minimal subtraction then the counterterms are gauge invariant. The reason is simple: since we will prove that we can

renormalize the theory gauge invariantly, the lowest-order counterterm that is not gauge invariant must be finite, after summing over all graphs of this order. But in minimal subtraction the only finite counterterm is zero. This method is of great use when renormalizing the complicated non-local operators that appear in generalized operator product expansions (Collins & Soper (1981)). We can renormalize the graphs without explicitly

investigating

the Ward identities. The disadvantage is that the method is

closely tied to



specific

renormalization

3. Non-invariant regulator any

regulator Piguet

There

non-invariant counterterms. One can use

adjust overall counterterms, if possible, to satisfy all the (Piguet & Rouet (1981), Symanzik (1970a), 't Hooft (1971a),

and

Ward identities and

plus

prescription.

& Sibold (1982a, b, c)).

are two

be used. There

different forms of the Ward identities, either of which may the Ward identities derived in Section 2.13 for Green's

are

functions, and there

are the ones for the 1PI graphs, as derived by Lee Our will use BRS the identities for Green's functions (1976). approach a with combination of and 2. It is based on the 1 together approaches

treatment of Brandt (1976). Most other treatments have used the identities for the 1PI graphs. We will restrict general

cases



our

simplest theories, like QCD. More supersymmetries are not treated here. See

attention to the

with chiral

Chapter 13 for references.

or



296

Renormalization of gauge theories 12.1 Statement of results

simplicity we will mainly treat one case: a theory of a gauge field coupled to a Dirac field, with the gauge-fixing term being the usual one, £(3-Aa)2/(2£). We assume that the gauge group is simple (in the For





physically, this implies that there is only one independent gauge coupling. With a 1/(1) gauge group and one or more Dirac fields, this theory is quantum electrodynamics. If the gauge group is SU(3) and the matter fields are in the triplet representation, then we have quantum chromodynamics. mathematical sense);

The result to be proved is that the Green's functions are made finite by renormalizing the values of all the parameters in the basic Lagrangian

(2.11.7). These parameters

are

the gauge

coupling,

the fermion

masses

the field strength renormalizations, and the gauge-fixing parameter. The resulting Lagrangian expressed in terms of renormalized fields se







\zzg% + £ zftjip (l/2{)5- Aal





M, is

Mgu (12.1.1)

Z3/flD"Cfl.

We have allowed the fermion fields to be in several irreducible representations of the gauge group labelled by i. There are separate field-strength renormalizations Z(2° and bare masses M{£ for each representation. The

covariant derivative is



where g0 and

(dfl + igRXZ-ltaAl)il/,

gR are the bare and renormalized

dimensional

regularization Following Lee (1976)

we

couplings. (With

fi2 ̄d/2g, with gR write the bare coupling as write

we

(12.1.2)



dimensionless.)





0o so

that the

coupling

Ga



Zdfd'c. + gRXCabe(df)<*A?.

—7-l/2rTa

V°



MS



GkXZ-'c^AI.

Observe that in accordance with theTesults

gauge-fixing

(12.1.4)

tensor is



the

(12-L3>

2^T720R>

of the gauge field to the ghost is XgR:

ZdfD>ca The field strength



term

(d-A)2

to be

proved,

(12.1.5) the coefficient of

is finite, when the renormalized field is used.

12.1 Statement This

gives



gauge-fixing term



of results

297

(d'A0)2/(2£0\ when expressed in terms of

the bare field, with

?o



(121-6)

Z3£

The main theorem to be proved is: Theorem 1. The renormalizations

Afg>, can

be chosen

X, Z,

Zf,

and

that Green's functions of A, ij/,

so

(12.1.7)

Z3

{/},

c, and c are finite.

To prove this result we will use the Ward identities for BRS invariance. These involve a number of

composite fields, which we also composite operators are

finite. The counterterms for these

need to prove related to the

basic counterterms (12.1.7). We will prove: Theorem 2.

(5BRS(renormalized field)/(5/lR is finite. That is, its Green's functions with any number of renormalized fields are finite. We define the renormalized BRS transformation SBRS/dlR as

follows: (1) Let the BRS transformation 8BRS be defined by (2.13.1), (2.13.2) with g and { replaced by g0 and f0 and with the fields replaced by

unrenormalized fields (i.e., A -+A0, etc). Then the BRS invariant.

(2) Define 6AR





MS

V K^

Lagrangian (12.1.1) is

6XZ ̄ 1/2Z"1/2. Then







<Ws<A/<^r

<W"/<^r Smsca/S?.R







igRxtail/ca,

^caZ + gRcabccbAlX, hgRXcahcchcc \gRXc









a c,



(12.1.8)

1.



sBRSca/SAR=-d'Aa.



Renormalized Ward identities follow from the unrenormalized The operators

multiplication by 8X/8ar. because of Theorem 2. Certain

auxiliary operators are useful proving Theorems 1 and 2.

appearing

for

reasons

in them

by finite,

ones

are

which only become

apparent in





Z □ c* +



(Z/X



4?



d^A^Xc^,

l)d^/gR + cabccbA^

c a c

The operator (9a is

zero

by

the



cabccbcc.

ghost equations

(12.1.9)

(12.1.10) (12.1.11)

of motion. We will prove:

298

Renormalization

Theorem 3. Green's functions of (Da,

fields

are

finite.

Theorem 4. Green's functions of

than zero) of basic fields

are

(bigger

¥afl, and c

theories

a c

with any number of basic

dR<f> and any number (greater finite, where 0 is any basic field.

31^

Theorem 5. Green's functions of

number

of gauge

with

c a c

with

than zero) of basic fields

are

one

or two

<5R</>'s

and any

finite.

The last few results have

no intuitive appeal. They will be needed as part inductive proof of the important Theorem 1. We will also find it convenient to use CPT invariance of the theory (after dimensional

of

an

regularization). Now reversal of one time and one space component is equivalent to reversal of components 0,1,2, and 3 with a spatial rotation. So to obtain the TP part of CPT, we need only consider reversal of the 0 and 1 components only. Therefore,

we

define

if/x v 0or l,if^v>2, 0, otherwise.

f-1,

^=jI



Let the fields transform under CPT





1,

2)

as

</,(x)^y1tT(0x),'} 4a(x)^Aa(9xl

(12.1.13)

c"(x)-*<f(0x),



?"(*)- *c"(fe). Then the theory is CPT invariant. (We

y0T



y°*5 and yMt



regarded

as

y-matrices

in which

y°



y^) Notice that the ghost field, cfl, transforms to itself,

antighost field, ca, complex conjugate fields.

rather than to the

use

even

though

these fields might be

12.2 Proof of renormalizability 12.2.1 Preliminaries

theory provide relations between different mostly relate Green's functions of elementary fields to Green's functions containing the composite fields listed in (12.1.8)—(12.1.11). However, to prove renormalizability, we actually need relations between Green's functions of elementary fields only. Consequently proofs tend to be rather indirect and long. The following references contain a representative selection of the proofs The Ward identities of



gauge

Green's functions. However, the identities

12.2 Proof of renormalizability

299

in the literature: 't Hooft (1971a, b), Taylor (1971), Slavnov (1972), 't Hooft & Veltman (1972b), Lee & Zinn-Justin (1972), Becchi, Rouet & Stora (1975),

Lee (1976), Itzykson & Zuber (1980), and

Piguet & Rouet (1981). A great introduced the simplification by discovery by Becchi et al. of their the symmetry. However, proofs still are mostly rather inexplicit. The proof was

given in this section gives all the steps needed to go from the basic Ward identities to the relations between the counterterms. The method

to be

given by Brandt (1976) and Cvitanovic (1977). A point at which proofs became rather inexplicit turns out in this method to be the

follows that many

point at which the operator ¥ail (defined It is

in (12.1.10)) makes its appearance. unobvious operator to use, but its use is essential to completing the that the gauge coupling to matter fields is the same as the self-

an

proof coupling of the gauge field. The proof is by induction on the number, JV, of loops. We assume that all graphs with less than N loops have been successfully renormalized by counterterms of the form implied by the Lagrangian (12.1.1). We also require that Green's functions of the composite operators considered in Theorems 2 to 5 are also finite up to N 1 loops if the indicated counterterms are used. The induction starts with tree graphs, which need no —

counterterms.

Our strategy is

as

follows:

(1) At each order below N loops we have values of the five independent renormalizations Z, Z2, Z3, M0, and X. For each 1PI Green's function with

overall divergence

an

computed



value of the overall counterterm is hence

at each order less than N. Partition this counterterm into a

set of counterterms to cancel the overall

graphs

divergences

of the individual

for the Green's function.

(2) Compute N-loop contributions to Z, Z2, Z3, M0, and X by imposing renormalization conditions on certain Green's functions. The Ward identities will be true, but it is not immediate that the many other 1PI graphs are finite. This is done at step (3). (3) Using these Ward identities show that Theorems 1 to 5 hold for N-loop

graphs. technical but important. Its use is that, in order to say that the only divergence of an N-loop graph is the overall divergence, we must have subtracted off its subdivergences. However, for individual graphs the

Step (1) is

imposed by gauge invariance do not hold. Consider, for the two-Joop graph, Fig. 12.2.1. Its self-energy subgraph needs a example, counterterm of the form A {g^k2 Bxkjiv. constraints



300

Renormalization of gauge theories

Fig. 12.2.1.

Fig. 12.2.2.

Similarly the subgraph of Fig. 12.2.2 has As

B2k^kv.

will see, the Ward identities

we



counterterm

imply

A2g^k2

that the total self-

energy counterterm is obtained from the term £(Z3 l)(d^Av in JS?. It follows that Ax + A2 Bx + B2. However, the relation is -





the separate graphs, i.e.,

we

have





dvAJ2

false for

Al^B1, A2^ B2.

At steps (2) and (3), where we discuss the N-loop counterterms, we then know that the only divergences are overall ones. Moreover, we know this without

having different graphs.

to check

on

intricate series

of cancellations

between

But for the purposes of finding the constraints imposed by on the AMoop counterterms, it is convenient to consider

the Ward identities

single overall counterterm for the sum of all AMoop graphs for a given 1 PI Having obtained these constraints, we decompose the N- loop counterterms into individual portions for each graph; this enables us a

Green's function.

to

continue the induction at the next order. If we

use

minimal subtraction, the counterterms

without

by-graph

worrying about

can

the constraints

be obtained graph- imposed by Ward

identities. These constraints will be satisfied automatically. For a

in

general renormalization prescription the Figs. 12.2.1 and 12.2.2 have the form

counterterms to

example, in the subgraphs

ff2k-/^"4) + fl;], ^-^[Vtf-^ + ftj],





finite and depend on the renormalization prescription. The Ward identities tell us that we can renormalize the divergences by a

where

a\

and

b\

are

transverse counterterm

(a,



a2)/(d



4)



a\ +a'2= (bx



b2)/(d



4)



b[



b'2.

Evidently ax + a2 bx + b2, and a\ +a'2 b[ + b2. The first equation always be satisfied, the second must be imposed by choice of renormalization prescription. Minimal subtraction with a\ =a'2 b\ b2 0 always satisfies these equations. =



must







12.2.2 Choice We now assume that set

of counterterms

step (1) has been carried out. The next step is to pick



of 1PI Green's functions to fix Z, Z2, Z3, M0, and X. This is somewhat

12.2

301

Proof of renormalizability

arbitrary, but our choice will determine the form of the rest of the proof. We following set:

choose the

(1) The fermion self-energy has divergences proportional to p and to These are cancelled by counterterms in Z2 and M0. (2) Z is chosen to cancel the p2 divergence in the ghost self-energy.

(3)

X is chosen to make the

ghost-gluon coupling

finite

part is concerned. (4) Z3 is chosen to cancel the part of the divergence + /cM/cv. energy that is proportional to -

Next

we

check for

of

as

far

the

as

1.

the cabc

gluon's self-

g^vk2

will examine the Green's functions used in Theorems 1 to 5 to

possible divergences

lower order

are

at

N-loop

order. Since all divergences at

cancelled, the possible remaining divergences

are

overall

of Af-loop 1PI Green's functions. These are Green's functions with either elementary external lines or with insertions of the various composite operators we use. The dimension of a 1 PI Green's function must

divergences

non-negative degree of divergence potentially divergent. The contributions to such a Green's function are (a) AMoop basic graphs, (b) graphs with counterterms to cancel be zero or greater in order that it have



and thus be

subdivergences, (c) an overall N-loop knowledge of Z, Z2, Z3, M0, and X. We

counterterm derived from

must prove that the

sum

our

of these

contributions is finite.

12.2.3 Graphs with external derivatives There is



derivative

on a

ghost line where it exits from an interaction. Thus

the 1PI graphs of Fig. 12.2.3 have negative degree of divergence

Fig.

contributing

to

<0| Tcccc\0}

contributing

to

(0\TccAA\0)

contributing

to

<0|7c accc|0>

contributing

to

<0\T&aftAc\0y

even

12.2.3. Graphs with negative degree of divergence and non-negative dimension.

302

Renormalization of gauge theories

though their dimension is zero. counterterms

are

present,

so

Subdivergences are cancelled and no overall corresponding IPI Green's functions are

the

finite.

argument shows that the ghost self-energy needs counterterm, but only a field-strength renormalization Z. The

same

no mass

12.2.4 Graphs finite by equations of motion Consider Green's functions of (9a with basic fields. These

are

finite. For

example,

<0|T^(x)^(y)|0>



Znx<0|Tcfl(x)cb(y)|0> +





gKXcadc^(0\ Tcd(x)Ac?(x)cb(y)\0> idahd{d)(x



y).

The only graphs for the left-hand side with

(12.2.1)

subgraph are of N-loop the form Fig. 12.2.4(a) and (b). The graph (a) has a ghost self-energy made finite by its wave-function counterterm. Graph (b) needs a counterterm in Oa proportional to □ ca. Such a counterterm is graph (c), which has the N- loop contribution to the Z[Jca term in (9a. Finiteness of (12.2.1) shows that this is the correct counterterm. Similarly, the other potentially divergent Green's function of &a, viz., IPI

an

(0\TOacbAc\0) is finite.

(a)

(c)

(b)

Fig. 12.2.4. Overall-divergent graphs for Green's functions of 0



12.2.5 Gluon selfenergy We have the Ward 0







identity

(5<0| Td-Aa(x)cb(y)\0)/8lR (l/O<0|T3-^(x)5^5(y)|0>-<0|T^(x)cb(y)|0>

\£-^<0\TA^(x)Ab\y)\0}



i&*(x -y\

The second term involves Oa because of our definition of line we used (12.2.1) on Ga9 and remembered that Green's have the derivatives outside the

(12.2.2)

SKAa^.

In the last

functions of d definition. time-ordering, by



303

Proof of renormalizability

12.2

Fig.

12.2.5. Gluon self-energy.

The only possible divergences in <0|T,4jJ(x),4j();)|0> are in the AMoop self-energy graphs (Fig. 12.2.5), and they consist of terms proportional to g^M2 or to kjkv. Remember that we used Z3 to cancel any divergence proportional to g^k2 kjiv. Both the remaining divergences are absent, by (12.2.2). Thus no renormalization of the gauge parameter is needed, and no gluon mass is needed. The form of (12.2.2) is just as in QED. It implies that the gluon self-energy is purely transverse, so that the longitudinal part of the propagator is unchanged by higher-order correction. Hence, we can —

write:

ddxei?;t<0|T^^x)^*(0)|0> ^

X I A  ̄\ 111 I 11 >

= —;=



&??,



q2 + ie |_

1 +

n(<z2) (12.2.3)

12.2.6 BRS transformation of A". Consider the Ward identity 0





The first

5(0\TAl(x)cb(y)\0}/5XR (l/^)<0|r^(x)a^6(y)|0>



<0|TMflW^)|0>.

(12.2.4)

just proved to be finite. The only potentially for the second term are shown in Fig. 12.2.6. Equation divergent graphs shows their sum be to finite. (12.2.4) The c(y)A(z) Green's function (Fig. 12.2.7) of dRA is also possibly divergent (logarithmically). But we have: term we have

d dx"

(0\TdRA"ix)cb(y)Acv(z)\0>







<0\TO°(x)c*(y)Acv(z)\0> -i^(x-y)^<0|^v(z)|0> 0.

t N-lOOp

Z^AMoop

Fig. 12.2.6. Divergent graphs for right- hand side of (12.2.4).

(12.2.5)





r Fig. 12.2.7. Graphs for 0>.

<0|T<^(x)c1'(>K.W|

304

Renormalization of gauge theories

Applying d/dx* is equivalent to multiplying by k^ in momentum space. Since the divergence is at most logarithmic, (12.2.5) shows that there is no divergence

at all.

12.2.7

Gluon

self-interaction

Now







5<0\TA;(x)Abv(y)?(z)\0}/SlR <0\T5KAl(x)Abv(y)?(z)\0} -<0\TA;(x)dRAb(y)c<(z)\0} -<0\TAl(x)A'Xy)5tp(z)\0> 1



where

we

divergent,

used the and

lowest-order

we

finite

previous

have

r)

+-_<0|r^4^|0>,

(12.2.6)

result. The three-gluon vertex is

equal (Z3XZ ̄l possible left-over divergence to

a counterterm

vertex. There is no



linearly

1) times the that satisfies

(12.2.6).

Similarly

—<0\TAK(w)A,(x)A^ymz)\Oy=rimte. Here the

only potentially divergent AMoop

1PI

(12.2.7)

subgraphs

are

as

in

Fig. 12.2.8. We have just seen that the divergences in Fig. 12.2.8(6) are cancelled by the counterterm for the triple gluon coupling. Since the divergence in graph (a) is logarithmic, (12.2.7) proves that by the counterterm in the four-gluon interaction.

it is exactly

cancelled

crossings (a)

(b)

Fig. 12.2.8. Potentially divergent graphs for four-point Green's function of gluon.

12.2.8 SRc The only Green's function of 5Rc that could be divergent is

<0|TV"|0>.

(12.2.8)

12.2

But 0





305

Proof of renormalizability

d(0\Tca(x)cb{y)cc{z)\0} (0\TdRca(x)cb(y)cc(z)\0} -(l/£K0\Tc?(x)d'Ab(y)?(z)\0> + (l/£) < 0| Tca(x)cb(y)d-Ac(z)\0y,

(12.2.9)

finiteness of (12.2.8) follows from finiteness of the ghost-gluon vertex, which is a renormalization condition. so

12.2.9 Quark-gluon interaction,

dip, d\J/;

introduction of 98afl

Consider the Ward identity 0





d<0\TiP(x)ij,(y)c(z)\0>/dlR <0| Td^clOy <0| T^^c|0> + <0| Ti/,¥(d-A/Z)\0>. -

Finiteness of the last term follows if we can (Note that Now X

dR\j/

is related to

dR\// by the

prove

dRil/





igRXta\j/ca finite.

CPT transformation of (12.1.13).)

defined by requiring the ghost-gluon vertex to be finite. An proof, which we now give, brings in all the remaining operators was

explicit listed in (12.1.9)—(12.1.11), and in particular &afl. In Fig. 12.2.9 we list all Green's functions still to be proved finite. Observe that dRc d-A/l;9 so that it is finite if the gauge field is. Also Green's or c a c with SR(j) and any number greater than zero of functions of 98 =



finite, if the Green's functions of Figs. 12.2.7 to 12.2.9 are (Here dR<f) is the BRS variation of any elementary field <j>) This proves

basic fields finite.

are

Theorems 4 and 5, so it remains to prove finiteness of the Green's functions illustrated in Fig. 12.2.9. The idea behind the proof is to examine the right-most vertex on the

ghost line in Fig 12.2.9(a). This interaction

comes

from the

following

term in the

Lagrangian:

LbcgR(d^ca)cbA^x.

(12.2.10)

(a)

Z^+x%¥&^-, logarithmically divergent, contributing to

<0|T<5Ri/^c|0>

{b)

JL4fcJ¥gg-+-i logarithmically divergent, contributing

<0| T<%aflc\0)

to

Fig. 12.2.9. Green's functions not yet proved finite.

Renormalization

306

The derivative is on the line

parts,

we see

that the vertex



of gauge

entering equals

theories

the vertex

graph. By integrating by

0Rc.J d4x[F(3McV" + c°<?d-A'l

(12.2.11)

loop-momentum integrals, so Factoring out the field d^c* on

In the first term the derivative is outside the

degree of divergence is the external line gives the

the

reduced by one. basic vertex

cabccbActl

in the operator

{%a)1.

The

second term in (12.2.11) contains d-A which we shall relate to something else by use of Ward identities. We first prove finiteness of ^a/J, by formalizing the argument leading to

(12.2.11). This is done in

an

unobvious way:

□,<o|raRc-(x)cty)?(z)|o> (l/X){Znz<0|T^Rcac*cc|0> + gRXccde(0\ TVa<*M^d)|0>} + gR<0\TdRc°cb(ccdec'ld-A°)\0} =

-0R^<°|rV-°^(z)|O> =



iflM^Ol TC"(x)c?(y)|0>^?(x +igR<0\TSRc°cbdR(ccdec<>c°)\Oy



finite



z)

i0R<O| TdR<fd-Ab{c a c)c|0>

-0r^<°ItV^W|O>. by the antighost equation of motion. identity plus the nilpotence property

The next-to-last line follows line

uses a

Ward

on

hand side of (12.2.12) is

The last

(12.2.13)

<5R[<Vfl]=0. Now the second term

(12.2.12)

the last line is finite (by Theorem 5), and the left-

finite.

So the last term

divergence is of the possible finiteness of ¥afl follows. uncancelled

on

right is finite. The only Fig. 12.2.10, from which

the

form of



Fig. 12.2.10. Only possibly divergent graph for (12.2.12).

12.3 More general theories

SRij/ follow by

Finiteness of

the

same

307

manipulations applied

<0|r,5R#c|0>: □ ,<O|7A<Kx)#(30?(z)|O> =0R<o|rtv(x)#oo|o>a*V-*) -igR<0\T5^3^(0 a c)\0}

to

(12.2.14)

-9K^\TSRiP(x)il}(y)^(z)\0y. 12.3 More general theories

proved renormalizable the simplest gauge theories: gauge group with a single component and with fermion matter fields. More general cases can easily be treated by the same methods. The general result is that renormalizations are needed for each independent coupling in the basic Lagrangian and for the field strength for each irreducible field multiplet. Let us examine some specific generalizations. In the last section

that is, those with

we



12.3.1 Bigger gauge group The gauge group can in general be a product of several components: G Y[ni i G{ ? ^U)v- Here each G{ is a simple compact non-abelian group =



(like SU(N)\

and there

factor there is

are v

abelian l/(l)'s. For each Gt and for each (7(1)

independent coupling g.m When we perform gauge fixing there will be a multiplet of ghost fields for each component of the gauge group. The proof of renormalizability will need no change. It will relate the an

renormalizations of the couplings within each multiplet. Thus for each of the n + v components of the gauge group there are renormalization factors Xi9Z3ii and Zh for the coupling, the gauge field and the ghost field. In addition there are the usual renormalizations for the matter fields. There

are some

special features

of the abelian

case

which

we

will treat in

Section 12.9.

12.3.2 Scalar matter The part of the

Lagrangian

for



scalar field coupled to gauge fields is

D^D^-V(<f>\<f>). Here Fis

DM



function of <\> and

(12.3.1)

¥ that is invariant under the gauge group, and

9ovariant derivative. example, consider an SU(2) gauge theory

is the usual

For

in which ¥ is



doublet

308

Renormalization

of gauge

under the gauge group. Then the most

theories

general renormalizable form of Kis {1232)

V=m2<p4>+\k(<p4>)2,

while

D^D<j>





Wave-function and

\d^\2 |3M0|2

IgA'^ffy + g2AlAb^tatb<f> \gAl4?f3"4> + ig2Aa2^<t>.





(12.3.3)

are used to make the propagator is cf)fd(f)A coupling proved finite as for a fermion. The self- of the coupling 0-field is made finite by a renormalization of the Xft 02 term, which is the only four-point coupling invariant under global SU(2) transformations. One further Green's function, viz., <0| TAA(f)f(f)\0y has a mass

renormalization

finite, and the

potential logarithmic divergence.



It is

proved finite by the Ward identity:



(l/^)<0|Ta-^^(/) (/)|0>-h<0|Tc^(/)t(/)|0> + <0|Tc^^V|0> + <0|Tc^(/)t^|0>.

(12.3.4)

12.3.3 Spontaneous symmetry breaking Consider the abelian Higgs model as an easy example. is

^basu



-i^2v + (3?



ieAJjUP + ieA?)<t>



Its basic

Lagrangian

|AVtf> -/2/2A2)2- (12.3.5)

The symmetry is spontaneously broken with < 01 </> 10 > =//(A we write </> [f//. + (</>l + i</>2)]/

order. So

^basic



Jl:





i^2v (e2f2/2l2)A2 +

ief/VA^fa + h^\ + \ty\ -/2<A2/4 MM* + 4>\)IA +

^(4>\ + </?l)2/16 + eAli<pid"4>2 + e2A2(<}>2 -

If

we

quantize with



the

*J2), in lowest

simple



<j>2 + 2/>1/A)/2.

gauge

fixing

(12.3.6)

term

(12-3.7)

**=—¥? A?,

by the discussion of Section 9.2, where we breaking in a non-gauge theory. We first renormalize in the unbroken theory (i.e., with/2 f2). We need wave- then renormalization is covered

treated spontaneous symmetry

-?



function renormalizations Z2 and Z3, coupling renormalization A2 -? /£ Y9 and mass renormalization (which is effectively f2-+f2Zm). Since the Faddeev-Popov ghost is a free field, the gauge-coupling renormalization is 1. After spontaneous symmetry breaking the same renormalizations X =

12A Gauge

dependence of counter terms

make the Green's functions finite.

Lagrangian makes use

309

Unfortunately the AJj*¥2 term in the free

the Feynman rules rather messy, so it is convenient to another gauge condition. We will discuss this in Section 12.5.

12.4 Gauge dependence of counterterms To

quantize and renormalize a gauge theory, we choose to fix the gauge in a particular way. It is important to show that physical quantities are independent of this choice. On the other hand, Green's functions of the elementary fields can certainly be gauge dependent. Since it is the Green's functions with which

we

work when renormalizing the theory,

we must

understand their gauge dependence and its effect on the renormalizations. One important class of physical quantities which we will treat is the set of

gauge-invariant operators. Also important is the S- physical states. In a spontaneously broken theory there are particle states that couple to the elementary fields of the theory. Some such states are physical, and we must prove that their S-matrix is gauge independent. But other states, like the Faddeev-Popov ghost, are unphysi- cal. We will not attempt a complete treatment here. Green's functions of

matrix for

When a gauge symmetry is unbroken there may even be no states

elementary fields. Indeed, it is commonly expected that in QCD colored states are confined. Certainly, within perturbation theory there are severe infra-red problems in obtaining the S-matrix for quarks and gluons. However, S-matrix elements of hadrons can be obtained from

coupled

to the

gauge-invariant Green's functions. Consider, for example:

<0| ^K(w)/5A(x)/5/i(>;)/5v(z)|0>,

(12.4.1)

wherej^ is the axial isospin current. Application of the LSZ formalism will give the 5-matrix for %n -? %n scattering. Similarly the gauge-invariant field *l/i*l/j*l/k£ijk is an interpolating field for baryons. Here ^ is a quark field and i its color label. There are two

sorts

of gauge-dependence that we will consider. The first is

change the gauge-fixing function Fa to a different function of the specific application of our general results for this case will be given Section 12.5 for the i^-gauge. The second type of gauge-dependence is a

where

we

fields. A

in

variation of the parameter <!;. Although variation of i by a factor X is equivalent to variation of Fa by a factor I  ̄1/2, there are a number of special simplifications that will lead us to treat this second case first. The

use

of IJRS invariance will make

dependence very

simple.

our

computation

of gauge

310

Renormalization of gauge theories 12.4.1

Change oft

The Green's functions and the renormalization factors in general depend on <!;. We

can use

the action

Green's function with

no

principle

<0|TAT|0>.

explicit dependence

to compute the

Here X denotes any

dependence on ^ of a product of local fields

<!;. Then

on

^<0|rX|0>=ild4K0|r:—{y):X\0> +

^td*y<0\T:F2a(yy.X\0>.

In the first term we use Y to denote any of the

factors. vacuum

As usual, we

expectation

use

the

symbol:

value. Next



we use

(12.4.2)

independent renormalization

to

indicate subtraction of the

the form of the BRS

transformations to write

(12.4.3)

(l/Z)F2a=d(caFa)-cadFa9 and substitute for

F\

in

(12.4.2).

We

can use

d(0\TcFX\0} and the

equation of



the Ward identity 0

motion

<0\TcadFa(y)X\0y= ̄i^(0\TcJX/dca(y)\0y a

to

give

|<0|TX|0>=iE|JdV0|y:i^|0> -^fd4>;<0|T:caFa:^|0>



(Nc/20<0|TX|0>. (12.4.4)

Here

Nc is

the number of factors of the

Suppose first that

we

choose to

use

the

ghost field

in X.

same counterterms

for all values

particular value, say £ elementary fields will not be finite except at this special value. But if we take a Green's function of gauge-in variant operators then (12.4.4) indicates that the Green's function is gauge independent, for such operators contain no ghosts. of £

as

at

some



0. Then Green's functions of

12.4

Gauge dependence of counterterms

Next consider a Green's function of

<0|7X|0>



elementary

311

fields:

<0|Tn^W)|0>.

(12.4.5)

i=l

are adjusted to keep Green's functions finite for (12.4.4) we find the ^-dependence of the counterterms

The renormalizations in <£ any value of f. Then in

by

requiring

the dYd^

\df dY'' term to be the sum of the counterterms needed to cancel the

divergences of

-:caFa:/(2a We consider the ultra-violet

divergences

of

<0\T:caFa(y):dX\0}.

(12.4.6)

Divergences in (12.4.6) interaction vertices

or

occur either when y coincides with some set of when it coincides with one of the x/s, which are the

positions of fields in X. We can renormalize Green's functions of :caFa: with elementary fields by adding to it an operator :D(y):. Since we choose X to be a product of elementary fields, the only remaining divergences in (12.4.6) are when y coincides with the position x of a BRS varied field d(f)(x) in SX. To renormalize this divergence we need counterterms to (12.4.6) of the form

<0|r:D(jO:W|0>+ £ aw(y *,)<0| TAT lO)^.^,,,, -

(12.4.7) The counterterm when y coincides with -

iE<p(x)d(y



x), there

normalization factor



are

possibly

d<j)(x)

has been written

as

derivatives of the ^-function, and the

i will be convenient later.

By use of Ward identities and equations of motion,

we can

write (12.4.7)

as

-<0|T:5i)(y):X|0>-X<0l7,:£^)^/^(y):X|0>.

(12.4.8)

This is the most general form of counterterm needed to keep the Green's function finite as £ varies. It is even the correct form for the change in counterterms caused by a change in the form of Fa9 as we will see in Section 12.4.4. To

proceed

any further we must do

which counterterms

to determine

occur.

restrict attention to the gauge condition Fa d-Aa. The only illustrated Green's functions by their elementary of:caFa:are

We now

divergent

actually

power-counting =

of gauge

Renormalization

312

theories

XT0" Fig. 12.4.1. Lowest-order graph for

Fig.

lowest-order

graph

has



cases

in

(0\TcA:caEa:\0)

Figs.

12.4.1 and 12.4.2. The

divergence

Since

integrate

over

to

only contribution caFa.

:caFa: with

12.4.3. Lowest-order graphs for

divergence proportional to fa as The



all y, the

operator D is proportional

Fig.

of the first

factor of the derivative at the ghost interaction. The (d^cJA* derivative caFa.

counterterm for Fig. 12.4.2 is proportional to we

graph for

12.4.2. Lowest-order

<0|rc:c.F.:|0>.

with



to the counterterm

BRS variation of

a BRS-varied field 6<j> is in the graphs of Fig. 12.4.3. So

logarithmic we

have



field.

and

£0oc0.

This

corresponds to a variation with £ of the wave-function renormalizations with g0 and M0 fixed. For example, dZ,d&

d£ dZ2 A

complication arises since

dZ3d& d£ dZ3

lainZ,

Hil/SS/Si)/ + il/dS/dij/)



d£

we

know £ is not renormalized:

ldlnZ, 2

;(S±F2a/2£)

l5A<

d£

+ F2J2Q). mnZ-3(A-gdS/SA d£



Putting

all this work together,

we

|<?-s(1+S

(12.4.10)

find

d*y(0\T:caFa(y):SX\0) ,

dlnZ 2

-<o|r*|o><;jv,-^





NA8lnZ3 ft

rain(zzn i

CL This is just the

(12.4.9)

ft



(12.4.11)

counterterm structure we need. The operator

caFa is

12.4

Gauge dependence of counter terms

313

multiplicatively renormalized by theory

no

a factor 1 + £d In Z3/d£. (In an abelian renormalization is needed since ca is a free field; so Z3 is

independent of £.) The other terms (where iV0 is the number of external <\> fields of X) come from renormalizations of 1PI graphs including both caFa and



6<j).

Z3/3^ occurs in two places. This comes from assuming that gauge-fixing term is d-A2/(2^) with no extra F2/(2£) renormalization factor. We proved this from the Ward identity: Note that d In

the















d(0\Tca(x)Fb(y)\0} (l/?<0\TFa(x)Fb(y)\Oy + (0\TcadFb\0} (\/ZK0\TFa(x)Fb(y)\O>-id(x-y)dabi

(12.4.12)

where the last line follows from the ghost equation of motion. Since Fa d-Aa, its Green's functions are finite, and so £ is finite. We can use (12.4.11) to prove gauge in variance of the S-matrix by picking out the residue when we go to the particle pole for each external line. If we use the correct, <!;-dependent counterterms in j£f, then a gauge- =

invariant operator like

(G°v)2



(d^Al



dvA; + QoZ¥cabcA\A<y

is actually <!;-dependent; from (12.4.11)

we see

that is

(12.4.13)

proportional

to

Z^i.

But the bare operator

(^(0)/iv) is gauge



independent.

(^i

 ̄ ̄

(0)v

°v^(0)fi + 9oCabc^{0)n^{0)v)

Both operators have UV

divergences,

so

that they

must be renormalized.

12.4.2 Change of Fa Let

us

assume

that Fa

interpolate between same

method

as

depends

two different

on



parameter

gauge-fixing

k.

This allows

us

to

conditions. We follow the

for the ^-dependence. The change in the

Lagrangian

is

d^=Y——--Fd-^ ̄-^ Ca dY 3k 3k

ydK





dK

So die

<0|TX|0>



ixgfd4y<0|T:^(y):AT|0> +



8F

i\d'y<0\T:ca-^:8X\0>.

(12.4.15)

314

Renormalization

of gauge

The counterterms for the second term have

in the

theories

exactly the same form as we met

Green's function, i.e., the form (12.4.8). Until the form of Fa is specified we can make no further statement about the necessary counterterms. We will examine an example of this in Section 12.5.

^-dependence

of



Green's function of gauge- invariant operators and omit any /c-dependence of the counterterms in <£. 0. Then the Green's functions are gauge-independent: <9<0| TX\0}/6k In the most general

case

we can

take





12.5

R^-gauge

mixing between the gauge field and the scalar field in broken spontaneously gauge theories one can use the i^-gauge devised by't To eliminate the

Hooft (1971b), Fujikawa, Lee & Sanda (1972), and Yao (1973). Let us use the Higgs model of (12.3.5) as an example. The i^-gauge is defined by using the

gauge-fixing Lagrangian J2V

We

use

the parameter

and the gauges

we





k to

(1/2£)(<M

interpolate



(12.5.1)

K^m<j>2)2.

between the

used in Section 12.2, where



i^-gauge, where



0. Also,

interacting Faddeev-Popov ghost is needed, even theory (which is the only case we will explicitly treat): An

j£?gc



d^cd^c



The BRS variations of the fields

6<f)2





(m

K^m2cc





dc





1,

ef/L

an

abelian

(12.5.2)

} F/£9\

(12.5.3)

are:



e^Jc,

(d'A-£Km(t)2)/£



0.

Observe first of all that the extra so

couplings relative to the gauge k



0 are

that the renormalizations of the dimensionless

super-renormalizable, couplings can remain unchanged. all

in



efynK<f>xcc.

Hi=e<l>2c, dc





These

are X

(which equals unity in

an

abelian theory), Y, Z2, and Z3. To determine what further renormalizations are needed, we use the method of Section 12.4.2, with dF/dK £m(/>2. =

Thus

we



need the counterterms to -ifm

\d*y(0\T:c<f)2(y):5X\0y.

(12.5.4)

The only divergent graphs contain c</>2, the BRS variation 6<f>u and other external lines. They have the form of

Fig. 12.5.1,

which is

no

logarithmi-

12.5

*■

^4>iV

315

Rr gauge 7^1



eC(t>2

Fig. 12.5.1. Divergent graphs for (12.5.4).

cally divergent. Notice that interactions on the ghost line would make the graph convergent. The counterterm to (12.5.4) therefore has the form

cfdV<0|T^/^1W|0> =

iC

(12.5.5)

d4KO|r:5S/501O'):A'|O>.

This counterterm is generated by



shift of <j>1 in the

Se(4>1+mlei4>2iA,cic)^£e(4>1



Lagrangian:

(12.5.6)

Bm/e9<f)29A9c9c)

with

?^ e

As far

as



(12.5.7)

C.

ok

renormalization is concerned, the effect of using the new gauge an ultra-violet divergence in the vacuum

condition is to generate

Although this might appear strange, it is permitted to happen since the field is not gauge invariant. Gauge-in variant Green's functions are unchanged, of course, as is the S- matrix. The ghost field, 02, and the longitudinal part of A^ all couple to unphysical states. In this abelian theory, F^ is a gauge-invariant field which expectation value of the scalar field.

couples to the transverse part of the gluon, while

ZmZ2^4>





ZmZ2l(4>1



ZmZ2(Bm/e)2

Bm/e)2 +



tf\

^ [2(ftn/e)ZmZ2]









is gauge invariant and renormalized, and couples to the (f)1 -particle. The Lagrangian can be written out in terms of the fields, to exhibit the interactions. It is rather fearsome-looking. We set k 1 and find =

£e



-iZ3F2^ (l/2£)d-A2 + \m2B2Z2Al + dcdc -



m2Blcc + AJd^2m(Z2B



1)

iZ2(d(t>Ud(t>22)-i(t>2m2^e-2(3B2Y-Zm) -h<t>Wte + \A2e-\B2Y-Zm)-] -^A2y(^ + ^)2-^2me-15y01(^ + (/)22) -iA2m3e"3B(B2y-ZJ01 + eZ2A¥xH2 +\e2Z2Al{ct>2 + <j>22) +



emBZ2<\>xAl



im<\>xcc.

(12.5.8)

Renormalization of gauge theories

316 The

same

methods

complications

can

be

applied

theory with extra Fujikawa, Lee & Sanda

to a non-abelian

in the structure of the vertices



see

(1972). 12.6 Renormalization of gauge-invariant operators

applications such as to deep-inelastic scattering we need the operator product expansion of gauge-invariant operators. It is natural to assume that only gauge-invariant operators appear in the expansion. Moreover, in using the operator product expansion, we need the anomalous dimensions In

of the operators. To compute these, we need the expression for the renormalized operators in terms of the bare operators. Again, it is natural to that only gauge-invariant operators are needed. Both assumptions, taken literally, are false (Dixon & Taylor (1974), and Kluberg-Stern & Zuber (1975)). We will first treat the renormalization

assume

problem, where certain gauge-variant operators mix with gauge-invariant operators. As we will see, the gauge-variant operators vanish in physical matrix elements (Joglekar & Lee (1976), Joglekar (1977a and b)). We will simplify many parts of the proof by working in coordinate space. Then we will apply the same methods to find the operators that appear in the operator

product expansion. ordinary global symmetry (like

In the case of an

Lorentz invariance

or

isospin), operator that is invariant under the symmetry mixes only with invariant operators. In the case of a gauge theory, we break the invariance an

of the action

by gauge-fixing, leaving only



BRS invariance. But

invariant operator even mixes with operators that How can this be?

To

answer

this question,

we

consider

an

are not

a gauge- BRS invariant.

operator 0 that is invariant

under some given transformation. Let C be the sum of its counterterms, and assume

the action is invariant. Then the Green's functions

under the transformation, o







Now C is defined finite. For

<5<o|ta:(0



c)|o>

io\Tsx{o



c)\oy



io\TXS(o

(0\TdX(V



C)\0}



(0\TX 5C\0>.

so

are

invariant

so



Q\oy (12-61)

that Green's functions of 0 + C with the variation of an

elementary fields elementary field is

ordinary symmetry, elementary field, so the first term on the right of (12.6.1) is finite. Hence the other term, an arbitrary Green's function of SC, is finite. If we use

are

again

an

an

minimal subtraction, the counterterms are powers of 1/e, so SC is finite only if it is zero. Thus counterterms to an invariant operator are invariant.

12.6 Renormalization But when 6 represents

of gauge-invariant operators

317

BRS transformation the variation of



an

elementary field is composite. So finiteness of(0\TSX(G + C)\Q} does not follow from finiteness of the elementary Green's functions of 0 + C. Hence the counterterms to a gauge-invariant operator need not be BRS invariant, let alone gauge invariant. This argument also shows us how to handle the problem. Let X be a (t)1(x1)...(f)N(xN) of local fields. Then the only divergences

product

< 01T5X ((9 (y) the position of



of

C(y)) 10 > that lack a counterterm are when y coincides with

a BRS-varied operator <5</>f(x.) in SX. Hence, provided y is not equal to any xi9 (0\TSX((9{y) + C(y))|0> is finite. From (12.6.1) we then see that <0| TXdC\0} is then finite. But if we use minimal subtraction this

means

that it is

zero.

Hence

an

arbitrary

Green's function

SC with elementary fields is zero. Thus, as counterterms to a gauge-invariant operator

an are

<0| TX5C\Q}

operator, <5BRSC



of

0: tne

BRS invariant.

try verifying this theorem by explicit calculations (e.g. Kluberg-Stern & Zuber (1975)). These apparently contradict the theorem. However, the non-invariant counterterms must correspond to operators that vanish by the equations of motion. As usual, the treatment of One

can

derivatives of fields in covariant perturbation theory implicitly generates commutator terms. After Fourier transformation into momentum space the non-invariant operators are not manifestly zero.

actually interested in are gauge-invariant operators rather BRS-invariant merely operators, for it is only the gauge-invariant that have physical significance independently of the method of operators gauge-fixing. So we must show how it is that the gauge-variant operators that mix with (9 do not enter into physical quantities. Let 0u...,0n be the set of gauge-invariant operators that mix with (9. These are the operators which have the same transformations under global symmetries (e.g. Lorentz, isospin) as 0, and whose dimension is at most that of (9. Choose this set so that it is linearly independent (after use of the equations of motion). There are three other classes of operators that mix What

we are

than

with (!): Class A These operators, Ai9

At



are

the BRS variation of

operator:

5At

Class B These operators, Class C

Any

Bi9 vanish by the equations of motion.

other operators that mix with (9, and that

combinations of

At\ B(%

are not

linear

and #.'s.

nilpotence of BRS transformations vanishing by the equations of motion.

The

some

ensures that

SAt



0 up to terms

318

Renormalization

of gauge

theories

The most obvious classes of BRS-invariant operators are the gauge- invariant operators 0t and the BRS transformation of operators bA{. The first result to prove is that there are no others; i.e., there are no operators of

class C. The second result is that the renormalization matrix expressing the renormalized operators [#], \A\ and \B\ in terms of unrenormalized

triangular:

M [A]

^7 =













Zaa

Zab\





\°\

operators is



ZBB\ w

It is easy to prove this second result. The only operators that

counterterms to

(12.6.2) can

be

class B operator must themselves vanish by the equations of motion, i.e., they are of class B. A Green's function of a class A operator SA.



be written

can

as

<k0\T5Al{y)X\0y=-(0\TAt5X\0). Thus if the

counterterm to

A{

is

C(At)

(12.6.3)

then

<0|T[^ + CM.)]<5X|0> is finite if y equals

SC(At)9

of the

xt\ Hence the vanishing by the equations

none

modulo terms

counterterm to

SA( is

of motion. That is, the

counterterms to At SA- are class A and class B only. The proof that there are no class C operators is somewhat complicated. It =

essentially a mathematical exercise in homology theory. We refer the reader to Joglekar (1977a, b) and Joglekar & Lee (1976) for proofs. The importance of these results is as follows: both class A and class B

is

operators vanish in physical matrix elements. For class B this is because of the equations of motion. For class A we obtain the matrix element by the

LSZ reduction formula from a Green's function < 01TA tX10 >. But there is

Ward



identity

<0|T<M>|0>= -<0|7\3.(5A:|0>.

(12.6.4)

Whenever X is gauge invariant we get zero. But we may also use elementary field operators in X. A term in 6X with a varied field operator 6<f>(x) has no physical particle pole for this line. This can be seen by examining the possible Feynman graphs, and observing that 6<j> contains a ghost field. If the renormalization matrix in (12.6.2) were not triangular then renormalized operators of classes A and B would be non-zero on-shell, even though the unrenormalized operators are zero. The triangularity ensures that for

physical

matrix elements

\P\ Thus

we can

disregard

we =

have

ZecQ.

the non-invariant operators.

(12.6.5)

319

12.7 Renormalization-group equation 12.6.1 Caveat

practical calculations with Feynman graphs one must beware of taking (12.6.5) too glibly. Consider the calculation of [G^] in QCD. The only gauge-invariant operator to mix with it is {J/ij/, so we can write (12.6.5) as In

(&}) (*? ZA(G^\

(12.6.6)



The Z21-coefficient is

zero

because

{j/i//

has lower dimension than G2.

Fig. 12.6.1. One-loop graph for We compute the

There

are

one-loop

term in

proportional

graph

of Fig. 12.6.1.

0vvP2



Ga2x

to

Tensors that vanish on-shell are, for

simplify

from the

various tensor structures for the counterterms. The

counterterm is

To

Zn

<0|T^[G;Jv]|0>.

example,

9Hv(P + <l)2

Pflv

the calculations it is



(P

tempting



<l\(P + ?)v

to set q



0. But then these

equal and it is not possible to separate the three coefficients. If one sets p2 0 (p + q)2 so that p and q are on-shell, and then multiplies the graph by polarization vectors e^e,v satisfying e-p 0 e'-(p + q\ then one has taken a physical matrix element and only the three structures

are









Ga2v

counterterm survives.

The calculations of

Kluberg-Stern

momentum, and they have to go to

12.7

some

& Zuber

effort to

(1975)

overcome

are at zero

the above

problems.

Renormalization-group equation

are derived in gauge theories just in as any other theory. One feature that is easy to overlook when making calculations is the variation of the gauge-fixing parameter under a

Renormalization-group equations

renormalization group transformation. We consider the theory (12.1.1) and let ^Nc.Nf.tfg ^e the renormalized Green's function with Ng external gluons, Nf fermions, Nf antifermions, Nc ghosts, and Nc antighosts. The corresponding bare Green's function is (7(0)

_

fNcyNfy]Sg/2rT

Renormalization of gauge theories

320

and it is RG invariant:

-l"£ Here

/?,

?S-'""'5=-?S>cM-



yM and 6 are obtained

by requiring

#0,

(12.7.1)

M0, and q0



Z3£ to

be RG

invariant. In the minimal subtraction scheme we have

g0



V2-d/2g\l+ fa^)/(4-^r L

n=\

(12.7.2) n= 1

2. Here



i+

I c,n(3,a/(4-rfr.

Zf stands for Z2, Z3, or Z. p





The

renormalization-group coefficients are

(d/2-2)g0/(dg0/dg) (d/2-2)g + ±gdaJdg,

yM=pdlnZM/dg=-lgdbl/dg^ y.







(12.7.3)

pd\nZi/dg= -\gecixldg, y3=

-jgdc3A/dg.

Similar formulae hold in any other subtraction scheme. Observe that, by the results of Section 12.4, g0 and ZM are independent of £ so that p and yM are independent of £. But Z2, Z3 and Z depend on ^, so that y2, y39 y and d also depend on <!;. As a special case, in an abelian theory, n2 ̄d/2gZ3112, so that Z3, y3 and 6 are independent of <!;. #0 =

The functions

renormalization-group equations

are

0-[fi^Ncy



N(y2+\Ngy3)G

&Zj£ Since the anomalous dimensions

equation

for renormalized Green's

then

are a

little

Landau gauge £



complicated.

are



Ncy + Nfy2

^-dependent,

computed /?

*Jty3 )G.

(12.7.4)

the solutions of the RG

The most convenient gauge to

use

is the

0, for then £ does not vary when a renormalization group

transformation is made. Tarasov Vladimirov to



three-loop order in this theory.

& Zharkov

(1980) have

12.8 12.8

Consider

321

Operator-product expansion

Operator-product expansion

the Green's function

M(y;xl9...9xN)



(0\Tj(y)j(0)X\0}.

(12.8.1)

where j is a renormalized gauge-invariant operator, and X is a product of local fields at points xl9...9xN. We have the Wilson expansion

M(>;;x1,...,xN)^XCI.(>;)<0|r[^(0)]X|0>.

(12.8.2)



all renormalized operators of the appropriate dimension. This expansion is proved in a gauge theory the same way as in any other

The

sum

is

over

theory.

application to deep-inelastic scattering in Chapter 13, we will take be a product of two gauge-invariant operators and use the LSZ

In the

X to

reduction formula to obtain



matrix element:

W=<P\Tj(y)j(0)\P> EQ<P|[^0)]|P>.

(12.8.3)





We would like to show that only gauge-invariant operators need be included in (12.8.3). Now the proof of the operator-product expansion (in Chapter 10) treats the right-hand side of the expansion in the same way as

renormalization counterterms. So the method normalization of

gauge-invariant operators

applied applies to

we

also

to the

re-

the Wilson

The operators G{ are either gauge invariant, are the BRS variation of something, or vanish by the equations of motion. Then if we

expansion. keep



=£ 0 in (12.8.3) (as is the

case

invariant operators. The renormalized operator j has,

in

applications)

according

we

only

to Section

need gauge-

12.6, the form

8r* + B,

(12.8.4)

where 7GI is gauge invariant and B vanishes Hence if we take a matrix element of j(y)j(0)9

by the equation of motion. like (12.8.3), (or if we take a

J=Jgi

gauge-invariant terms,

so

Green's function



ofj(y)j(0))9 then we can drop the

A and B

that

<P\Tj(y)j(0)\P>



<P\TjGl(y)jGl(0)\P>.

(12.8.5)

We now follow our proof in Section 12.6, starting with the Ward identity 0





Here

we

ShRS(0\Tjiy)j(0)X\0} <0\Tj(y)j(0)5X\0}.

assumed that the positions y



xifl9 and 0

(12.8.6) are

all distinct,

so

that

322

Renormalization of gauge theories

by the origin,

equations of motion. But if y so that:

<0\Tj(y)j(0)dX\0}





-?

0, then

none

of the

xifl9s

are at

the

XC^)<0|7[^.(0)]^|0> -

Xc,0)<o|r[^r,.(0)]x|o>.

(12.8.7)



But the left-hand side is zero, by (12.8.6), while the right-hand side is its leading behavior as y-> 0 and is therefore also zero. If an arbitrary Green's

function of

an

operator is zero, then the operator itself is zero, i.e.,

XQ [>',.]

=0.

(12.8.8)



Hence the operators needed in the expansion are those that are BRS invariant. As in Section 12.6, this means they are either gauge invariant or

of classes A and B. In physical matrix elements like (12.8.3)

we can

therefore

restrict the operators to be gauge invariant. The practical use of this result is to compute the Wilson coefficients by taking the state |P> to be a state of one on-shell quark or gluon. The

infrared divergences are regulated by going to d > 4. The Wilson coefficients and the gauge-invariant renormalization counterterms of the tt/s can be unambiguously obtained provided care is taken to separate the IR from the UV divergences (both of which appear

as

poles

at d



4).

12.9 Abelian theories: with and without photon

mass

Consider QED with a possible mass term for the photon. The Lagrangian can be expressed in terms of unrenormalized or renormalized fields:



Here

we

have





single

iZ3Flv + k<Z,Al + ZJW vector field A

Z>>







and

we



M0)f

(12.9.1)

let

(^-ie04>A (f^-ieoZ^>.

(12.9.3)

A^ the photon. Without a mass term for the photon, Lagrangian (12.9.1) is invariant under the gauge transformation We will call

A? A? + d^io,  ̄j > ?A-+tAexp(ie0Z.W ij} ij) exp( ie0Z\!2a)).)

the

->?

-*



(12-9-4)

12.9 Abelian theories: with and without With

a mass term

photon

323

mass

be

the Lagrangian is not invariant, but the theory as we will see.

can

non-abelian theory with

is not

consistently treated, In

contrast,



consistent. For example, there

are

gluon



unphysical

states



term

mass

the

Faddeev-Popov

which violate the spin-statistics theorem, and the negative metric gluon states. In the massless theory these cancel in sums over intermediate states, but in a massive non-abelian theory they do not cancel. These

ghosts



problems do not

occur

in the abelian theory with



massive photon.

12.9.1 BRS treatment of massive photon Even with we use

a mass term

for the photon, the Lagrangian is BRS invariant, if fixing. We write

the standard gauge <£



j£'.nv



(l/2£)d-A2



d^c



(12.9.5)

£m2Z3cc.

and define BRS variations

SA^



aMc, Sij/ (5c





ie0ZV2oA> 5¥ 0,



ie0Z^c9

dc= -d-A/l

Then if is BRS invariant, aside from the Faddeev-Popov ghost is



an

(12.9.6)

irrelevant total divergence. Since

free field, it can be omitted without affecting physics (except gravity). We may use the general methods of Section 12.2, with the result that the theory is renormalizable. Since the ghost is a free field, the renormalization a

any

factors X and Z

are

both unity. Hence

(12.9.7)

e0=eRZ;1'2

e is gauge independent. (If we use minimal write eR ja2 ̄d/2e.) We now find that e0A{?] eRA^ and that the covariant derivative and gauge transformations simplify:

and the renormalization of

subtraction

we





(12.9.8)

(d^ieRA^, xjj^^xjj.

D^



Using this method we also find that ml finite

mass.

Then the

Lagrangian

^=

the

1m2, so that the ghost field has

(12.9.10)

Z2iA(if? + e4- M0)ip,

ghost ignored, since it is

counterterm

ZJ

is

-iZ3Flv + ¥m2Al-(l/2£)d-A2 +

with



(12.9.9)

for the photon



free field. Note that there is

no

mass.

Similar methods apply in theories where the gauge group is



product of a

324

Renormalization

of gauge

U(l) factors.

non-abelian group and one or more

theory

of weak and

gauge group

12.9.2

electromagnetic 5(7(2) ?(1).

Elementary

treatment

theories

Weinberg-Salam simple case with

The

interactions is





of abelian theory with photon

mass

sophisticated than necessary for the the theory by elementary methods. First we write the free photon propagator for the perturbation theory of the The BRS methods abelian theory,

are

so we

much

will

more

now treat

Lagrangian (12.9.10): i Duv Mv



 ̄n











0, then

we

+ Guv y"v



r_ 9^

cannot

remove

m2 + ie

If



k2-m2+i£K ,

' kukjm2) " v/



k k

t^—:f-|—- m2k2-£m2+ie j

(i-g)Mv im2 + ie



(12.9.11)

k2- the

gauge-fixing

term, for then the

propagator does not exist. (This is the same as taking the m is non-zero, then the propagator exists when the

ik^kjk2m29

is removed. However it behaves like

limit ^ oo.) But if gauge-fixing term

rather than

-?

1/k2.

So the

has worse divergences theory with m 0 and £ manifestly renormalizable. Even so, physical quantities are independent of £ and the theory is renormalizable if £ is finite, as we will see. Hence enough cancellations are present as £ that the theory remains renormalizable if =



oo

-?

we

than usual and is not

oo

only compute physical quantities.

Our treatment makes extensive

use

of the equations of motion:

AS =



Z3m2A^ + Z3(D^"



0,



d?d-A)



d*d-A/£



e0Zj/2#ytyZ2

AS

<?+



|^



Z2¥(



i 0-

M0)



0.

(12.9.12)

We have not assumed the relation (12.9.7). gauge-field equation of motion gives 0

Here,

we



(n/Z



Taking

m2)d'A.

the divergence of the

(12.9.13)

have used the invariance of the theory under the

transformations, which

are

(12.9.4) with

global

constant co. The Noether current

12.9 Abelian theories: with and without photon for this in variance is the

electromagnetic

current:

(12.9.14)

jr=-e0zy2Z2¥W. This is conserved, by the electron's equation of motion. The result (12.9.13) that d-A is a free field is important for three

(1)

It would otherwise be difficult to

interpret

the

which has

(2)

The

325

mass

reasons:

longitudinal part of A^

unphysical properties. <!;-dependent part of the propagator

is confined to the kjiv term, which only contributes to d-A. Then the ^-dependence decouples from physics if d-A is free.

(3) Similarly the bad ultra-violet behavior when ^

-?

oo

is decoupled from

physics. To make these statements

We will prove

(1)

Ward identities,

(2)

m2,

(3)

no counterterms are



precise we will examine the Green's functions.

directly:

Z^lm2,

so

that

no mass counterterm

is needed for the photon,

needed for the d-A2 term,

eRZz 1/2, with eR finite, but not Z3 depends on ^, (5) Z2 (6) the S-matrix is independent of <!;.

(4) e0



We will also compute the subtraction is used, and

we

exact

^-dependence of Z2 when minimal ^-dependence of the pole.

will compute the exact

residue of the electron's propagator

12.9.3 Ward identities Green's functions of &A9

<£^ and S£^ are non-zero, since the derivatives in

are implicitly taken outside the time-ordering. To obtain identity corresponding to (12.9.13), we use identities like

these operators the Ward

<0|T^(x)X|0>=i<0|7^^|0>.

Now

(Ult



m2)d-A



d^All



iej^i + ieJS?^.

(12.9.15)

(12.9.16)



We choose e0 eRZ^1/2, m^ Z J m2. This gives a Lagrangian of the form (12.9.10). There are no counterterms for A2 and d-A2, and the counterterm 1 for ^^. We will prove for {J/4^ is proportional to the counterterm Z2 =





later that this is correct. But for the moment we will choose to have

Lagrangian

our

in this form. Our gambit is that if extra counterterms should be

326

Renormalization

needed, then

they will

of gauge

theories

not be available and the Green's functions will

diverge. (12.9.16)

From

(nx/Z =



m2K0\Td'A(x)X\0y

i-^(0\TAX/AA^(x)\Oy +



it follows that

eR(0\TiI/(x)AX/A{l/(x)\0} -eR(0\Ti//(x)AX/Ail/(x)\Oy

i|£ ̄£T<0|T^|0>,

(12.9.17)

where the

gauge variation is computed according to (12.9.4). A convenient way to write these results is as follows:

(1) Let 0 be

(2)

Let



(5(field)

free scalar boson field of =

mass

£1/2m.

gauge variation of 'field' with parameter 0, i.e.,

Sa^=drf, 6\jj

dij/

(3)



ieR(f)\l/,





\eR<f>^j.

Then (12.9.17) is

(i/ZKO\Td-Ax\oy The field <j> is



<o\T<f>(x)sx\oy.

similar to the Faddeev-Popov ghost except for

(12.9.18)

being a boson

(this will be important later). In a non-abelian theory the nilpotence of the BRS transformation is crucial to proving Ward identities and is proved using the anticommutation of c with itself. A field introduced in the same way has to be the ghost field and must be to be a boson.



fermion. In (12.9.18)

we can

as



choose 0

12.9.4 Counterterms proportional to A2 and d-A2 Let

us

apply (12.9.17)

to the case X



Av{y):

(aj^m2)-^<0\TA^(x)Av(y)\0y -i-JL^x-y). This equation implies that d-A2

no

counterterms

proportional

to

(12.9.19) A2

or

to

needed. For suppose otherwise. Then consider the lowest order in which there is a divergence in the photon propagator. The divergence comes are

from the insertion of



divergent self-energy graph

in the free photon

propagator. The divergence has the form of Fig. 12.9.1, where the cross denotes the counterterm to this divergence. Power-counting indicates that

12.9 Abelian theories: with and without

photon

327

mass

Fig. 12.9.1. Counterterm to photon's self-energy.

only divergences are proportional to kjkv and g^v. A divergence g^k2 kjcv is cancelled by the wave-function renormalization Z3 for the photon. Insertion of the extra counterterms in the propagator gives a divergent

the —

contribution to the left-hand side of (12.9.19). But the left-hand side is finite. So there are in fact no divergences. Hence m\ Z^ lm2. It also follows that =

photon self-energy is transverse. Applying(12.9.17)toA: ^M(^v(z)andto^K(w)^A(>;)^v(z)showsthatno counterterms cubic or quartic in A are needed. the



12.9.5 Relation between e0 and

Apply (12.9.17)

to X

(D^ =

Z3

<A0#(z):





m2)^<0|T^(x)^|0>

e(0\ rMO0(z)|O> [8(x -y)- S(x



(12.9.20)

z)].

We assumed e0 Z^ l/2eR. If this is not the correct counterterm, then let the \l/-\]/-A 1 PI vertex first diverge at AMoops. Then the left-hand side of (12.9.20) =

diverges

at order

make the

e2iN+1i

right-hand

while there is

a counterterm

Z2



1 available to

side finite. Hence the left-hand side does not

diverge.

12.9.6 Gauge dependence The gauge variation of

Green's function



<0| TX\0}

is

^olralo-^j d4x<o|r[a-^2(x)-<o|a-^2|o>]x|o> ^jd4x<0|T^X|0> .

8 In Z,

+ i-



^-

d4x < 01T8 A(x)<j>(x)5X 10 > ■

-N,<0\TX\0>dhiZ2/dZ

=^|d4x<0|T[?/,2(x)-<0|</.2|0>]l2X|0> -iv^,<o|TA:|o>ainz2/a^

(12.9.21)

Here N^ is the number of i/f fields in X (which equals the number of ¥ fields).

328

Renormalization

of gauge





<0|7V4?/1J0>=



ft



theories

lowest order





^tx

( ) 51nZ2 ^J



^<0|T^/i|0>= ^^-+-^-^ {) ainZ.



ln{

Fig. 12.9.2. Examples of (12.9.21).

The

disappearance

of the factor

\ in

the last equality is due to the identical <O|02|O> appears, rather than 02,

nature of the two <j> fields. Also, <j>2 because the pure vacuum graphs never



occur on

the left-hand side. (A better

derivation uses two different fields <j> t and <j> 2 for each application ofthe Ward identities.) Notice that we cannot derive this equation if <j> is a fermion, for

then (f>2 the

non-abelian theory, the derivation must be replaced by complicated one which leads to the weaker (12.4.4). The graphical



more

0. So, in



(12.9.21) for a few simple cases is shown in Fig. 12.9.2. From (12.9.21) it is immediate that all Green's functions of gauge-in variant

structure of

fields

^-independent. only graphs with UV divergences are those with the simple loop (Fig. 12.9.3) attached to a ¥ or a \j/ vertex together with those with the c In Z2/d£ factor. We therefore find that are

The

dlnZ,



"£ .







l—nt

(12-9.22)

(2nY(k2-im2)2

is UV finite. Hence

ainZ22 d£

e2

/£m2 V'2"2

167i2jr(2-d/2)(7—j \4w



_



+ finite.



Fig. 12.9.3. All UV divergent graphs contributing

to

(12.9.21) contain this loop.

12.9 Abelian theories: with and without photon If

we use

329

mass

minimal subtraction then

Z2=Z2(£ =0)exp

(12.9.23)

%n2(d-4)_

^-dependence of the S-matrix, we recall the LSZ formula. It consider the corresponding Green's function, and pick out the

To obtain the

tells

us to

poles in its external momenta. Since we use transverse polarization (e-k 0) for photons this picks out graphs with the loop Fig. 12.9.3 and the d In Z2/d£ =

terms. Then



S-

Fermions

From

physical

Fig.

(^p*)

masses

photons

<0| Ti/^|0>

12.9.2 for

(t^)<0|rX|0>.



and for

(12.9.24)

<0| T^v|0>, we see that the

and the residue of the photon pole

are

^-independent, while

the gauge dependence of z2 exactly cancels the gauge dependence of the particle pole coefficient of <0(7X|0>.

Wecancomputeexplicitlythe^-dependenceoftheresidue,z2,ofthepoleof

the fermion propagator, if

we use







i.e., z2



z2(£



minimal subtraction:

lim (12.9.22)

^?[y ln{^)] +

0) exp i

^p \y

Since £0 ^Z3, e\ Z^le2, and ml £m2 in (12.9.25) are RG invariant. =



12.9.7





1 + In

Z^m2,

(-^ v47r/r

the combinations

(12.9.25)

e2£

and

Renormalization-group equation

Using our knowledge of the renormalization of e0, £0, and ml we find the RG equation for a Green's function of Nf\J/'s, N {iji% and NA A's to be

0=(4+pi y?MiL+y>m2ib ^|+N^+>N^)G'  ̄



(12.9.26)

with P



iey3 + (d/2-2)e.

If we use minimal subtraction then the only ^-dependent coefficient is y2, and

330

Renormalization

from

(12.9.23)

we

of gauge

theories

find that 72



y2?



(12.9.27)

0) + ^2/8tt2.

Theresults(12.9.23),(12.9.25),and(12.9.27)werederivedbyLautrup(1976) and

by Collins (1975a). 12.10

Unitary

The unitary gauge is the limit £

operators

-?

gauge for massive oo.

photon

Since the S-matrix and gauge-invariant

<!;-independent, this limit exists for them. But for gauge-variant there are severe ultra-violet divergences. Thus the limits £-? oo,

are

operators d -? 4 do not commute.

regulated theory. The resulting UV divergences by extra counterterms beyond those that we have already considered. Since the S-matrix is gauge-independent, all We may take £->oo first, in the at d



4 may be cancelled

these counterterms must vanish by the equations of motion.

13 Anomalies

A characteristic feature of relativistic quantum field theories is that symmetries of the classical theory are not always present after quantization.

We do not

mean

non-invariant

here the spontaneous breaking that is characterized by a and by the presence of the Goldstone bosons. Rather

vacuum

situation where there is

we mean a

no

conserved current for the symmetry

despite the absence of any terms in the action that appear to break the symmetry. Such breaking of a symmetry is called anomalous. If the classical action is invariant, then theorem gives symmetry

us



breaking.



naive

application

of Noether's

conserved current. That is, there is no anomalous What prevents the argument from being correct is the

presence of UV divergences. The current is a composite operator, i.e., a product of elementary fields at the same point, and to define it, some kind of

regularization and renormalization is needed. The renormalization may invalidate the equations used to prove Noether's theorem. For simplicity, we will consider only global symmetries, as opposed to

local,

or

gauge,

symmetries.

The

simplest

cases

of

global symmetries

were

considered in Chapter 9. These could be treated by using an ultra-violet regulator that preserved the symmetry. The proof of Noether's theorem can then be made in the cut-off counterterms are needed.

theory. We showed that only symmetric Consequently the symmetry remains good after the

cut-off is removed. However, not all symmetries can be preserved after regularization. The case which we will treat in this chapter is that of chiral symmetries. These are transformations that act independently on the left- and right-handed components of Dirac fields. These are particularly interesting because sometimes the anomalous breaking of chiral symmetries cancels. Indeed

theorem, first proved by Adler & Bardeen (1969), that if anomalous breaking of a chiral symmetry is zero to one-loop order then it is

there is

zero to



all orders.

Our treatment will

valid in the

use

dimensional

physical space-time

regularization. Chiral symmetry is 4, but not when d =£ 4. The

dimension d



anomaly in, say, an equation of current conservation will be an operator with

332

Anomalies

in effect



4, where d is the space-time 4, were it not for the existence of

coefficient proportional to d

dimension. This would vanish

at d





ultraviolet divergences which allow the anomaly operator to have a pole at d 4 so that a non-zero anomaly results at d 4. We will see explicitly how this =



works. There

key issues.

are two

The first is to derive

simple forms

for the

anomaly. Of these the most dramatic is the Adler-Bardeen theorem that tells us that in some cases there is complete cancellation of the anomaly. The second issue is to derive the results in a form that is applicable to the physical theory, i.e., after renormalization and removal of the cut-off. Even though a particular derivation depends on the choice of a particular

regularization scheme,

the final results must be

Note that to prove existence of an

independent

anomaly, it is not

of this choice.

sufficient to say that

symmetry in question is broken in the regulated theory. The breaking may go away after removal of the cut-off. For example, if one uses a lattice the

cut-off, then Poincare invariance is lost. However,

one must

prove that

Poincare invariance is restored in the renormalized continuum limit, if the theory is to agree with real world

phenomena. Aside from the case of chiral transformations, there are a number of other important situations where there are anomalies. One of the simplest is that of dilatations. These

space-time: x"->axm. A classical Lagrangian is scale-invariant if it contains no dimensional parameters, like a mass scale. But to cut-off ultra-violet divergences we necessarily introduce a mass scale. The symmetry is necessarily broken in the regulated theory and the question arises of whether the symmetry remains broken after the answer

scale transformations

theory

on

is renormalized and the cut-off is removed. This

is, in general, yes, if the theory has interactions. The

there is, in fact, when

are

we

a mass

scale hidden in the renormalized

reason

theory,

is that

as we saw

discussed dimensional transmutation and the renormalization

group, in Chapter 7. Detailed treatments can be found in the literature (see Collins (1976), Brown (1980) and references therein). The simple form for the Ward identity is known as the Callan-Symanzik equation (Callan (1970) and

Symanzik (1970b)). The information contained in the Callan-Symanzik equation is in fact also contained in the renormalization-group equation that we studied in Chapter 7. Other situations which we will not treat include the following: chiral gauge

Piguet & Rouet (1981)), conformal transformations (Sarkar (1974)) and supersymmetries (Piguet & Rouet (1981), Clark, Piguet & Sibold (1979, 1980), and Piguet & Sibold (1982a, b, c)). theories (see Costa et al (1977), Bandelloni et al. (1980),

13.1 Chiral

transformations

333

13.1 Chiral transformations

QCD with two flavors of quark: up and down. (We could flavors, but no essentially new ideas would be needed.) The Lagrangian is (2.11.7). If the quarks were all massless then the classical theory is invariant under the following transformations of the quark fields: We will consider

have

more

i/,



exp

{i[|(l



y5)K + com + i(l



y5)K + <*")] ty.

(13.1.1)

Here the matrices ta are the generators of the isospin group acting on the flavor indices. The transformations (13.1.1) form a group that we will call

U(1)L? U(\)R? SU(2)L?SU(2)R. The symmetry of the QCD Lagrangian under these transformations is broken by mass terms for the quarks. Since the of the u and d quarks are small, the chiral symmetries are only weakly broken. These symmetries and their breaking were understood well before the

masses

advent of QCD-see, for example, Treiman, Gross & Jackiw (1972).

symmetries in the light of QCD can be found in Marciano & Pagels (1978) and in Llewellyn-Smith (1980). Thus it is unnecessary to go into details here. What we will emphasize is how the Treatments of chiral

potential

for anomalies arises.

Since these transformations involve y5, they

spin space-time,

the

we

imposition of

can

an

in

coupled to symmetries of

some sense

to the

expect trouble when the theory is regulated, for

ultra-violet cut-off must alter the

Notice that there is a

transformations have coL

correct. The

are

theory. Since spin is related

structure of the

space-time structure. involving y5; these subgroup them the treatment of Chapter 9 is

U(1)?SU(2) coR. For



corresponding

Noether currents

ra



not

are

(13.1.2)

Z¥y?ty.

The conserved charge derived from j*1 is the conserved quark number, while the transformations generated by fa are just ordinary isospin

transformations. For the other generators of chiral transformations, let

us

define axial

currents

y£5 In four dimensions

y^y5







y 5yM,

so

that we could have written /y5 in place

of the commutator in (13.1.3). However, because

regularization, currents are hermitian.

we must use

(13.1.3)

^l[^y5]rV. we

will

the form (13.1.3) in order to

use

dimensional

ensure

that the

334

Anomalies

To define the

this

we

will

use

theory we must regulate its ultra-violet divergences, and for dimensional regularization. We will

theory is not invariant under the transformations

see

that the regulated

generated by

the axial

currents;^ and;^a. This allows the possibility of an anomaly. When thecut-off is removed we will find that we can arrange for the non-singlet currents)^ to be without anomaly.

Although we will not demonstrate it, it is true that the

singlet current necessarily has an anomalous divergence. Our remarks above were addressed to the case that all the quarks mass less.

generalize our discussion. The vector singlet current /* is the quark number i.e., 1/3 of baryon number so this remains an

so we now

current for exact





symmetry. The SU(2) symmetry given by the vector currents is broken

by quark

mass

differences:

(13.1.4)

fy'J =-¥[*-, Mty, where M is the quark small

are

But in the real world, there are quark mass terms in the Lagrangian,

we

have

an

mass

matrix. Since the masses of the u and d quarks are

approximate isospin symmetry of strong interactions.

the theory given in Chapter 9, the currents;^ and/,

finite, since the breaking is from The axial symmetries

mass terms

as

(Symanzik (1970a)).

broken by the anomaly

are

By

defined by (13.1.2), are

as

well

as

by the quark

mass terms:

dj%





2\\jjy5M\l/



anomaly,

Wa5=-ih5{t\M}il,. The

and

(13.1.5)

approximate rf-quark light enough give SU(2) ? SU(2) symmetry. The axial part appears to be spontaneously broken as well as explicitly broken. The abnormally light mass of the pion is usually w-

taken to

mean

Marciano &

to

masses are

that it would be



us an

Goldstone boson if mu & Leutwyler (1982) for



Pagels (1978) and Gasser

md



more

0. See

details.

13.2 Definition of 7 s In order to formulate consistently the dimensional

regularization of theories

with fermions we had to define an infinite set of matrices yM, {jx we saw

in

Chapter 4, they satisfied

the

{/jWI. Our task is

now to

find





0,1,2,...). As

algebra (13.2.1)

trl=4.

generalization of the matrix which at d



4 is called

y5. Its four dimensional definition is ys



iy<y y.

(13.2.2)

13.2 Definition

335

ofy5

and it satisfies the anticommutation relation

{y*,y5}=0 (rf



4).

(13.2.3)

It would be natural to assume that this relation (13.2.3) can be maintained for arbitrary values of d.

show, when

now

we

Unfortunately

inconsistency arises, as we will trace of y5 with a product of the

an

wish to compute the

ordinary yM's. The ultimate result will be that stated

in Section 4.6, where

we

used the definition (13.2.2) of y5 for all values of d. Then y5 has mixed commutation and anticommutation relations, (4.6.3). We now demonstrate the inconsistency in trace calculations, starting with

try5: dtr(y5)









In the first and last lines

tr(yMy5yM) -tr(y5yM/) (13.2.4)

-dtv(y5).

used

we

y,y"

tr(y5y?yj



Hy^}



g^=d.

In the second line we used cyclicity of a trace, and in the third line we assumed

(13.2.3). From (13.2.4)

we see

that tr y5



0 except at d

apply dimensional regularization we wish to obtain meromorphic function of d. Hence we must have try5 Similarly











Here

we

same

a =

we

result that is



0 for all d.

tr7A75V/v7A -tr75/V/v7A 2Ktr ysyJx + 2#v tr 75 Va d tr y5y^yy 2tr75{7?,7v} + (4 d)try5y^yv (4-d)try5Vv. -







used the result tr 75

Now, the

0. Now when



technique



we find that (d be used to prove that

0. Hence

can

(4-^tr(757KWv)





0.

(13.2.5)

2) tr 757M7V is zero. (13-2-6)

4 this equation permits the usual non-zero trace of y5 with four other Dirac matrices. However, if the trace is to be meromorphic in d9 equation (13.2.6) shows that it must be zero at d 4, and we can therefore not obtain

At d





normal physics at d section.



4. This is the inconsistency referred to at the start of this

336

Anomalies

We are therefore forced to drop one of the hypotheses that led to (13.2.6). Candidate hypotheses for removal include:

(1) the anticommutation relation (13.2.3) ('t Hooft & Veltman (1972a)), (2) the use of dimensional regularization for fermion loops (Bardeen (1972), and Chanowitz, Furman & Hinchliffe (1979)).

These last authors have shown how to calculate with a totally anticommuting ys. Bardeen chooses to use a regulator other than dimensional regularization for all fermion loops. On the other hand, Chanowitz et al. regulate fermion loops with an even number of y5's dimensionally. Their procedure is useful for low-order graphs, since the Ward identities are preserved for graphs without fermion loops. However, we then lose the use of dimensional regularization as a complete regulator; the theorems that we derived in Section 6.6 and in Chapter 12 no longer apply. The details of higher-order calculations by this method have not been spelled out. Therefore let

us

follow 't Hooft & Veltman (1972a) and Breitenlohner & change the anticommutation relation. In fact, we may

Maison (1977a) and

the definition (13.2.2) for all values of d9 just as stated in Section 4.6. Our definition is, of course, not completely Lorentz covariant, since the first four dimensions are picked out as special. But this is not an overwhelming use

objection, for

actual

our

physics

is confined to these dimensions. An

important advantage of the definition is that it gives of y5. We are therefore guaranteed consistency. One notational inconvenience arises. We have \i



0,1,2,.... Usually

we

only

refer

explicitly

a set

construction

of matrices yM for

to the first

four; the

rest are

collectively. But there is / with ja 5. It is not the same as y5 by (13.2.2). However, we bow to standard usage and use y5 to denote

referred to

defined

a concrete



the matrix defined in (13.2.2). Confusion should be rare. The y5 so defined has mixed commutation and anticommutation relations

that follow from (13.2.2) and the properties of /. These were stated in (4.6.3). The derivation of (13.2.4)-( 13.2.6) now fails, because y5 does not anticom- mute with all of the

trace of

an

y^s. We can derive the correct relations easily. Since the

odd number of y^'s is zero, tr (/y5)



we

have

tr{y5yKyY)

0,



0,

(13.2.7)

etc.

We may read off the trace of y5 with an even number of /'s from (4.5.13) and its relatives with

more than

four

y-matrices. Thus

trW tr(?5yV)





°'l o.J

we

have

(13.2.8)

13.2 Definition

ofy5

337

Finally ^(y5yKy,y,yv)



(13.2.9)

^K^

where eK^v is given by (4.6.2). The relations (4.6.3) mean that thechiral transformations (13.l.l)no longer

generate a symmetry if d is not equal to 4. To discuss the resulting problems, the following notation is useful (Breitenlohner & Maison (1977a)): ^

\guv9



if n and



v are



or

larger,]



otherwise;

(1321°)

V"=g"vVv. Here V is any vector. Then g^ is a projector onto the unphysical dimensions.

Thus, for example,

{V,y5}= I We may also define

projections

W,y5}=2f?5 2?5f.J =

onto the four

\g"v if ix and v are ^0 otherwise; W

The

physical dimensions:

less than 4 J >



g>"Vv.



following results

are

elementary,

but will prove useful:

?fi=(6-<*)f, ffi=(4-<of, ffy?=-2f, 4f Tfiu =

Let

define a>fl

us







(13.2.12)

(13.2.13)





<w£)/2 and a>"5



(- a£



a£)/2.

Then the

variation of if under the chiral transformations (13.1.1) is b<£



i#(2o>2y5 + a)"[M,f] + a>a5y5{M,rfl})i^ -2i}>y5p(co05 + o/5tfl)"A- (13.2.14)

Hence the

divergences

SJia





dji



of the axial currents are

ihsP* M)* + fast?* -i¥y5{t°,M}^ + fa5{t°,P}il,/2, 2#y5Aty + #y5^. -



(13.2.15) (13.2.16)

The second term in each equation can potentially give an anomaly when we letd^4.

Anomalies

338 13.3 There

are a

Properties

of axial currents

number of somewhat different situations in which axial currents

appear. The original papers on chiral anomalies primarily addressed anomalies in Ward identities of the chiral currents of strong interactions. More general

have since been worked out, with corresponding generalizations of the Adler-Bardeen theorem. In this section we will list the various

cases

cases

and state what is known. 13.3.1 Non-anomalous currents

The following properties apply, for example, to the non-singlet axial currents in

QCD:

(1) If there is no anomaly to one-loop order (as for the current)^), then there is no anomaly to all orders. The Ward identities of the current with elementary fields have no anomalies. (2) Under the same condition as in (1) there is no anomaly in the two-current Ward identities.

(3) Under the same condition there is an anomaly in a three-current Ward identity (like the one for dK(0\ TjKa5(x)j£(y)jvc(z)\0y). However, the only non-zero term in the anomaly is the one-loop contribution. The theorem that the

complete anomaly in these cases is determined

one-loop value is due to Adler & Bardeen (1969). The lack of anomalies in the Ward identities of

elementary fields is essential if the currents

one

current

generate the

are to

by the

with

correct

transformation law for the fields. These transformations imply commutation

relations for the currents. Since these commutators

are

also given

by

the

Ward identities with two currents, the two-current identities must be

anomaly-free. No such consistency requirement applies to the three-current Ward identity. The value of its anomaly is related to the decay rate for n° 2y, and -?

the lack of higher-order corrections enables

successful prediction to be made easily. (See Adler (1970); for reviews from the point of view of QCD, see Marciano & Pagels (1978) and



Llewellyn-Smith (1980).)

13.3.2 Anomalous currents The

singlet

current y 5 has an anomalous

fy'S



divergence.

C(g)GlvGa^ + mass

where G"v is the gluon field strength

g?Mv



terms,

tensor and

fii*vKAG?

It has the form

(13.3.1)

G"v is its dual: (13.3.2)

13.4 Ward identity for bare axial current

339

The Adler-Bardeen theorem asserts that the coefficient of C{g) is equal to its

one-loop

value

However,

the operators^ and GG need renormalization, and there is not

C(g)=Nng2/(32n2). an

obvious natural renormalization condition. So the value of C is susceptible to change by redefinition of the renormalization prescription. We will not treat this case here.

13.3.3 Chiral gauge theories

Weinberg-Salam theory of weak interactions have a gauged chiral symmetry. It is essential that there be no anomaly, for otherwise the theory is not renormalizable and loses other important properties (Gross & Jackiw (1972) and Korthals Altes & Perrottet (1972)). A generalization of the Adler-Bardeen theorem is that there is no anomaly to any order of perturbation theory if there is none to one-loop order. Proofs have been given by Becchi, Rouet & Stora (1976), and by Costa et al. (1977). Theories like the

13.3.4 Supersymmetric theories Supersymmetric theories have potential anomalies similar to the chiral completely general treatment has not yet been given, but many cases have been treated see Piguet & Rouet (1981), Piguet & particular Sibold (1982a, b, c), Clark, Piguet & Sibold (1979,1980), and Jones & Leveille (1982). anomalies. A



13.4 Ward identity for bare axial current Without

use

of the

equations

of motion the

divergence

of the

non-singlet

current)^ is

dJi5=&2fa5ta(ty-M0)il, + hc + iZ2<A{M0, f}y5il, + &¥?{% y5ty ^L The first







equation

term we will call the

03A1)

^nonv

of motion term. When inserted in

Green's function with elementary fields it

gives

<o\TDaejx)i\AU^yi)U^j)\o> i



=E#*-:vxo|Tn^n*n*io>,?,--,,., i



340

Anomalies

The variations of the fields

i. (We set cor





a>l



just their chiral transformations multiplied by

are

of5

in (13.1.1) to obtain these transformations.) This

equation (13.4.2) has the expected right-hand side for the Ward identity in the absence of anomalies. The

mass term on

the right of (13.4.1) is expected; it is the non-anomalous

breaking. The anomaly term is D°nom. If we insert it in a Green's function with no divergences of any kind, then we may set d 4. The result must be zero =

Danom vanishes if only the first four y-matrices are used (for then 0). But if the Green's function has an overall divergence, or if it has a divergent subgraph that contains the vertex for Danom, then the result may be because

{p, y 5}



non-zero at





4,

as we

will

see.

—<0|T^5(x)n^n,/'n,/'l0>



The full Ward identity reads

right-hand

side of

(13.4.2) +

mass term

+<o|rr>aflnoinwn^n'Ar['/;io>-

(13.4.3)

Recall that in the case of a symmetry such as the isospin SU(2) of QCD that has

anomaly, we used its Ward identity to prove the current finite. The only possible counterterm for the current is proportional to itself, so finiteness of the divergence of the current, d 7, implies finiteness of the current itself. The Ward identities imply that the divergence of the current is finite. no

However, for the axial currents the extra term in (13.4.3) prevents this argument from

being

made.

13.4.1 Renormalization of operators in Ward identities Our aim will be to construct



finite current j%a5 that at d



4 satisfies

a non-

anomalous Ward identity:

A<0|T7L5(x)n^n^n^l°>=right-hand side of (13.4.2). (13.4.4) The first step is to observe that the only counterterm to fa5 is itself. No other operators have the correct dimension and quantum numbers. So we can

define



minimally

subtracted operator



12Z5Z2{p[y",y5]t^.

(13.4.5)

Throughout this section we will use square brackets to indicate minimal subtraction. So the renormalization factor Z5 has the form

Z5



1 -I-

poles

at d



4.

13.4 Ward To obtain the

identity for bare axial

operator jRa59

we

341

current

will later show that

we

have to invoke



further finite renormalization. We will show that the Ward

by

the

minimally

subtracted

identity (13.4.4) is not correct ttjRa5 is replaced

[ja5].

renormalization to obtain a Ward Jro5





where

z5(g)

is



Rather

we must

make



further finite

identity without anomalies. Thus we have

Z5Ua5j

(13A6)

iz5Z5Z2(A[/,y5]^,

finite factor.

The anomaly operator I>anom is a dimension four scalar quantity, so there are several operators with which it can mix. It is proportional to the e-tensor

AK?fiV. Since eKAfiV appears nowhere in the Lagrangian, the tensor A*;/iV is invariant under all Lorentz transformations. Given these restrictions, a complete list of operators that times a fourth rank tensor, which we will call

mix with D°nom is

^5,^?om^{M0,rfl}y5(A. involving only ghost and gluon fields can be constructed such that it mixes with D°nom. The linearity in eKXflv implies the presence of four factors of vector objects (derivatives or ,4-fields). Therefore the coefficient of No operator

the operator is independent of mass, by our usual results. Gauge invariance of a restricted set of gauge variant counterterms (see Chapter 12) of which have low enough dimension to appear. Since ja5 is even under

^anom allows none

so is D°nom. Therefore the only allowed counterterm to proportional quark mass has a flavor factor {M,ta} rather than the commutator [M, **]; this gives us the operator DaM that appears in (13.4.1). We can therefore write the minimally subtracted operator corresponding

charge

to

conjugation,

[Daflnom]

as

WnonJ



ZaITanom



Za5d-ja5

The equation of motion operator is finite



by itself

(13.4.7)

ZamDaM.



see

(13.4.2)



so we

have

(13.4.8)

[^m]=^m.

Note that the definition of Datm includes some counterterms, but these are manufactured from the wave-function, mass, and coupling renormalizations

in the Lagrangian. We also need the renormalization of the

lDaM ̄\



mass

operator

Z5MD°M.

It is somewhat uQobvious that the renormalization is

(13.4.9) multiplicative.

The

easiest method is to examine the y-matrix structure of self-energy graphs with

Anomalies

342 an

insertion of

¥tay5\lj

or

\j/y5ij/.

multiplicatively renormalizable

These operators can then be shown to be with a common factor.

of (13.4.6)—(13.4.9) we can express the equation (13.4.1), for the divergence of the bare axial current, in terms of renormalized operators to

By

use

find

[D*em]=Z-l(l-Za5Z^)d^fa5-] -

[DaM]Z^(l



ZaMZa *)



Z-1

[^nom].

(13.4.10)

on the right are all linearly independent, so the linear combination of them can equal the finite left-hand side is for

The renormalized operators

only

way a

the coefficients to be finite. Since

we use

ZaM

za5





minimal subtraction this implies:

i-z5.

We therefore have the renormalized Ward

^<0|T[;^]n^n^n^l°>

(13.4.11)

Z5M-l,} identity:

=r-hs. of (13A2)

+<o|T(ra+[^nom])n^n^n^io>, (13.4.12)

which apparently still has an anomaly. Before showing how the anomaly in disappears, let us examine some low-order graphs.

fact

13.5

One-loop calculations

The tree approximation for the two-point Green'sfunctionoffD^] is given by Fig.

13.5.1. To

save

algebra we will set quark masses to zero. The graph's

value is

l^2(-W p)y5ta%. P

(13.5.1)





If we let the external momenta be

physical, i.e., in the first four dimensions,

then this vanishes. The vertex for D%nom has the property we define as evanescence: it vanishes when the cut-offis removed and we go to the physical

renormalized theory. We will formulate

later, when

we



precise definition

of evanescence

have understood the subtleties associated with

inserting

the

vertex inside loops. a

Danom >





) p

Fig. 13.5.1. Tree approximation for two-point Green's function of [££nom].

One-loop calculations

13.5

343

P'



to

(b)

(c)

Fig. 13.5.2. Graphs up to order g2 for two-point Green's function of;£5. Next let us examine the graphs for the two-point function of the current;'^, as

given in Fig.

13.5.2. Graph (a) has the value

ysrvS^. Taking the divergence is the



When

we set

lowest-order





case

p'





same as

y5t°iplp2

0,

multiplying by i(p'



p)M. The result is

i(f/p'2)y5f + (13.5.1).

to get the four-dimensional

result,

(13.5.3) we

obtain the

of the chiral Ward

identity. Fig. 13.5.2(b). Its

The next order graph is propagators



(13-5.2)

value, with the external

amputated, is \4-d

r"=76^(27l/i)

k\p'



k)\p + k)2

This is evidently divergent. It is easy to calculate the pole at d

pole (r26)









X)

4:

g2fCF f 1 pole < -^-j- —^y^yy

g2tacF

l^yy5(d-6)2J

;-32^-pole)4-rf

C¥f,y?y*.

%n2A-d Here

we

(13.5.5)

have twice used the result that

yKfy¥f







yKfysT + yKfysyK 2f?5



if f ys

(6-dWy5.

(13.5.6)

344

Anomalies

There

is



^2^7^75ffl,A>

in

implicit

graph

counterterm

the

definition

w^tri trie quark wave-function renormalization

Z2



g2CF



0(g4).

(13.5.7)

graph Fig. 13.5.2(c) therefore cancels

the UV

lr l-S%n2{4-d) ?

The resulting counterterm

j%5 given by



divergence of graph (b), leaving a finite result. No additional renormalization is needed :

Z5 Let

us

next

take the

(13.5.8)

l+O(04).



of (13.5.5) plus its counterterm, by

divergence

multiplying by i(p' p)^ It is left as an exercise for the reader to verify that the Ward identity (13.4.3) holds at this order. What we will do is examine the —

graphs

g2

of order

for the Green's function of Dlnom. These

Fig. 13.5.3. Note that the definition includes



The \[fAy5\jj

gives rise

term P

Z2i0y5^ to the





are

listed in

covariant derivative:

2ig0Z2^0y5ta^

(13.5.9)

graphs (b) and (c).

p'

-Tzr-

x^r

t7



(a)

(b)

(c)

(d)

Fig. 13.5.3. Graphs of order g2 for two-point Green's function of [fi°nom]- Graph (<j) equals r3a

t**i-yW 4* Crf2f(2^)*-" 16*4 J







ig2cF

| dx| ^-3-^(2^)*-" Jo Jo x



Mfi



t + 2h^ + ft)y

k2{p>+k)2{p + k)2

dyx

{r(3-d/2)Dd!2-3yv[t'(l -y)-*x]tf'(l -2y) + p\\ -2x)] ir(2

rf/2)^'2 2yv[2(/>'(l -y)- tx)yKy5f -





2yKfy5W-x)-fy)



7K(PV-2y) + kl-2x))y5yKy},

(13.5.10)

where D=

If it

were

not

-p2x(l-x)-p'2y(l-y) + 2p-p'xy.

that there is

an

(13.5.11)

ultra-violet divergence given by the

345

13.5 One-loop calculations

d/2)9 we could let p and p' be four-dimensional, and then set d 4 to obtain zero. (Note that yKyK yKyK d 4.) The pole prevents this argument from being made. First of all, notice that the pole is T(2











po's(r,.)--^IIffpoie|5--jH.(^ _^V£,!£±? 87T 4

The manipulations

following

y^sf







(d



yKyKfy5

are easy to



do incorrectly, so let us

iv/^f

yKyKfy5 (10-d)f<%.

we

(13.5.12,

result

yKfy5yK

split yK 4)-dimensional piece y*.

In the first line

fe!vV





the Dirac matrices

on

be careful. We need the







if y5

(13.5.13)

four-dimensional piece y* and a Then in the second line we used the

into



or anticommutators of yK and yK with yM and y5. The graphs of Fig. 13.5.3(6) and (c) may be evaluated similarly. The sum of the pole terms for all three graphs is

commutators

poie(r3a



r3b



r3C)=



g^2



4_rf

(13.5.14)

which is cancelled by thecounterterm Fig. 13.5.3(d). Thisisin agreement with our general result (13.4.11).

ready to compute the value at d 4 of the sum of the graphs of Fig. 13.5.3. Considerable simplification occurs. Since p and p' are now zero, the term in (13.5.10) that multiplies T(3 d/2) vanishes. Similarly the last The term multiplying T(2 zero. d/2) gives remaining two terms have a We

are now







factor yKyK



d-4, which cancels the pole

r3flU^=p.



to leave a finite result:

o=10L^'-^5-

(13-5.15)

Similarly T3b and r3c give (r3fc



r3JU4,^^0

Effectively Fig. 13.5.3

d-ja5. fc

can



to

^tv-^,

g2CF/(An2)

times

(13.5.16) the

vertex

for

why the result should be of this form. loop integration, the vertex for D%nom vanishes when

It is easy to understand

Without =

sums





get

p'

the =

0. When we include the integration over the components of k, we

a non-zero

value for the

graphs

even

if p



p'



0. However, the

Anomalies

346

property of the basic vertex implies that it has effectively a factor d-4. We only get a finite result by multiplying by an ultra-violet evanescence

4 is of a local operator. divergence so the effect at d A general theory of evanescent operators can be worked out. The results simply generalize what we have learnt from examples: =



(1) We define an evanescent

graph when

tree

we set



vertex as one



that is finite and that vanishes in



4 and when all momenta and polarizations are

four-dimensional. (2) A Green's function d

or a

graph or an operator is evanescent if it is finite at polarizations are

4 and if it vanishes when its external momenta and



four-dimensional. (3) Consider



graph containing

an evanescent

vertex.

If the graph is

completely finite then it is evanescent. ('Completely finite' means that the graph and all its subgraphs have negative degree of divergence.)

(4)

A renormalized operator

[£]

whose basic vertex is evanescent has the

following expansion:

[£] The

sum

is

over



£C£K[F]

+ evanescent

(13.5.17)

operators.



operators V whose basic vertices

are non-evanescent.

The only operators that are needed are the ones that according to the usual power-counting and symmetry requirements will mix with E under renormalization. The general proof is left as an exercise to the

reader. 13.6 Non-singlet axial current has

no

anomaly

13.6.1 Reduction of anomaly The only operator that can appear on the right-hand side of (13.5.17) for the case





Dlnom

is

d-\_ja5].

The restrictions

are

that it be pseudoscalar,

iso vector, gauge invariant and of dimension at most four. (If we had non-zero

quark

masses, then the operator

Wnom]



[£>£,]

could also

C(g)d'lJi5]

appear.)

+ evanescent,

So

we

have

(13.6.1)

which, when substituted into the renormalized identity (13.4.12), gives

^<o|T(i-Q[;s5]n^n^n*io> =

r.h.s. of (13.4.2) + evanescent.

(13.6.2)

13.6

So

we

Non-singlet axial

current has no

347

anomaly

should define the renormalized current i{.3





(i-c)[A] (l-C)Z;1ra5i

(13.6.3)

which is (13.4.6) with z5 1 C. For the physical four-dimensional theory this implies that j^a5 has Ward identities with no anomaly, viz. (13.4.4), as we =



wished to prove. From our calculations of Fig. 13.5.3 C



we see

(13.6.4)

g2CF/(4n2) + 0(g4).

general theory of evanescent idea, however, is simple. in the physical theory is when a

proof has been long and involves

Our

that



operators summarized in Section 13.5. The basic

The only way

an

anomaly

can appear

divergence cancels an effective factor of d 4 for the evanescence of an anomaly. The anomaly in the four-dimensional theory is a local operator, —

and the only possible operators are those which power-counting would allow as counterterms to

d-j.

In the case of our iso vector current, the only such operator is d -j itself. So

finite renormalization (13.6.3)

serves to

eliminate the anomaly at d

13.6.2 Renormalized current has

no

us

fx— right-hand side



djj.

right-hand side {Z



A.

anomalous dimension

apply the renormalization-group operator /xd/dju identity (13.4.4). For the right-hand side we get Let



to the Ward

anomalous dimensions of fields}.

To get the same result for the left-hand side, the current j£a5 must have zero

anomalous dimension (when d



4):

dja This is a sensible result: the current j^a5 is a physical object, and it should not on how we parametrize the theory by a renormalized coupling. Useful consequences follow, for the minimally subtracted current does have an anomalous dimension:

depend

nd[U/dn=-y5[ja5l

(13.6.6)

where

y5(g)





nd\n(Z5)/dn

[(2-d/2)g + p(g)]d\n(Z5)/dg.

(13.6.7)

348

Anomalies

Now, the coefficient C is

presence of

function of g; it is dimensionless and



masses cannot

depend

factor Z5. We also define it to

on

have no cut-off dependence,

between renormalized operators at d

even

in the

them, just like the renormalization =

since it is a factor

4.

We therefore have 0





A*d/afl5/d^

-P(dC/dg)[ja5-]-(\-C)y5[ja5l

that

so

/W(l-q/30



(13.6.8)

y5(l-C).

It follows that 1-C

where

used

we

as a

expansion starting

expj exp



boundary condition

g2.

at order

convergent, the order

<13-6-9)

dg' dg'y5(gVp{0')h the fact that C has

In order that the



integral

perturbation in (13.6.9) be

g2 term in y5 must vanish. (13.6.7) it must be that Z5 has no order g2 term; this

So from the definition

by explicit calculation. Moreover, we loop value of C, so that

we

know

y5









(13.6.4) the one-

A,g*C¥l{2n2) + 0(g%

where the one-loop term in /? is A tg3. Hence of the leading divergence in Z5:

z3

know from

i-, Z¥\, 4n2(4-d)



(13.6.10)

(13.6.10) gives us a prediction

o{g6).

(13.6.11)

The reader is invited to check this by Feynman graph calculations. We may use the techniques of Chapter 7 to sum the divergences. We find the full renormalization factor of y'^a5 to be if1

^ J



(i_c)z5-1

a5







exp{J9od,W)[^-

y5(g%d/2-2)g'

dg' exp\ a"* /?(9')[(<*/2-W+ /?(<?')]. l +

->1

Evidently

(d/2-2)g' + p(g').

(rf-4)O[ln(02/(4-<*))] (13.6.12)

as<^4.

the Noether current is finite in the

complete theory.

13 J Three-current Ward

identity; the triangle anomaly

13.7 Three-current

identity;

Ward

13.7.1 General form

349

the triangle anomaly

of anomaly

We consider the Green's function D&(P,P')

jd*xdVp-x+i"'-y<0| Tj*a5(0)j&x)j;(y)\0>.



(13.7.1)

Only connected graphs contribute; Lorentz in variance forces the vacuum expectation value of a current to be zero. The currents are all renormalized currents, so all subdivergences are cancelled by counterterms, and the only possible infinity in (13.7.1) is an overall divergence. In fact there is divergence, as we will now show.

4+k *

r^ *



\^k J>'

no

overall







Fig. 13.7.1. Lowest-order graph for (13.7.1). Individual graphs for (13.7.1) have a linear divergence, as can be seen from, say, Fig. 13.7.1. Any divergence must be linear in external momenta and

proportional

to the e-tensor. The

only possibility is

^[a{d)pK



(13.7.2)

b{d)p'Kl

There is also the constraint of conservation of the vector current. This is expressed by constructing



Ward

identity in the dimensionally regularized

theory. Consider

PM0AMV=-ifd'^^ By

use

of the result

i^Z^y^ty and the

(13.7.3)



Z2{j}ta{ip

equations of motion,

p^



we



M0)\jj



Z¥(



ilj)



M0)fty

find

jdW+''>-<0| T/ifl5(0)Z2<A[Y, r>]/<AW|0> dVij,''y<O|TZ20[^f*]y^;^)|O>.



(13.7.4)

In these equations we assumed that the currents;^ andyj are conserved. Each of the terms in (13.7.4) is



Green's function of a

vector

and



pseudovector

350

Anomalies

current

Parity invariance forces

them to be zero,

PfiD^v

Similarly

pfvD^v

so



0.

(13.7.5a)



0.

(13.7.56)

The counterterm (13.7.2) must therefore give zero when multiplied by p^ or p'v. This forces the whole counterterm to be zero; the Green's function (13.7.1) is finite

as

it stands.

In the regulated theory the axial current is not conserved, so we cannot 0 by the same manipulations. Indeed prove the Ward identity (p + p\DXflv =

we

have

(P +

P')xD^v



commutator terms +

d4xd4yei(px

+ p'y)



x<0\TEa(0)rb(x)rc(y)\0}. (13.7.6) Here E is the evanescent operator in E





(13.6.2):

lDlBom]-C8Ua5] [^nom]-C(l-C)-1a7fia5.

The commutator terms in

(13.7.6) vanish,

as

(13.7.7)

in the Ward

identity

for the

vector currents. Hence, finiteness of the left-hand side implies that the Green's function of £ with/£ and) is finite. Even though graphs for it are quadratically *

divergent, the divergences cancel. Now, E is an evanescent operator. This means that its Green's functions with elementary fields vanish in the four-dimensional theory. The general theory of evanescent operators, which we summarized at the end of Section 13.5, then tells us that the only way that the right-hand side of (13.7.6) will fail to vanish is for E to be part of a graph or subgraph with overall degree

of divergence at least zero. Now, the definition of E has ensured that these subgraphs are all evanescent. Hence we are left with the complete graphs. So we have (at d 4) =

(p

The

tensor



p,hD^

structure

dimensionless at d



4,

is so

A(g)e^pp^Eabc -i^jip+p'np-py1**.



the

it

can

only one possible. The coefficient A is only be a function of g. Note that the right-

hand side of (13.7.8) obeys vector current conservation,

P,(P + P\D^ as

should be.

(l3-7-8)



0,

so

that

13.7 Three-current Ward

351

identity; the triangle anomaly

13.7.2 One-loop value The lowest-order value of A is easily computed from the graphs of Fig. 13.7.1. a

"va"P





* = 3

f d"/c

tr

?J (2n)d +

[tf ¥' + 2|)ys(/> + JQy'ME -

k\p



k)2(p'



/>')]

k)2



charge conjugate.

(13.7.9)

The factor 3 is the number of quark colors. To evaluate this, notice that

/r(YV) 0\ 1 =

(13-7.10)

Since eKXfiV is restricted to the first four dimensions, it follows that the trace of y5 with four Dirac matrices is zero if one of the matrices is a y:

tr(y5fyVV) Let

us

the

commute

whenever possible. tr

-p'



ft in (13.7.9) p'

We will also set p





now

an

left, and

to the =

use

(13.7.11)

0. The result is

use

(13.7.12)

of (13.7.11). The charge conjugate

equal contribution.

have

^<WV This is now

(13.7.11)

-tr[2ty5U> + my'lP"f'] + 0 -tr[2|)-5#TVy]+0 -te[2fttys^y] + 0-

The terms indicated by '0' vanish by We



0.

[tf -}' + 2%)y5(j> + JWM -f + m



term gives







'd'k

tr[#y5#*y]

}{2nfk2{p'-kf{p

(13.7.13)

kf



only a logarithmically divergent integral. After use of Feynman

parameters the standard result (4.4.14) gives

A^apP'P

"£J>.

1-JC

dy±T(2



dfl)





tr[i/75/W] [





*(1 -x)- p'2y(l -y) + Ipp'xy]

+0(d-4)

-"g^W^"- The

evanescence

of the vertex has effectively given

(13.7.14) a

factor of d



4 which

Anomalies

352 cancels the UV pole to leave



finite result

(13.7.15)

A=-3/{in2) + 0(g2). This result for the anomaly in the axial Ward identity for (13.7.1) found by Adler (1969), and Bell & Jackiw (1969).

was

first

13.7.3 Higher orders There are, in fact, (Adler & Bardeen (1969)) no higher-order corrections to the

anomalyfor(13.7.1).WewillfollowtheproofduetoZee(1972).Thebasicidea is simple. Each of the currents in (13.7.1) is RG invariant, and there is

no

overall counterterm. Therefore this Green's function is invariant when

we

RG transformation. The anomaly must therefore be invariant also. But the anomaly coefficient A (g) depends on the coupling g and on no

make

an

other parameter of the

theory.

We

can

change g arbitrarily by changing the

mass /x. Hence A is independent of g. This proof may easily be written out. Renormalization-group invariance of D*j£ is the equation

renormalization

"dp* Hence

^AgKbc^Pj'p Since

jud/d^u



\idjc\i



fid/dg,



(p

independent of

g, so it





V^^iC



°-

this gives fidA/dg

Then A is

+ p





0.

equals its lowest-order

value:

(13.7.16)

3/(8tt2) exactly.

This is a very striking result. The proof we have given is very simple, but the reader should not suppose it is not



deep result. The whole

power of the

renormalization apparatus is needed for its derivation. We first had to show that there is no anomaly in the divergence of the axial current. Then we had to show that there was no counterterm needed to make

Definite. These results

involved considerable cancelation of UV infinities. Since the anomaly in d-ja5

disappears when the U V cut-off is removed, it can affect a Ward identity only by being enhanced by a UV infinity which has not made its appearance earlier. Thus the anomaly is associated with a UV pole implicit in the Feynman graphs. It is precisely for this reason that it must have the dependence on the

13.7 Three-current Ward identity; the triangle anomaly

353

parameters of the theory and on the external momenta that is characteristic of a renormalization counterterm. In particular, it is polynomial in the

momenta and masses of the degree determined by U V power-counting. Once this is clear, the most general possible form of the anomaly is (13.7.8). The final step to show that A is independent of g is trivial. An important phenomenological consequence of the anomaly is a calculation of the decay rate for n° -> 2y (see Marciano & Pagels (1978) and

Llewellyn-Smith (1980)). The amplitude is proportional to the number of quark colors, so the decay rate is proportional to the square of this number. The measured rate in fact agrees with the standard theory that there are three colors.

14

Deep-inelastic scattering In this chapter we will show how the operator-product expansion can be used

to compute the cross-section for calculation is fairly

deep-inelastic scattering. Since the this straightforward, process is one of the classic tests of

quantum chromodynamics (QCD). The process is the scattering of a lepton of momentum P momentum pM, with the

only observed particle in

on a

the final state

hadron of

being the

lepton: / + N -? /' +

practice hadrons in the target

In

one uses a

there

are

anything.

(14.0.1)

beam of electrons, muons,

are

nucleons. There

experimental data,

the

cases

-?





->

fx +

are a

or

neutrinos, and the

number of cases for which

being distinguished by

the types of

lepton involved: e +



jx + N

The basic

v +

N -?

v +



->





(e

anything,

anything, anything,

or

jj)





(14.0.2)

anything. J

measuring the cross-section for these processes is to (quarks and gluons) of the hadron. we have a of a Suppose scattering lepton on a small-sized constituent, and that the momentum transfer is large, so that the scattering happens over a small time-scale. The weak and electromagnetic interactions have a small coupling, so to a good approximation the lepton does not interact again. reason for

study the fundamental constituents

Moreover the interactions that turn the final state, involving the struck quark, happen on a much longer time-scale. So these interactions do

into hadrons

not interfere with the basic Born graph.

Hence, we should be able to calculate

the cross-section for the process (14.0.1) rather simply in terms of the cross- section for lepton-quark scattering (Fig. 14.0.1). The approximation in which final-state interactions of the hadrons are ignored is called the impulse

approximation. over

We can use it because

all hadronic final states. 354

we

choose to

sum

the cross-section

355

14.1 Kinematics, etc.

Fig. 14.0.1. Parton model for deep-inelastic scattering. A more mathematical formulation of the same idea is the parton model,

in Section 14.2. There are, however, many weak points in the intuitive discussion just given, and we must remedy them. In the remainder of

explained

this chapter we will compute the cross-section from the theory of strong interactions (QCD). The parton model will in fact give a qualitatively correct

approximation to the real cross-section. Our intuitive argument shows why deep-inelastic scattering is simple enough to permit calculations. Our treatment in this chapter is based on material to be found, among other places, in Gross (1976) and Treiman, Jackiw & Gross (1972). 14.1 Kinematics, etc. We will compute the cross-section to lowest order in weak and electromagnetic interactions. Then the amplitudes contributing to the process (14.0.1) have the form of Fig. 14.1.1 where a boson is exchanged between the lepton and the hadron. The boson can be a photon, a

Wot a

Z. At high enough

energy, it is also necessary to include Higgs boson exchange. Higher-order weak and electromagnetic corrections do not need to be included, except for

photon effects. We will ignore the soft photon corrections here, and will concentrate on understanding the strong-interaction corrections. soft



Fig. 14.1.1. Amplitude for



?ot>* process

contributing

to

deep-inelastic scattering.

We first review the kinematics of the process. The

two

independent

Lorentz invariants of the hadron system are chosen to be

Q2=-q%, v



(14.1.1)

p-q,

where q* is the momentum transfer from the leptons. The

mass

of the final

hadron state X is then

ml



(p



q)2



m£ + 2v



Q2.

(14.1.2)

356

Deep-inelastic scattering

In the

laboratory frame, where the initial hadron is at rest, we can express Q2

and v in terms of the initialand final lepton energies E and £', and of the lepton

scattering angle

6:

neglected lepton inequalities hold We have

<22







2££'(1



cos

0).

(14.1.3)

mN(E-E').

masses

with respect to E and £'. The

following

(14.L4)

v>Q2/2>0.

The region we will investigate is where both Q2 and m\ get large in a fixed ratio. We define the Bjorken scaling variable x Q2/(2v). Then we let Q2 get with fixed. the x This is called large Bjorken limit or the deep-inelastic region. =

In this

region the missing

mass mx

m2



is large:

(14.1.5)

Q2(l/x-l\

(where we neglect m^ compared with Q2). In order that m\ be positive we must have 0 <

x <

equivalent variable to x that is The cross-section is given by E'

dc

d3//

applies only if x is not at its endpoints. An sometimes used is co 1/x.

1. The Bjorken limit





dc

InE'dE'dcosd n

4ME

^ZK'1//plO<^|;riiv>D^)|2^4K-^-^) x

4k) ^^^^^^^^^^(A^ /ept

j**d

(14.1.6)

exchanged vector boson couples, gw is its coupling, and Dx\q) is its propagator. The lepton tensor LKA is easily computed in the tree approximation for Here

and

are

the currents to which the

LKX The hadron

Wltv(p,q)

tensor is





<l\f;?(0)\l'><l'\j)^ (0)|/>.

equal

(14.1.7)

to

^-YJ<p\jtl(Oy\X>(X\jv(0)\PXW#*)(P"x "P" 47T x

4ttJ dVr^p|;>)Vv(0)|p>,



<?")

(14.1.8)

where the normalization is the standard convention.

Deep-inelastic scattering is the region where Q gets large with x fixed. This treated in Chapter 10. There we took all fixed ratio, so that Q^> co with Q/p-q fixed,

is not the short-distance limit

we

components of q" to infinity in



357

14.1 Kinematics, etc. i.e., x is proportional tog. This is not in the physical region 0 <

inelastic scattering. Luckily

x <

1 for deep-

relation (Christ, Hasslacher & Mueller (1972)) to relate the deep-inelastic limit of W^ to the we

can

use



dispersion

short-distance limit of the time-ordered matrix element

T,v



[d V?<P|

77>)VV(0)|0>.

(14.1.9)

perform this analysis in Section 14.3. decompose W^ into tensors with scalar coefficients. We will assume that the hadron is unpolarized, i.e., that we average over its spin states. Then the most general form of W^ is We will

Let

us now

Ww



Wd

0,v +



Wjl2) + W2(Plt



q,v/q2)(pv

-i^3W?V/(2M2)+ W^qJM2 + WJ/(2M2) + W5(P??, -iW6(pfa The scalar coefficients W.(Q2,v)

are





qvv/q2)

q,pv)/(2M2).

(14.1.10)

called the structure functions. The

normalizations are such that they are dimensionless. A number of properties

follow from symmetries and basic quantum mechanics (Treiman, Jackiw, & Gross (1972)): (1) Each Wi is real. (2) Time reversal invariance of strong interactions implies that W6 In the

of

case

conserved,

so

purely electromagnetic



that

qllWllv

electromagnetic current is invariant. So W3 0.



process, the current

0. Hence W4= W5





jh^d

0.

is

0. Moreover the

a pure vector, and strong interactions are

parity



only conserved if quark masses are zero. In that case the only non-zero structure functions are Wx, W2 and W3. It turns out that when the masses are non-zero the contributions of W± and W5 to the cross-section are suppressed by a factor of order m^JQ2, where ml and mq are lepton and quark masses. We will discuss this further in In the case of neutrino scattering, the currents

are

Section 14.8. When

we

compute the structure functions in the Bjorken limit,

Q2

-?

oo, x

fixed, we will find that they behave as a certain power of Q2 times logarithms.

The power is (Q2)0 for

Wl9

l/g2for W2and W3iznd l/<24for W4and W5. So it

anticipate these results and define scaling structure functions FifaQ2) which depend only logarithmically on Q2. The standard is convenient to

definitions are:

F1(x,Q2)=Wl(Q2iv Q2/(2x))i Fi(xiQ2) vWi/M29 for i 2 and 3, F.(x9 Q2) v2 WJM\ for i 4 and 5. =











")





(14.1.11)

Deep-inelastic scattering

358

14.2

Parton model

discuss the true theoretical predictions for deep-inelastic let scattering, us discuss the parton model (Bjorken & Paschos (1969)). This is Before

we

the simplest model for the process, and contains the essence of the correct physics. One considers the initial state hadron in the overall center-of-mass for the whole scattering. Now time dilation slows down processes in the hadron so that they typically occur on a rather long time-scale T of order Q/m^. Then the scattering happens on a much shorter time-scale T ̄ 1/Q. This implies that the hadron may be regarded as an assembly of non- interacting point-likeconstituents. These are what were originally (Feynman (1972)) called partons. (We now identify them as quarks and gluons.) Since the hadron in this frame is ultra-relativistic, we must regard the partons as massless and as each moving parallel to the hadron with a certain fraction z of its momentum. The structure functions of the hadron are obtained

partons'

structure functions in tree

approximation

by computing the

and by then summing

all partons weighted by their number density. Letfa/N(z)dz be the number of partons of type a in hadron N with fraction z

over

to z +

dz of its momentum. Then, for example, the electromagnetic structure

functions

are:

2F7n(*)



* ̄1*Tl(*) +

The relation F2



heavy quark

2xF1

(14.2.1)

terms.

spin-^ of the quarks (Callan independent of Q2. This is the

is characteristic of the

& Gross (1969)). Notice that Fj and F2 are property known as scaling. Experimentally,

Fl

and F2 obey this property

approximately. It is

common

in the literature to

use

the terms 'structure function' and

because of the parton model relation between them (14.2.1), which is also approximately true in QCD-see Section 14.7. However, it is important to distinguish the two terms. Structure 'quark distribution'

functions

are

interchangeably,

coefficients of certain tensors in

a current

correlation function.

be defined for other processes, e.g., the Drell-Yan process (Lam & They Tung (1978)). On the other hand the quark distribution functions are exactly can

what their

name

implies: probability distributions

of

quarks

in



hadron.

The parton model is correct in a super-renormalizable theory (Drell,

Levy

and Yan (1969,1970a, b, c)). However, in a renormalizable theory like QCD, there are processes inside a hadron that happen significantly on all time scales

143 Dispersion relations and moments

359

assumption of the parton model does not hold. cause the problems allows the to come into operator-product expansion play to solve the problem. There are two approaches, essentially equivalent for deep-inelastic scattering:

down to zero. Then the basic

However, the fact that it is short times that

(1)

Use a dispersion relation to show that moments ofthe structure functions

(i.e., M

FiAQ)



(14.2.2)

Axx^-'F^Q2))

be directly computed by the Wilson expansion. This approach was initiated by Christ, Hasslacher & Mueller (1972), and it is the one we will can

use.

(2) One

can

generalize

the derivation of the

operator-product expansion

(Amati, Petronzio & Veneziano (1978), Ellis et al. (1979), Libby & Sterman (1978), Stirling (1978)). For deep-inelastic scattering, the result is equivalent to the first method without the taking of moments.

14.3 Dispersion relations and moments Consider the time-ordered Green's function

TMV defined by (14.1.9). It can be

expanded in scalar structure functions Tl9T29T3i T4, T5, T6,just like W^. We will only need Tl9 T2i and T3. The operator-product expansion derived in Chapter 10 can be applied to T^ when Q2 and v get large with Q2/v2 fixed. As we have seen, this is not the scaling region, for we have x-> oo instead of x fixed. However, we will relate T^v to W^ by a dispersion relation. Then we will

W^v in the physical region can be obtained from the operator-product expansion for T^v. If Q2 is fixed and positive then each T- is analytic in the v-plane. There are cuts going out to infinity from the thresholds v ± Q2/2. See Fig. 14.3.1. see

that information

on



(This

is



Fig.

standard

14.3.1.

property.

Analyticity of



It

can

be

proved

and contour to derive

by expressing

(14.3.1).

360

Deep-inelastic scattering

<N | Tj^yVjlO)| N > in terms of W^ and noting that The discontinuities across the cut are:

T^P,q)\\^ TJP,q)\;;'£ is

W^

W^v



] > 0), J

^WJptq), (if v>0).





where



W^ is zero if | v|< Q2/2.)

4tt Wv?(p,q),

(if

v <

(14.3.1)

replaced by its hermitian conjugate /:

with ;'

K=^*yeiqy<p\My)jM\p>- For the

electromagnetic

hermitian,

so

By

Cauchy's

neutral-current processes the current is

W^ W^. W^v is for charged-current neutrino the for antineutrino scattering. structure functions W^ gives

that

scattering then

or

(14.3.2)

But if



theorem

we

have

where C is any contour enclosing v, as shown in Fig. 14.3.1. We will be able to compute the

Ti(Q2, v) in

T/s in the short-distance limit v/Q2



power series in 1/x

Tr=

vT,/M2=



Tln(Q2)x-\



T,.n(Q2)x-" (i

n = 0

expand

2v/Q2:



n=0

0. So suppose we

-?



2or3).

(14.3.4)

(Weexpandvr./M2(ifns2or3)inanalogywith(14.1.11).)Thenfrom(14.3.3) i

r dvve2 ,

where

a.



0 if i



convergence at | v' |

1 and =

a.



,,

1 if i









T,(22,v'Mv7M2r,

or

we



is large enough to give

oo, then we can deform the contour and pick up only the

contribution from the discontinuity of T{

Finally

3. If

(14.3.5)

across

the cuts:

write the right-hand side in terms of the scaling functions

F;(x,22): 7y<22)= =

-a -

f1dx'x'"-i[Fi(x',e2)+(- ir+"^',G2)]

2i\_FtJQ2) + (-

1)"+" W)]-

(14.3.7)

14.4 Expansion

for

scalar current

361

This is the dispersion relation referred to earlier. It relates the power series expansion of 7^ about v 0 to the moments of the F's, which are defined by =

PiAQ2)=

fdx^-'F^G2).



(14.3.8)



applies if the integral is convergent. For small it of n values enough diverges. If vai Wi behaves like vp as v -? oo, then we have convergence only if n is greater than p. Now the limit v -? oo at fixed Q2 is a The relation (14.3.7) only

Regge limit (elastic scattering of a virtual boson off a hadron at energy m \). So there is fori





1 standard expectation (Treiman, Jackiw & Gross (1972)) that p Ofori 2. This is equivalent to FvF2/x9F3 all behaving =

1 or 3 and p





1/x behavior for quark distributions, and is roughly what is measured experimentally. Our theoretical predictions will give all the terms in the series expansions of the 7)'s, e.g. (14.3.4). Those coefficients for which (14.3.7) does not apply will like 1/x as x -? 0. In the parton model this would correspond to a

not have any direct

implications

14.4 To

for

Expansion

deep-inelastic cross-sections.

for scalar current

explain without a slew of indices the method for computing moments of functions, let us first work out the case where ;M is replaced by a

structure

hermitian scalar operator;. For example,; might be ZZmqtqi9 appropriate to the coupling of a scalar boson to a particular flavor i of quark. We have a

single scalar

structure function:

F(x9Q2)=W(viQ2)



(l/4n)

dAy^<p\j(y)j(0)\Py9

(14.4.1)

while the time-ordered function is: T(v, Q2)



IdV?

y(p\ Tj(y)j(0)\p}.

(14.4.2)

(Notice that we choose j to be a renormalized operator. As is the case for the

cf)2 operator for a scalar field, the renormalization factor for qq is the same as the mass renormalization factor, so [qq ̄\ ZZmqq Zmg0g0, where g0 is the =



bare

quark field.) dispersion-relation argument applied

The

T(Q2/(2x\Q2)=



n = 0

to the series

Tn(Q)x-?

(14.4.3)

gives UQ)





2i[Fn(Q) + (



1)"F?(G)].

(14.4.4)

362

Deep-inelastic scattering

Since the current is hermitian, r(Q) nW) Regge

theory suggests





have F(x, Q)

{-4iF?(e)' \ 0,

1/x



we

as x -?

0,



F(x, Q)9

so

that

if"iseven> if n is odd. so

this

[(14 4 5)'

equation is

valid

only if n is

bigger than 1. We now apply the operator-product expansion. The results of Chapter 10

apply in the Now

limit

Q2

-?

oo

with

THx-*

v2/Q2 fixed, i.e., with x

Q2

(14.4.5) is

not

Q.

(14-4-6)

0 term are non-leading by a power a have reliable only prediction for n 0. But the relation

that it would appear that all but the n

of

we

constant times

Q-"[TH(2v/Qrl



so

and that







expected

to hold unless n > 1.

Fig.

14.4.1. T(v, Q).

remedy this problem, we must find an object for which x ̄nTn contains This is done by making a partial wave Q in the t-channel. That is, we treat T(v, Q) (Fig. 14.4.1) as we decomposition would treat a scattering amplitude, and decompose it in terms of angular To

the leading-power behavior as

-?

oo.

momenta:

<p|r/GO./(0)|0>=



£ =



(plVjlpyM^yq/q2).

(14.4.7)

Here<p| Vj\p} is the reduced matrix element of some operator 7JflflofspinJ,

appropriate polynomials in yq/q2. (The operator V3 is not necessarily local.) Now we perform an operator-product expansion of T7'(y);(0), keeping the leading-power behavior for each spin:

and the M3

are

<p\Tj(y)j(0)\p>





2i

£ (- i)\pK^JpyCJa(y2)y^...y^

J,a

(14.4.8) Here the operator

(9^



is



different operators of the same

local operator of spin J. The label V denotes spin. Only the symmetric part of the operator

is relevant, and in order that it be of definite

spin, it

must be trace less. Since

14.4 Expansion

363

scalar current

for

the hadron is unpolarized the matrix element has the form

(14.4.9) <pKl..Jp> <p\&Ja\p>(p?r..p,j -traces), (p\(9Ja\p} is the reduced matrix element-a scalar quantity. The =

where

adjusted for later convenience. leading y^O is obtained by using operators of the minimum possible dimension. In QCD these are normalizations in The

have been

power of y as

^J(^...plj ?i9

(14.4.8)



= u

21

 ̄JqytliDfl2... iDHj q, symmetrized minus traces,

-23"JGUiViD ...iD?

Gl



symmetrized minus

traces.

(14.4.10) In accordance with the results of Section 12.8

we

have kept only gauge-

invariant operators. (We use D^ to denote the co variant derivative, and q to quark of flavor q. Sums over color indices are implicit in

denote the field of a

(14.4.10). Only hermitian operators are needed.) Our usual power-counting arguments imply that the behavior of the scalar coefficient

CJa(y2)

in (14.4.8) is

(^2)"dim(C) times

logarithms,

with dim(C)





2dim(;)-dim(0£ J+ J

2dim0)-dim[<p|^[p>/<p|p>].

The dimension minus the spin of

(14.4.11)

operator is evidently the important quantity here; it is called the twist of the operator. The leading twist is two, for the operators of (14.4.10), and for them C(y)  ̄y ̄* modulo logarithms. an

Fourier transformation of (14.4.8)

now

T(v,g2)= -2i£ <p\VJa\p>CJ*(Q)x-J J,a

gives +

non-leading

powers of ft (14.4.12)

where

we

define

e?-<mwjdv?c?

(.4.4,3,

Perturbative calculations of T(v, Q2) will give us CJa in (14.4.13). We can then obtain moments of F(x, Q) from the dispersion relation:

2Fj(Q)=Z<p\&Ja\p>CJ%Q)[l a

if J is greater than

one.



0(l/Q)l

(14.4.14)

364

Deep-inelastic scattering 14.5 Calculation of Wilson coefficients

There are two parts to the calculations. The first part is to do low-order calculations of the Wilson coefficients

where

now explicitly indicated the dependence on all parameters. corrections have logarithms of Q//z, so the second part of the Higher-order calculation is to compute the anomalous dimensions and then to do a we

have

renormalization-group transformation

to set \i to be of order

Q. Thus

we

write

2FJLQ)

I <p\eJtt\p>wCJa\g(Q),i)MatAg(Q),Q/ri-



(14.5.1)

a,a'

The subscript {jx) on the matrix element denotes renormalization with unit of The matrix M is obtained

by solving the renormalization group and for the CJa is C, equation (g(Q\ 1) well approximated by its lowest-order mass ju.

term in perturbation theory. Measurements of deep-inelastic scattering at one value of Q are enough to give <p|0Jfl|p>, and then (14.5.1) predicts the moments of the structure functions at other values of Q.

14.5.1 Lowest-order Wilson coefficients The Wilson coefficients

independent of the target, so we may calculate by a quark state. In tree approximation we have the expansion sketched in Fig. 14.5.1. The scalar 'current'; is the renormalized operator ZZJq^ [^g,], for a particular quark flavor i. Since the Wilson coefficients are independent of mass, we set quark masses to zero. Then from the graphs of Fig. 14.5.1 we find are

them with the hadron state replaced =

itr[>(J + ftl 2 (p + q)2

r=





4)] (p-q)2

i_ tr W 2



1/x2 ',, 2. '1-1/x2

(14.5.2)

-2i,

The factor \ comes from averaging over the spin of the quark. For a quark of any other flavor than i,

p + q

or

for



gluon,

we

have T= 0 to this order.

p-q

Fig. 14.5.1. Wilson expansion of

<0|T//|0>

to lowest order.

of Wilson coefficients

14.5 Calculation

To compute the Wilson coefficients

operators defined by (14.4.10). In

we

also need the matrix elements ofthe

quark



365

state with flavor i we have

<p^i;'...,Jpo=i^tr(^i^2...^) =

(14.5.3)

2dirp^...pMji

in tree approximation, with the label i' denoting a quark flavor or a gluon. For a

gluon

state and

gluon operator

<H°^...,.,N>=2/v-/v All other matrix elements elements

are zero.

Hence the non-zero reduced matrix

all equal to 2. Comparison of (14.5.2)-(14.5.4) with the

(14.4.12)

(14-5-4)

are

operator-product expansion

shows that

0{g2\

CJi



1 +

CJi'



0{g2\

if

if

J > 1 and is even,

i'j= U

or

iff is



gluon.

"] I



(14.5.5)

14.5.2 Anomalous dimensions The anomalous dimension of the operator ZZ^^ is ym:

^(ZZmqq) and

we

let

yJaa\g)



ym(g)ZZmqqi

(14.5.6)

be the anomalous dimension matrix of the (9Jais:

^Ja



Iyaa(g)eJa'-

(14.5.7)

(Operators ofdifferent spin do not mix.) As shown in Section 10.5 in a simpler case, the Wilson expansion then implies a renormalization-group equation for the Wilson coefficients:





CJa(9, QM





X ya.aCJa' + ymCJa a'

(14.5.8)

This equation would be trivial to solve were it not that it is a matrix equation. The lowest-order counterterms for the operators are generated by the graphs of Fig. 14.5.2. It is evident that the different operators mix. To solve

diagonalize the anomalous dimension matrix at order g2. The first step is to recall that the counterterms are independent of quark masses. So the renormalizations respect the SU(nn) symmetry of the flavor space. Therefore let us now choose a new basis for the twist-2

the RG

equation

operators.

we must

366

Deep-inelastic scattering

"A



'*\



AAA AA _rs

CJg

Fig. 14.5.2. Lowest-order divergences of the operators in the Wilson expansion of <0|T/y|0>.

There is



multiplet

of non-singlet operators:

GJ,?s,a

where the /lfl's There

are two

are

the nfl





2i

-^yMiLDM2... ^

(14.5.9)

singlet operators: 0", defined by &'s=

(14.4.10),'



q*\



flav,i The

;.>,

n{l matrices that generate the flavor symmetry.

renormalizations

(14.5.10)

are

[0JS]

\[e>JNS'\



U21 \

z22





|U? 1,

(14.5.11)



(14.5.12)



from which the anomalous dimension matrix 7u

7i2



?22 0

^=|?21



14.6 OPE for vector and axial currents is obtained

367

by

Z^^ A calculation of the

divergences

i?|-Zar



of the

(14.5.13) of Fig. 14.5.2

graphs

gives (Gross

(1976)) r,2

I1









1)+\?2*J'

12

Vn=^T2 8tt y?

J(J

J(J





12

1)

(J



l)(J + 2)









2nfl (J2 + J + 2) .zi 4jt2 J(J + l)(J + 2)'

r;2._il<4±i±*

(14.5,4,

14.5.3 Solution of RG equation



non-singlet

The Wilson coefficients of the non-singlet operators evolve

tt^CJNS(g(n), Qln) An

approximate solution

can





be found



very

yis + ym)CJNS.

by taking gives

simply:

(14.5.15)

the one-loop

approximation for the anomalous dimensions. This

CJNS(9(?\Q/li)



 ̄ln?2/A)

CJNS(9(Qll)\\Mfi/A)]

(14.5.16)

y^ 7 and y^ denote the coefficients of g2/4n2 in yJNS(g) and ym{g\ and At is the coefficient of g3/(4n2)inp(g). We may replace CJNS{g{Q)9 l)byits

where -

value in tree approximation. The accuracy of (14.5.16) may be systematically

improved by taking more terms in the perturbation expansions of /?, yNS9 ymi and C The singlet coefficients may be obtained by diagonalizing the 2 x 2 matrix of anomalous dimensions. Then there are two linear combinations of singlet

coefficients that have simple behavior like (14.5.16).

14.6 OPE for vector and axial currents We will now apply the operator-product expansion to the structure functions of T for a weak >or

electromagnetic

current. The

argument is



simple

368

Deep-inelastic scattering

generalization of the treatment in Sections scattering with a scalar current. We have

14.4 and 14.5 of

deep-inelastic

dV*,yX<H^...JP>(-i)Jx

rMV(p,*) ̄-2i

i_2

x>"-H-^-^CM vdy2 oy"3yy

-^???^^-^)(^-^-?W) -i/J.../'V^C?(y2)|-

(146.1)

acting on the Cffl,s are arranged to be in combinations with divergence, to correspond to the condition q* 7^v 0. Fourier transformation as at (14.4.13), with suitable factors of Q2, gives

The derivatives zero



^v--2il<p|^j?x J,a

*{(-0,,v+v?vA?2)* ̄JCia(e) + (l/v)(p? q?v/q2)(pv qvv/q2)x<  ̄JC{°{Q) i<wV(l/2v)x-JC^(0}. -





(14.6.2)

The set of leading twist operators is the same as in Section 14.4, and we have arranged normalizations so that the C/s are dimensionless in leading twist. Hence the moments of the structure functions satisfy

FltJ(Q) + (- 1)JFM(0= £Cf(0<p|^"|p>+correction, a

F2,j-1(Q) + (-1)JF2.j-i(Q) f



lCJ2"(QKp\&Ja\p) + concction, a

3,j(6) + (- l)J 1^3,j(e)= Z Cia(Q)<P\?Ja\P> +

+ correction.

(14.6.3) These equations are valid for J times logarithms. 14.6.1

Wilson



1, and the corrections

coefficients



electromagnetic

are

of order

1/Q2

case

To compute the lowest-order Wilson coefficient we consider deep-inelastic scattering pn a quark target. The graphs are the same as in the case of a scalar

current, Fig. 14.5.1, except that the current operator is

now

the elec-

14.7 Parton interpretation

tromagnetic t-1

of Wilson expansion

369

current. We find

(lowest order)

-4K^M*£&]} Xe2A 9 ) (- 0?v







Expanding about 1/x





Vfiv

U>,



<7?<7v/<Z2h),1/X2

+<

,2

liilvll

qj>/q2)(py







<?vv/<?2)

<>. --^- 1/x







(14.6.4)

gives



r1=-2ie2

(i/x)2"+2



(l/x)2n+1



o(02),

n=0

vT2/M2=

,2 -4ie2 y £



The Wilson coefficients [ficients



are



therefore therefor







(14.6.5)

O(02).

n /^2n+i _l n^2







if J is

even

and > 2,

C3=0, C^ The relation



(14.6.6)

0+O(^2).

C2 2Ct corresponds to the Callan-Gross relation F2 =2xFt in the parton model. Since the renormalization-group equation is the same

for

both

C\a

corrections



CJ2a9 of 0(g(Q)2). and

the Callan-Gross relation is true in QCD with These corrections

are

from the Wilson coefficient,

and have been calculated. See Buras (1981) for an up-to-date list of references.

The renormalization-group equations except that ym is replaced by conserved current is zero.

14.7 Parton

zero

are

the

same as

in the scalar case,

since the anomalous dimension of

interpretation

of Wilson



expansion

The

use of moments in comparing theory and experiment is not very convenient, since cross-sections are needed outside the range in which they are measured. A more convenient form can be derived in which an expansion

is obtained for the structure functions themselves. We will just summarize the results. More details

can

again

be found in Buras

It is sufficient to examine the case of



(1981).

scalar current for which the

expansion is F(x, Q)





p.

-falN(z,n)Ca(z/x;Q/n,g(n)).

(14.7.1)

370

Deep-inelastic scattering

The sum is over all species of parton, i.e., flavors of quark and

antiquark, and gluon.ThegeneralizedWilsoncoefficientCaiseffectivelyastructurefunction deep-inelastic scattering of a parton state, while falN(z, n) is a parton distribution. Lowest-order calculations reproduce the parton-model result,

for

so

that

Cq

C-=eq2S(z/x-l)/2y



Cg

0.



The renormalization group

A*;p/U(*>A0

(14.7.2)

equation





for / has the form

-jybaWx,g)fbiN(z>i*)-

This is called the Altarelli-Parisi (1977) equation

(14-7-3)

it was first derived by these authors in leading logarithmic approximation from an heuristic argument. Later derivations (Collins & Soper (1982a), and Curci, Furmanski

& Petronzio (1980))



complete. Integro-differential equations like (14.7.3) are not particularly easy to work with. One mathematical simplification that can be made is to take moments, with the result that the operator-product expansion of are more

Sections 14.5 and 14.6 is recovered. This will enable

us to see

that the two

essentially equivalent. However, the mathematically more complicated method using convolutions, as in (14.7.3), gives more physical insight, can be extended to other processes (Buras (1981)), and can be used without knowing structure functions at small x. To see the equivalence of the two methods, let us define the following methods

are

moments:

/&fo)= C?



y£>= Then (14.7.1)

fdzz"-1/,,^), P Wz)(xlzrlCa{z,x-,QltL,(m, {

d(x/z)(x/zr-lyba(z,x;g(fx)).

(14.7.4)

implies Fn(Q)



TS?Ui*)C"\g,

Qlfr

(14.7.5)



This expansion has the the

same as

same

form

as

(14.4.14). In fact, moments of the/'s are

the matrix elements of the twist-2 operators:

(N\0Ja\N}



P dxxJ ̄ Va/jvM +fm(*)l

Jo

<14-7-6)

14.8

W4

and

371

W5

a contribution from each type of parton and antiparton. This equation can be proved (Collins & Soper (1982a), and Curci, Furmanski & Petronzio (1980)) provided only that the same renormalization prescription

where there is

(e.g., minimal subtraction) is used for both the operators and for the parton distributions.

14.8 W4 and W5 So far

we

have

ignored F4

and

F5. They

are zero

conserved. But the weak-interaction currents

if the current/ in

are not

W^ is

conserved if quark

masses are non-zero:

^OA/a



?5u?=^([>aa] =

Here m is the quark

mass



y5{m^a}w (14.8.1)

iDa/2.

matrix. It follows that W4 and W5 are non-zero in we should regard the ratio m/Q as setting the

neutrino scattering. However, scale for their effect

the cross-section (Jaffe & Llewellyn-Smith (1973) and Llewellyn-Smith (1972)). They therefore give a small contribution to the deep-inelastic cross-section. Since W4 and W5 are therefore non-leading in on

the Bjorken limit, they are difficult to compute directly. A convenient technique to compute W4 and W5 is to consider

q?W^v



(q2W4 + p-qW5/2)qv/M2 + q2pvW5/2M2.

(14.8.2)

We have the operator formula

Wv so wecan

7M(y)Z)(0).



Jd *yeiq'y<p\Uy)D(0)\p},

(14.8.3)

compute W4and W5 by making an operator-product expansion for

Since there is an explicit factor of quark mass



in the expression for

Z), this operator behaves as a dimension 3 rather than a dimension 4 operator.

Q2. Since,in(14.1.10),thetensorsmultiplyingW/4andPV5havea^or^fvinthem,a

This suppresses the Wilson coefficients for W4 and W5 by



power of

similar suppression occurs because the lepton masses are much less than Q, as is seen by contracting q^ or qv with the lepton tensor (14.1.7). The result is that

W4 and W5 make

a negligible contribution to the cross-section at large Q. A detailed treatment of W4 and W5 within QCD can be made but has not yet been published. It generalizes the results of Jaffe & Llewellyn-Smith

(1973), who worked within the parton model.

References

Adler, S. (1969). Phys. Rev. 177, 2426. Adler, S. (1970). In Lectures on Elementary Particles and Quantum Field Theory, eds. S. Deser et al. (M.I.T., Cambridge, Mass.). Adler, S. & Bardeen, W. (1969). Phys. Rev. 182, 1517. Aizenman, M. (1981). Phys. Rev. Lett. 47, 1. Altarelli, G. & Parisi, G. (1977). Nucl. Phys. B126, 298. Amati, D., Petronzio, R. & Veneziano, G. (1978). Nucl. Phys. B146, 29. Appelquist, T. & Carazzone, J. (1975). Phys. Rev. Dll, 2856. Ashmore, J. F. (1972). Nuovo Cimento Lett. 4, 289. Bandelloni, G., Becchi, C, Blasi, A. & Collina, R. (1980). Comm. Math. Phys. 72, 239. Bardeen, W. A. (1972). In Proceedings XVI International Conference on High Energy Physics (National Accelerator Laboratory, Batavia, IL). Bardeen, W. A, Buras, A. J., Duke, D. W. & Muta, T. (1978). Phys. Rev. D18, 3998. Baym, G. (1960). Phys. Rev. 117, 886. Becchi, C, Rouet, A. & Stora, R. (1975). Comm. Math. Phys. 42, 127. Becchi, C, Rouet, A. & Stora, R. (1976). Ann. Phys. (N.Y.) 98, 287. Beg, M. A. B. & Sirlin, A. (1982). Phys. Rep. 88, 1. Bell, J. S. & Jackiw, R. (1969). Nuovo Cimento 51, 47. Bernard, C. (1974). Phys. Rev. D9, 3312.

Bjorken, Bjorken,

J. D. (1982). Physica Scripta 25, 69. J. D. & Drell, S. D. (1966). Relativistic Quantum Fields (McGraw-Hill. New York). Bjorken, J. D. & Paschos, E. A. (1969). Phys. Rev. 185, 1975. Bogoliubov, N. N. & Parasiuk, O. (1957). Acta Math. 97, 227. Bogoliubov, N. N. & Shirkov, D. V. (1955/6). Usp. Fiz. Nauk 55, 149 (1955).

(German trans.: Fortschr. Phys. 4, 438 (1956).) Bogoliubov, N. N. & Shirkov, D. V. (1980). Introduction to the Theory of Quantized Fields (3rd edn) (Wiley, New York). Bollini, C G. & Giambiagi, J. J. (1972). Phys. Lett. 40B, 566. Brandt, R. A. (1967). Ann. Phys. (N.Y.) 44, 221. Brandt, R. A. (1976). Nucl. Phys. B116, 413. Breitenlohner, P. & Maison, D. (1977a). Comm. Math. Phys. 52, 11. Breitenlohner, P. & Maison, D. (1977b). Comm. Math. Phys. 52, 39. Breitenlohner, P. & Maison, D. (1977c). Comm. Math. Phys. 52, 55. Brown, L. S. (1980). Ann. Phys. (N.Y.) 126, 135. Brown, L. S. & Collins, J. C. (1980). Ann. Phys. (N.Y.) 130, 215. Buras, A. J. (1981). In Proceedings of Lepton-Photon Conference (Bonn). Buras, A. J., Floratos, E. G., Ross, D. A. & Sachrajda, C. T. (1977). Nucl. Phys. B131, 308.

References

373

Callan, C. G. (1970). Phys. Rev. D2, 1541. Callan, C. G., Coleman, S. & Jackiw, R. (1970). Ann. Phys. {N.Y.) 59, 42. Callan, C. G. & Gross, D. (1969). Phys. Rev. Lett. 22, 156. Caswell, W. E. & Kennedy, A. D. (1982). Phys. Rev. D25, 392. Caswell, W. E. & Zanon, D. (1981). Nucl. Phys. B182, 125. Celmaster, W. & Gonsalves, R. J. (1979). Phys. Rev. D20, 1420. Chanowitz, M, Furman, M. & Hinchliffe, I. (1979). Nucl. Phys. B159, 225. Cheng, T. P., Eichten, E. & Li, L.F. (1974). Phys. Rev. D9, 2259. Chetyrkin, K. G, Kataev, A. L. & Tkachov, F. V. (1981). Phys. Lett. 99B, 147, and erratum. Chetyrkin, K. G. & Tkachov, F. V. (1981). Nucl. Phys. B192, 159. Christ, N., Hasslacher, B. & Mueller, A. (1972). Phys. Rev. D6, 3543. Cicuta, G. M. & Montaldi, E. (1972). Nuovo Cimento Lett. 4, 329. Clark, T. E., Piguet, O. & Sibold, K. (1979). Nucl. Phys. B159, 1. Clark, T. E., Piguet, O. & Sibold, K. (1980). Nucl. Phys. B172, 201. Coleman, S. (1973). Comm. Math. Phys. 31, 259. Coleman, S. & Gross, D. J. (1973). Phys. Rev. Lett. 31, 851. Coleman, S. & Jackiw, R. (1971). Ann. Phys. (N.Y.) 67, 552. Coleman, S. & Weinberg, E. (1973). Phys. Rev. D7, 1888. Collins, J. C. (1974). Nucl. Phys. B80, 341. Collins, J. C. (1975a). PhD. Thesis: Cambridge University. Collins, J. C. (1975b). Nucl. Phys. B92, 477. Collins, J. C. [\916).Phys. Rev. D14, 1965. Collins, J. C. (1980). Phys. Rev. D22, 1478. Collins, J. C. & Soper, D. E. (1981). Nucl. Phys. B193, 381. Collins, J. C. & Soper, D. E. (1982a). Nucl. Phys. B194, 445. Collins, J. C. & Soper, D. E. (1982b). Nucl. Phys. B197, 446. Collins, J. C, Wilczek, F. & Zee, A. (1978). Phys. Rev. D18, 242. Cornwall, J. M. & Norton, R. E. (1973). Phys. Rev, D8, 3338. Costa, G., Marinucci, T., Tonin, M. & Julve, J. (1977). Nuovo Cimento 38A, 373. Creutz, M. (1980). Phys. Rev. D21, 2308. Creutz, M. (1983). Quarks, Gluons and Lattices (Cambridge University Press). Creutz, M. & Moriarty, K. J. M. (1982). Phys. Rev. D26, 2166. Curci, G., Furmanski, W. & Petronzio, R. (1980). Nucl. Phys. B175, 27. Cvitanovic, P. (1977). Nucl. Phys. B130, 114. Dirac, P. A. M. (1981). In Perturbative Quantum Chromodynamics (Tallahassee, 1981), ed. Duke, D. W. & Owens, J. F. (AIP, New York). Dixon, J. A. & Taylor, J. C (1974). Nucl Phys. B78, 552. Drell, S. D., Levy, D. & Yan, T. M. (1969). Phys. Rev. 187, 2159. Drell, S. D., Levy, D. & Yan, T. M. (1970a). Phys. Rev. Dl, 1035. Drell, S. D., Levy, D. & Yan, T. M. (1970b). Phys. Rev. Dl, 1617. Drell, S. D., Levy, D. & Yan, T. M. (1970c). Phys. Rev. Dl, 2404. Dyson, F. J. (1952). Phys. Rev. 85, 631. Ellis, R. K., Georgi, H, Machacek, M, Politzer, H. D. & Ross, G. G. (1979). Nucl. Phys. B152, 285. Faddeev, L. D. & Popov, V. M. (1967). Phys. Lett. 25B, 29. Fayet, P. & Ferrara, S. (1977). Phys. Rep. 32, 249. Fetter, A. L. & Walecka, J. D. (1971). Quantum Theory of Many-Body Systems (McGraw-Hill, New York). Feynman, R. P. (1972). Photon-Hadron Interactions (Benjamin, New York). Fisher, M. E. & daunt, D. S. (1964). Phys. Rev. 133, A224.

Freedman, D. Z., Muzinich, I. J. & Weinberg, E. J. (1974) Ann. Phys. (N.Y.) 87, 95.

374

References

Fujikawa, K., Lee, B. W. & Sanda, A. I. (1972). Phys. Rev. D6, 2923. Gasser, J. & Leutwyler, H. (1982). Phys. Rep. 87, 77. Gastmans, R. & Meuldermans, R. (1973). Nucl. Phys. B63, 277. Gastmans, R., Verwaest, J. & Meuldermans, R. (1976). Nucl. Phys. B105, 454. Gell-Mann, M. & Low, F. E. (1951). Phys. Rev. 84, 350. Gell-Mann, M. & Low, F. E. (1954). Phys. Rev. 95, 1300. Georgi, H. & Politzer, H. D. (1976). Phys. Rev. D14, 1829. Gildener, E. & Weinberg, S. (1976). Phys. Rev. D13, 3333. Glimm, J. & Jaffe, A. (1981). Quantum Physics: A Functional Integral Point of View (Springer-Verlag, New York). Goldstone, J., Salam, A. & Weinberg, S. (1962). Phys. Rev. 127, 965. Gross, D. (1976). In Methods in Field Theory, ed. Balian, R. & Zinn Justin, J. (North-Holland, Amsterdam). Gross, D. & Jackiw, R. (1972). Phys. Rev. D6, 477. Gross, D. & Neveu, A. (1974). Phys. Rev. D10, 3235. Hahn, Y. & Zimmermann, W. (1968). Comm. Math. Phys. 10, 330. Hasenfratz, A. & Hasenfratz, P. (1980). Phys. Lett. 93B, 165. Hawking, S. W. & Israel, W. (1979). General Relativity (Cambridge University Press). Hepp, K. (1966). Comm. Math. Phys. 2, 301. Hohenberg, P. (1967). Phys. Rev. 158, 383. Itzykson, C. & Zuber, J.-B. (1980). Quantum Field Theory (McGraw-Hill, New York). Jackiw, R. (1968). Ann. Phys. (N.Y.) 48, 292. Jackiw, R. & Johnson, K. (1973). Phys. Rev. D8, 2386. Jaffe, R. & Llewellyn-Smith, C. H. (1973). Phys. Rev. D7, 2506. Joglekar, S. (1976). Ann. Phys. (N.Y.) 100, 395. Joglekar, S. (1977a). Ann. Phys. (N.Y.) 108, 233. Joglekar, S. (1977b). Ann. Phys. (N.Y.) 109, 210. Joglekar, S. & Lee, B. W. (1976). Ann. Phys. (N.Y.) 97, 160. Jones, D. R. T. & Leveille, J. P. (1982). Nucl. Phys. B206, 473. Kazama, Y. & Yao, Y.-P. (1982). Phys. Rev. D25, 1605. Kluberg-Stern, H. & Zuber, J. B. (1975). Phys. Rev. D12, 467. Kogut, J. (1983). Rev. Mod. Phys. 55, 775. Korthals Altes, C. P. & Perrottet, M. (1972). Phys. Lett. 39B, 546. Lam, C. S. & Tung, W.-K. (1978). Phys. Rev. D18, 2447. Lam, Y.-M. P. (1972). Phys. Rev. D6, 2145. Landau, L. D. & Pomeranchuk, I. (1955). Dokl. Akad. Nauk USSR 102, 489. Langacker, P. (1981). Phys. Rep. 72, 185. LaRue, G. S., Phillips, J. D. & Fairbank, W. M. (1981). Phys. Rev. Lett. 46, 967. Lautrup, B. (1976). Nucl. Phys. B105, 23. Lee, B. W. (1976), In Methods in Field Theory, ed. Balian, R. & Zinn-Justin, J. (North-Holland, Amsterdam). Lee, B. W. & Zinn-Justin, J. (1972). Phys. Rev. D5, 3121. Libby, S. & Sterman, G. (1978). Phys. Rev. D18, 3252. Llewellyn-Smith, C. H. (1972). Phys. Rep. 3, 261. Llewellyn-Smith, C. H. (1980). In Quantum Flavordynamics, Quantum Chromodynamics, and Unified Theories, ed. Mahanthappa, K. T. & Randa, J. (Plenum, New York). Lowenstein, J. H. (1971). Phys. Rev. D4, 2281. Lowenstein, J. H., Weinstein, M. & Zimmermann, W. (1974a). Phys. Rev. D10, 1854.

375

References Lowenstein,

J. H., Weinstein, M. & Zimmermann, W. (1974b). Phys. Rev. DIO,

2500. Lurie, D. (1968). Particles and Fields (Interscience, New York). Macfarlane, A. J. & Woo, G. (1974). Nucl Phys. B77, 91; B86, 548(E). Mack, G. (1972). Kaiserslautern Lectures. Springer Lecture Notes in Physics, vol. 17, ed. Ruhl, W. & Vancura, A. Marciano, W. & Pagels, H. (1978). Phys. Rep. 36, 138. Marciano, W. & Sirlin, A. (1975). Nucl. Phys. B88, 86. Mermin, N. D. & Wagner, H. (1966). Phys. Rev. Lett. 17, 1133. Mueller, A. H. (1978). Phys. Rev. D18, 3705. Mueller, A. H. (1979). Phys. Rev. D20, 2037. Mueller, A. H. (1981). Phys. Rep. 73, 237. Oppenheimer, J. R. (1930). Phys. Rev. 35, 461. Ovrut, B. A. & Schnitzer, H. J. (1980). Phys. Rev. D21, 3369. Pauli, W. & Villars, F. (1949). Rev. Mod. Phys. 21, 434. Piguet, O. & Rouet, A. (1981). Phys. Rep. 76, 1.

Piguet, O. & Sibold, K. (1982a). Nucl. Phys. B197, 247. Piguet, O. & Sibold, K. (1982b). Nucl. Phys. B197, 257. Piguet, O. & Sibold, K. (1982c). Nucl. Phys. B197, 272. Ramond, P. (1981). Field Theory: A Modern Primer (Benjamin, New York). Ross, G. G. (1981). Rep. Prog. Phys. 44, 655.

Sarkar, S. (1974). Nucl. Phys. B83, 108. Schwinger, J. (1951). Phys. Rev. 82, 664. Shaw, R. (1955). Cambridge University Thesis. Slavnov, A. (1972). Theor. Math. Phys. 10, 153 (Russian); 10, translation). Speer, E. R. (1974). J. Math. Phys. 15, 1.

99

(English

Speer, E. R. & Westwater, M. J. (1971). Ann. Inst. Henri Poincare A14, 1. Stirling, W. J. (1978). Nucl Phys. B145, 477. Streater, R. F. & Wightman, A. S. (1978). PCT, Spin and Statistics, and All That

(Benjamin, New York). Stueckelberg, E. C. G. & Petermann,

A. (1953). Helv. Phys. Acta 26, 499. Sudakov, V. (1956). Zh. Eksp. Teor. Fiz. 30, 87 (EngUsh trans.: Sov. Phys. JETP 3, 65). Symanzik, K. (1970a). Comm. Math. Phys. 16, 48. Symanzik, K. (1970b). Comm. Math. Phys. 18, 227. Symanzik, K. (1973). Comm. Math. Phys. 34, 7.

Symanzik, K. (1982). In Proceedings XXII International Conference on High Energy Physics, ed. Petian, P. & Porneuf, M. (Editions de Physique, Les Ulis,

France). Tarasov, O. V, Vladimirov,

A. A. & Zharkov, A. Yu. (1980). Phys. Lett. 93B,

429. Taylor, J. C. (1971). Nucl. Phys. B33, 436. 't Hooft, G. (1971a). Nucl Phys. B33, 173. 't Hooft, G. (1971b). Nucl Phys. B35, 167.

(1973). Nucl. Phys. B61, 455. 't Hooft, G. & Veltman M. (1972a). Nucl Phys. B44, 189. 't Hooft, G. & Veltman M. (1972b). Nucl. Phys. B50, 318. Tkachov, F. V. (1981). Phys. Lett. 100B, 65. Toussaint, D. (1978). Phys. Rev. D18, 1628. Treiman, S. B., Jackiw, R. & Gross, D. J. (1972). Lectures and Its Applications (Princeton University Press). 't Hooft, G.

on Current

Algebra

376

References

Veltman, M. (1977). Nucl. Phys. B123, 89. Vladimirov, A. A., Kazakov, D. I. & Tarasov, O. V. (1979). Zh. Eksp. Teor, Fiz. 11, 1035 (English trans.: Sov. Phys. JETP 50, 521). Ward, J. C. (1950). Phys. Rev. 78, 182. Weinberg, S. (1960). Phys. Rev. 118, 838. Weinberg, S. (1973). Phys. Rev. D8, 3497. Weinberg, S. (1974). In Proceedings XVII International Conference on High

Energy Physics (London, 1974), ed. Smith, J. R. (Rutherford Lab., Chilton, Didcot, Berkshire, England). Weinberg, S. (1976). In Gauge Theories and Modern Field Theory, eds. Arnowitt, R. & Pran Nath (M.I.T., Cambridge, Mass.). Weinberg, S. (1980). Phys. Lett. 91B, 51. Wilson, K. G. (1969). Phvs. Rev. 179, 1699. Wilson, K. G. (1973). Phys. Rev. D7, 2911. Wilson, K. G. (1974). Phys. Rev. D10, 2445. Wilson, K. G. & Fisher, M. E. (1972). Phys. Rev. Lett. 28, 240. Wilson, K. G. & Kogut, J. (1974). Phys. Rep. 12, 75. Witten, E. (1976). Nucl Phys. B104, 445.

Yang, C. N. & Mills, R. (1954). Phys. Rev. Yao, Y. P. (1973). Phys. Rev. D7, 1647. Zee, A. (1972). Phys. Rev. Lett. 29, 1198.

96, 191.

Zimmermann, W. (1968). Comm. Math. Phys. 11, 1. Zimmermann, W. (1969). Comm. Math. Phys. 15, 208. Zimmermann, W. (1970). In Lectures on Elementary Particles and Quantum Field Theory, eds. S. Deser et al. (M.I.T., Cambridge, Mass).

Zimmermann, W. (1973a). Ann. Phys. (N.Y.) 77, 536. Zimmermann, W. (1973b). Ann. Phys. (N.Y.) 11, 570.

Index

e-expansion 62 <j>3 theory 36 one-loop calculations 38 A parameter of QCD 188 Adler-Bardeen theorem 338 Altarelli-Parisi equation 370 anomalies 331 Adler-Bardeen theorem 338 not in non-singlet axial current 346 one-loop calculation 342 triangle 349 anomalously broken symmetries 18, 331 asymptotic freedom 187 bare parameters definition 10 renormalization group calculation 209 dimensional regularization 217 basic graph 90 basic Lagrangian 89 definition 11 BPH 53 see also renormalization prescriptions BPHZ 53, 133 see also renormalization prescriptions broken symmetries 18 anomalous 18 dynamically 19 explicitly 18 spontaneously 18 not at d 2 26 =

perturbation theory 26, 247 renormalization 247, 314 BRS, see gauge theories canonical momentum field 5 causality 6 charge conjugation symmetry, for gauge theories 35 chiral symmetry 17, 333 and anomalies 338 commutation relations canonical equal-time 5 from Green's functions 6

composite operators

138

definition 146 equations of motion 155, 158 mixing under renormalization 149 properties 152 renormalization, example 142 tensor operators 150 twist 363 Ward identities 160 see also operator product expansion continuum limit 8 see also renormalization coordinate space, renormalization in 59, 277 cosmological term, and renormalization 118 counterterm Lagrangian 11, 89 cut-off dimensional regularization 13, 53, 62 for infra-red 56 lattice 8 Pauli-Villars 12 see also individual entries decoupling theorem 222 examples 226 exceptions 223 manifest decoupling 230 proof 233 renormalization group 240 deep inelastic scattering 354 Altarelli-Parisi equation 370 calculation of Wilson coefficients 364 operator product expansion 361 partons and operator product expansion

369 W± and W5 357, 371 degree of divergence 45 differentiation with respect to parameters in Lagrangian 163 dimensional regularization 13, 53, 62 axioms 64 definition 67 convergence 68 Dirac matrices in 83, 334 list of integrals 81 properties 73

378

Index

reduced to ordinary integration 67 uniqueness 65 dimensional transmutation 203 Dirac matrices y5 at 4^4 86, 334 definition in d dimensions 83

dynamically broken symmetries 19

unbroken, invariance of counterterms 245 Goldstone bosons, see Nambu-Goldstone bosons Goldstone model 247 Grassmann algebra 27 Green's functions

momentum space 15

definition 6 equations of motion 13 functional integral 7 generating functional 7

operator 5

momentum space 15

equations of motion Green's function 6,13

Euclidean

space-time 8.

evanescent operators 346

explicit symmetry breaking 18 factorization theorems 1, 187 Faddeev-Popov ghost, see gauge theories

Feynman rules for gauge theories 34 see also perturbation theory fixed points 190 forest formula 109 3-loop example 106 functional integral for bosons 6 for fermions 27 Monte-Carlo calculation 13

gauge dependence of counterterms 309, 327 gauge invariant operators 316 gauge theories 28 BRS invariance 32

Hamiltonian, field theory 5 infra-red behavior, and renormalization group 187, 191 Landau ghost 196 large mass expansion 222 large momentum behavior of Green's functions 185 lattice 8 leading logarithms 193 non-renormalization group 196 list of integrals 81 low-energy behavior 191 mass-shell renormalization prescription 48 massive photon, renormalization 324 minimal subtraction 56, 135 see also renormalization prescriptions Monte-Carlo 13

charge conjugation symmetry 35 Faddeev-Popov ghost 30 Feynman rules 34 gauge invariance v. gauge independence 31

Nambu-Goldstone bosons 21, 252 Noether's theorem 18 non-renormalizability 49

gauge-fixing 30 operator product expansion 321 proof of renormalizability 298 quantization 29

normal products,

renormalization 293 renormalization group 319 Slavnov-Taylor identities 32 Gell-Mann-Low formula 22 General Relativity as example of gauge theory 17 cosmological term and renormalization 118 generating functional for Green's functions 7 global symmetries renormalization 244 spontaneously broken, invariance of counterterms 247

in real physics 122 non-renormalization of currents 162 see

composite operators

operator product expansion 257 and parton model 369 and renormalization group 274 deep inelastic scattering 361 examples 139,258 gauge theories 321 in coordinate- and momentum-space 257 proof 266 to define composite operators 258 overlapping divergence 99 oversubtraction 130 on 1 PR graphs 131

partition function, functional integral as 8 parton model 355, 358, 369 Pauli-Villars cut-off 12

379

Index perturbation theory 21

saddle point expansion 26 Feynman rules for gauge theories 34 maximum accuracy with asymptotic freedom 189 not convergent 24, 189 perturbative QCD 1 Poincare symmetry 17 propagator 22 short-distance behavior 278 thermal 280 as

QCD A parameter 188 asymptotic freedom 188 Feynman rules 34 perturbative 1 see also gauge

theories

QED gauge dependence of counterterms 327 in external field 291 renormalization 293, 322 renormalization group 329 Ward identities 326 see also gauge theories see QCD quantum electrodynamics, see QED quantum field theory analog of statistical mechanics 8 axioms 6 review 4 standard properties 6 standard theories 36 see also, <£3 theory; QCD; QED quantum mechanics

quantum chromodynamics,

as

field theory at d

/^-gauge

314

regulator,

see



1 6

reconstruction theorem 6

cut-off

renormalizability 49, 116 renormalization 9, 88 and continuum limit 9 axial current 340 coordinate space 59, 277 examples 281 examples 38 for massless theories 59 forest formula 109 gauge dependence of counterterms 309, 327 gauge invariant operators 316 gauge theories 293 with symmetry breaking 308, 314 in external field 291 locality of counterterms 97, 125 non-zero temperature 286

oversubtraction 130 prescriptions 52 QED 293, 322 with photon mass 323 renormalization group 50, 168 A parameter 188 asymptotic freedom 187 change of renormalization prescription 200 coefficients 180 composite operators 219 computation of bare parameters 209 dimensional regularization 217 cut-off procedure, dependence on 206 decoupling theorem 240 dimensional transmutation 203 equation 183 examples 50, 172 fixed points 190 gauge theories 319 infra-red behavior 187, 191 Landau ghost 196 large momentum behavior 185 leading logarithms 193 minimal subtraction 56, 181 several couplings 197 solution of RG equation 184 renormalization prescriptions 52 BPH 53 BPHZ 53, 133 mass-shell 48 minimal subtraction 135 at one loop 56

spontaneously broken symmetries 253 renormalized field 10 renormalized mass, and physical mass 51

Slavnov-Taylor identities 32 spontaneously broken symmetries 18 absence of IR divergences 252 choice of renormalization prescription 253 renormalization 247 renormalization of current 251 super-renormalizability, definition 118 supersymmetries 17, 339 symmetries 15, 244 BRS, see gauge theories charge conjugation, see gauge theories gauge 28 Ward identities 19, 160 see also broken symmetries; global symmetries thermodynamic limit of functional integral 8 triangle anomaly 349

380 twist 363 ultra-violet regulators 12 see- also cut-off; dimensional regularization unitary gauge 324, 330 Ward identities 16, 19, 160 axial current 339. 349 OED 326

Index renormalization 246-256 Weinberg's theorem 125 Wick rotation 8, 39 Wightman axioms 6 Wilson expansion, see operator product expansion Yang-Mills theory, theories

see gauge