T H E M A N Y FAC E T S O F G E O M E T RY

This page intentionally left blank

The Many Facets of Geometry
A tribute to Nigel Hitchin

Edited by
Oscar Garcı́a-Prada
Jean Pierre Bourguignon
Simon Salamon

1
 3
Great Clarendon Street, Oxford ox2 6dp
Oxford University Press is a department of the University of Oxford.
It furthers the University’s objective of excellence in research, scholarship,
and education by publishing worldwide in
Oxford New York
Auckland Cape Town Dar es Salaam Hong Kong Karachi
Kuala Lumpur Madrid Melbourne Mexico City Nairobi
New Delhi Shanghai Taipei Toronto
With offices in
Argentina Austria Brazil Chile Czech Republic France Greece
Guatemala Hungary Italy Japan Poland Portugal Singapore
South Korea Switzerland Thailand Turkey Ukraine Vietnam
Oxford is a registered trade mark of Oxford University Press
in the UK and in certain other countries
Published in the United States
by Oxford University Press Inc., New York


c Oxford University Press 2010
The moral rights of the authors have been asserted
Database right Oxford University Press (maker)
First published 2010
All rights reserved. No part of this publication may be reproduced,
stored in a retrieval system, or transmitted, in any form or by any means,
without the prior permission in writing of Oxford University Press,
or as expressly permitted by law, or under terms agreed with the appropriate
reprographics rights organization. Enquiries concerning reproduction
outside the scope of the above should be sent to the Rights Department,
Oxford University Press, at the address above
You must not circulate this book in any other binding or cover
and you must impose the same condition on any acquirer
British Library Cataloguing in Publication Data
Data available
Library of Congress Cataloging in Publication Data
Data available
Typeset by SPI Publisher Services, Pondicherry, India
Printed in Great Britain
on acid-free paper by
CPI Antony Rowe

ISBN 978–0–19–953492–0

1 3 5 7 9 10 8 6 4 2
To Nigel Hitchin
This page intentionally left blank
 SHORT CONTENTS

FULL CONTENTS ix
PREFACE xvi
LIST OF EDITORS AND CONTRIBUTORS xvii

1 Geometry and physics: a personal view 1

Nigel Hitchin

2 Mathematical work of Nigel Hitchin 11

Sir Michael Atiyah

3 The Einstein–Maxwell equations, extremal Kähler metrics,

and Seiberg–Witten theory 17
Claude LeBrun

4 The Nahm transform for calorons 34

Benoit Charbonneau and Jacques Hurtubise

5 Nahm’s equations and free-boundary problems 71

Simon K. Donaldson

6 Some aspects of the theory of Higgs pairs 92

S. Ramanan

7 Mirror symmetry, Hitchin’s equations, and Langlands

duality 113
Edward Witten

8 Higgs bundles and geometric structures on surfaces 129

William M. Goldman

9 Locality of holomorphic bundles, and locality in quantum

field theory 164
Graeme Segal
viii Short Contents

10 Toeplitz operators and Hitchin’s projectively flat

connection 177
Jørgen Ellegaard Andersen

11 Towards a non-linear Schwarz’s list 210

Philip Boalch

12 An introduction to bundle gerbes 237

Michael K. Murray

13 Projective linking and boundaries of positive holomorphic

chains in projective manifolds, part I 261
F. Reese Harvey and H. Blaine Lawson, Jr.

14 Skyrmions and nuclei 281

Nicholas S. Manton

15 Mirror symmetry of Fourier–Mukai transformation for

elliptic Calabi–Yau manifolds 299
Naichung Conan Leung and Shing-Tung Yau

16 S -duality in hyperkähler Hodge theory 324

Tamás Hausel

17 Non-embedding and non-extension results in special

holonomy 346
Robert L. Bryant

18 Branes on Poisson varieties 368

Marco Gualtieri

19 Consistent orientation of moduli spaces 395

Daniel S. Freed, Michael J. Hopkins, and Constantin Teleman

INDEX 421
 FULL CONTENTS

PREFACE xvi
LIST OF EDITORS AND CONTRIBUTORS xvii

1 Geometry and physics: a personal view 1

Nigel Hitchin

2 Mathematical work of Nigel Hitchin 11

Sir Michael Atiyah

3 The Einstein–Maxwell equations, extremal Kähler metrics,

and Seiberg–Witten theory 17
Claude LeBrun

4 The Nahm transform for calorons 34

Benoit Charbonneau and Jacques Hurtubise
4.1 Introduction 34
4.2 The work of Nye and Singer 35
4.2.1 Two types of invariant self-dual gauge fields on R4 35
4.2.2 The Nahm transform 37
4.2.3 Involutivity of the transforms 39
4.3 Twistor transform for calorons/Kač–Moody monopoles 39
4.3.1 Upstairs: twistor transform for calorons 39
4.3.2 Downstairs: caloron as a Kač–Moody monopole 40
4.4 From Nahm’s equations to spectral data, and back 46
4.4.1 Flows of sheaves 46
4.4.2 Boundary conditions 48
4.5 Closing the circle 55
4.5.1 Starting with a caloron 55
4.5.2 Starting with a solution to Nahm’s equation 58
4.5.3 From Nahm to caloron to Nahm to caloron 64
4.6 Moduli 66
5 Nahm’s equations and free-boundary problems 71
Simon K. Donaldson
5.1 Introduction 71
5.2 An infinite-dimensional Riemannian manifold 74
x Full Contents

5.3 Three equivalent problems 78

5.3.1 θ equation ⇒ φ equation 79
5.3.2 Φ equation ⇒ U equation 81
5.3.3 U equation ⇒ θ equation 81
5.4 Existence results and discussion 82
5.4.1 Monge–Ampère and the results of Chen 82
5.4.2 Comparison with the free-boundary literature 84
5.4.3 Degenerate case 86
5.5 Relation with Nahm’s equations 86
6 Some aspects of the theory of Higgs pairs 92
S. Ramanan
6.1 Moduli of vector bundles 92
6.2 Hecke correspondence 94
6.3 Moduli of Higgs pairs 95
6.4 Higgs pairs and the fundamental group 97
6.5 Non-abelian Hodge theory 98
6.6 Hitchin morphism 99
6.7 Quantization 101
6.8 Hecke transformation and Hitchin discriminant 102
6.9 Hitchin component 103
6.10 Reductive groups and principal bundles 105
6.11 Reductive groups and Higgs pairs 106
6.12 Real forms 108
7 Mirror symmetry, Hitchin’s equations, and Langlands
duality 113
Edward Witten
7.1 A-model and B-model 113
7.2 Mirror symmetry and Hitchin’s equations 114
7.3 Hitchin fibration 116
7.3.1 A few hints 117
7.4 Ramification 118
7.5 Wild ramification 120
7.6 Four-dimensional gauge theory and stacks 123
7.6.1 Stacks 125
8 Higgs bundles and geometric structures on surfaces 129
William M. Goldman
8.1 Introduction 129
8.2 Representations of the fundamental group 130
8.2.1 Closed surface groups 130
8.2.2 Representation variety 130
8.2.3 Symmetries 132
8.2.4 Deformation space 132
 Full Contents xi

8.3 Abelian groups and rank 1 Higgs bundles 133

8.3.1 Symplectic vector spaces 133
8.3.2 Multiplicative characters: G = C∗ 133
8.3.3 Jacobi variety of a Riemann surface 134
8.4 Stable vector bundles and Higgs bundles 135
8.5 Hyperbolic geometry: G = PSL(2, R) 137
8.5.1 Geometric structures 137
8.5.2 Relation to the fundamental group 138
8.5.3 Examples of hyperbolic structures 138
8.6 Moduli of hyperbolic structures and representations 142
8.6.1 Deformation spaces of geometric structures 142
8.6.2 Fuchsian components of Hom(π, G)/G 143
8.6.3 Characteristic classes and maximal representations 144
8.6.4 Quasi-Fuchsian representations: G = PSL(2, C) 145
8.6.5 Teichmüller space: marked conformal structures 148
8.6.6 Holomorphic vector bundles and uniformization 149
8.7 Rank 2 Higgs bundles 150
8.7.1 Harmonic metrics 150
8.7.2 Higgs pairs and branched hyperbolic structures 151
8.7.3 Uniformization with singularities 152
8.8 Split R-forms and Hitchin’s Teichmüller component 153
8.8.1 Convex RP2 -structures: G = SL(3, R) 153
8.8.2 Higgs bundles and affine spheres 154
8.8.3 Hyperconvex curves 155
8.9 Hermitian symmetric spaces: maximal representations 156
8.9.1 Unitary groups U(p, q) 156
8.9.2 Symplectic groups Sp(n, R) 157
8.9.3 Geometric structures associated to Hitchin
representations 158
9 Locality of holomorphic bundles, and locality in quantum
field theory 164
Graeme Segal
9.1 Noncommutative geometry 165
9.2 Algebraic structures up to homotopy 169
9.3 Quantum field theory 171
10 Toeplitz operators and Hitchin’s projectively flat
connection 177
Jørgen Ellegaard Andersen
10.1 Introduction 177
(n)
10.1.1 Geometric construction of Zk 179
10.1.2 Asymptotic faithfulness 181
10.1.3 Kashdan’s property (T) and the mapping class group 182
10.1.4 Geometric construction of the curve operators 184
xii Full Contents

10.1.5 The Nielsen–Thurston classification of mapping classes

is determined by TQFT 185
10.1.6 General setting 188
10.2 The Hitchin connection 188
10.3 Toeplitz operators and Berezin–Toeplitz deformation
quantization 194
10.4 The formal Hitchin connection 198
10.5 Asymptotics flatness of Toeplitz operators 201
10.6 General asymptotic faithfulness 204
11 Towards a non-linear Schwarz’s list 210
Philip Boalch
11.1 Introduction 210
11.1.1 Naive generalizations 211
11.1.2 Non-linear analogue: the Painlevé VI equation 213
11.2 What is Painlevé VI? 214
11.2.1 Conceptual approach to Painlevé VI 216
11.2.2 Explicit non-linear equations 217
11.2.3 Monodromy of Painlevé VI 218
11.3 Algebraic solutions from finite subgroups of SL2 (C) 219
11.3.1 What exactly is an algebraic solution? 219
11.3.2 (A) → (C) 219
11.4 Beyond Platonic Painlevé VI solutions 221
11.4.1 Construction 222
11.4.2 Relating connections (A) and (B) 223
11.4.3 New solutions 226
11.5 Pullbacks 228
11.6 Final steps 231
11.6.1 Up to degree 24 231
11.6.2 Quadratic/Landen/folding transformations 232
11.7 Conclusion 233
12 An introduction to bundle gerbes 237
Michael K. Murray
12.1 Introduction 238
12.2 Background 242
12.3 Bundle gerbes 243
12.3.1 Pullback 245
12.3.2 Dual and product 245
12.3.3 Characteristic class 245
12.3.4 Connective structure 246
12.4 Triviality 247
12.4.1 Holonomy 250
12.4.2 Obstructions to certain kinds of Y → M 251
 Full Contents xiii

12.5 Examples of bundle gerbes 251

12.5.1 Lifting bundle gerbe 252
12.5.2 Projective bundles 253
12.5.3 Bundle gerbes on Lie groups 254
12.6 Applications of bundle gerbes 255
12.6.1 Wess–Zumino–Witten term 255
12.6.2 Faddeev–Mickelsson anomaly 256
12.6.3 String structures 257
12.7 Other matters 257
13 Projective linking and boundaries of positive holomorphic
chains in projective manifolds, part I 261
F. Reese Harvey and H. Blaine Lawson, Jr.
13.1 Introduction 261
13.2 Projective hulls 264
13.3 Projective linking and projective winding numbers 267
13.4 Quasi-plurisubharmonic functions 269
13.5 Boundaries of positive holomorphic chains 270
13.6 Projective Alexander–Wermer theorem for curves 271
13.7 Theorems for general projective manifolds 275
13.8 Relative holomorphic cycles 277
14 Skyrmions and nuclei 281
Nicholas S. Manton
14.1 Skyrmions 281
14.2 Rational map ansatz 283
14.3 Skyrmions and α-particles 286
14.3.1 B = 8 286
14.3.2 B = 12 287
14.3.3 B = 16 289
14.3.4 B = 24 290
14.3.5 B = 28 291
14.3.6 B = 32 291
14.4 Quantization and calibration 292
14.5 Conclusions 294
Appendix: Double rational map ansatz 295
15 Mirror symmetry of Fourier–Mukai transformation for
elliptic Calabi–Yau manifolds 299
Naichung Conan Leung and Shing-Tung Yau
15.1 Introduction 299
15.2 Mirror symmetry and SYZ transformation 300
15.2.1 Geometry of Calabi–Yau manifolds 300
15.2.2 Mirror symmetry conjectures 300
xiv Full Contents

15.2.3 SYZ transform 301

15.2.4 Mirror of elliptic fibrations 302
15.3 Fourier–Mukai transform and elliptic CY manifolds 304
15.3.1 General Fourier–Mukai transform 304
15.3.2 Elliptic fibrations and their duals 304
15.3.3 FM transform and spectral cover construction 305
15.4 Symplectic FM transform and twin Lagrangian fibrations 306
15.4.1 Symplectic Fourier–Mukai transform 306
15.4.2 Twin Lagrangian fibrations 307
15.4.3 Dual twin Lagrangian fibration and its FM transform 312
15.4.4 Examples of twin Lagrangian fibrations 313
15.5 SYZ transformation of FM transform 315
15.5.1 SYZ transform of dual elliptic fibrations 315
15.5.2 SYZ transform of the universal Poincaré sheaf 317
15.5.3 SYZ transforms commute with FM transforms 320
15.6 Conclusions and discussions 321
16 S -duality in hyperkähler Hodge theory 324
Tamás Hausel
16.1 Introduction 324
16.2 Hyperkähler quotients 326
16.2.1 Moduli of Yang–Mills instantons on R4 326
16.2.2 Moduli space of magnetic monopoles on R3 327
16.2.3 Hitchin moduli space 328
16.3 Hodge theory 328
16.3.1 L2 harmonic forms on complete manifolds 328
16.3.2 Results on L2 harmonic forms 329
16.4 Mixed Hodge theory 331
16.4.1 Mixed Hodge structure of Deligne 331
16.4.2 Arithmetic and topological content of the E-polynomial 332
16.5 Applications of mixed Hodge theory 333
16.5.1 Nakajima quiver varieties 333
16.5.2 Spaces diffeomorphic to the Hitchin moduli space
M(C, PU (n) ) 336
16.5.3 Topological mirror test 337
16.5.4 Mirror symmetry for finite groups of Lie type 338
16.5.5 Conjectural answer 339
17 Non-embedding and non-extension results in special
holonomy 346
Robert L. Bryant
17.1 Introduction 346
17.2 Beginnings 349
17.2.1 Holonomy 349
 Full Contents xv

17.2.2 A non-trivial case 350

17.2.3 Hyper-Kähler viewpoint 351
17.3 Hyper-Kähler four-manifolds 353
17.3.1 An exterior differential systems proof 353
17.3.2 A sharper result 353
17.4 G2 -Manifolds 358
17.4.1 Hypersurfaces 359
17.4.2 Flow interpretation 362
17.5 Spin(7)-manifolds 362
17.5.1 Hypersurfaces 364
17.5.2 Flow interpretation 365
18 Branes on Poisson varieties 368
Marco Gualtieri
18.1 Introduction 368
18.2 Gerbe trivializations 369
18.3 Generalized connections 370
18.4 Generalized holomorphic bundles and branes 374
18.4.1 Generalized holomorphic bundles 375
18.4.2 Generalized complex branes 377
18.5 Multiple branes and holomorphic Poisson varieties 380
18.6 Relation to generalized Kähler geometry 385
18.7 Construction of generalized Kähler metrics 388
18.8 Relation to non-commutative algebraic geometry 391
19 Consistent orientation of moduli spaces 395
Daniel S. Freed, Michael J. Hopkins, and Constantin Teleman
19.1 Push–pull construction of TQFT 397
19.1.1 Quantum field theory 397
19.1.2 Topological construction 398
19.1.3 Remarks 400
19.2 Orientation and twisting 401
19.2.1 Ordinary cohomology 401
19.2.2 K -theory 403
19.3 Universal orientations and consistent orientations 405
19.3.1 Overview 405
19.3.2 Closed surfaces 405
19.3.3 Surfaces with boundary 408
19.3.4 The level 411
19.3.5 Pushforward maps 413
19.4 Families of surfaces, twistings, and anomalies 415

INDEX 421
 PREFACE

A conference to honour Professor Nigel Hitchin on the occasion of his 60th

birthday took place at the Consejo Superior de Investigaciones Cientı́ficas in
Madrid during the week of 4–8 September 2006. The meeting was devoted to
the many research topics covered by Professor Hitchin, and was a satellite of the
International Congress of Mathematicians, that had closed in the same city a
few days earlier.
This book was conceived at the time of the September conference, though it
took longer to acquire its present unity. It incorporates chapters from most of the
speakers, as well as chapters by Simon Donaldson, Michael Murray, and Edward
Witten, all of whom attended the conference ‘in spirit’. We are delighted too
that Nigel Hitchin himself agreed to write a personal view on ‘Geometry and
physics’ that places all the contributions in context.
Nigel Hitchin’s mathematical influence has been enormous, and his work is
frequently cited in widely different branches of mathematics. In accordance with
the book’s title, his work concerns not just differential geometry, but strikes
at the heart of algebraic geometry, complex analysis, Lie groups, as well as
abstract algebra and topology. The present contributions reflect his work on four-
manifolds, monopoles and instantons, Nahm’s equations, hyperkähler manifolds,
Higgs bundles, integrable systems, the geometrical structure of moduli spaces,
gerbes, harmonic forms, reduced holonomy, special Lagrangian submanifolds,
and generalized complex structures. Still this list is not exhaustive, and a number
of chapters address aspects of theoretical physics that underlie some of Hitchin’s
most significant results.
Seven of the contributors either obtained their doctorates under Nigel Hitchin’s
supervision (as did two of the editors), or studied actively with him during
their time as graduate students. It is likely that each of the remaining authors
would claim to have been a ‘student of Nigel’ at some point in their professional
career. Many of Nigel’s students have inherited the clear presentational style of
announcing deep results that we have become accustomed to in his own lectures
and seminars. This aspect was all too evident in the conference, and we feel that
it is well represented in this book.
The editors would like to extend special thanks to all the contributors and,
above all, to Nigel Hitchin himself.
Oscar Garcı́a-Prada
Jean Pierre Bourguignon
Simon Salamon
June 2009
 LIST OF EDITORS AND CONTRIBUTORS

Editors
Oscar Garcı́a-Prada (CSIC, Madrid, Spain)
Jean Pierre Bourguignon (IHES, France)
Simon Salamon (Politecnico di Torino, Italy)

Contributors
Jørgen Ellegaard Andersen (Aarhus Universitet, Denmark): Toeplitz opera-
tors and Hitchin’s projectively flat connection.
Sir Michael Atiyah (University of Edinburgh, UK): Mathematical work of
Nigel Hitchin.
Philip Boalch (ENS, Paris, France): Towards a non-linear Schwarz’s list.
Robert L. Bryant (MSRI, USA): Non-embedding and non-extension results in
special holonomy.
Benoit Charbonneau (Duke University, Durham, USA) and Jacques
Hurtubise (McGill University, Canada): The Nahm transform for calorons.
Simon K. Donaldson (Imperial College, London, UK): Nahm’s equations and
free-boundary problems.
Daniel S. Freed (University of Texas at Austin, USA), Michael J. Hopkins
(Harvard University), and Constantin Teleman (UC Berkeley, USA): Consis-
tent orientation of moduli spaces.
William M. Goldman (University of Maryland, USA): Higgs bundles and
geometric structures on surfaces.
Marco Gualtieri (MIT, USA): Branes on Poisson varieties.
F. Reese Harvey (Rice University, USA) and H. Blaine Lawson, Jr. (SUNY
at Stony Brook, USA): Projective linking and boundaries of positive holomorphic
chains in projective manifolds, part I.
Tamás Hausel (University of Oxford, UK): S-duality in hyperkähler Hodge
theory.
Nigel Hitchin (University of Oxford, UK): Geometry and Physics: a personal
view.
xviii List of Editors and Contributors

Claude LeBrun (SUNY at Stony Brook, USA): The Einstein–Maxwell equa-

tions, extremal Kähler metrics, and Seiberg–Witten theory.
Naichung Conan Leung (Chinese University of Hong Kong, Hong Kong) and
Shing-Tung Yau (Harvard University, USA): Mirror symmetry of Fourier–
Mukai transformation for elliptic Calabi–Yau manifolds.
Nicholas S. Manton (University of Cambridge, UK): Skyrmions and nuclei.
Michael K. Murray (University of Adelaide, Australia): An introduction to
bundle gerbes.
S. Ramanan (Chennai Mathematical Institute, India): Some aspects of the
theory of Higgs pairs.
Graeme Segal (University of Oxford, UK): Locality of holomorphic bundles,
and locality in quantum field theory.
Edward Witten (IAS, Princeton, USA): Mirror symmetry, Hitchin’s equations,
and Langlands duality.
 I
GEOMETRY AND PHYSICS: A PERSONAL VIEW

Nigel Hitchin

This excellent collection of chapters prompts me to recall here how their various
themes emerged in my own mathematical development. It may help the reader to
navigate a passage through them and also perhaps explain how the links between
physics and geometry which seem to underlie much of my own work came about.
I hope these background comments do not seem too self-indulgent, but they
might shed some light on the curious process of doing mathematical research –
a process which is largely hidden in the chapters we finally write.
I was never a serious student of physics. As a mathematics undergraduate
in Oxford from 1965 to 1968 I avoided the courses in relativity and quantum
mechanics that my peers were taking. Instead I would browse in the Jesus College
library trying to make sense of Weyl’s The Classical Groups or in particular
Harley Flanders’ Differential Forms with Applications to the Physical Sciences,
one which in a way defined my interests right from the start.
Although my DPhil thesis was about the Dirac operator, this was largely
in the context of the Atiyah–Singer index theorem and associated vanishing
theorems and it hardly occurred to me that physicists would be interested in
this Riemannian version. When I went to the Institute for Advanced Study in
1971–3 my neighbour in the next apartment was the physicist Andy Hanson
but we rarely spoke about our work. He once asked me what was so special
about 26 dimensions and all I could think of was that it was of the form 8k + 2
(where the mod 2 index exists). My main thoughts at that time were on the
two big unsolved problems in differential geometry: the Yamabe conjecture and
the Calabi conjecture. Yau’s presence at the Institute in the second year was
a clear influence here. My first paper, on four-dimensional Einstein manifolds
(Hitchin 1974), was heavily influenced by the Calabi conjecture and the K3
surface. There was a great lack of understanding of the Einstein equations
at the time amongst pure mathematicians, concerning both global and local
existence.
It was the conformal invariance of the Dirac operator which brought me
closer to physics. I spent the year 1973–4 in New York at the Courant Institute
and I began looking for a systematic way of finding all first-order conformally
invariant differential operators. I started to read papers of Penrose on the zero
rest-mass field equations; the conformal invariance of Maxwell’s equations on a
2 Geometry and physics

four-manifold gave an elliptic complex

Ω2± → Ω3 → Ω4
which intrigued me at the time. I remember rather nervously giving a talk one
cold December day at MIT to an audience of Atiyah, Kostant, Singer, and Irving
Segal about Szegö kernels for the Dirac operator and I became aware then of
Segal’s interest in conformal invariance, causality, and Mach’s principle.
I returned to Oxford in 1974 as Michael Atiyah’s research assistant. Roger
Penrose had also recently been appointed as the Rouse Ball Professor and I
began to learn from him and his students about twistor theory, in particular
the non-linear graviton construction. At the time everything was performed
over the complex numbers, or specialized to Minkowski space, and my interest
was still driven by the Calabi conjecture and whether a K3 surface, or some
sort of complexification, could be found using twistors. The first I heard of
the Yang–Mills equations was overhearing, while in the queue for tea in the
Mathematical Institute, Penrose describing his student Richard Ward’s new
twistorial construction: ‘You just turn the handle and out it comes . . .’ However
it was Singer’s visit in early 1977 that introduced us all to the Euclidean version
of the Yang–Mills equations.
For me this was a real eye-opener. First of all physicists like ‘t Hooft and
Jackiw had very concrete constructions of connections in Euclidean space which
satisfied the self-duality equations – the Yang–Mills instantons. At the same
time there were gravitational analogues – concrete four-dimensional Einstein
metrics produced by Hawking and Gibbons, and Eguchi and Hanson (my former
neighbour in Princeton). Their insight and intuition had achieved just the sort
of thing that pure mathematicians should have been able to do. But twistor
theory gave a link to algebraic geometry, so there was a chance that we could
contribute in reverse. While Atiyah, Ward, and Penrose were developing a version
of twistor theory to deal with the Euclidean case, I was recalling how my earlier
interests were appropriate for this particular problem. The inequality in my paper
(Hitchin 1974) could now be seen simply as the statement that the bundle Λ2+
on an Einstein manifold is self-dual; the conformal invariance of the Yang–Mills
equations meant that the four-sphere and vanishing theorems could be used; and
by turning around the elliptic complex which was my earlier interest, one could
get the deformation complex for self-dual connections
Ω0 (g) → Ω1 (g) → Ω2− (g).
The index theorem then showed that there were more solutions than the physi-
cists had yet found, which was one up for the mathematicians. Quite soon,
using differential geometry and recent results on the algebraic geometry of
vector bundles, Atiyah and I, and independently Manin and Drinfeld, found
a way to construct all of them. This experience was not only satisfying, but also
formative – precisely stated problems, originating in physics and unthought of
by pure mathematicians, yielding to the latest pure mathematical techniques.
 Geometry and physics 3

Chapter 3 by Claude LeBrun represents well this general area – he takes the
Einstein–Maxwell equation in four dimensions and links it with Kähler geometry
and topological obstructions. The work draws also on the more recent Seiberg–
Witten equations, one of the most important contributions to four-dimensional
geometry from the physicists. This of course followed on from Donaldson’s
own spectacular use of gauge theory in four-dimensional topology. I have to
admit that, when writing the 1978 paper with Atiyah and Singer, I thought
that self-dual connections on self-dual spaces was a cosy self-contained theory –
a sort of quaternionic version of Riemann surfaces. But then Taubes showed
how to construct self-dual connections in general and there was also clearly
a story to be told for algebraic surfaces. When Simon Donaldson became my
research student in 1980, I asked him to look into these questions, and the rest is
history.
Following on from the instantons, I began to look at related equations, and in
particular magnetic monopoles in Euclidean space. Here the physical intuition of
Nick Manton had acquired solid mathematical support in the existence proof of
Jaffe and Taubes, but the Bogomolny equations which describe the monopoles
were also clear cousins of the instanton equations and I wanted to use twistor
methods to describe the solutions more concretely. I began by asking what
was the natural background geometry in three dimensions to support these –
an analogue of the self-dual conformal structures in four dimensions. These
turned out to be what are known as Einstein–Weyl manifolds (after Weyl’s
early attempt to introduce the electromagnetic field into general relativity).
The important feature of these is that the space of geodesics has the structure
of a two-dimensional complex manifold, giving a version of twistor theory. I
then abandoned the general picture and concentrated on the Euclidean case and
came up with the spectral curve description of SU (2) monopoles. The solution
to the equations was determined by an algebraic curve which satisfied some
transcendental constraints. This was in some respects a geometrical version of
Corrigan and Goddard’s application of Ward’s method. I was about to move on
to something else when a preprint of Nahm fell on my desk (1982).
Nahm’s construction of SU (2) monopoles of charge k involved a system of
ordinary differential equations for three k × k matrices T1 , T2 , T3 : T1 = [T2 , T3 ],
etc. For k = 2 he showed that this could be solved with elliptic functions. Since
my spectral curve for charge 2 was elliptic there was a clear challenge to relate
directly the two approaches, which led to Hitchin (1983). In writing this, I learned
more about non-linear systems which can be linearized on the Jacobian of a
curve. Nahm’s equations – generalizations of the spinning top equations – are
examples of this type and the curve in question turns out to be the spectral
curve of the monopole as defined by the twistor approach. The Nahm approach,
however, has the specific advantage that one can see directly that the solutions
to the Bogomolny equations are nonsingular. One should note that singularity
issues dogged the early development of the theory – Ward’s charge 2 solution
had to be checked numerically and Forgács, Horváth, and Palla consigned
4 Geometry and physics

essentially the same solution to the waste bin because they thought it was
singular.
The Nahm method is the subject of Chapter 4 by Benoit Charbonneau and
Jacques Hurtubise, though not for monopoles – instead for the more sophisticated
calorons, which are instantons on S 1 × R3 . This is a more elaborate closing
of a circle of ideas than in Hitchin (1983). I am particularly pleased to see it
here since I began to make headway on the problem in 1983 and had every
intention of spending a sabbatical in Stony Brook working out the details, but
only got as far as giving a talk to Jacques and his colleagues in Montreal
when other issues intervened, of which more later. One of the most beautiful
applications of Nahm’s equations to monopoles was Simon Donaldson’s proof
(1984) that a circle bundle over the moduli space is naturally diffeomorphic
to the space of based rational maps from the projective line to itself. In his
contribution to this collection, Simon returns to Nahm’s equations but where
the matrices Ti are replaced by vector fields. He pointed out to me long ago that
Ashtekar’s Hamiltonian formulation of general relativity could be viewed this
way (I gave a minor extension to this in Hitchin (1998) in the context of hyper-
complex geometry). Donaldson’s chapter links Nahm’s equations to the physical
area of free boundary problems and their analogues on compact Riemannian
manifolds.
Arriving in Stony Brook in August 1983, I made contact with Martin Roček
in the Theoretical Physics Department. Earlier in the year Blaine Lawson had
visited Oxford and he was puzzled by a construction of hyperkähler manifolds
by Martin and his co-workers. ‘They take a group representation, turn a handle,
and you get a hyperkähler metric.’ My own interest at the time in hyperkähler
geometry was dominated by the twistor approach, so I was curious to learn what
was going on. We soon discovered that this was a straightforward generalization
of the symplectic quotient construction, which was very much in vogue in Oxford
at the time, for example, in the work of Atiyah and Bott on the Yang–Mills
equations on Riemann surfaces, and in Frances Kirwan’s work in algebraic
geometry. I worked hard contributing to the joint paper Hitchin et al. (1987a)
in trying to reconcile the physicists’ language of supersymmetric fields with the
twistor approach, but I think the reader can easily discern the borderline between
mathematics and physics. In fact Martin tried at the time to convince me to work
on another piece of geometry related to the supersymmetric sigma model, but I
wasn’t listening. It finally surfaced 25 years later in Marco Gualtieri’s 2003 thesis
as a generalized Kähler manifold.
On my return to Oxford in 1984, Atiyah and I set to work in evaluating the
hyperkähler metric on the moduli space of charge 2 monopoles – the hyperkähler
quotient construction when applied to a certain infinite-dimensional affine space
showed that every monopole moduli space has such a metric. I then began looking
at other moduli spaces with this property, where the gauge-theoretic equations
are hyperkähler moment maps. One of these could be defined on a Riemann
 Geometry and physics 5

surface and became the study of Higgs bundles (the terminology is due to Carlos
Simpson). At first (as in Hitchin 1986) I thought I was just getting a hyperkähler
metric on the cotangent bundle of the moduli space of stable bundles, and I
began explaining this on the blackboard to Simon Donaldson but got stuck.
Having decided that there must be a joint stability condition it was easy to see
what should happen because one knew the equations to be satisfied and the type
of vanishing theorem that would ensue. From then on, virtually every day a new
facet of the subject seemed to appear. The paper Hitchin (1987b) by no means
wrote itself but it certainly had an inner momentum. There was one aspect
however – the integrable system – which came a little later, because it was not
really part of the hyperkähler story. I was visiting Jacques Hurtubise in 1985 and
we went up to his family cottage in the Laurentian mountains. While he was doing
some maintenance work I sat on the porch with a pen and paper and suddenly
realized that we had a completely integrable Hamiltonian system where the fibres
were recognizable abelian varieties. Until then, my experience with such objects
was confined to Nahm’s equations but here was a vast natural source of such
things.
Chapter 6 by Ramanan gives a good survey of the subject from the point
of view of algebraic geometry, including the Hecke correspondence which links
up with Chapter 7 by Witten on the very active recent work on the geometric
Langlands programme from a physics point of view. This makes essential use of
both the gauge-theoretic equations and the integrable system. Chapter 8 by Bill
Goldman is again about Higgs bundles, but focuses on the interpretation as flat
bundles and in particular on representations of the fundamental group of the
surface into a real form of the group. One of the current problems in this area is
to give an understanding of the generalized Teichmüller spaces of Hitchin (1982)
as structures on the surface modulo diffeomorphism and Goldman discusses some
recent results in this direction.
From 1979 to 1989, I was a Fellow of St Catherine’s College, and apart from
my teaching duties there I also had multiple opportunities over lunch to talk
to Graeme Segal and Michael Atiyah, who were also Fellows. I never wrote a
joint paper with Graeme, but I valued greatly the questions and remarks which
came in our discussions. Our approaches were different – I would always want
to talk about stable bundles and a finite-dimensional moduli space whereas for
him stability was not an issue and, as he says in Chapter 9, the consideration of
the category or stack of bundles is actually closer to quantum field theory which
was always his prime motivation. In Autumn 1988 we had a seminar in Oxford
run by Atiyah, Segal, and Ruth Lawrence on quantum field theory and the Jones
polynomial. During that summer, and in particular at the International Congress
of Mathematical Physicists in Swansea, Witten, aided by Atiyah and Segal, had
seen how these invariants could be viewed via Chern–Simons theory. One aspect
which I liked very much was the geometric quantization of the moduli space of
flat unitary connections on a surface. As a research student I had learned a bit
6 Geometry and physics

about quantization from another student John Rawnsley, but I never understood
pairings between different polarizations. The interpretation here as a projectively
flat connection on a bundle over Teichmüller space was more appealing, and I
set to work with the aim of giving a talk at Michael Atiyah’s 60th birthday
meeting in 1989. For some reason, I chanced upon the paper of Welters (1983)
using a heat equation in the abelian case, and before long I saw that the Poisson-
commuting functions of my integrable system were instrumental in defining this
connection for non-abelian theta functions. My construction was a mixture of
algebraic geometry and differential geometry, but Chapter 10 by Jørgen Andersen
shows how much further one can go by a deeper analysis of the non-abelian heat
equation.
In January 1990 I left Oxford for Warwick, and began to learn about the
Painlevé equations. It was really the work of Paul Tod in Oxford on self-dual four-
manifolds with SO(3) symmetry that spurred my interest. The two-monopole
metric which Atiyah and I had calculated using complete elliptic integrals was one
such example, but one knew that the general solution of the Painlevé equation
involved genuinely new transcendental functions, so there was a question about
how far one could go in hoping to get closed-form expressions for such metrics.
On the other hand twistor theory said that these could all be described by
a complex three-manifold, so I tried to approach the problem that way by
looking at threefolds with an action of the complex group SO(3, C). This led
me to some algebraic solutions which I spoke about for the 60th birthdays of
Narasimhan and Seshadri (Hitchin 1996). It also led into the classical geometry of
Poncelet’s theorem, which I had learned about earlier from Atiyah in the context
of hyperbolic monopoles. The twistor spaces in these cases were equivariant
compactifications of SO(3, C) modulo a dihedral group and the obvious ques-
tion was to find solutions of Painlevé’s sixth equation for any finite subgroup.
Chapter 11 by Philip Boalch describes this subject and its general context well.
He has given a classification of icosahedral solutions, the most challenging case.
I am only now beginning to understand the twistorial interpretation of any of
these.
It was a visit of Jean-Luc Brylinski to Warwick that got me interested in
the subject of gerbes, an interest that was later reinforced by conversations
with Dan Freed. Chapter 12 by Michael Murray recalls the enthusiasm that I
conveyed, though it was more in terms of the potential of the subject rather
than any theorems. I felt that at the right time and place I would see the
correct role of gerbes and that I should look out for that in my research. Some
years later, after I moved to Cambridge, I noticed that there was a twistor
transform analogous to the Atiyah–Ward approach to the self-dual Yang–Mills
equations and I had my student David Chatterjee write his thesis on various
aspects of gerbes, including this one. Murray’s chapter carefully leads the reader
through his bundle gerbe approach, connecting it with my own viewpoint. One of
Chatterjee’s contributions was to use the harmonic theory of the current defined
by a codimension three submanifold to define a gerbe with connection – the direct
 Geometry and physics 7

analogue of a divisor of a holomorphic line bundle. This distributional approach

is central to Chapter 13 by Harvey and Lawson on holomorphic linking numbers,
a concept which is also present in twistor theory.
In October 1994 I took up the Rouse Ball Chair in Cambridge, and started
talking to Nick Manton again about monopoles. He and Michael Murray had
been working on monopoles which were symmetric under the isometric action
of a finite group. I joined in, with the aim of using Nahm’s equations. It was
clear that the quotient of the spectral curve by the finite group was elliptic and
so one expected to solve the Nahm equations with elliptic functions. This we
could do (and learned some important lessons about nonsingularity in doing
so), but we also recruited Paul Sutcliffe, who was a postdoc at the time, to do
some numerical and analytical work to plot energy densities of the resulting
monopoles. Manton’s Chapter 14 discusses his other interest in the area –
skyrmions. His pictures of energy densities reveal the remarkable similarities
in the qualitative behaviour of symmetric skyrmions and monopoles – as if the
actual equations matter less than their ability to encompass the non-linearities
of the fields. I should say that this experience taught me to appreciate more the
problems of even asking for explicit solutions. Working on charge 3 monopoles
using genus two theta functions, it became clear that it was far more efficient to
numerically solve the Nahm equations than evaluate the constrained functions
analytically.
The geometry seminar in Cambridge was always held in the Department of
Applied Mathematics and Theoretical Physics (in particular because Stephen
Hawking would usually attend). It also meant that we would choose topics of
common interest to discuss and one of these was the Strominger–Yau–Zaslow
(SYZ) approach to mirror symmetry. I found this a very attractive idea. It led
me to write a couple of papers (Hitchin 1997, 1999) about moduli spaces of
special Lagrangian submanifolds and to learn from Dan Freed about special
Kähler geometry on moduli spaces. It also led to my first application of the
theory of gerbes in lectures I gave at a meeting in Harvard in 1999 (Hitchin
2001a). Chapter 15 by Leung and Yau discusses the consequences of the SYZ
approach for an elliptically fibred Calabi–Yau manifold, and the mirror symmetry
in Witten’s chapter is again of this type. Tamás Hausel also writes about this
approach in Chapter 16.
Tamás became my student in Cambridge at the time of yet another question
about monopoles. (It seemed as if this subject would shadow me for the rest
of my life.) These were the remarkable conjectures of the physicist Ashoke Sen
about the L2 harmonic forms on the moduli space. A lot of progress had been
made by Graeme Segal and his student Alex Selby on this, and I asked Tamás
to look at the corresponding question on the hyperkähler moduli space of Higgs
bundles. His results, obtained by a number of different methods, form part of his
chapter.
In 1997 I returned to Oxford as the Savilian Professor of Geometry. My
interests were still directed towards trying to understand the special Kähler
8 Geometry and physics

geometry on the moduli spaces coming from physics. The moduli space of
complex structures on a Calabi–Yau threefold led me to the study of invariant
functionals on differential forms (Hitchin 2001b). This seemed to give a natural
setting for the central objects studied in string theory and M-theory, but also
a practical one when considering the associated flow equations which describe
the geometry in terms of an evolving hypersurface geometry. Chapter 17 by
Robert Bryant addresses the initial value problems associated with such flows,
using the methods of exterior differential systems. My approach to Calabi–Yau
complex structures was based on the fact that the action of GL(6, R) on Λ3 R6
has an open dense orbit; G2 structures appeared from a similar remark about
the group GL(7, R). In both instances I learned a lot about the background to
these facts from Bryant. I began to notice the occurrence of this type of situation
in a number of rather different places, most notably after a lecture in London by
Merkulov on the holonomy of affine connections, and I began wondering about
a geometrical use of SO(6, 6) in the spin representation.
In the Spring of 2001 I spent a sabbatical term in Madrid, visiting Oscar
Garcı́a-Prada at the Universidad Autónoma. I went with the intention of finishing
a book on hyperkähler manifolds (which has still not been written) but I
got distracted by thinking again about the common features of SL(6, R) and
SO(6, 6). The first week in May seemed like a succession of public holidays in
Madrid and it rained all the time that year but at the end of it I had the concept
of a generalized complex structure – a hybrid object that interpolates between
symplectic and complex geometry and should, I thought, have something to
contribute to mirror symmetry. Moreover the role of two-forms seemed to parallel
the physicists’ B-field. Even more, the natural setting seemed to require a gerbe in
the background. Marco Gualtieri, who wrote his thesis on the subject, addresses
this in Chapter 18, discussing the objects within this generalized area which
replace holomorphic vector bundles and submanifolds, for which he adopts the
physicists’ term brane. There are still many questions unanswered in this theory,
not least a tantalizing connection with noncommutative geometry. In this area,
the twisted de Rham complex – replacing the exterior derivative d by d + H
where H is a closed three-form – is a natural process. The K-theoretic analogue
of this uses the gerbe idea much more closely, and Chapter 19 by Dan Freed,
a continuation of a series with Mike Hopkins and Constantin Teleman, focuses
again on the equivariant twisted K-theory of a Lie group. I have much to learn
about twisted K-theory and the cycles that represent it, but it clearly fits into
the framework of the generalized geometry around which much of my current
research revolves.
If any conclusion can be drawn from this personal timeline, it is that it is
specific problems that have engaged me in research projects, and that theoretical
physics is perhaps the richest source of these. I have generally sought to learn
about an area of mathematics by seeing how it impinges on a particular problem,
which by itself might seem unimportant (the ‘Miss Marple’ approach). So
progress this way requires an ever-changing source of mathematical challenges,
 References 9

of the sort that string theory seems to provide in abundance. This is surely
one of the reasons for the currently active interface between geometry and
physics.

Acknowledgements
I would like to wholeheartedly thank the contributors to this book, including
some who were unable to be at the Madrid conference. It is a wonderfully rich
cross-section of articles which appeal to my own interests. I would like to thank
Oxford University Press for agreeing to publish the book, and the editors for
their hard work in assembling it. Above all I would like to express my immense
thanks to Oscar Garcı́a-Prada for instigating the whole project – the original
2006 conference and this book based on it.

References
Donaldson, S. K. (1984). Nahm’s equations and the classification of monopoles. Com-
mun. Math. Phys. 96, 387–407.
Hitchin, N. J. (1974). Compact four-dimensional Einstein manifolds. J. Differential
Geometry 9, 435–41.
Hitchin, N. J. with Atiyah, M. F. and Singer, I. M. (1978). Self-duality in four-
dimensional Riemannian geometry. Proc. R. Soc. London Ser A 362, 425–61.
Hitchin, N. J. (1983). On the construction of monopoles. Commun. Math. Phys. 89,
145–90.
Hitchin, N. J. (1986). Metrics on moduli spaces, The Lefschetz centennial conference,
Part I (Mexico City, 1984). Contemp. Math. 58, American Mathematical Society,
Providence, RI, 157–78.
Hitchin, N. J. with Karlhede, A. Lindström, U. and Roček, M. (1987a). Hyperkähler
metrics and supersymmetry. Commun. Math. Phys. 108, 535–89.
Hitchin, N. J. (1987b). The self-duality equations on a Riemann surface. Proc. London
Math. Soc. 55(3), 59–126.
Hitchin, N. J. (1992). Lie groups and Teichmüller space. Topology 31, 449–73.
Hitchin, N. J. (1996). Poncelet polygons and the Painlevé equations, in Geometry and
Analysis. (Bombay, 1992), Oxford University Press, Bombay, India, pp. 151–85.
Hitchin, N. J. (1997). The moduli space of special Lagrangian submanifolds. Ann.
Scuola Norm. Sup. Pisa Cl. Sci. 25, 503–15.
Hitchin, N. J. (1998). Hypercomplex manifolds and the space of framings, in The
Geometric Universe: Science, Geometry and the Work of Roger Penrose. S. A.
Huggett et al. (eds.), Oxford University Press, Oxford, pp. 9–30.
Hitchin, N. J. (1999). The moduli space of complex Lagrangian submanifolds. Asian J.
Math. 3, 77–92.
Hitchin, N. J. (2001a). Lectures on special Lagrangian submanifolds, in Winter School
on Mirror Symmetry, Vector Bundles and Lagrangian Submanifolds. C. Vafa and
S. T. Yau (eds.), Studies in Advanced Mathematics 23, AMS/International Press,
Providence, RI, pp. 151–82.
Hitchin, N. J. (2001b). Stable forms and special metrics, in Global Differential Geom-
etry: The Mathematical Legacy of Alfred Gray. M. Fernández and J.A. Wolf (eds.),
10 Geometry and physics

Contemporary Mathematics 288, American Mathematical Society, Providence, RI,

pp. 70–89.
Nahm, W. (1982). The construction of all self-dual monopoles by the ADHM method,
in Monopoles in Quantum Field Theory (Trieste, 1981). World Scientific Publishing,
Singapore, pp. 87–94.
Welters, G. E. (1983). Polarized abelian varieties and the heat equations. Compositio
Math. 49, 173–94.
 II
MATHEMATICAL WORK OF NIGEL HITCHIN

Sir Michael Atiyah

Dedicated to Nigel Hitchin

It was a great pleasure for me to take part in the 60th birthday celebration
for Nigel Hitchin. I have known him ever since he was a graduate student in
Oxford and subsequently as my assistant at the institute in Princeton and then
as a colleague and collaborator. I have watched him mature mathematically
over the years and gradually establish a unique niche for himself on the frontier
between differential geometry and theoretical physics. This is now a very popular
and active field but Nigel occupies a singular place in it by virtue of the many
important and beautiful ideas he has introduced.
It has been said that the real measure of a mathematician’s contribution is
how long it would have taken the community to discover these results in his
absence. In Nigel’s case it is clear that he would score highly by this criterion,
since many of his results have been somewhat neglected on their first appearance
and their significance has only become apparent some years later.
Nigel has a large output with many beautiful papers which deserve to be read
and re-read. Because of the constraints of time I will limit myself to half-a-dozen
topics which I particularly like, and which show Nigel’s real knack of finding
natural and elegant ideas.
I will omit reference to the papers we have written jointly, except tangentially,
but I would like to mention our work on instantons which became known
as the ADHM construction (Atiyah et al. 1978). One morning, after a long
struggle, we finally saw the light. We adjourned to lunch in St Catherine’s College
euphoric at our success, though post-prandial analysis often punctures premature
celebration. On this occasion there was no unseen error but instead there was a
letter from Manin informing us that he and Drinfeld had just reached the same
conclusion! This was the genesis of the four-author paper and it was some years
later before Nigel or I met Drinfeld.
My first topic, which dates back to 1982, is Nigel’s introduction of the spectral
curve of a magnetic monopole (Hitchin 1982). I remember his letter to me at the
time saying he had found something which he felt I would like. He was completely
right. The result was really simple, beautiful, and important. It was at the basis
of much subsequent work on monopoles by many people, as I will recall.
Let me first review the details.
12 Mathematical work of Nigel Hitchin

The Bogomolny equations

DA φ = ∗FA

are equations in R3 for a gauge-field (or G-connection) A, where DA is the

covariant derivative, φ (the Higgs field) is a section of the adjoint bundle, FA is
the curvature, and ∗ is the Hodge dual. Under appropriate decay conditions at
infinity the solutions describe magnetic monopoles. Up to gauge equivalence the
solutions are parametrized by a finite-dimensional moduli space. The case most
studied is when G = SU (2). Nigel introduced the notion of a spectral line of a
monopole. This is by definition an oriented line L for which the linear differential
equation on L

(DA − iφ)s = 0

(for a section s of the C2 -bundle) has square-integrable solutions. Since solu-

tions s have exponential behaviour at ±∞ this means that we have a solution
exponentially decaying at both ends.
Nigel’s brilliant observation is that the spectral lines of any monopole are
parametrized by a compact Riemann surface in the complex surface which
parametrizes all oriented lines in R3 . Moreover, the spectral curve completely
determines the monopole and so is a very effective tool in the study of monopoles.
Algebraic curves had previously appeared in the theory of integrable systems,
such as the KdV equation, but these were equations in one space and one time
dimension. The novelty in Hitchin’s case is that the Bogomolny equation is a
PDE in three dimensions. Moreover, the spectral curve has a beautifully simple
geometric origin.
Among the many applications of the spectral curve I will single out just two:

1. The study of symmetric monopoles by Manton and collaborators (Manton

and Sutcliffe 2004). These correspond to symmetric spectral curves which
are easy to study.
2. The metric on the (relative) moduli space for magnetic charge 2. This metric
was explicitly described in (Atiyah and Hitchin 1988) and turned out to be
asymptotic to a Taub-NUT metric with negative mass-parameter. A purely
formal observation, this was many years later given an interesting M-theory
interpretation in terms of ‘anti-branes’ by Seiberg and Witten (1996).

My second example is the introduction (Hitchin 1987) by Nigel of ‘Higgs

bundles’ in 1987. For SU (2) these are solutions of the equations, on a compact
Riemann surface X,

FA + [φ, φ∗ ] = 0
dA φ = 0
 Mathematical work of Nigel Hitchin 13

where A is a connection on a rank 2 vector bundle E and φ is a section of

End (E) ⊗ K, with K the canonical line-bundle. This leads to a moduli space
M with an extremely rich structure. Here are some of its key properties:
1. M is hyperkähler of dimension 4(3g − 3) where g is the genus of X.
2. The determinant gives a map
M −→ H 0 (X, K 2 )
whose generic fibre is a complex torus (Prym variety).
3. M contains as a dense open set the cotangent bundle of stable SU (2)
bundles on X.
4. For all complex structures except two (±I) M can be identified with the
moduli space of flat SL(2, C) bundles on X.
5. M contains a copy of Teichmüller space given by pairs (A, φ) with E =
K 1/2 ⊕ K −1/2


0a
φ= , a ∈ H 0 (X, K 2 ).
10
When I first saw this large and meaty paper I expected that it would have
significant applications, but it appeared not to have had much impact at the time.
An exception is the use Witten (1991) made of it in his study of SL(2, C) Chern–
Simons theory. But, very recently, after nearly 20 years, the moduli space of Higgs
bundles plays a central role in the geometric Langlands program (Kapustin and
Witten 2007).
In this the theory on a Riemann surface X is obtained by dimensional reduc-
tion from a super-symmetric Yang–Mills theory on the four-manifold R2 × X.
A key fact is that this theory is a ‘twisted’ version of more physical theories,
corresponding to the twisting (tensoring with K) introduced by Nigel. Although
such twistings have become familiar over the past 20 years it appears that Nigel
was the first to introduce one (in the two-dimensional version).
One potential application of Nigel’s work would be the abelianization pro-
gramme put foward in Atiyah (1990) where the aim would be to identify the
space of ‘non-abelian theta-functions’ as the subspace of a space of ordinary
abelian theta-functions invariant under an appropriate discrete group.
My third example is the work of Nigel and his collaborators (Hitchin et al.
1987) on the hyperkähler quotient construction. This arose from the work
of physicists in super-symmetry, but it fits well into conventional symplectic
and complex geometry. What it does is to produce a hyperkähler manifold of
dimension 4(n−d) from the action of a compact Lie group G of dimension d on
a hyperkähler manifold of dimension 4n.
In particular one can start with a linear action of G on a quaternion vector
space H n . This produces a host of interesting hyperkähler manifolds as quotients.
Even better we can start with an infinite-dimensional linear space over H and an
action of an infinite-dimensional Lie group. Suitable examples occur naturally in
14 Mathematical work of Nigel Hitchin

the physics of gauge theories and the formal quotienting process can be rigorously
justified.
Cases where this programme has been successfully used include the work
of Kronheimer (1989) on ALE spaces, moduli spaces of instantons on R4 and
monopoles or Higgs bundles on R3 .
A more recent application has been (Atiyah and Bielawski 2002) the construc-
tion of hyperkähler metrics on manifolds associated with nilpotent conjugacy
classes in Lie groups. The sub-regular class leads to the four-dimensional man-
ifolds studied by Kronheimer (extending the work of Brieskorn), while other
classes lead to higher dimensional generalizations.
All of this has totally transformed the subject of hyperkähler manifolds.
While compact manifolds are rare, non-compact ones are numerous and of great
interest. Before this general theory was developed Nigel wrote an elegant little
paper (Hitchin 1979) which has always been one of my favourites. It deals with
the Gibbons–Hawking gravitational instantons (Gibbons and Hawking 1978).
These are the special cases of ALE manifolds associated to the An singularities,
that is, C2 /Γ where Γ is cyclic. What Nigel did was to show how the Gibbons–
Hawking metric could be simply derived by twistor methods. These connected
naturally to the algebraic geometry used by Brieskorn and the outcome is the
hyperkähler extension of the Brieskorn theory, which includes the metric as well
as the complex structure.
Two things should be noted about this paper of Nigel’s. First it eventually
led to Kronheimer’s complete treatment of all the ADE singularities. Second
it showed in an explicit way how twistor theory could be harnessed in such
problems. In fact twistor theory lies behind much of Nigel’s work and he was
one of the first to successfully add twistor theory to other techniques in the
differential geometric study of moduli spaces. This was illustrated again in our
joint work on monopoles (Atiyah and Hitchin 1988) as well, of course, as in the
better known work on instantons (Atiyah et al. 1978). It is interesting that a
new link between twistor theory and gauge theory has been initiated by Witten
(2004).
Many mathematicians reaching the age of 60 may be slowing down, at least
in terms of producing really novel ideas. This is not the case with Nigel, so let
me end with two examples of more recent work, each of which illustrates Nigel’s
knack of identifying fruitful new geometric ideas. In both cases physicists have
been quick to latch on to Nigel’s work and exploit it.
The first is his use of stable three forms in connection with special metrics
(Hitchin 2001). A three-form ρ on a linear space V is said to be stable if it is in
an open orbit of GL(V ). There are two interesting special cases:

1. dim V = 6 with stabilizer SL(3) × SL(3)

2. dim V = 7 with stabilizer G2

Both preserve a volume element φ(ρ).

 Mathematical work of Nigel Hitchin 15

If now M is a compact-oriented manifold of dimension 6 or 7 with an every-

where stable three-form ρ then the total volume

V (ρ) = φ(ρ)

is a functional of ρ. If we consider the variational problem for V (ρ) restricted to

closed forms ρ in a given cohomology class in H 3 (M, R), then critical points of
V (ρ) give, in the two cases:
1. dim 6 is a complex three-manifold with trivial canonical bundle (i.e. a
Calabi–Yau manifold).
2. dim 7 is a Riemannian manifold with holonomy G2 .
These are of interest to physicists for string theory in dimension 6 and for
M-theory in dimension 7.
One of the merits of Hitchin’s approach in dimension 6 is that explains why the
local moduli space is given by H 3 (M, R). The Hitchin functional in dimension 7
turns out to be related to 11-dimensional supergravity, which is the large volume
limit of M-theory.
My final example (Hitchin 2003), is Nigel’s introduction of the notion of
a generalized complex structure. This helps to unify symplectic and complex
geometry and should be helpful in understanding mirror symmetry.
The definition is quite simple. We consider a manifold X with tangent bundle
T X and a linear automorphism τ of
T X ⊕ T ∗X
with τ 2 = −1, and satisfying an integrability condition. Two special cases are
1. A complex structure J leads to a generalized complex structure with


J 0
τ=
0 −J ∗
2. A symplectic structure ω : T X −→ T ∗ X gives a generalized complex struc-
ture with


0 −ω−1
τ=
ω 0
Already this notion of generalized complex structure is being used by physicists
as in (Kapustin and Witten 2007), where it fits in very naturally.
If we look back over these examples of Nigel’s work we see that he always seems
able to identify some natural geometric concepts, which have a good motivation
and elegant definitions. It is not surprising that these lead to fruitful applications
in geometry and physics.
Finally, I should emphasize how clear and readable all Nigel’s papers are.
Based on simple novel ideas the papers almost seem to write themselves, but
this is the hallmark of real mathematics.
16 Mathematical work of Nigel Hitchin

References
Atiyah, M. F. (1990). The Geometry and Physics of Knots. Cambridge University Press.
Atiyah M. F. and Bielawski, R. (2002). Nahm’s equations, configuration spaces and
flag manifolds. Bull. Braz. Math. Soc., New Ser. 33(2), 157–76.
Atiyah, M. F., Drinfeld, V. G., Hitchin, N.J. and Manin, Yu. I. (1978). Construction
of instantons. Phys. Lett. A 65, 185–7.
Atiyah, M. F. and Hitchin, N. J. (1988). The Geometry and Dynamics of Magnetic
Monopoles. Princeton University Press.
Gibbons, G. W. and Hawking, S. W. (1978). Grativational multi-instantons. Phys.
Lett. B 78, 430.
Hitchin, N. J. (1979). Polygons and gravitons. Math. Proc. Camb. Phil. Soc. 85, 465–76.
Hitchin, N. J. (1982). Monopoles and geodesics. Commun. Math. Phys. 83, 579–602.
Hitchin, N. J. (1987). The self-duality equations on a Riemann surface. Proc. London
Math. Soc. 55(3), 59–126.
Hitchin, N. J. (2001). Stable forms and special metrics. Contemp. Math. 288, 70–89.
Hitchin, N. J. (2003). Generalized Calabi-Yau manifolds. Q. J. Math. 54, 281–308.
Hitchin, N. J., Karlhede, A., Lindström, U. and Roček, M. (1987). Hyperkähler metrics
and supersymmetry. Commun. Math. Phys. 108, 535–89.
Kapustin, A. and Witten, E. (2007). Electro-magnetic duality and the geometric
Langlands program, hep-th/0604151.
Kronheimer, P. B. (1989). The construction of ALE spaces as hyper-Kahler quotients.
J. Diff. Geom. 29, 665–83.
Manton, N. J. and Sutcliffe, P. M. (2004). Topological Solitons. Cambridge University
Press, Cambridge.
Seiberg, N. and Witten, E. (1996). Gauge Dynamics and Compactification to Three-
dimensions in The Mathematical Beauty of Physics: A Memorial Volume for Claude
Itzykson Saclay, France, June 5–7 (Advanced Series in Mathematical Physics, Vol.
24) by Claude Itzykson, Jean Bernard Zuber, and Jean-Michel Drouffe, pp. 333–66.
Witten, E. (1991). Quantization of Chern-Simons gauge theory with complex gauge
group. Commun. Math. Phys. 137, 29–66.
Witten, E. (2004). Perturbative gauge theory as a string theory in twistor space.
Commun. Math. Phys. 252, 189–258.
 III
THE EINSTEIN–MAXWELL EQUATIONS, EXTREMAL
KÄHLER METRICS, AND SEIBERG–WITTEN THEORY

Claude LeBrun
Dedicated to Nigel Hitchin on the occasion of his 60th birthday

Nigel Hitchin has played a key role in the exploration of four-dimensional

Riemannian geometry, and in particular has made foundational contributions
to the theory of self-dual manifolds, four-dimensional Einstein manifolds, spinc
structures, and Kähler geometry (Atiyah, Hitchin, and Singer 1978; Hitchin
1974a; 1974b; 1975; 1979; 1981; Hitchin, Karlhede, Lindström, and Roček 1987).
In the process, he has often alerted the rest of us to the profound mathematical
interest of beautiful geometric problems that had previously only been considered
by physicists. I would therefore like to use the occasion of Nigel’s 60th birthday as
an opportunity to draw the attention of an audience of geometers and physicists
to an interesting relationship between the four-dimensional Einstein–Maxwell
equations and Kähler geometry, and point out some fascinating open problems
that directly impinge on this relationship.
Let us begin by recalling that a two-form F on an oriented pseudo-Riemannian
n-manifold (M, g) is said to satisfy Maxwell’s equations (in vacuo) iff

dF = 0
dF = 0

where  is the Hodge star operator. If M is compact and g is Riemannian, these

equations of course just mean that F is a harmonic two-form, and Hodge theory
thus asserts that there is in fact exactly one solution in each de Rham cohomology
class [F ] ∈ H 2 (M, R). This solution may be found by minimizing the action

F −→ |F |2g dµg
M

among all closed forms F ∈ [F ]. In this context, dimension 4 enjoys a somewhat

privileged status, because it is precisely when n = 4 that both the action and
the solutions themselves become conformally invariant, in the sense that they
are unaltered by replacing g with any conformally related metric g̃ = u2 g.
When n > 2, coupling these equations to the gravitational field (Hawking
and Ellis 1973; Misner, Thorne, and Wheeler 1973) gives rise to the so-called
18 Einstein–Maxwell equations

Einstein–Maxwell equations (with cosmological constant)

dF = 0
dF = 0
 
r+F ◦F = 0
0

where r is the Ricci tensor of g, (F ◦ F )jk = Fj
F
k is obtained by composing

F with itself as an endomorphism of T M , and [ ]0 indicates the trace-free part
with respect to g. In the compact Riemannian setting, these equations may be
understood as the Euler–Lagrange equations of the functional

 
(g, F ) −→ sg + |F |2g dµg
M

where F is again allowed to vary over all closed two-forms in a given de Rham
class, and g is allowed to vary over all Riemannian metrics of some fixed total
volume V . The privileged status of dimension 4 becomes more pronounced in
this context, for it is only when n = 4 that the Einstein–Maxwell equations
imply that the scalar curvature s of g is constant. Indeed, this just reflects
Yamabe’s observation (Yamabe 1960) that a Riemannian metric has constant
scalar curvature
 iff it is a critical point of the restriction of the Einstein–Hilbert
action sdµ to the space of volumeV metrics in a fixed conformal class. When
n = 4, the conformal invariance of |F |2 dµ thus implies that critical points of
the above functional must have constant scalar curvature; but when n = 4, by
contrast, the scalar curvature turns out to be constant only when F has constant
norm.
We have just observed that the Einstein–Maxwell equations on a four-manifold
imply that the scalar curvature is constant. But what happens in the converse
direction is far more remarkable: namely, any constant-scalar-curvature Kähler
metric on a four-manifold may be interpreted as a solution of the Einstein–
Maxwell equations. Indeed, suppose that (M 4 , g, J) is a Kähler surface, with
Kähler form ω = g(J·, ·) and Ricci form ρ = r(J·, ·). Let
s
ρ̊ = ρ − ω
4
denote the primitive part of the Ricci form, corresponding to the trace-free Ricci
tensor
s
r̊ := [r]0 = r − g.
4
Suppose that the scalar curvature s of (M, g) is constant, and set
ρ̊
F =ω+ .
2
Then (g, F ) automatically solves the Einstein–Maxwell equations. This general-
izes an observation due to Flaherty (1978) concerning the scalar-flat case.
 Einstein–Maxwell equations 19

On a purely calculational level, this fact is certainly easy enough to check.

Indeed, on any oriented Riemannian four-manifold, the two-forms canonically
decompose
Λ2 = Λ+ ⊕ Λ−
into self-dual and anti-self-dual parts, and it is then easily shown that for any
two-form
F = F+ + F−
one has
 
F ◦F = 2F + ◦ F − .
0

Since in our special case we have F + = ω and F − = ρ̊/2, it therefore follows that
 
F ◦ F = −r̊,
0

so that
 
r+F ◦F =0
0

as required. Moreover, since ρ is automatically closed, and its self-dual part sω/4
is closed if s is assumed to be constant, we conclude that F is indeed harmonic
on a constant-scalar-curvature Kähler surface, exactly as claimed.
But there is actually a great deal more going on here. Recall that Calabi (1982)
defined an extremal Kähler metric on a compact complex manifold (M 2m , J) to
be a Kähler metric which is a critical point of the Riemannian functional

g −→ s2 dµ
M

restricted to a fixed Kähler class [ω] ∈ H 2 (M ). The associated Euler–Lagrange

equations then amount to requiring that the gradient ∇s of the scalar curvature
be the real part of a holomorphic vector field. In particular, any constant-scalar-
curvature Kähler metric is extremal in this sense. In fact, as was recently
 proved
by Chen (2006), extremal Kähler metrics actually always minimize s2 dµ in
their Kähler classes. In the constant-scalar-curvature case, this was long ago
shown by Calabi, using a simple but elegant argument. Indeed, if (M 2m , g, J) is
a compact Kähler manifold of complex dimension m, then
 
4π
sg dµg = ρ ∧ ω = c1 · [ω]m−1
M M (m − 1)!
and
 
ω ∧m 1
1dµ = = [ω]m
M M m! m!
20 Einstein–Maxwell equations

so that the Cauchy–Schwarz inequality tells us that

 
( sdµ)2 16π 2 m (c1 · [ω]m−1 )2
s dµ ≥ 
2
= (3.1)
M 1dµ (m − 1)! [ω]m
with equality iff s is constant.
Calabi (1982) also considered the Riemannian functionals

g −→ |r|2g dµg
M

and

g −→ |R|2g dµg
M
2
obtained by squaring the L norms of the Ricci curvature r and the full Riemann
curvature R. Here, his observation was that the restriction of either of these
functionals to the space of Kähler metrics can be rewritten in the form

g −→ a + b s2 dµ
M

where a and b > 0 are constants depending only on the Kähler class. For example,
 
1 8π2
|r|g dµg =
2
s2g dµg − c2 · [ω]m−2 ,
M 2 M (m − 2)! 1
so that

8π 2 m (c1 · [ω]m−1 )2
|r|2g dµg ≥ − c21 · [ω]m−2 , (3.2)
M (m − 2)! m − 1 [ω]m
with equality iff s is constant. Thus extremal Kähler metrics turn out to minimize
these functionals too.
I would now like to point out an interesting Riemannian analog of Calabi’s
variational problem that leads to the Einstein–Maxwell equations on a smooth
compact four-manifold. To this end, first notice that the Kähler form of a Kähler
surface is self-dual and harmonic. Let us therefore introduce the following notion:
Definition 3.1 Let M be smooth compact oriented four-manifold, and let [ω] ∈
H 2 (M, R) be a deRham class with [ω]2 > 0. We will then say that a Riemannian
metric g is adapted to [ω] if the harmonic form ω representing [ω] with respect
to g is self-dual.
This allows us to introduce the Riemannian analog of a Kähler class:
Definition 3.2 Let M be smooth compact-oriented four-manifold, and let [ω] ∈
H 2 (M, R) be a deRham class with [ω]2 > 0. We then set

G[ω] = smooth metrics g on M which are adapted to [ω] .

 Einstein–Maxwell equations 21

In particular, if ω is the Kähler form of a metric g on M which is Kähler with

respect to some complex structure J, then G[ω] contains the entire Kähler class
of ω on (M, J). Of course, however, G[ω] is vastly larger than a Kähler class. In
particular, if g belongs to G[ω] , so does every conformally related metric g̃ = u2 g.
It is also worth noticing that
Gλ[ω] = G[ω]
for any non-zero real constant λ.
It is now germane to ask precisely how large G[ω] really is relative to

G = smooth metrics g on M ,

and to ponder the dependence of G[ω] ⊂ G on [ω] as we allow this cohomology

class to vary through the open cone

C = [ω] ∈ H 2 (M, R) [ω]2 > 0 .

Proposition 3.1 (Donaldson/Gay–Kirby) Let M be any smooth compact

four-manifold with b+ (M ) = 0. For any [ω] ∈ C, G[ω] ⊂ G is a Fréchet submani-
fold of finite codimension b− (M ). Moreover, G[ω] = ∅ for all [ω] belonging to an
open dense subset of C.
Indeed, if g ∈ G[ω] and if ω ∈ [ω] is the harmonic representative, then
Donaldson (1986, p. 336) has shown that Tg G[ω] is precisely the L2 orthogonal
of the b− (M )-dimensional subspace
{ω ◦ ϕ | ϕ ∈ Hg− } ⊂ Γ(2 T ∗ M ),
where Hg− is the space of anti-self-dual harmonic two-forms with respect to g;
moreover, his proof also shows that the subset of [ω] ∈ C for which G[ω] = ∅ is
necessarily open. On the other hand, Gay and Kirby (2004) found an essentially
explicit way of constructing a metric g adapted to any [ω] ∈ C ∩ H 2 (M, Z), so
that G[ω] = ∅ for any [ω] in the dense subset [C ∩ H 2 (M, Q)] ⊂ C.
We now consider the natural generalization of Calabi’s variational problem to
this broader context.
Proposition 3.2 An [ω]-adapted metric g is a critical point of the Riemannian
functional

g −→ s2g dµg
M

restricted to G[ω] iff either

r g is a solution of the Einstein–Maxwell equations, in conjunction with a
unique harmonic form F with F + = ω; or else
r g is scalar-flat (s ≡ 0).
22 Einstein–Maxwell equations

Proof. Consider a one-parameter family of metrics

gt := g + th + O(t2 )
in G[ω] . By Donaldson’s result, we know that h can be taken to be any smooth
symmetric tensor field which satisfies

h, ω ◦ ϕdµ = 0
M

for all harmonic forms ϕ ∈ Γ(Λ− ), where ω is the g-harmonic representative of

[ω]. On the other hand, a standard calculation Besse (1987) shows that
d
s = ∆haa + ∇a ∇b hab − hab rab ,
dt t=0

and
d 1 a
[dµ] = h dµ,
dt t=0 2 a
so that
  
d 2 ˙
s dµ = 2sṡdµ + s2 dµ
dt M t=0 M M

 
= 2s ∆haa + ∇a ∇b hab − hab r̊ab dµ

where r̊ again denotes the trace-free part of the Ricci tensor.

Let us now ask when a metric is critical within its conformal class. This
corresponds to setting h = vg for some smooth function v. We then have
  
d
s2 dµ = 2s(3∆v)dµ = 6 ds, dvdµ,
dt M M

so the derivative is zero for all such variations iff s is constant.

We may thus assume henceforth that s is constant. We then have
 
d
s dµ = 2s (∆haa + ∇a ∇b hab − hab r̊ab )dµ
2
dt M

= −2s h,r̊dµ.
M

If s ≡ 0, this obviously vanishes for every h, and g is a critical point. Otherwise,

g will be critical iff r̊ belongs to the L2 orthogonal complement of Tg G[ω] . But we
already have seen that this orthogonal complement precisely consists of tensors
of the form ω ◦ ϕ, ϕ ∈ Hg− . Thus, when s ≡ 0, g is a critical point iff s is constant
and r̊ = ω ◦ ϕ for some ϕ ∈ Hg− . But, setting
ϕ
F =ω+ ,
2
 Einstein–Maxwell equations 23

this is in turn equivalent to saying that (g, F ) solves the Einstein–Maxwell

equations, as claimed.

 2
So, why are constant-scalar-curvature Kähler metrics critical points of s dµ
restricted to G[ω] ? Well, we will now see that they typically turn out not only
to be critical points, but actually to be minima. Indeed, the following result
(LeBrun 1995, 2001, 2003) may be thought of as a Riemannian generalization of
Calabi’s inequalities (3.1 and 3.2):

Theorem 3.1 Let (M 4 , J) be a compact complex surface, and suppose that [ω]
is a Kähler class with c1 · [ω] ≤ 0. Then any Riemannian metric g ∈ G[ω] satisfies
the inequalities

(c1 · [ω])2
s2 dµ ≥ 32π 2 (3.3)
[ω]2
  (c · [ω])2 
1
|r|2 dµ ≥ 8π 2 2 − c2
1 (3.4)
[ω]2
with equality iff g is constant-scalar-curvature Kähler.

In the equality case, the complex structure J˜ with respect to which g is Kähler
will typically be different from J, but must have the same first Chern class c1 ,
while its Kähler class must be a positive multiple of [ω].
We also remark that if (M, J) is not rational or ruled, the hypothesis that
c1 · [ω] ≤ 0 holds automatically, and that in this setting a Kähler metric is
extremal iff it has constant scalar curvature. In this context, the relevant constant
is of course necessarily non-positive.
By contrast, if (M, J) is rational or ruled, there will always be Kähler classes
for which c1 · [ω] > 0. When this happens, the above generalization (3.3) of (3.1)
turns out definitely not to hold for arbitrary Riemannian metrics. Instead, the
correct generalization (LeBrun 1997) is that
4π c1 · [ω]
Y[g] ≤  , (3.5)
[ω]2 /2
where the Yamabe constant Y[g] is obtained by minimizing the Einstein–Hilbert
action sdµ over all unit-volume metrics g̃ = u2 g conformal to g. Moreover, the
inequality is strict unless the Yamabe minimizer is a constant-scalar-curvature
Kähler metric, so that (3.3) is in fact violated by an appropriate conformal
rescaling of any generic Riemannian metric of positive scalar curvature.
It is also worth remarking that no sharp lower
 bound in the spirit of Theorem
3.1 is currently known for the square-norm |R|2 dµ of the Riemann curvature
tensor. Deriving one would be extremely interesting and potentially very useful,
but, for reasons I will now explain, the technical obstacles to doing so seem
formidable.
24 Einstein–Maxwell equations

Recall that, by raising an index, the Riemann curvature tensor may be

reinterpreted as a linear map Λ2 → Λ2 , called the curvature operator. The decom-
position Λ2 = Λ+ ⊕ Λ− thus allows one to view this linear map as consisting of
four blocks:
⎛ ⎞
⎜ W+ + s
r̊ ⎟
⎜ 12 ⎟
⎜ ⎟
R=⎜
⎜
⎟.
⎟ (3.6)
⎜ ⎟
⎝ r̊ W− + s ⎠
12

Here W± are the trace-free pieces of the appropriate blocks, and are called the
self-dual and anti-self-dual Weyl curvatures, respectively. The scalar curvature
s is understood to act by scalar multiplication, whereas the trace-free Ricci
curvature r̊ = r − 4s g acts on two-forms by ϕab → 2ϕ[a c r̊b]c .
When (M, g) happens to be Kähler, Λ2,0 ⊂ ker R, and the entire upper-left-
hand block is therefore entirely determined by the scalar curvature s. For Kähler
metrics, one thus obtains the identity
s2
|W+ |2 ≡ ,
24
and Gauss–Bonnet-type formulae like

2 
1 s |r̊|2
(2χ + 3τ )(M ) = + 2|W+ | −
2
dµ
4π 2 M 24 2
reduce many questions about square norms of curvature to questions about the
scalar curvature alone. But for general Riemannian metrics, the norms of s and
W+ are utterly independent quantities, so if one wants to use the identity
 
2 
s 1
|r|2 dµ = −8π 2 (2χ + 3τ )(M ) + 8 + |W+ |2 dµ (3.7)
24 2
to prove a generalization of (3.2) for Riemannian metrics, information must be
obtained concerning not only the scalar curvature, but also concerning the self-
dual Weyl curvature as well.
The curvature estimates of Theorem 3.1 are derived by means of Seiberg–
Witten theory (Witten 1994), making it clear that this really is an essentially
four-dimensional story. The complex structure J determines a spinc structure on
M with twisted spin bundles S± ⊗ L1/2 , where L−1 is the canonical line bundle
Λ2,0 of (M, J). For simplicity, suppose that c1 · [ω] < 0. For each metric g ∈ G[ω] ,
one then considers the Seiberg–Witten equations
DA Φ = 0
1
FA+ = − Φ  Φ̄
2
 Einstein–Maxwell equations 25

where the unknowns are a unitary connection A on the line-bundle L → M

and a twisted spinor Φ ∈ Γ(S+ ⊗ L1/2 ); here DA : Γ(S+ ⊗ L1/2 ) → Γ(S− ⊗ L1/2 )
denotes the twisted Dirac operator associated with A, and FA+ is the self-dual
part of the curvature of A. One then shows that there must be at least one
solution for each g ∈ G[ω] by establishing a count of solutions modulo gauge
equivalence which is independent of the metric and which is obviously non-zero
for a Kähler metric.
However, the Seiberg–Witten equations can be shown to imply various curva-
ture estimates via Weitzenböck formulae. In particular, the existence of at least
one solution for each metric g̃ = u2 g conformal to g is enough to guarantee that
the curvature of g satisfies

s2 dµg ≥ 32π 2 [c1 (L)+ ]2
M
 √
(s − 6|W+ |)2 dµg ≥ 72π 2 [c1 (L)+ ]2
M

where [c1 (L)] is the orthogonal projection of c1 (L) ∈ H 2 (M, R) = Hg+ ⊕ Hg−
+

into the space Hg+ of harmonic self-dual two-forms, defined with respect to g.
Since ω is assumed to be self-dual with respect to g, we therefore have
(c1 · [ω])2
[c1 (L)+ ]2 ≥
[ω]2
by the Cauchy–Schwarz inequality. Inequality (3.3) follows. Since yet another
Cauchy–Schwarz argument shows that

2   √
s 1 1
+ |W+ | dµg ≥
2
(s − 6|W+ |)2 dµg ,
24 2 36
the second inequality and (3.7) together imply (3.4). The fact that only Kähler
metrics can saturate (3.3) or (3.4) is then deduced by examining the relevant
Weitzenböck formulae.
One might be tempted to expect the story to be similar for the norm of the
full Riemann tensor. After all, the identity
 
2 
s
|R| dµ = −8π (χ + 3τ )(M ) + 2
2 2
+ 2|W+ | dµg
2
M M 24
certainly provides a good analog of (3.7). In the Kähler case, one has
s2
= |W+ |2 ,
24
so this simplifies to become
 
1
|R|2 dµ = 8π 2 (c2 − c21 ) + s2 dµg ,
M 4 M
26 Einstein–Maxwell equations

and applying (3.1) we therefore obtain Calabi’s inequality


(c1 · [ω])2
|R|2 dµ ≥ 8π 2 + c2 − c21 (3.8)
M [ω]2
for any Kähler metric. In light of Theorem 3.1, it might therefore seem reasonable
to hope that one could simply extend this inequality to general Riemannian
metrics by means of Seiberg–Witten theory. However, we will now show that
this cannot work. The key idea is to examine certain extremal Kähler metrics
from the vantage point of their reversed orientations.

Theorem 3.2 Calabi’s inequality (3.8) cannot possibly be extended to general

Riemannian metrics by means of Seiberg–Witten theory. Indeed, there actually
exist smooth compact-oriented Riemannian four-manifolds (M, g) which admit a
spinc structure of almost-complex type with non-zero Seiberg–Witten invariant,
but such that

(c1 · [ω])2
|R|2 dµ < 8π 2 − (χ + 3τ )(M )
M [ω]2
for some self-dual harmonic two-form ω on (M, g).

Proof. A Kodaira-fibered complex surface is by definition a compact complex

surface X equipped with a holomorphic submersion : X → B onto a compact
complex curve, such that the base B and fiber Fz = −1 (z) both have genus
≥ 2. The product B × F of two complex curves of genus ≥ 2 is certainly Kodaira
fibered, but such a product signature τ = 0. However, one can also construct
examples (Atiyah 1969; Kodaira 1967) with τ > 0 by taking branched covers of
products.
Let X be any such Kodaira-fibered surface with τ (X) > 0, and let : X → B
be its Kodaira fibration. Let p denote the the genus of the base B, and let q
denote the genus of some fiber F of . A beautiful result of Fine 2004 then
asserts that X actually admits a family of extremal Kähler metrics; namely, for
any sufficiently small > 0,

[ω ] = 2(p − 1)F − c1

is a Kähler class on X which is represented by a Kähler metric g of constant

scalar curvature.
These metrics, being Kähler, have total scalar curvature

sg
dµg
= 4πc1 · [ω ] = −4π(χ + c21 )(X)

and total volume


[ω ]2
dµg
= = (2χ + c21 )(X).
2 2
 Einstein–Maxwell equations 27

Since sg
is constant, it follows that


32π 2 (χ + c21 )2
s2g
dµg
= .
2χ + c21
These metrics therefore satisfy
 
1
|R|g
dµg
= 8π (c2 − c1 ) +
2 2 2
s2 dµg
X 4 X

(χ + c21 )2
= 8π 2 −(χ + 3τ )(X) +
(2χ + c21 )
On the other hand, there are symplectic forms on X which are compatible with
the non-standard orientation of X; for example, the cohomology class F + εc1 is
represented by such forms if ε is sufficiently small. A celebrated theorem of
Taubes (1994) therefore tells us that the reverse-oriented version M = X of X
has a non-trivial Seiberg–Witten invariant (Kotschick 1998; Leung 1996). The
relevant spinc structure on X is of almost-complex type, and its first Chern class,
which we will denote by c̄1 , is given by

c̄1 = c1 + 4(p − 1)F.

Since this a (1, 1) class, one has

c̄1 · [ω ] (χ + 3 τ )
(c̄1 )+ = ω = − ω ,
[ω ]2 [ω ]2

relative to the Kähler metric g , so that

(χ + 3 τ )2 (χ + 3 τ )2
|(c̄1 )+ |2 = = .
[ω ] 2 (2χ + c21 )

Now since c̄1 arises from an almost-complex structure on X, we have

|(c̄1 )− |2 − |(c̄1 )+ |2 = 2χ − 3τ,

so that
(χ + 3 τ )2
|(c̄1 )− |2 = 2χ − 3τ + ,
(2χ + c21 )
and
(χ + 3 τ )2
|(c̄1 )− |2 − (χ − 3τ )(X) = χ(X) + ,
(2χ + c21 )
28 Einstein–Maxwell equations

where, for example, τ indicates τ (X). But it therefore follows that

  
1 − 2
|R|2
g
dµg
− |(c̄1 ) | − (χ − 3τ ) = −(2χ + 3τ )
8π 2 X
(χ + c21 )2 − (χ + 3 τ )2
+
(2χ + c21 )
4χ (χ + 3 τ ) + 4χ2 2
= −(2χ + 3τ ) +
(2χ + c21 )
2χ + c21 + 3 τ
= −(2χ + 3τ ) + 2χ
2χ + c21
2χ
= −3τ (X) 1 − ,
2χ + c21
which is negative for any sufficiently small . The result therefore follows once we
take “ω” to be the anti-self-dual harmonic form (c̄1 )− , which becomes self-dual
on M = X.

Similarly, careful examination of these examples also shows that, for any
constant t > 1, the Seiberg–Witten equations cannot imply an estimate of
 √
(s − t 6|W+ |)2 dµ

which is saturated by constant-scalar-curvature Kähler metrics. Of course, the

Seiberg–Witten equations still imply lower bounds for such quantities, but they
are simply never as sharp as those obtained for t ∈ [0, 1].
In this chapter, we have seen that constant-scalar-curvature Kähler metrics
occupy a privileged position in four-dimensional Riemannian geometry. I would
therefore like to conclude this discussion by indicating a bit of what we now
know concerning their existence.
There are several ways to phrase the problem. From the Riemannian point
of view, one might want to fix a smooth compact-oriented four-manifold M ,
and simply ask whether there exists an extremal Kähler metric g, where the
associated complex structure is not specified as part of the problem. Since M
must in particular admit a Kähler metric, two necessary conditions are that M
must admit a complex structure and have even first Betti number. Provided
these desiderata are fulfilled, Shu (2006) has then shown that an extremal
metric g always exists. For all but two diffeotypes, moreover, one can actually
arrange for the extremal Kähler metric g to have constant scalar curvature.
However, these two exceptional diffeotypes are CP2 #CP2 , and CP2 #2CP2 , and
it is now known (Chen, LeBrun and Weber 2008) that both these manifolds carry
Einstein metrics—indeed, even Einstein metrics which are conformal rescalings
of extremal Kähler metrics! Since, in conjunction with F = 0, any Einstein metric
of course satisfies the Einstein–Maxwell equations, we thus immediately deduce
the following:
 Einstein–Maxwell equations 29

Theorem 3.3 Let M be the underlying four-manifold of any compact complex

surface of Kähler type. Then M admits a Riemannian solution (g, F ) of the
Einstein–Maxwell equations.

While the above formulation of Shu’s result certainly suffices to imply

Theorem 3.3, it unfortunately also obfuscates the nature of the proof, which
involves constructing extremal Kähler metrics compatible with some fixed com-
plex structure in each possible deformation class. The key tool used for this
purpose is due to Arezzo and Pacard (2006, 2009) who have shown that
constant-scalar-curvature Kähler metrics can be constructed on blow-ups and
desingularizations of constant-scalar-curvature Kähler orbifolds, under only very
mild assumptions on the complex automorphism group; similar results moreover
have even been proved concerning the strictly extremal case (Arezzo, Pac-
ard, and Singer 2007). These gluing results represents a vast generalization
of earlier work by the present author and his collaborators (Kim, LeBrun
and Pontecorvo 1997; LeBrun 1991; LeBrun and Singer 1993) regarding the
limited realm of scalar-flat Kähler surfaces. In fact, by invoking the theory of
Kähler–Einstein metrics (Aubin 1976; Yau 1977), Arezzo and Pacard (2006)
had already shown that every Kähler-type complex surface of Kodaira dimen-
sion 0 or 2 admits compatible constant-scalar-curvature Kähler metrics. Shu’s
results concerning the remaining cases of Kodaira dimensions −∞ and 1 are
much less robust, but still easily produce enough examples to imply most of
Theorem 3.3.
Of course, one ultimately does not want to settle for mere existence statements;
we would really like to completely understand the moduli space of solutions!
From this point of view, the first question to ask is whether there can only be
one solution for any given complex structure and Kähler class. Modulo complex
automorphisms’ uniqueness always holds in this setting, as was proved in a series
of a fundamental papers by Donaldson (2001), Mabuchi (2004), and Chen and
Tian (2005). For a fixed complex structure, one also knows that the Kähler classes
of extremal Kähler metrics sweep out an open subset of the Kähler cone (LeBrun
and Simanca 1993), and somewhat weaker results are also available regarding
deformations of complex structure (Fujiki and Schumacher 1990; LeBrun and
Simanca 1994). However, it turns out that the set of Kähler classes which are
representable by extremal Kähler metrics may sometimes be a proper non-empty
open subset of the Kähler cone (Apostolov and Tønnesen-Friedman 2006; Ross
2006). The latter phenomenon is related to algebro-geometric stability problems
(Mabuchi 2005; Ross and Thomas 2006) in a manner which is still only partly
understood, but there is reason to hope that a definitive understanding of such
issues may result from the incredible ferment of research currently being carried
out the field.
As we saw in Theorem 3.3, Kähler geometry supplies a natural and beautiful
way of constructing solutions of the Einstein–Maxwell equations on many com-
pact four-manifolds. In the opposite direction, we also have the following easy
but rather suggestive result:
30 Einstein–Maxwell equations

Proposition 3.3 Let M be the underlying four-manifold of any compact

complex surface of non-Kähler type with vanishing geometric genus. Then M
does not carry any Riemannian solution of the Einstein–Maxwell equations.

Proof. Let us begin by remembering the remarkable fact (Barth, Peters, and
Van de Ven 1984; Buchdahl 1999; Siu 1983) that a compact complex surface is
of Kähler type iff it has b1 even. Consequently, for any non-Kähler-type complex
surface M , b1 is odd, and b+ = 2pg , where pg = h2,0 is the geometric genus
(Barth, Peters, and Van de Ven, 1984, Theorem IV.2.6). Since the latter is
assumed to vanish, M then has negative-definite intersection form, and Hodge
theory tells us that there are no non-trivial self-dual harmonic two-forms for any
metric g on M .
Now suppose that (g, F ) is a Riemannian solution of the Einstein–Maxwell
equations on M . Then the harmonic two-form F satisfies F + = 0, and hence

r̊ = −[F ◦ F ]0 = −2F + ◦ F − = 0.

The metric g must therefore be Einstein. But the negative-definiteness of the

intersection form also tells us that (2χ + 3τ )(M ) = c21 (M ) ≤ 0. The Hitchin–
Thorpe inequality for Einstein manifolds (Hitchin 1974b) therefore guarantees
that M has a finite normal cover M̃ which is diffeomorphic to either K3 or
T 4 , and so, in particular, has b1 even. Pulling back the complex structure J
of M to this cover, we therefore obtain a complex surface (M̃ , J) ˜ of Kähler
˜
type. Averaging an arbitrary Kähler metric h on (M̃ , J) over the finite group of
deck transformation of M̃ → M then gives us a Kähler metric which descends
to (M, J). Thus (M, J) is of Kähler type, in contradiction to our hypotheses.
Our supposition was therefore false, and M thus cannot carry any Riemannian
solution of the Einstein–Maxwell equations.

This chapter has endeavored to convince the reader that the four-dimensional
Einstein–Maxwell equations represent a beautiful and natural generalization of
the constant-scalar-curvature Kähler condition. However, it still remains to be
seen whether solutions exist on many compact four-manifolds other than complex
surfaces of Kähler type. In this direction, my guess is that Proposition 3.3 will
actually prove to be rather misleading. For example, there are ALE Riemannian
solutions of the Einstein–Maxwell equations, constructed (Yuille 1987) via the
Israel–Wilson ansatz, on manifolds very unlike any complex surface. It thus
seems reasonable to conjecture that there are plenty of compact solutions that
in no sense arise from Kähler geometry. Perhaps some interested reader will feel
inspired to go out and find some!

Acknowledgments
The author would like to thank Maciej Dunajski, Caner Koca, and Galliano
Valent for their interesting comments and queries regarding an earlier version
 References 31

of this chapter, as some of these resulted in important improvements to this

chapter. This chapter was supported in part by NSF grant DMS-0604735.

References
Apostolov, V. and Tønnesen-Friedman, C. (2006). A remark on Kähler metrics of
constant scalar curvature on ruled complex surfaces. Bull. London Math. Soc. 38(3),
494–500.
Arezzo, C. and Pacard, F. (2006). Blowing up and desingularizing constant scalar
curvature Kähler manifolds. Acta Math. 196(2), 179–228.
Arezzo, C. and Pacard, F. (2009). Blowing up and desingularizing Kähler orbifolds of
constant scalar curvature II. Ann. Math. 170, 685–738. E-print math.DG/0504115.
Arezzo, C., Pacard, F. and Singer, M. (2007). Extremal metrics on blow ups. E-print
math.DG/0701028.
Atiyah, M. F. (1969). The signature of fibre-bundles, in Global Analysis (Papers in
Honor of K. Kodaira), University Tokyo Press, Tokyo, pp. 73–84.
Atiyah, M. F., Hitchin, N. J. and Singer, I. M. (1978). Self-duality in four-dimensional
Riemannian geometry. Proc. R. Soc. London Ser. A 362, 425–61.
Aubin, T. (1976). Equations du type Monge-Ampère sur les variétés kählériennes
compactes. C. R. Acad. Sci. Paris A 283, 119–21.
Barth, W., Peters, C. and Van de Ven, A. (1984). Compact Complex Surfaces. Springer-
Verlag, Berlin, Heidelberg, New York, Tokyo.
Besse, A. L. (1987). Einstein manifolds, Vol. 10 of Ergebnisse der Mathematik und ihrer
Grenzgebiete (3) [Results in Mathematics and Related Areas (3)], Springer-Verlag,
Berlin.
Buchdahl, N. (1999). On compact Kähler surfaces. Ann. Inst. Fourier (Grenoble)
49(1), 287–302.
Calabi, E. (1982). Extremal Kähler metrics, in Seminar on Differential Geometry, Vol.
102 of Ann. of Math. Stud., Princeton University Press, Princeton, NJ, pp. 259–90.
Chen, X. X. (2006). Space of Kähler metrics III on the lower bound of the Calabi energy
and geodesic distance. E-print math.DG/0606228.
Chen, X. X., LeBrun, C. and Weber, B. (2008). On conformally Kähler, Einstein
manifolds. J. Am. Math. Soc. 21(4), 1137–1168. E-print arXiv:0705.0710.
Chen, X. X. and Tian, G. (2005). Uniqueness of extremal Kähler metrics. C. R. Math.
Acad. Sci. Paris 340(4), 287–90.
Donaldson, S. K. (1986). Connections, cohomology and the intersection forms of 4-
manifolds. J. Differential Geom. 24(3), 275–341.
Donaldson, S. K. (2001). Scalar curvature and projective embeddings. I. J. Differential
Geom. 59(3), 479–522.
Fine, J. (2004). Constant scalar curvature Kähler metrics on fibred complex surfaces.
J. Differential Geom. 68(3), 397–432.
Flaherty, E. J., Jr. (1978). The nonlinear graviton in interaction with a photon. Gen.
Relativity Gravit. 9(11), 961–78.
Fujiki, A. and Schumacher, G. (1990). The moduli space of extremal compact Kähler
manifolds and generalized Weil-Petersson metrics. Publ. Res. Inst. Math. Sci.
26(1), 101–83.
Gay, D. T. and Kirby, R. (2004). Constructing symplectic forms on 4-manifolds which
vanish on circles. Geom. Topol. 8, 743–77 (electronic).
32 Einstein–Maxwell equations

Hawking, S. W. and Ellis, G. F. R. (1973). The Large Scale Structure of Space-Time,

Cambridge University Press, London.
Hitchin, N. J. (1974a). Harmonic spinors. Adv. Math. 14, 1–55.
Hitchin, N. J. (1974b). On compact four-dimensional Einstein manifolds. J. Differential
Geom. 9, 435–42.
Hitchin, N. J. (1975). On the curvature of rational surfaces, in Differential Geometry
(Proceedings of the Symposium on Pure Mathematics, Vol. XXVII, Part 2, Stanford
University, Stanford, CA, 1973), American Mathematical Society, Providence, RI,
pp. 65–80.
Hitchin, N. J. (1979). Polygons and gravitons. Math. Proc. Cambridge Philos. Soc.
85(3), 465–76.
Hitchin, N. J. (1981). Kählerian twistor spaces. Proc. London Math. Soc. (3) 43(1),
133–50.
Hitchin, N. J., Karlhede, A., Lindström, U. and Roček, M. (1987). Hyper-Kähler metrics
and supersymmetry. Commun. Math. Phys. 108(4), 535–89.
Kim, J., LeBrun, C. and Pontecorvo, M. (1997). Scalar-flat Kähler surfaces of all genera.
J. Reine Angew. Math. 486, 69–95.
Kodaira, K. (1967). A certain type of irregular algebraic surfaces. J. Analyse Math.
19, 207–15.
Kotschick, D. (1998). Signatures, monopoles and mapping class groups. Math. Res.
Lett. 5(1–2), 227–34.
LeBrun, C. (1991). Scalar-flat Kähler metrics on blown-up ruled surfaces. J. Reine
Angew. Math. 420, 161–77.
LeBrun, C. (1995). Polarized 4-manifolds, extremal Kähler metrics, and Seiberg-Witten
theory. Math. Res. Lett. 2, 653–62.
LeBrun, C. (1997). Yamabe constants and the perturbed Seiberg-Witten equations.
Commun. Anal. Geom. 5, 535–53.
LeBrun, C. (2001). Ricci curvature, minimal volumes, and Seiberg-Witten theory.
Invent. Math. 145, 279–316.
LeBrun, C. (2003). Einstein metrics, four-manifolds, and differential topology, in Sur-
veys in Differential Geometry, Vol. VIII (Boston, MA, 2002), Surveys in Differential
Geometry, VIII, International Press, Somerville, MA, pp. 235–55.
LeBrun, C. and Simanca, S. R. (1993). On the Kähler classes of extremal metrics, in
Geometry and Global Analysis (Sendai, Japan, 1993), Tohoku University, Sendai,
Japan, pp. 255–71.
LeBrun, C. and Simanca, S. R. (1994). Extremal Kähler metrics and complex defor-
mation theory. Geom. Funct. Anal. 4(3), 298–336.
LeBrun, C. and Singer, M. (1993). Existence and deformation theory for scalar-
flat Kähler metrics on compact complex surfaces. Invent. Math. 112(2),
273–313.
Leung, N. C. (1996). Seiberg Witten invariants and uniformizations. Math. Ann.
306(1), 31–46.
Mabuchi, T. (2004). Uniqueness of extremal Kähler metrics for an integral Kähler class.
Int. J. Math. 15(6), 531–46.
Mabuchi, T. (2005). An energy-theoretic approach to the Hitchin-Kobayashi correspon-
dence for manifolds. I. Invent. Math. 159(2), 225–43.
Misner, C. W., Thorne, K. S. and Wheeler, J. A. (1973). Gravitation, W. H. Freeman
and Co., San Francisco, CA.
 References 33

Ross, J. (2006). Unstable products of smooth curves. Invent. Math. 165(1), 153–62.
Ross, J. and Thomas, R. (2006). An obstruction to the existence of constant scalar
curvature Kähler metrics. J. Differential Geom. 72(3), 429–66.
Shu, Y. (2006). Compact complex surfaces and constant scalar curvature Kähler
metrics. E-print math.DG/0612013.
Siu, Y. (1983). Every K3 surface is Kähler. Invent. Math. 73, 139–50.
Taubes, C. H. (1994). The Seiberg-Witten invariants and symplectic forms. Math. Res.
Lett. 1, 809–22.
Witten, E. (1994). Monopoles and four-manifolds. Math. Res. Lett. 1, 809–22.
Yamabe, H. (1960). On a deformation of Riemannian structures on compact manifolds.
Osaka Math. J. 12, 21–37.
Yau, S. T. (1977). Calabi’s conjecture and some new results in algebraic geometry.
Proc. Natl. Acad. USA 74, 1789–99.
Yuille, A. L. (1987). Israel-Wilson metrics in the Euclidean regime. Classical Quantum
Gravity 4(5), 1409–26.
 IV
THE NAHM TRANSFORM FOR CALORONS

Benoit Charbonneau and Jacques Hurtubise

Dedicated to Nigel Hitchin on the occasion of his 60th birthday

4.1 Introduction
One rather mysterious feature of the self-duality equations on R4 is the exis-
tence of a quite remarkable non-linear transform, the Nahm transform. It maps
solutions to the self-duality equations on R4 invariant under a closed translation
group G ⊂ R4 to solutions to the self-duality equations on (R4 )∗ invariant under
the dual group G∗ . This transform uses spaces of solutions to the Dirac equation,
it is quite sensitive to boundary conditions, which must be defined with care,
and it is not straightforward: for example, it tends to interchange rank and
degree.
The transform was introduced by Nahm as an adaptation of the original
ADHM construction of instantons (Atiyah et al. 1978) having the advantage that
it can be generalized. The series of papers (Corrigan and Goddard 1984; Nahm
1983, 1984) details Nahm’s original transform. In the monopole case, G = R,
and the transform takes the monopole to a solution to some non-linear matrix
valued ordinary differential equations, Nahm’s equations on an interval, and there
is an inverse transform giving back the monopole. Much of the mathematical
development of this transform is due to Nigel Hitchin (1983), who showed for the
SU (2) monopole how the monopole and the corresponding solution to Nahm’s
equations are both encoded, quite remarkably, in the same algebraic curve, the
spectral curve of the monopole. This work was extended to the cases of monopoles
for the other classical groups by Hurtubise and Murray (1989). This extension
illustrates just how odd the transform can be: one gets solutions to the Nahm
equations on a chain of intervals, but the size of the matrices jump from interval
to interval.
The Nahm transform for other cases of G invariance has been studied by
various authors (among others, Braam and van Baal 1989; Charbonneau 2004;
Cherkis and Kapustin 2001; Jardim 2001, 2002a, 2002b; Nye 2001; Schenk 1988;
Szabó 2005); a nice survey can be found in Jardim (2004). Here, we study the
case of ‘minimal’ invariance, under Z; the fields are referred to as calorons, and
the Nahm transform sends them to solutions of Nahm’s equations over the circle.
This case is very close to the monopole one, and indeed calorons can be considered
as Kač–Moody monopoles (Garland and Murray 1988).
 The work of Nye and Singer 35

Calorons have been studied from different angles by various authors, including
Bruckmann and van Baal (2002); Bruckmann et al. (2003, 2004), Chakrabarti
(1987), Kraan (2000), Kraan and van Baal (1998a; 1998b; 1998c), Lee (1998),
Lee and Lu (1998), Lee and Yi (2003), Nógrádi (2005), Norbury (2000), and
Ward (2004). In particular, many explicit solutions have been found.
This project started while we were studying the works of Nye (2001) and
Nye and Singer (2000) on calorons; they consider the Nahm transform directly,
and do most of the work required to show that the transform is involutive. The
missing ingredients turn out to lie in complex geometry, in precisely the same
way Hitchin’s extra ingredient of a spectral curve complements Nahm’s. The
complex geometry allows us to complete the equivalence, which can then be used
to compute the moduli. It thus seems to us quite appropriate to consider this
problem in a volume dedicated to Nigel Hitchin, as it allows us to revisit some
of his beautiful mathematics, including some on calorons which has remained
unpublished.
In Section 4.2, we summarize the work of Nye and Singer towards showing
that the Nahm transform is an equivalence between calorons and appropriate
solutions to Nahm’s equations. In Section 4.3, following in large part on the work
of Garland and Murray (1988), we describe the complex geometry (‘spectral
data’) that encodes a caloron. In Section 4.4, we study the process by which
spectral data also correspond to solutions to Nahm’s equations. In Section 4.5,
we close the circle, showing the two Nahm transforms are inverses. In Section
4.6, we give a description of moduli, expounded in Charbonneau and Hurtubise
(2007).

4.2 The work of Nye and Singer

4.2.1 Two types of invariant self-dual gauge fields on R4
Nye and Singer study the Nahm transform between the following two self-dual
gauge fields (we restrict ourselves to SU (2), though they study the more general
case of SU (N )):
4.2.1.1 SU (2) Calorons of charge (k, j)
SU (2) calorons are self-dual SU (2) connections on S 1 × R3 , satisfying appro-
priate boundary conditions. We view S 1 × R3 as the quotient of the standard
Euclidean R4 by the time translation (t, x) → (t + 2π/µ0 , x). Let A be such a
connection, defined over a rank 2 vector bundle V equipped with a unitary
structure; we write it in coordinates over R4 as

A = φdt + Ai dxi ,
i=1,2,3

with associated covariant derivatives

 
∂  
∂
∇= + φ dt + + Ai dxi ≡ ∇t dt + ∇i dxi .
∂t i=1,2,3
∂xi i=1,2,3
36 The Nahm transform for calorons

We require that the L2 norm of the curvature of A be finite, and that in suitable
gauges, the functions Ai be O(|x|−2 ) as |x| → ∞, and that φ be conjugate to
diag(i(µ1 − j/2|x|), i(−µ1 + j/2|x|)) + O(|x|−2 ) for a positive real constant µ1
and a positive integer j (the monopole charge). We also have bounds on the
derivatives of these fields.
The boundary conditions tell us in essence that in a suitable way the connec-
tion extends to the two-sphere at infinity times S 1 ; furthermore, one can show
that the extension is to a fixed connection, which involves fixing a trivialization
at infinity; there is thus a second invariant we can define, the relative second
Chern class, which we represent by a (positive) integer k. There are then two
integer charges for our caloron, k and j, the instanton and monopole charges,
respectively.
There is a suitable group of gauge transformations acting on these fields (Nye
2001): Nye’s approach is to compactify to S 1 × B̄ 3 with a fixed trivilization over
the boundary S 1 × S 2 . The gauge transformations are those extending smoothly
to the identity on the boundary. Two solutions are considered to be equivalent
if they are in the same orbit under this group.
4.2.1.2 Solutions to Nahm’s equations on the circle
The second class of objects we consider are skew adjoint matrix-valued functions
Ti (z), i = 0, . . . , 3, of size (k + j) × (k + j) over the interval (−µ1 , µ1 ), and of
size (k) × (k) over the interval (µ1 , µ0 − µ1 ) (hence µ1 < µ0 − µ1 , so we impose
that condition) that are solutions to Nahm’s equations
dTσ(1)
+ [T0 , Tσ(1) ] = [Tσ(2) , Tσ(3) ], for σ even permutations of (123), (4.1)
dz
on the circle R/(z → z + µ0 ). These equations are reductions of the self-duality
equations to one dimension, and are invariant under a group of gauge transfor-
d
mations under which the dz + T0 transforms as a connection.
We think of the Ti as sections of the endomorphisms of a vector bundle K
whose rank jumps at the two boundary points; at these points ±µ1 we need
boundary conditions.
Case 1: j = 0. At each of the boundary points, there is a large side, with a rank
(k + j) bundle, and a small side, with a rank k bundle. We attach the two at
the boundary point using an injection ι from the small side into the large side,
and a surjection π going the other way, with π · ι the identity. One asks, from
the small side, that the Ti have well-defined limits at the boundary point. From
the large side, one has a decomposition near the boundary points of the bundle
Ck+j × [−µ1 , µ1 ] into an orthogonal sum of subbundles of rank k and j invariant
d
with respect to the connection dz + T0 and compatible with the maps ι, π. With
respect to this decomposition the Ti have the form
   
T̂i O ẑ (j−1)/2
Ti = , (4.2)
O(ẑ (j−1)/2 ) Ri
 The work of Nye and Singer 37

in a basis where T0 is gauged to zero and for a choice of coordinate ẑ :=

z − (±µ1 ). At the boundary point ẑ = 0, each T̂i has a well-defined limit that
coincides using π, ι with the limit from the small side. The Ri have simple poles,
and we ask that their residues form an irreducible representation of su(2).
Case 2: j = 0. Here the boundary conditions are different. We have at the
boundary, identification of the fibres from both sides. With this identification,
we ask that, in a gauge where T0 is zero, the Ti have well-defined limits (Ti )±
from both sides, and that, setting
A(ζ) = (T1 + iT2 ) − 2iT3 ζ + (T1 − iT2 )ζ 2 , (4.3)
one has that taking the limits from both sides,
A(ζ)+ − A(ζ)− = (α0 + α1 ζ)(ᾱ1 − ᾱ0 ζ)T (4.4)
for vectors α0 , α1 in Ck . In particular, A(ζ)+ − A(ζ)− is of rank at most one for
all ζ.
In both cases, there is a symmetry condition, that, in a suitable gauge,
Ti (0) = Ti (0)T
and, finally, an irreducibility condition, that there be no covariant constant
sections of the bundle that are left invariant under the matrices Ti . Solutions for
which there are such covariant constant sections should correspond to calorons
where instanton charge has bubbled off, leaving behind a caloron of lower charge.

4.2.2 The Nahm transform

From caloron to a solution to Nahm’s equations.
Let z be a real parameter. Using the Pauli matrices ei , we define the Dirac
operators Dz acting on sections of the tensor product of the vector bundle V
with the spin bundles S±  C2 × S 1 × R3 by
Dz : Γ(V ⊗ S+ ) → Γ(V ⊗ S− )

s → (∇t + iz)s + (ei ∇i )s.
i
 
The Weitzenbock formula Dz Dz∗ = ∇∗ ∇ + ρ FA−zguarantees that Dz∗ is injec-
tive. Nye and Singer (2000) show that for z not in the set of lifts ±µ1 + nµ0
(n ∈ Z) of the boundary point ±µ1 from the circle to R, and with suitable choices
of function spaces, the operator Dz is Fredholm, and its index away from these
points is k + j for z lying in the intervals (−µ1 + nµ0 , µ1 + nµ0 ) and is k in the
intervals (µ1 + nµ0 , −µ1 + (n + 1)µ0 ). There is thus a bundle over R whose rank
jumps at ±µ1 + nµ0 (n ∈ Z), with fibre ker(Dz ) at z. There is a natural way of
shifting by µ0 , identifying ker(Dz ) with ker(Dz+µ0 ), giving a bundle K over the
circle.
38 The Nahm transform for calorons

Over each interval, this bundle sits inside the trivial bundle whose fibre is the
space of L2 sections of V ⊗ S− . Let P be the orthogonal projection from this
trivial bundle onto K. As elements of K decay exponentially, the operation Xi
of multiplying by the coordinate xi can be used to define operators on sections
of K by
d d
+ T0 = P · ,
dz dz
Ti = P · Xi .

Theorem 4.1 (Nye 2001, section 4.1.2, p. 108) The operators defined in
this way satisfy Nahm’s equations.

This theorem and Theorem 4.2 below fall in line with the general Nahm
transform heuristic philosophy: the curvature of the transformed object, seen
as an invariant connection on R4 , is always composed of a self-dual piece and
another piece depending on the behaviour at infinity of the harmonic spinors.
Since the latter are decaying exponentially, that other piece is zero and the
transformed object is self-dual, or equivalently once we reduce, it satisfies Nahm’s
equations (see for instance Charbonneau 2006, section 3).

From a solution to Nahm’s equations to a caloron.

For the inverse transform, we proceed in the same way: from a solution to Nahm’s
equations, we define a family of auxiliary operators parameterized by points of
R4 acting on sections of K ⊗ C2 by
⎛ ⎞
d 
Dx,t = i ⎝ + T0 − it + ej (Tj − ixj )⎠ .
dz j=1,2,3

One must again distinguish two cases, j > 0 and j = 0.

Case 1: j > 0. We define the space W of L21 sections of K ⊗ C2 such that at

±µ1 the values of the sections coincide: for the sections s1 on the small side and
s2 on the large side, we need ι(s1 ) = s2 . Let X be the space of L2 sections of
K ⊗ C2 over the circle. Set

D̂x,t := Dx,t : W → X (4.5)

Case 2: j = 0. We have as above the space W of sections. We define in addition a

two-dimensional space U associated with the end points ±µ1 . The jump condition
given by (4.4) at µ1 imply that there is a vector u+ ∈ Kµ1 ⊗ C2 such that, as
elements of End(Kµ1 × C2 ),


3
   
(Tj )+ − (Tj )− ⊗ ej = u+ ⊗ u∗+ 0 .
j=1
 Twistor transform for calorons/Kač–Moody monopoles 39

Here the subscript ‘0’ signifies taking the trace-free End(K) ⊗ Sl(2, C) compo-
nent inside End(K) ⊗ End(C2 )  End(K ⊗ C2 ).
One has a similar vector u− at −µ1 . Let U be the vector space spanned by
u+ , u− . Let Π : Kµ1 ⊕ K−µ1 → U be the orthogonal projection, let X be the sum
of the space of L2 sections of K with the space U , and set
D̂x,t := (Dx,t , Π) : W → X (4.6)
The kernel of this operator consists of sections s in the kernel of Dx,t , with values
at ±µ1 lying in U ⊥ ; the cokernel consists of triples (s, cµ1 uµ1 , c−µ1 u−µ1 ), where
∗
s is a section of K ⊗ C2 , lying in the kernel of Dx,t , with jump discontinuity
c+ u+ at µ1 and c− u− at −µ1 .
Theorem 4.2 (Nye 2001, p. 104) The operator D̂x,t has trivial kernel for
all (x, t), and has a cokernel of rank 2, defining a rank 2 vector bundle over
R4 , with natural time periodicity which allows one to build a vector bundle V
over S 1 × R3 , defined locally as a subbundle of the infinite-dimensional bundle
X × S 1 × R3 . Let P denote the orthogonal projection from X to V . Setting, on
sections of V ,
∂
∇i = P · , i = 1, 2, 3
∂xi
∂
∇t = P · ,
∂t
defines an SU (2) caloron.

4.2.3 Involutivity of the transforms

We would like these two transforms to be inverses of each other, so that they
define a one-to-one correspondence between gauge equivalence classes of calorons
and gauge equivalence classes of solutions of Nahm’s equations on the circle.
The results of Nye and Singer get us most of the way there. What is missing
is a proof that the solutions to Nahm’s equations one obtains from a caloron
satisfy the correct boundary and irreducibility conditions and then that the two
constructions are inverses to one another.
The easiest way to do so, as Hitchin did in his original paper on SU (2)
monopoles, is to exploit the regularity given by a third equivalent set of data,
which involves complex geometry. To do this, we examine some work of Garland
and Murray, building on unpublished work of Hitchin.

4.3 Twistor transform for calorons/Kač–Moody monopoles

4.3.1 Upstairs: twistor transform for calorons
Like all self-dual gauge fields on R4 or its quotients R4 /G, calorons admit a
twistor transform, translating the gauge fields into holomorphic vector bundles
on an auxiliary space, the twistor space associated to R4 /G. Following Garland
40 The Nahm transform for calorons

and Murray (1988, section 2), we summarize the construction, again for SU (2)
only.
It is convenient first to recall from Hitchin (1982, section 3) the twistor space
for R3 . It can be interpreted as the space of oriented lines in R3 and it is
isomorphic to T P1 , the tangent bundle of the Riemann sphere. Let ζ be the
standard affine coordinate on P1 , and η be the fibre coordinate in T P1 associated
to the basis vector ∂/∂ζ. The incidence relation between the standard coordinates
on R3 and (ζ, η) is given by

η = (ix1 − x2 ) + 2x3 ζ + (ix1 + x2 )ζ 2 .

The space T P1 comes equipped with standard line bundles O(k), lifted from P1 ,
and line bundles Lt , parameterized by t ∈ C, with transition function exp(tη/ζ)
from the open set ζ = ∞ to the open set ζ = 0. For t ∈ R, these line bundles
correspond to the standard U (1) monopoles on R3 given by a flat connection
and constant Higgs field it. Let Lt (k) := Lt ⊗ O(k). These bundles are in some
sense the building blocks for monopoles for higher gauge groups.
According to Garland and Murray, the twistor space T for S 1 × R3 parame-
trizes pairs consisting of a point in S 1 × R3 and a unit vector in R3 . Such a pair
gives a cylinder along which to integrate, and that projects to an oriented line in
R3 . There is then a C∗ fibration π : T → T P1 , which is in fact holomorphic and
the complement of the zero section in Lµ0 . It has a natural fibrewise compactifica-
tion T̃ = P(O ⊕ Lµ0 ) compactifying the cylinders into spheres. The complement
T̃ \T is a sum of two disjoint divisors T 0 and T ∞ mapping isomorphically to
T P1 .
In the monopole case, the holomorphic vector bundle was obtained by inte-
grating ∇s − iφ along real lines using a metric coordinate s along the line, and
the Higgs field φ. Here, the analogous operation is integrating ∇s − i∇t over the
cylinders to obtain a vector bundle E over T . The boundary conditions allow
us to extend the bundle to the compactification T̃ ; we denote this extension
again by E. The boundary conditions also give line subbundles over these
divisors, E10 = L−µ1 (−j) over T 0 , corresponding to solutions of ∇s − iφ which
decay as one goes to plus infinity on the cylinder, and E1∞ = Lµ1 (−j) over T ∞ ,
corresponding to solutions which decay as one goes to minus infinity.
The twistor space has a real structure τ , which acts on the bundle by τ ∗ (E) =
E , and maps the line subbundle L−µ1 (−j) over T 0 to the annihilator of the
∗

subbundle Lµ1 (−j) over T ∞ .

4.3.2 Downstairs: caloron as a Kač–Moody monopole

As Garland and Murray (1988, section 6) show, there is a very nice way of
thinking of the caloron as a monopole over R3 , with values in a Kač–Moody
algebra, extending a loop algebra: the fourth variable t becomes the internal loop
algebra variable. This way of thinking goes over to the twistor space picture.
 Twistor transform for calorons/Kač–Moody monopoles 41

Indeed, the twistor transform for the caloron gives us a bundle E over T .
Taking a direct image (restricting to sections with poles of finite order along 0, ∞)
gives one an infinite-dimensional vector bundle F over T P1 . If w is a standard
fibre coordinate on T vanishing over T 0 and with a simple pole at T ∞ , it induces
an endomorphism W of F , and quotienting F by the O-module generated by the
image of W − w0 gives us the restriction of E to the section w = w0 , so that E
and (F, W ) are equivalent. One has more, however; the fact that the bundle E
extends to T̃ gives a subbundle F 0 of sections in the direct image which extend
over T 0 , and a subbundle F ∞ of sections in the direct image which extend
over T ∞ . One can go further and use the flags 0 = E00 ⊂ E10 ⊂ E20 = E over T 0 ,
0 = E0∞ ⊂ E1∞ ⊂ E2∞ = E over T ∞ to define for p ∈ Z and q = 0, 1 subbundles
0 ∞
Fp,q , Fp,q of F as
 
0
F−p,q = s ∈ F | w−p s finite at T 0 with value in Eq0 ,
∞
 
Fp,q = s ∈ F | w−p s finite at T ∞ with value in Eq∞ .

We now have infinite flags

· · · ⊂ F−1,0
0
⊂ F−1,1
0
⊂ F0,0
0
⊂ F0,1
0
⊂ F1,0
0
⊂ F1,1
0
⊂ ··· ,
∞ ∞ ∞ ∞ ∞ ∞
· · · ⊃ F2,0 ⊃ F1,1 ⊃ F1,0 ⊃ F0,1 ⊃ F0,0 ⊃ F−1,1 ⊃ ··· . (4.7)
∞ ∞
Note that W · Fp,q
0 0
= Fp+1,q and W · Fp,q = Fp−1,q . Garland and Murray show
r Fp,0
0 ∞
and F−p+1,0 have zero intersection and sum to F away from a compact
curve S0 lying in the linear system |O(2k)|.
r F 0 and F ∞ have zero intersection and sum to F away from a compact
p,1 −p,1
curve S1 lying in the linear system |O(2k + 2j)|.
r Quotients F 0 /F 0 (p−1)µ0 +µ1
p,0 p−1,1 are line bundles isomorphic to L (j).
r Quotients F 0 /F 0 are line bundles isomorphic to Lpµ0 −µ1 (−j).
p,1 p,0
r Quotients F ∞ /F ∞ are line bundles isomorphic to L−(p−1)µ0 −µ1 (j).
p,0 p−1,1
r Quotients F ∞ /F ∞ are line bundles isomorphic to L−pµ0 +µ1 (−j).
p,1 p,0

It is worthwhile stepping back now and seeing what we have from a group
theoretic point of view. On T P1 , we have a bundle with structure group
 C) of Gl(2, C)-valued loops. The flags that we have found give two
G = Gl(2,
reductions R0 , R∞ to the opposite Borel subgroups B0 , B∞ (of loops extending
to 0, ∞, respectively, in P1 , and preserving flags over these points) in G. As for
finite-dimensional groups, we have exact sequences relating the Borel subgroups,
their unipotent subgroups, and the maximal torus:

0 → U0 → B0 → T → 0,
(4.8)
0 → U∞ → B∞ → T → 0.
42 The Nahm transform for calorons

In a suitable basis, B0 and B∞ consist, respectively, of upper and lower triangular

matrices, and T of diagonal matrices.
The two reductions are generically transverse, and fail to be transverse
over the spectral curves of the caloron. This failure of transversality gives
us geometrical data which encodes the bundle and hence the caloron. This
approach is expounded in Hurtubise and Murray (1990) for monopoles, but
can be extended to calorons in a straightforward fashion. We summarize the
construction here. The vector bundle F can be thought of as defining an element
f of the cohomology set H 1 (T P1 , G), and so a principal bundle Pf ; one simply
thinks of f as the transition functions for Pf , in terms of Čech cohomology. The
reductions to B0 , B∞ can also be thought of as elements f0 , f∞ of H 1 (T P1 , B0 ),
H 1 (T P1 , B∞ ), respectively, or as principal bundles Pf0 , Pf∞ , or, since they are
reductions of Pf , as sections R0 of Pf ×G G/B0 = Pf∞ ×B∞ G/B0 and R∞ of
P ×G G/B∞ = Pf0 ×B0 G/B∞ .
We note also that the elements f0 , f∞ when projected to H 1 (T P1 , T ) give
fixed elements α0 , α∞ , in the sense that they are independent of the caloron,
or rather depend only on the charges and the asymptotics, which we presume
fixed. This is a consequence of the fact that the successive quotients in sequence
(4.7) depend only on the µi and the charges. Let A0 , A∞ denote the principal T
bundles associated to α0 , α∞ .
Let A0 (U0 ) be the sheaf of sections of the U0 bundle associated to A0 by the
action of T on U0 . The cohomology set H 1 (T P1 , A0 (U0 )) describes (Hurtubise
and Murray 1990, section 3) the set of B0 bundles mapping to α0 . To find the
bundle f0 , and so f , then, one simply needs to have the appropriate class in
H 1 (T P1 , A0 (U0 )).
Note that U0 also serves as the big cell in the homogeneous space G/B∞ .
Following Gravesen (1989), we think of G/B∞ as adding an infinity to U0 , so if
U0 denotes the sheaf of holomorphic maps into U0 , the sheaf M of holomorphic
maps into G/B∞ can be thought of as meromorphic maps into U0 . Hence there
is an sequence of sheaves of pointed sets (base points are chosen compatibly in
both U0 and G/B∞ )

0 → U0 → M → Pr → 0

defining the sheaf Pr of principal parts. As B0 , T act on these sheaves, we can

build the twisted versions

φ
Pf0 (U0 ) → Pf0 (M) −−→ Pf0 (Pr),

A0 (U0 ) → A0 (M)−−→A0 (Pr).

As we are quotienting out the action of U0 , we have A0 (Pr)  Pf0 (Pr).

 Twistor transform for calorons/Kač–Moody monopoles 43

We are now ready to define the principal part data of the caloron. The principal
part data of the caloron is the image under φ of the class of the reduction given
as a section R∞ of Pf0 ×B0 G/B∞ = Pf0 (M):

φ(R∞ ) ∈ H 0 (A0 (Pr)).

To get back the caloron bundle from the principal part data, one takes the
coboundary δ(φ(R∞ )), in the obvious Čech sense, for the second sequence,
obtaining a class in H 1 (T P1 , A0 (U0 )). One checks that this is precisely the
class corresponding to the bundle P0 ; while this seems a bit surprising, the
construction is fairly tautological, and is done in detail in (Hurtubise and Murray
1990, section 3). In our infinite-dimensional context, the bundles are quite special,
as they, and their reductions, are invariant under the shift operator W .
Of course, this construction does not tell us what the mysterious class φ(R∞ )
actually corresponds to, but it turns out that, over a generic set of calorons, it is
quite tractable. Indeed, the principal part data, as a sheaf, is supported over the
spectral curves, and, if the curves have no common components and no multiple
components, then the principal part data amounts to the following spectral data
(Garland and Murray 1988):
r The two spectral curves, S0 and S1
r The ideal IS0 ∩S1 , decomposing as IS01 · IS10 ; the real structure interchanges
the two factors
r An isomorphism of line bundles O[−S10 ] ⊗ Lµ0 −2µ1  O[−S01 ] over S0
r An isomorphism of line bundles O[−S01 ] ⊗ L2µ1  O[−S10 ] over S1
r A real structure on the line bundle O(2k + j − 1)[−S01 ] ⊗ Lµ1 |S1 , lifting the
real involution τ on T P1

The structure of this data can be understood in terms of the Schubert structure
of G/B0 . The sheaf of principal parts, by definition, lives in the complement of
U0 , where the two flags cease to be transverse. In the case that concerns us here
(remember the invariance under W ), there are essentially two codimension one
varieties at infinity to U0 that we consider, whose pull-backs give the two spectral
curves. The decomposition of S0 ∩ S1 into two pieces is simply a pull-back of the
Schubert structure from G/B0 . For the line bundles, the basic idea is that in
codimension one, the principal parts can be understood using embeddings of P1
into G/B0 given by principal Sl(2, C)’s in G (see for instance Boyer et al. 1994,
section 4), reducing the question at least locally to understanding principal parts
for maps into P1 , a classical subject. For simple poles, the principal part of a map
into P1 is encoded by its residue. Here, the poles are over the spectral curves,
and globally, the principal part is then encoded as a section of a line bundle over
each curve; we identify these below. A way of seeing that the spectral data splits
into components localized over each curve is that we can choose appropriately
44 The Nahm transform for calorons

two parabolic subgroups P ar0 , P ar1 and project from G/B0 to G/P ar0 and
G/P ar1 .
To see what the ‘residues’ correspond to here, we exploit a description of the
bundle first given in Hurtubise and Murray (1989, proposition 1.12), the short
exact sequence
.. ..
. .
 0 
∞
F/ Fp−1,1 + F−p+1,0  ⊕ 
⊕ 0 ∞
 0  F/ F p,0 + F−(p−1),0
∞
0 −→ F −→ F/ Fp,0 + F−p,1 −→ ⊕ −→ 0. (4.9)
 0 
 0 ⊕  F/ Fp,1 + F−p,1 ∞
∞
F/ Fp,1 + F−p,0 ⊕
.. ..
. .
0
The quotients at the right are supported over the spectral curves (F/(Fp,0 +
∞ 0 ∞
F−(p−1),0 ) over S0 , and F/(Fp,1 + F−p,1 ) over S1 ) and those in the middle are
generically line bundles.
The shift operator W moves the quotients in the middle and last columns
two steps down. The quotients are in fact direct image sheaves R1 π̃∗ of sheaves
on T̃ derived from E; for example, let Ep,0,−(p−1),∞ be the sheaf of sections
of E over T with
 0 poles∞ of order p at T0 and a zero1 of order p − 1 at T∞ .
The sheaf F/ Fp,0 + F−p,1 can be identified with R π̃∗ (Ep,0,−(p−1),∞ ). One
can make similar identifications for the other sheaves. In a way that parallels
Hitchin (1983), we have the following result:
Proposition 4.3
0
1. The spectral curves S0 and S1 are real (τ -invariant); the quotient F/(Fp,1 +
∞
F−p,1 ) over S1 inherits from E a quaternionic structure, lifting τ .
0 ∞ 0 ∞
2. The sheaves F/(Fp,0 + F−(p−1),0 ) and F/(Fp,1 + F−p,1 ) satisfy a vanishing
theorem:
   
∞
H 0 T P1 , F/ Fp,0
0
+ F−(p−1),0 ⊗ L−z (−2) = 0,

for z ∈ [(p − 1)µ0 + µ1 , pµ0 − µ1 ], and

  0 ∞
 
H 0 T P1 , F/ Fp,1 + F−p,1 ⊗ L−z (−2) = 0,
for z ∈ (pµ0 − µ1 , pµ0 + µ1 ).
The vanishing holds because these cohomology groups encode L2 solutions to
the Laplace equation in the caloron background, which must be zero (see Hitchin
1983, theorem 3.7, and Hurtubise and Murray 1989, theorem 1.17).
For generic spectral curves, following a line of argument of Hurtubise and Mur-
ray (1989), Garland and Murray show directly that the spectral data determines
the caloron. To this end, they identify in Garland and Murray (1988, section 6)
 Twistor transform for calorons/Kač–Moody monopoles 45

the quotients

 
∞
0
F/ Fp,0 + F−(p−1),0 = Lpµ0 −µ1 (2k + j)[−S10 ]|S0

= L(p−1)µ0 +µ1 (2k + j)[−S01 ]|S0 ,

 0 ∞

F/ Fp,1 + F−p,1 = Lpµ0 +µ1 (2k + j)[−S01 ]|S1

= Lpµ0 −µ1 (2k + j)[−S10 ]|S1 ,

 0 ∞

F/ Fp,0 + F−p,1 = Lpµ0 −µ1 (2k + j) ⊗ IS10 ,
 0 ∞

F/ Fp,1 + F−p,0 = Lpµ0 +µ1 (2k + j) ⊗ IS01 .

These identifications realized in exact sequence (4.9) give

.. ..
. .
L(p−1)µ0 +µ1
(2k + j) ⊗ IS01 L(p−1)µ0 +µ1
(2k + j)[−S01 ]|S0
⊕ = Lpµ0 −µ1 (2k + j)[−S10 ]|S0
F → Lpµ0 −µ1 (2k + j) ⊗ IS10 −→ ⊕ −→ 0.
⊕ Lpµ0 +µ1 (2k + j)[−S01 ]|S1
Lpµ0 +µ1 (2k + j) ⊗ IS01 = Lpµ0 −µ1 (2k + j)[−S10 ]|S1
.. ..
. .
(4.10)

The ‘residues’ are then the sheaves on the right, supported over the spectral
curves; the maps from the middle to the right-hand sheaves give the various
isomorphisms. This diagram shows that F , and hence the caloron, is encoded in
the spectral data.
More generally, we can define the spectral data composed of the curves S0 , S1 ,
the pull-backs to T P1 of the Schubert varieties in G/B0 , and the sheaves on
the right-hand side of the sequence (4.9), with isomorphisms similar to the
ones for generic spectral data given by maps from the sheaves in the middle
column.
To summarize, we have that the caloron determines principal part data, a
section of a sheaf of principal parts supported over the spectral curve; this
data determines the caloron in turn. In the generic case, the principal part
data equivalent to spectral data can be described in terms of two curves, their
intersections, and sections of line bundles over these curves. We note that these
generic calorons exist; indeed, we already know that a caloron is determined by
a solution to Nahm’s equations, and it is easy to see that generic solutions to
Nahm’s equations exist, as we shall see in Section 4.6.
46 The Nahm transform for calorons

4.4 From Nahm’s equations to spectral data, and back

4.4.1 Flows of sheaves
The solutions to Nahm’s equations we consider also determine equivalent spectral
data, as we shall see. To begin, note that by setting A(ζ, z) as in (4.3), and
A+ (ζ, z) = −iT3 (z) + (T1 − iT2 )(z)ζ,
Nahm’s equations (4.1) are equivalent to the Lax form
dA
+ [A+ , A] = 0. (4.11)
dz
The evolution of A is by conjugation, so the spectral curve given by
det (A(ζ, z) − ηI) = 0 (4.12)
in T P1 is an obvious invariant of the flow (4.11); if the matrices are k × k, the
curve is a k-fold branched cover of P1 . We define for each z a sheaf Lz over the
curve via the exact sequence
A(ζ,z)−ηI
0 → O(−2)k −−−−−−→ O k → Lz → 0. (4.13)
This correspondence taking a solution to Nahm’s equations to a curve and a flow
of line bundles over the curve is fundamental to the theory of Lax equations (see
Adams et al. 1990). In our case, we have an equivalence:
Proposition 4.4 There is an equivalence between
1. Solutions to Nahm’s equations on an interval (a, b), given by k × k matrices
with reality condition built from skew-Hermitian matrices Ti as in (4.3)
2. Spectral curves S that are compact and lie in the linear system |O(2k)|
lying in T P1 , and flows Lz , for z ∈ (a, b), of sheaves supported on S, such
that
a. H 0 (T P1 , Lz (−1)) = H 1 (T P1 , Lz (−1)) = 0.

b. Lz = Lz ⊗ Lz−z , for z, z  ∈ (a, b).
c. The curve S is real, that is, invariant under the antiholomorphic involu-
tion τ (η, ζ) = (−η̄/ζ̄ 2 , −1/ζ̄) corresponding to reversal of lines in T P1 .
d. There is a linear form µ on H 0 (S, Lz ⊗ τ ∗ (Lz )) vanishing on H 0 (S, Lz ⊗
τ ∗ (Lz )(−C)) for all fibres C of T P1 → P1 and inducing a positive defi-
nite Hermitian metric (σ1 , σ2 ) → µ(σ1 τ ∗ (σ2 )) on H 0 (S ∩ C, Lz ).
We start with the passage from (1) to (2). For the relation H 0 (T P1 , Lz (−1)) =
A(ζ,z)−ηI
H 1 (T P1 , Lz (−1)) = 0 to hold, we need O(−3)k −−−−−−→ O(−1)k to induce an
isomorphism on H 1 . As shown in Hurtubise and Murray (1989, lemma 1.2), the
groups H 1 (T P1 , O(p)) are infinite-dimensional spaces nicely filtered by finite-
dimensional pieces, corresponding to powers of η in the cocycle. A basis for
H 1 (T P1 , O(−3)) is 1/ζ, 1/ζ 2 , η/ζ, . . . , η/ζ 4 , η 2 /ζ, . . . , η 2 /ζ 6 , . . ., and a basis for
 From Nahm’s equations to spectral data, and back 47

H 1 (T P1 , O(−1)) is obtained from it by multiplying by η. Multiplication by

η then induces an isomorphism, and so multiplication by A(ζ, z) − ηI gives
an isomorphism H 1 (T P1 , O(−3)k ) → H 1 (T P1 , O(−1)k ). For the relation Lz =

Lz ⊗ Lz−z , see Hurtubise and Murray (1989, section 2.6) or Adams et al. (1990)
for the general theory of the Lax flows. The reality of the curve follows from
the fact that the Ti are skew Hermitian. For the final property, the positive
definite inner product on Ok with respect to which the Ti are skew Hermitian
induces one on H 0 (T P1 , Lz ), by passing to global sections in sequence (4.13). The
inner product is a linear map H 0 (S, Lz ) ⊗ H 0 (S, τ ∗ (Lz )) → C. As Lz represents,
in essence, the dual to the eigenspaces of the A(ζ)T , this linear map factors
through H 0 (S, Lz ⊗ τ ∗ (Lz )), and it must be zero on sections that vanish on
fibres.

Now that the passage from (1) to (2) is established, note that since τ ∗ Lz−z =
z  −z ∗
L , the product Lz ⊗ τ (Lz ) is constant. The last condition of (2) imposes
severe constraints on Lz : when S is smooth, and Lz a line bundle, it tells us
that
Lz ⊗ τ ∗ (Lz )  KS (2C). (4.14)
Indeed, in that case the vanishing of the cohomology of Lz (−1) tells us
deg(Lz ) = g + k − 1 and deg(Lz ⊗ τ ∗ (Lz )) = 2g + 2k − 2. The space of sections
H 0 (S, Lz ⊗ τ ∗ (Lz )) is of dimension g + 2k − 1, while H 0 (S, Lz ⊗ τ ∗ (Lz )(−C))
and H 0 (S, Lz ⊗ τ ∗ (Lz )(−C  )) are of dimension g + k − 1, and dim H 0 (S, Lz ⊗
τ ∗ (Lz )(−C − C  )) = g − 1, unless it is the canonical bundle, in which case
the dimension is g. There cannot be a non-zero form on H 0 (S, Lz ⊗ τ ∗ (Lz ))
vanishing on H 0 (S, Lz ⊗ τ ∗ (Lz )(−C)) + H 0 (S, Lz ⊗ τ ∗ (Lz )(−C  )) unless the
intersection H 0 (S, Lz ⊗ τ ∗ (Lz )(−C − C  )) is of dimension g, in which case
Lz ⊗ τ ∗ (Lz )(−C − C  ) is the canonical bundle.
Now we deal with the passage (2) to (1). In essence, it has been treated
by Hitchin (1983), but we give a more invariant construction of A suitable
for our purposes. The fibre product F P of T P1 with itself over P1 is the
vector bundle O(2) ⊕ O(2), with fibre coordinates η, η  , and projections π, π 
to T P1 ; the lift of O(2) to this vector bundle has global sections 1, ζ, ζ 2 , η, η  ,
and the diagonal ∆ is cut out by η − η  . Hence over F P we have the exact
sequence
η−η 
0 → O(−2) −−−−→ O → O∆ → 0. (4.15)
Lifting Lz via π, tensoring it with this sequence, and pushing it down via π 
gives
A(ζ)−η 
0 → V (−2) −−−−→ V → Lz → 0 (4.16)
for some rank k vector bundle V , which over (ζ, η) = (ζ0 , η0 ) is the space of
sections of L in the fibre of T P1 over ζ = ζ0 , and so is independent of η: V is
lifted from P1 . We then identify V . The fibre product F P = O(2) ⊕ O(2) lives
48 The Nahm transform for calorons

in the product T P1 × T P1 , and is cut out by ζ − ζ  , and one has the exact
sequence
ζ−ζ 
0 → O(−1, −1) −−−−→ O → OF P → 0. (4.17)
Lifting Lz to T P1 × T P1 , tensoring it with this sequence and taking direct image
on the other factor gives the long exact sequence
0 → H 0 (T P1 , Lz (−1)) ⊗ O(−1) → H 0 (T P1 , Lz ) ⊗ O → V
→ H 1 (T P1 , Lz (−1)) ⊗ O(−1) · · · .
The vanishing of the cohomolology of Lz (−1) shows that V  Ok .
We then have given A(ζ, z) as a map of bundles; the process of turning it into
a matrix-valued function A(ζ, z), and of showing that it evolves according to
Nahm’s equations as one tensors it by Lt , is given in Hitchin (1983, proposition
4.16). Similarly, the definition of an inner product on H 0 (T P1 , Lz ) and the proof
that the Ti are skew adjoint with respect to the inner product follow the line
given in Hitchin (1983, section 6).

4.4.2 Boundary conditions

Now suppose that we have a solution to Nahm’s equations satisfying the bound-
ary conditions for an SU (2) caloron, as given above. We shall show that the
boundary behaviour allows us to identify the initial conditions for the flow, in
other words to say what the sheaves Lz are. As there are two intervals, there are
two spectral curves: S0 , which is a k-fold cover of P1 , and S1 , which is k + j-fold.
We suppose the curves are generic, in that they have no common components
and no multiple components. A solution to Nahm’s equations is called generic if
its curves are.
Proposition 4.5 There is an equivalence between
1. Generic solutions A(ζ, z) to Nahm’s equations over the circle, denoted
A0 (ζ, z) on (µ1 , −µ1 + µ0 ) and A1 (ζ, z) on (−µ1 , µ1 ), satisfying the condi-
tions of Section 4.1.2.2
2. Generic spectral curves S0 in T P1 of degree k over P1 that are the support
of sheaves L0z , z ∈ [µ1 , µ0 − µ1 ], and spectral curves S1 in T P1 of degree
k + j over P1 that are the support of sheaves L1z , z ∈ [−µ1 , µ1 ], with the
following properties:
r L0z , for z ∈ [µ1 , µ0 − µ1 ], L1z , for z ∈ (−µ1 , µ1 ), S0 , and S1 have the
properties of Proposition 4.4.
r The intersection S0 ∩ S1 decomposes as a sum of two divisors S01 and
S10 , interchanged by τ .
r At −µ1 = µ0 − µ1 (on the circle), L0−µ = O(2k + j − 1)[−S10 ]|S0 and
1
L1−µ1 = O(2k + j − 1)[−S10 ]|S1 .
r At µ1 , L0µ = O(2k + j − 1)[−S01 ]|S0 and L1µ = O(2k + j − 1)[−S01 ]|S1 .
r There is a1 real structure on L10 lifting τ . 1
 From Nahm’s equations to spectral data, and back 49

As a consequence there are isomorphisms of line bundles O[−S01 ] ⊗ Lµ0 −2µ1 

O[−S10 ] over S0 and O[−S10 ] ⊗ L2µ1  O[−S01 ] over S1 .
Let us begin with a discussion of the case k = 0, studied in Hitchin (1983).
He showed that the solution to Nahm’s equations for an SU (2) monopole corre-
sponded to the flow of line bundles Lt+µ1 (j − 1), t ∈ [−µ1 , µ1 ] on the monopole’s
spectral curve, with the flow being regular in the middle of the interval, and
having simple poles at the ends of the interval, with residues giving an irreducible
representation of SU (2).
In the construction given in Proposition 4.4, we obtain for this flow a bundle V
over P1 × [−µ1 , µ1 ], and a regular section of Hom(V (−2), V ) over this interval.
For t ∈ (−µ1 , µ1 ) the bundle V is trivial on P1 × {t}. The singularity at the end
of the interval is caused by a jump in the holomorphic structure in V at the
ends. (Both ends are identical, because of the identification O = L2µ1 over the
spectral curve.)
To understand the structure of V at L−µ1 = O(j − 1), we note that in the
natural trivializations lifted from P1 , local sections of O(j − 1) over T P1 are
filtered by the order of vanishing along the zero section:

O(j − 1) ⊃ ηO(j − 3) ⊃ η 2 O(j − 5) ⊃ · · · ;

this filtration can be turned onto a sum, as we have the subsheaves Li of sections
η i s, with s lifted from P1 . The construction, by limiting to a curve which is a
j-sheeted cover of P1 , essentially says that the degrees of vanishing of interest
are at most j − 1, as one is taking the remainder by division by the equation of
the curve. The bundle V over P1 × {µ1 } decomposes as a sum

V = O(j − 1) ⊕ O(j − 3) ⊕ O(j − 5) ⊕ · · · ⊕ O(−j + 1).

More generally, for later use, set

def
Vk,k−2
= O(k) ⊕ O(k − 2) ⊕ O(k − 4) ⊕ · · · ⊕ O(k − 2),

so V = Vj−1,−j+1 . Writing 0 = η j + η j−1 p1 (ζ) + · · · + pj (ζ) for the equation of

the spectral curve, with polynomials pi of degree 2i, we can write the induced
automorphism A(ζ) − ηI in this decomposition as
⎛ ⎞
−η 0 0 . . . 0 −pj (ζ)
⎜ 1 −η 0 . . . 0 −pj−1 (ζ) ⎟
⎜ ⎟
⎜ 0 1 −η . . . 0 −pj−2 (ζ) ⎟
⎜ ⎟.
⎜ .. .. .. .. .. .. ⎟
⎝ . . . . . . ⎠
0 0 0 . . . 1 −η − p1 (ζ)

There is another way of understanding how the pole of Nahm’s equations

generates a non-trivial V . We first restrict to ζ = 0, and vary z. For a moment,
suppose by translation that z = 0 is the point where the structure jumps. From
50 The Nahm transform for calorons

the boundary conditions, the residues at 0 are given in a suitable basis by


−(j − 1) 2 − (j − 1) (j − 1)
Res(A+ ) = diag , ,..., , (4.18)
2 2 2
⎛ ⎞
..
0 0 . 0 0
⎜ ⎟
⎜ ⎟
⎜ 1 0 . . . 0 0⎟
⎜ ⎟
Res(A) = ⎜⎜0 1 . . . 0 0⎟ .
⎟ (4.19)
⎜ ⎟
⎜. . . . .⎟
⎝ .. . . . . . . .. ⎠
0 0 0 1 0
ds
As in Hurtubise (1989, proposition 1.15), we solve dz + A+ s = 0 along ζ = 0, for
j−1
s of the form z 2 ((1, 0, . . . , 0) + z · holomorphic). The sections Aj s also solve
the equation, and, using these sections as a basis, one conjugates A − ηI to
⎛ ⎞
−η 0 0 . . . 0 −pj (0)
⎜ 1 −η 0 . . . 0 −pj−1 (0) ⎟
⎜ ⎟
⎜ 0 1 −η . . . 0 −pj−2 (0) ⎟
⎜ ⎟ (4.20)
⎜ .. .. . . .. .. ⎟
⎝ . . . . . ⎠
0 0 0 . . . 1 −η − p1 (0)
using a matrix of the form M (z) = (Holomorphic in z)N (z), with
 −(j−1) 2−(j−1) (j−1)

N (z) := diag z 2 , z 2 , . . . , z 2 . (4.21)

This process can be applied over any point ζ, as the matrices A(ζ), A+ (ζ) have
residues (in z) conjugate to (4.18) and (4.19). Indeed, at ζ = 0, they represent
standard generators of a representation of Sl(2), and moving away from ζ = 0
simply amounts to a change of basis. Explicitly, conjugating the residues of A(ζ),
A+ (ζ) by N (ζ) takes them to the residues of ζA(1), A+ (1). Conjugating again
by a matrix T takes them to the residues of ζA(0), A+ (0), and then by N (ζ)−1
to the residues of A(0), A+ (0). Thus, for a suitable
M (ζ, z) = (Holomorphic in z, ζ) · N (ζ −1 z)T N (ζ) (4.22)
one has
⎛ ⎞
−η 0 0. . . 0 −pj (ζ)
⎜ 1 −η 0. . . 0 −pj−1 (ζ) ⎟
⎜ ⎟
⎜ −η. . . 0 −pj−2 (ζ) ⎟
M (ζ, z)(A(ζ, z) − ηI)M (ζ, z)−1 =⎜ 0 1 ⎟. (4.23)
⎜ .. .. .. .. .. .. ⎟
⎝ . . . . . . ⎠
0 0 0 . . . 1 −η − p1 (ζ)
At z = 0, it is a holomorphic section of Hom(V (−2), V ) in standard trivializa-
tions.
 From Nahm’s equations to spectral data, and back 51

ds
The same procedures work when k = 0. Indeed, the singular solution to dz +
A+ s has the same behaviour. Starting from a solution to Nahm’s equations, and
integrating the connection as in Hurtubise (1989), we find that, near µ1 on the
interval (−µ1 , µ1 ), a change of basis of the form
M (ζ, z) = (Holomorphic in z, ζ) · diag(Ik×k , N (ζ −1 z)T N (ζ)) (4.24)
conjugates A1 (ζ, z) − ηI to the constant (in z) matrix
⎛ ⎞
a11 (ζ) − η a12 (ζ) . . . a1k (ζ) 0 0 0 ... 0 g1 (ζ)
⎜ a21 (ζ) a22 (ζ) − η . . . a1k (ζ) 0 0 0 ... 0 g2 (ζ) ⎟
⎜ ⎟
⎜ .. .. .. .. .. .. .. .. .. .. ⎟
⎜ . . . . . . . . . . ⎟
⎜ ⎟
⎜ ak1 (ζ) a (ζ) . . . a (ζ) − η 0 0 0 ... 0 gk (ζ) ⎟
⎜ k2 kk ⎟
⎜ f1 (ζ) f2 (ζ) . . . fk (ζ) −η 0 0 . . . 0 −pj (ζ) ⎟
⎜ ⎟. (4.25)
⎜ 0 0 ... 0 1 −η 0 . . . 0 −pj−1 (ζ) ⎟
⎜ ⎟
⎜ 0 0 . . . 0 0 1 −η . . . 0 −pj−2 (ζ) ⎟
⎜ ⎟
⎜ .. .. .. .. .. .. .. .. .. .. ⎟
⎝ . . . . . . . . . . ⎠
0 0 ... 0 0 0 0 . . . 1 −η − p1 (ζ)
Setting C(p(ζ)) to denote the companion matrix of p(ζ), we write this matrix
schematically as
⎛ 0   ⎞
A
(ζ) −ηI 0 G(ζ)
A1 (ζ) − ηI = ⎝ F (ζ) ⎠. (4.26)
C(p(ζ)) − ηI
0
Let Madj denote the classical adjoint of M , so that Madj M = det(M )I. Then
  
det(A1 (ζ) − ηI) = det(A0 (ζ) − ηI) η j + η j−i pi (ζ)
 
+ (−1)j F A0 (ζ) − ηI adj G. (4.27)

The matrix A0 (ζ) is equal to the limit A0 (ζ, µ1 ). At the boundary point µ1 ,
the bundle Vµ01 for S0 is trivial, since the solution on S0 is smooth at that point;
one has
A0 (ζ)−ηI
0 → O(−2)k −−−−−−→ O k → L0µ1 → 0. (4.28)
The limit bundle Vµ11 for S1 is

Vµ11 = Ok ⊕ Vj−1,−j+1 (4.29)

with
A1 (ζ)−ηI
0 → Vµ11 (−2) −−−−−−→ Vµ11 → L1µ1 → 0. (4.30)
We now want to identify the limit bundles L0µ1 , L1µ1 , and the gluing between
them. We want to show that there is a divisor D contained in the intersection S0 ∩
52 The Nahm transform for calorons

S1 such that L0µ1  O(2k + j − 1)[−D]|S0 and L1µ1  O(2k + j − 1)[−D]|S1 ,

and that the correspondence between the matrices is mediated by the maps
O(2k + j − 1)[−D]|S0 ← O(2k + j − 1) ⊗ ID → O(2k + j − 1)[−D]|S1 .
To prove its existence, we need to understand how to lift and push down
O(2k + j − 1) ⊗ ID (here ID is the sheaf of ideals of D on T P1 ) and so
O(2k + j − 1) through the sequence of (4.15). To do so, we compactify T P1 by
embedding it into T = P(O(2) ⊕ O), adding a divisor at infinity P∞ . Let C be
the fibre of the projection T → P1 , then P∞ + 2C is linearly equivalent to the
zero section P0 of T P1 . Our fibre product now compactifies to a P1 × P1 bundle
over P1 . Similarly, the sequence of (4.15) compactifies to
 η−η 
0 → O (−2C − P∞ − P∞ ) −−→ O → O∆ → 0. (4.31)
The line bundle O(mC + nP∞ ) over T has over each fibre C a (n + 1)-
dimensional space of sections, and these sections are graded by the order in
η, as before. Define an  × ( − 1) matrix S(, η) by
⎛ ⎞
−η 0 · · · · · · 0
⎜ 1 −η . . . .. ⎟
⎜ . ⎟
⎜ ⎟
⎜ 0 1 . . . . . . ... ⎟
S(, η) = ⎜
⎜ .. . . . .
⎟.
⎟ (4.32)
⎜ . . . −η 0 ⎟
⎜ . ⎟
⎝ . .. ⎠
. . 1 −η
0 ··· ··· 0 1

Lemma 4.6 Let m, n > 0. Lifting O(mC + nP∞ ) to the fibre product and
tensoring with the sequence (4.28), then pushing down, we obtain an exact
sequence

n−1  
0 −→ O (m − 2i − 2) C + (n − i − 1)P∞
i=0
↓ S(n, η)

n   (1,η,η2 ,...,ηn )
O (m − 2i)C+ (n − i)P∞ −−−−−−−−−−→ O(mC + nP∞ ) → 0.
i=0

Over T P , it becomes
1

S(n,η) (1,η,η 2 ,...,η n )

0 → Vm−2,m−2n −−−−−−→ Vm,m−2n −−−−−−−−→ O(m) → 0.

The proof is straightforward. We now want to define a subsheaf of O(2k + j).

Set
⎛ 0 ⎞

−
A (ζ)  ηI 0
R(ζ, η) = ⎝ F ⎠.
S(j, η)
0
 From Nahm’s equations to spectral data, and back 53

We define a vector of polynomial functions in (ζ, η)

(φ1 , . . . , φk ) = −(−1)k F (ζ)(A0 (ζ) − ηI)adj .

We have
   
(φ1 , . . . , φk ) A0 (ζ) − ηI + F (ζ)(−1)k det A0 (ζ) − ηI = 0. (4.33)
 k−1
Write φi = k−1
j=0 φji (ζ)η , and (−1) det(A (ζ) − ηI) = η +
j k 0 k j
j=0 hj (ζ)η .
Decomposing (4.33) into different powers of η, we obtain
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
0 ··· 0 φ01 · · · φ0k h0
⎜ φ01 · · · φ0k ⎟ ⎜ ..
⎟ ⎜
.. .. ⎟ ⎜ .. ⎟
⎜
−⎜ .. .. .. ⎟ + ⎜ . . . ⎟ ⎜ ⎟
⎟ A0 (ζ) + ⎜ . ⎟ F = 0. (4.34)
⎝ . . . ⎠ ⎝φk−1,1 · · · φk−1,k ⎠ ⎝hk−1 ⎠
φk−1,1 · · · φk−1,k 0 ··· 0 1

Let Mn be the (k + n) × (k + n) matrix:

⎛ ⎞
φ01 . . . φ0k h0 0 0
⎜ .. .. .. .. . . ⎟
⎜ . . . . . 0 ⎟
⎜ ⎟
⎜φk−1,1 . . . φk−1,k hk−1 h0 ⎟
⎜ ⎟
Mn = ⎜ .. . ⎟. (4.35)
⎜ 0 ··· 0 1 . .. ⎟
⎜ ⎟
⎜ . .. .. .. ⎟
⎝ .. . . 0 . hk−1 ⎠
0 ··· 0 0 0 1

Using (4.34), we have the commuting diagram

R(ζ,η)
O(−2)⊕k ⊕ Vj−3,−j+1 / O⊕k ⊕ Vj−1,−j+1 /R
Mj−1 Mj
 S(k+j,η)  (1,η,...,η k+j−1 ) 
V2k+j−3,−j+1 / V2k+j−1,−j+1 / O(2k + j − 1).

It defines a rank 1 sheaf R, and embeds it in O(2k + j − 1). This embedding

fails to be surjective when (1, η, . . . , ηk+j−1 )Mj = 0, or equivalently when
 
F (A0 (ζ) − ηI)adj = 0 and det A0 (ζ) − ηI = 0.

These conditions can only be met on S0 , and, because of (4.27), on S1 . In short,

there is a subvariety D of S0 ∩ S1 , with R = O(2k + j − 1) ⊗ ID . There are
natural surjective maps from R to L0µ1 , L1µ1 : for the projection Πm+n,n on a sum
with m + n summands onto the first n summands and the injection In,n+m into
54 The Nahm transform for calorons

the first n summands, we have that the diagram of exact sequences

A0 (ζ)−ηI
O(−2)⊕k / O⊕k / L0µ
O O O1
Πk+j−1,k Πk+j,k
R(ζ,η)
O(−2)⊕k ⊕ Vj−3,−j+1 / O⊕k ⊕ Vj−1,−j+1 /R

Ik+j−1,k+j I
 A1 (ζ)−ηI  
⊕k
O(−2) ⊕ Vj−3,−j−1 / O⊕k ⊕ Vj−1,−j+1 / L1µ
1

commutes. This diagram identifies L0µ1 , L1µ1

as O(2k + j − 1)[−D] over their
respective curves.
A similar analysis for −µ1 shows that L0−µ1 , L1−µ1 are O(2k + j − 1)[−D  ], for
some D . To relate D to D  , note that as in Hitchin (1983, section 6), there is a
real structure on L0 lifting the real involution τ . Hence Ti (0) = Ti (0)T . On the
other hand, for a solution A(z) to Nahm’s equation (4.11),
dA(−z)T
+ [A+ (−z)T , A(−z)T ] = 0. (4.36)
dz
As solutions to the same differential equation with the same initial condition, all
the way around the circle,
A(−z) = AT (z). (4.37)
This symmetry tells us that in terms of the matrix A1 (ζ), A0 (ζ) at µ1 , the
equation for D is
0 = (A0 (ζ) − ηI)adj G. (4.38)
 0 
Using (4.27), we see that along
 S0 where  det A (ζ) − ηI = 0, the intersection
S0 ∩ S1 is cut out by 0 = F A0 (ζ) − ηI adj G. Generically along S0 , the matrix
(A0 (ζ) − ηI) has corank 1, and so (A0 (ζ) − ηI)adj has rank 1. We can thus write
it as the product U V of a column vector and a row vector. The equation for
the intersection is then 0 = (F U )(V G), the product of the defining relations
for D and D  . If the two curves are smooth, intersecting transversally, then
S0 ∩ S1 = D + D . This property holds independently of whether the matrices
arise from calorons or not. As one has the result for generic intersections, varying
continuously gives S0 ∩ S1 = D + D even for non-generic curves.
The skew Hermitian property of the Ti implies that A(ζ, z) = −ζ 2 Ā(−1/ζ̄, z)T .
This symmetry transforms the equation 0 = F (A0 (−1/ζ̄) − ηI)adj into 0 =
(A0 (ζ) − ηI)adj G, showing that τ (D) = D .
We have now obtained our complete generic spectral data from our generic
solution to Nahm’s equations. The converse, starting with the spectral data, is
essentially done above, apart from the boundary behaviour. This last piece is
dealt with in Hurtubise and Murray (1989, section 2); the conditions on the
 Closing the circle 55

intersections of the spectral curves used here are more general, but the proof
goes through unchanged.

4.5 Closing the circle

We have two different types of gauge fields, with a transform relating them; both
have complex data associated to them, which encodes them; this data, we have
seen (at leastin the generic case), satisfies
 the same conditions,
 with, for
 example,
∞ ∞
the sheaves F/(Fp,0 0
+ F−(p−1),0 ) ⊗ Lz (−1) and F/(Fp,1 0
+ F−p,1 ) ⊗ Lz (−1)
obtained from the caloron satisfying the same conditions as the Lt obtained from
the solution to Nahm’s equation for appropriate values of z and t. We now want
to check that the complex data associated to the object is the same as that
associated to the object’s transform.

4.5.1 Starting with a caloron

Starting with a caloron, one can define, as above, the spectral curves and sheaves
over them:
r The curve S0 is the locus where Fp,0
0 ∞
and F−p+1,0 have non-zero intersection.
r The curve S1 is the locus where F 0 and F ∞ have non-zero intersection.
−p,1
r The quotient F/(Fp,0
p,1
0
+ F∞ ), supported over S0 , is generically
−(p−1),0
isomorphic to the line bundle Lpµ0 −µ1 (2k + j)[−S10 ]|S0 = L(p−1)µ0 +µ1 (2k +
j)[−S01 ]|S0 .
r The quotient F/(F 0 + F ∞ ), supported over S1 , is generically isomorphic
p,1 −p,1
to the line bundle Lpµ0 +µ1 (2k + j)[−S01 ]|S1 = Lpµ0 −µ1 (2k + j)[−S10 ]|S1 .
We can ‘shift’ the caloron by the U (1) monopole with constant Higgs field
iz. Let us consider the direction in R3 , corresponding to ζ = 0: the positive
x3 direction. If one does this, for z ∈ ((p − 1)µ0 + µ1 , pµ0 − µ1 ), Fp,0
0
represents
the sections in the kernel of ∇3 − i∇0 + z along each cylinder that decay at
∞
infinity in the positive direction; similarly, F−(p−1),0 represents the sections
along each cylinder that decay at infinity in the negative direction. For z ∈
(pµ0 − µ1 , pµ0 + µ1 ), Fp,1
0
represents the sections along each cylinder that decay
∞
at infinity in the positive direction; similarly, F−p,1 represents the sections along
each cylinder that decay at infinity in the negative direction. With these shifts,
the spectral curves represent the lines for which the solutions that decay at
0
one of the ends do not sum to the whole bundle. The quotient F/(Fp,0 +
∞
F−(p−1),0 ) represents this failure and therefore the existence of a solution decay-
ing at both ends. Suppose that z ∈ (pµ0 − µ1 , pµ0 + µ1 ). One has the same for
0 ∞
F/(Fp,1 + F−p,1 ).
Following the Nahm transform heuristic, starting with the caloron, we solve
the Dirac equation for the family of connections shifted by z, and obtain in doing
so a bundle over an interval and a solution to Nahm’s equations on this bundle.
The spectral curve, from this point of view, over ζ = 0, consist of the eigenvalues
56 The Nahm transform for calorons

of A(0, z) = T1 + iT2 (z); there is a sheaf over the curve whose fibre over the point
η in the spectral curve given by the cokernel of A(0, z) − ηI.
A priori, it is not evident what link there is between eigenvalues on a space
of solutions to an equation defined over all of space, and the behaviour of
solutions to an equation along a single line. The link is provided by a remarkable
formulation of the Dirac equation (see for instance Donaldson and Kronheimer
1990 or Cherkis and Kapustin 2001. The idea, roughly, is to write the Dirac
equation as the equations for the harmonic elements of the complex
⎛ ⎞
∇3 + i∇0 + z ⎠
D1 = ⎝  
−∇1 − i∇2 ⊕2
D2 = ∇1 + i∇2 ∇3 + i∇0 + z
L (V ) −−−−−−−−−−−−−−−−−→ L (V ) −−−−−−−−−−−−−−−−−−−−−−−−→ L2 (V ).
2 2

(4.39)
Proposition 4.7 The operators D1 , D2 commute with multiplication by w =
x1 + ix2 , and for self-dual connections, D2 D1 = 0. The kernel Kz = ker(D2 ) ∩
ker(D1∗ ) of the Dirac operator Dz∗ is naturally isomorphic to the cohomology
ker(D2 )/Im(D1 ) of the complex.
On a compact manifold, this equivalence is part of standard elliptic theory.
In the non-compact case, the analysis must be done with care. The simplest
way, for us, is to adapt Nye and Singer (2000, section 4). Using the techniques
developed by Mazzeo and Melrose (1998), they show that the Dirac operator Dz∗
is Fredholm if and only if an ‘operator on fibres along the boundary’ P = (∇0 −
z)∞ + i(η1 e1 + η2 e2 + η3 e3 ) is invertible along all the circles S 1 in S 1 × S∞
2
=
∂(S × R ). Again, the e1 are the Pauli matrices, and the constraint must hold
1 3

for all choices of constants (η1 , η2 , η3 ). Nye and Singer show that it does as long
as z does not have values ±µ1 + nµ0 . In the elliptic complex, the constraint gets
replaced by the essentially equivalent condition that the complex
⎛ ⎞
η
⎝ 3
+ i(∇0 )∞ + z ⎠  
−η1 − iη2 η1 + iη2 , η3 + i(∇0 )∞ + z
⊕2
L (V |S 1 ) −−−−−−−−−−−−−−−→ L (V |S 1 ) −−−−−−−−−−−−−−−−−−−−−−→ L2 (V |S 1 )
2 2

defined over the circle be exact. The equivalence between the Euler characteristic
of the complex and the index of the Dirac operator then goes through establishing
the isomorphism.
We now turn to analysing the way a solution to Nahm’s equations is extracted
from this complex. The operator A(0, z) on the kernel of Dzx is defined by
multiplication by x1 + ix2 followed by the projection on that kernel. On the
cohomology, it is simply multiplication by w = x1 + ix2 , simplifying matters
considerably.
Suppose η is an eigenvalue of A(0, z); hence there is a representative v of
the cohomology class such that (x1 + ix2 − η)(v) = D1 (s) for some s in L2 . In
particular, along w = η in S 1 × R3 , we have (∇3 + i∇0 − z)s = 0, and so (η, ζ) =
(w, 0) belongs to the spectral curve. Conversely, if (∇3 + i∇0 − z)s = 0 along
 Closing the circle 57

w = η, we can build a cohomology class v by extending s to a neighbourhood so

that it is compactly supported in the x1 , x2 directions and satisfies (∇1 + i∇2 )s =
D1 s
0 along w = η; one then sets v = w−η . The spectral curves are thus identified.
0 ∞
We can further identify sections of the quotients F/(Fp,1 + F−p,1 ) (for the
caloron twisted by z) over S1 ∩ {ζ = 0} with sections of the cokernel of A1 (0, z) −
ηI as both are supported over S1 ∩ {ζ = 0}. Indeed, let σ(w) and ρ(x3 ) be smooth
functions such that
 
1, |w| < ; 0, x3 < 0;
σ(w) = and ρ(x3 ) =
0, |w| > ; 1, x3 > 1.

Notice that the derivative (∇1 + i∇2 )(σ(w)) is supported on the annulus <
|w| < 2 .
Consider a ball B centred at η in the fibre above ζ = 0 of radius 2 chosen such
that B ∩ S1 = {η}. The ball parameterizes cylinders S 1 × {w} × R, for w ∈ B,
and thus a section φ ∈ H 0 (B, F ) is in fact a section of V on S 1 × B × R satisfying
(∇3 + i∇0 + z)φ = 0. The holomorphicity condition is then (∇1 + i∇2 )φ = 0.
Let φ0 ∈ H 0 (B, Fp,1
0
|(0,η) ). Then σ(w − η)ρ(x3 )φ0 lies in L2 (S 1 × R3 , V ). Thus
the section D1 (σ(w − η)ρ(x3 )φ0 ) is a coboundary for the complex (4.39). Sim-
∞
ilarly, if φ∞ lies in F−p,1 , then σ(w − η)(ρ(x3 ) − 1)φ∞ also lies in L2 , and its
image by D1 is also a coboundary.
Suppose for simplicity S1 ∩ {ζ = 0} contains only points of multiplicity 1.
Consider now a general section φ ∈ H 0 (B, F ). Away from the spectral curve,
φ decomposes into a sum φ0 + φ∞ , and combining the two constructions above
gives a coboundary K(φ) for the complex (4.39). This decomposition has a pole
0 ∞
at η if φ(η) represents a non-trivial element in the quotient F/(Fp,1 + F−p,1 ).
However, the section
 
K(φ) = D1 σ (w − η)ρ(x3 )φ − ρ∞

  
σ (w − η) ∇3
+ i∇0 + z ρ(x3 )φ − ρ∞
=
− ∇1 + i∇2 σ (w − η)ρ(x3 )φ − ρ∞

is L2 . Because of the pole of φ0 + φ∞ , it is not in D1 (L2 (V )) but by construction

it is definitely in the kernel of D2 . It thus represents a non-trivial cohomology
class for the complex (4.39).
Doing this for each point of S1 ∩ {ζ = 0} expresses Kz as a sum of classes
localized along the lines in S 1 × R3 corresponding to the intersection of the
0 ∞
spectral curve with ζ = 0, identifying Kz with sections of F/(Fp,1 + F−p,1 ) over
0 ∞
ζ = 0. Let Kz (φ) correspond to a non-zero element φ of F/(Fp,1 + F−p,1 ) over
(η, 0).
This class projects non-trivially to the cokernel of A(0, z) − ηI. Indeed,
the action of A(0, z) on Kz is multiplication by w = x1 + ix2 , so if Kz (φ) =
(A(0, z) − ηI)Kz (ψ) = (w − η)Kz (ψ) for some ψ with φ = (w − η)ψ, then the
decomposition φ = φ0 + φ1 is holomorphic and so Kz (φ) = 0.
58 The Nahm transform for calorons

For intersections of higher multiplicity, the case presents only notational diffi-
culties, but is conceptually identical. The choice of ζ made above just simplified
the analysis. Varying it we obtain the following result:

Proposition 4.8 For z ∈ (pµ0 − µ1 , pµ0 + µ1 ), the curves S1 associated to the

caloron and S1 associated to its Nahm transform are identical; there is a natural
0 ∞
isomorphism between the sheaf F/(Fp,1 + F−p,1 ) for the caloron shifted by z and
the cokernel sheaf of A (ζ, z) − η.
1

Similarly, for z ∈ ((p − 1)µ0 + µ1 , pµ0 − µ1 ), the curves S0 associated to the

caloron and S0 associated to its Nahm transform are identical; there is a natural
0 ∞
isomorphism between the sheaf F/(Fp,0 + F−(p−1),0 ) for the caloron shifted by z
and the cokernel sheaf of A0 (ζ, z) − η.

We remark, in particular, for generic solutions, that flowing to a boundary

point gives the divisor D in the intersection of the two curves, and so an
identification of the spectral data.

4.5.2 Starting with a solution to Nahm’s equation

We now want to identify the spectral data for a solution to Nahm’s equations
and the caloron it produces via the Nahm transform.

Case 1: j > 0. Again, as for the opposite direction, it is more convenient to

give a cohomological interpretation of the Dirac equation. This interpretation is
governed by choosing a particular direction in R3 , and so a particular equivalence
of S 1 × R3 with C∗ × C. We choose the direction corresponding to ζ = 0, so
set β = T1 + iT2 , w = x1 + ix2 , α = T0 − iT3 , y = µ0 x0 − ix3 , and consider the
complex
⎛ ⎞
i(dz + α − iy)⎠ 
D1 =⎝ 
−β + iw  2 ⊕2 D2 = β − iwi(dz + α − iy)
L̃2 −−−−−−−−−−−−−−−−−→ L̃1
2
−−−−−−−−−−−−−−−−−−−−−−−−→ L̃2 . (4.40)

For this complex to be defined, we must display a bit of care in our choice of
function spaces:
r L̃22 is the space of sections s of K that are L22 over each interval and such that,
(subscripts denoting the limits on the appropriates sides of the jump points)
π(slarge ) = ssmall , π(dz slarge ) = dz ssmall when j = 1 and slarge = ι(ssmall ),
dz slarge = ι(dz ssmall ) when j > 1.
r (L̃2 )⊕2 is the space of sections (s1 , s2 ) of K ⊕ K that are in L2 over each
1 1
interval, with π(slarge ) = ssmall at the jump points when j = 1 and slarge =
ι(ssmall ) when j > 1.
r L̃2 is the space of L2 sections of K.
 Closing the circle 59

Proposition 4.9 For solutions to Nahm’s equations, D2 D1 = 0, and

 2 
3
D1∗ D1 = − dz + (T0 − iµ0 x0 ) − (Tj − ixj )2 . (4.41)
j=1

So for v = 0 satisfying D1∗ D1 (v) = 0, there is a point-wise relation

  
3
d2z v2 = 2 dz + (T0 − iµ0 x0 ) v + 22
(Tj − ixj )v2  0. (4.42)
j=1

Hence if v is a regular solution vanishing at two points, v = 0.

We look at solutions to the equations D1∗ D1 (v) = 0 near a boundary point ±µ1 ;
one has on the interval (−µ1 , µ1 ) near ±µ1 , a decomposition Ck+j = Ck ⊕ Cj :
the 2k-dimensional space of solutions with boundary values (and derivatives)
in Ck continue outside the interval, and solutions with boundary values in Cj
are governed by the theory of regular singular points for o.d.e.s. As the sum
of the squares of the residues at the boundary points of the Ti is half of the
Casimir element in the enveloping algebra of su(2), its value is −(j − 1)2 /4. The
theory of regular singular o.d.e. then gives, for ρ one of (1 ± j)/2, j-dimensional
spaces of solutions of the form ẑ ρ f (ẑ) to D1∗ D1 (v) = 0, where f (ẑ) is analytic,
at each of the boundary points. Here ẑ is a coordinate whose origin is at the
boundary point.
We can use this knowledge to build a Green’s function. This operation is
somewhat complicated by the poles of the Ti , as we now see in a series of lemmas.
Lemma 4.10 Let x ∈ (−µ1 , µ1 ), and u ∈ Ck+j . There is a unique solution to
D1∗ D1 (v) = δx u on the circle.
Proof. Such a solution must be continuous, smooth on the circle but for a
single jump in the derivative at x, of value u. Let U be the 2k-dimensional space
of solutions on the small side, outside (−µ1 , µ1 ). Those solutions propagate into
the interval from both ends, and for f ∈ U , let f− be the continuation into the
interval (−µ1 , µ1 ) from the −µ1 side, and f+ from the µ1 side. Let V+ and V−
be the j-dimensional spaces of solutions of the form ẑ (1+j)/2 f (ẑ) born at µ1 and
−µ1 , respectively. Consider the map
Rx : U ⊕ V+ ⊕ V− → Ck+j ⊕ Ck+j
!
f− (x) + g− (x) − f+ (x) − g+ (x)
(f, g+ , g− ) →    
.
f− (x) + g− (x) − f+ (x) − g+ (x)

This map must be injective (thus bijective) because of the convexity property of
solutions given by (4.42). If Rx−1 (0, u) = (f, g+ , g− ), the desired Green’s function
v can be chosen by taking f outside of (−µ1 , µ1 ), f + g− on (−µ1 , x), and f + g+
on (x, µ1 ).

60 The Nahm transform for calorons

Lemma 4.11 Let u+ , u− ∈ Ck . There is a unique solution to D1∗ D1 (v) = 0 on

(−µ1 , µ1 ) with values u+ , u− at µ1 , −µ1 , respectively.
Proof. Let W− and W+ be the (k + j)-dimensional affine spaces of solutions
with boundary value u− at −µ1 , and u+ at µ1 , respectively. Consider at an
intermediate point x the affine map
Rx : W+ ⊕ W− → Ck+j ⊕ Ck+j
  

(g+ , g− ) → g− (x) − g+ (x), g− (x) − g+ (x) .
It is injective, otherwise one has, taking differences, an element of ker (D1∗ D1 )
vanishing at both ends; it is thus surjective, and the inverse image of zero gives
solutions whose values and derivatives match at x.

A similar result holds on the other interval:
Lemma 4.12 Let x ∈ (µ1 , µ0 − µ1 ), u+ , u− ∈ Ck , and u ∈ Ck . There is a
unique solution to D1∗ D1 (v) = δx u on the interval, with value u+ at µ1 , and
value u− at µ0 − µ1 .
Combining the two previous lemmas, we obtain the Green’s function for the
small side.
Lemma 4.13 Let x ∈ (µ1 , µ0 − µ1 ) and u ∈ Ck . There is a unique solution to
D1∗ D1 (v) = δx u on the circle.
Proof. Again, one has, varying the boundary values at ±µ1 , an affine 2k-
dimensional space of continuous solutions, with the right jump in derivatives at
x, and possibly extra jumps in the derivatives at ±µ1 . One considers the affine
map from this space to Ck ⊕ Ck taking the jumps in the derivatives at ±µ1 ;
it has to be injective (convexity) and so is surjective, allowing us to match the
derivatives.

By the usual ellipticity argument, the Green’s function solution to D1∗ D1 (v) =
δx u is smooth away from x. By convexity, dz v is increasing everywhere away
from x. Since we are on a circle, both v and dz v must attain their maximum
at x. In addition, integrating over the circle, we get
|u|2  dz <v, v> (x)− − dz <v, v> (x)+

= d2z <v, v>
 
= 2(∇0 − iµ0 x0 )v2 + 2 (Tj − ixj )v2
j

C (∇0 − iµ0 x0 )v2 + v2 .
 Closing the circle 61

The last inequality follows from the fact that the solution to Nahm’s equation is
irreducible. This L21 bound on v ensures continuity, with vL∞  CvL21 , and
hence the L∞ norm of the Green’s function is bounded by a constant times the
norm of u.

Case 2: j = 0. We can again use a cohomological version of the Dirac oper-

ator, but we must modify the function spaces a little bit to account for the
jump at µT1 in the solution
 to Nahm’s equations given by ∆+ (A(ζ)) = (α+,0 +
α+,1 ζ) ᾱ+,1 − ᾱ+,0
T
ζ for suitable column vectors α+,i , and the similar jump
 T 
∆− (A(ζ)) = (α−,0 + α−,1 ζ) ᾱ−,1 − ᾱ−,0
T
ζ at −µ1 . The corresponding jumps for
the matrices Ti are
i 
∆± (T3 ) = T
α±,1 ᾱ±,1 − α±,0 ᾱ±,0
T
,
2
T
∆± (T1 + iT2 ) = α±,0 ᾱ±,1 .

We want solutions to the Dirac equations D1∗ (s1 , s2 ) = D2 (s1 , s2 ) = 0 with

jumps that are multiples of v± := (iα±,1 , α±,0 ) at ±µ1 , so we modify the function
spaces for the complex (4.40):
r L̃22 is the space of sections s of K that are L22 over each interval, continuous at
the jumping points, and with a jump discontinuity ∆± (dz s) = 12 (α±,1 ᾱ±,1
T
+
α±,0 ᾱ±,0 )s(±µ1 ) in derivatives at ±µ1 .
T

r (L̃2 )⊕2 is the space of sections (s1 , s2 ) of K ⊕ K that are L2 over the
1 1
intervals, but with a jump discontinuity ∆± (s1 , s2 ) = c± v± at ±µ1 .
r L̃2 is the space of L2 sections of K.

Proposition 4.14 For solutions to Nahm’s equations and for the complex
defined using those function spaces just defined, Proposition 4.9 holds, and
furthermore, at the jump points,

∆(dz (<v, v>)) = <∆(dz v), v> + <v, ∆(dz v)>

 T  (4.43)
= <ᾱ±,0 T
v, ᾱ±,0 T
v> + <ᾱ±,1 T
v, ᾱ±,1 v>  0

for v = 0 satisfying D1∗ D1 (v) = 0; hence, if v is a regular solution vanishing at

two points, then v = 0.

Proceeding as for j > 0, one can then build a Green’s function. Using
the Green’s function in both cases j > 0 and j = 0, we have the analog of
Proposition 4.7.

Proposition 4.15 The cohomology ker(D2 )/Im(D1 ) of the complex (4.40) is

isomorphic to the space of L2 solutions to the Dirac equation D2 (v) = 0 = D1∗ (v),
which are the harmonic representatives in the cohomology classes. As such, they
have minimal norm in the class.
62 The Nahm transform for calorons

Proof. The cohomology class of a solution to the Dirac equation is non-zero

as the convexity property ensures that if D1∗ D1 (v) = 0 over the whole circle
then v = 0. On the other hand, from a solution to D2 (u) = 0, one uses the
Green’s function to solve D1∗ D1 (v) = D1∗ (u) and gets a harmonic representative
u + D1 (v).

We can now turn to identifying the spectral data of a solution to Nahm’s
equations with the data of the caloron it induces, beginning with the spectral
curves. Suppose we are in the generic situation of spectral curves that are
reduced and with no common components. Let us work over ζ = 0, and suppose
that the intersection of ζ = 0 with the curves is generic, so distinct points of
multiplicity 1.
We turn first to analysing the kernel of the matrices β(z) − iw = T1 (z) +
iT2 (z) − ix1 + x2 , at points w where det(β(z) − iw) = 0 so that we are on
the Nahm spectral curve, let us say for the interval (−µ1 , µ1 ). We note that
since [β, dz + α] = 0, the spectrum is constant along the intervals (−µ1 , µ1 ) or
(µ1 , µ0 − µ1 ), and indeed, if (β(z0 ) − iw)f0 = 0, solving (dz + α)f = 0, f (z0 ) =
f0 gives (β(z) − iw)f = 0 over the interval.
For such an f and a bump function ρ, setting
(s1 , s2 ) = (ρf, 0) (4.44)
defines a cohomology class of the complex (4.40). This class is zero only when
the integral of ρ is zero over the interval as then for some σ, we have (dz +
α)(σf ) = ((dz σ)f, 0) = (ρf, 0). Note that we can modify by a boundary and do
ρ → ρ + dz σ to place the bump anywhere on the interval. This fact is quite useful.
Given this class at a fixed y = 0, one can extend it to other y by taking
(s1 , s2 ) = (eiyz ρf, 0). (4.45)
We can think of this family of cohomology classes as a section over S 1 × R3
of the bundle that Nahm’s construction produces from the solution to Nahm’s
equations. This section has exponential decay as x3 → ±∞. To see this, we
exploit the fact that we can move the bump function ρ around to give differ-
ent representatives supported on (−µ1 , µ1 ) for the class (s1 , s2 ) = (eiyz ρf, 0) =
(e−x3 z eiµ0 x0 ρf ). Recall that the norm of any representative in the cohomology
class is at least the norm of the harmonic (Dirac) representative. As we take x3
to +∞, we move the bump towards µ1 , giving up to a polynomial the bound
exp(−x3 µ1 ) on the L2 norm of the harmonic representative; similarly, as we go
to −∞, we move the bump towards −µ1 , and get a bound of the form exp(x3 µ1 ).
The condition on a section of E over C∗ = {eiµ0 x0 +x3 } to extend over 0, ∞
is given in Garland and Murray (1988, section 3), and amounts to a growth
condition. More generally, there are a whole family of growth conditions, giving
us not only our bundle E but also a family of associated sheaves E(p,q,0)(p ,q ,∞) ,
the sheaf of sections of E over T with poles of order p at T0 , with leading
term in Eq0 , and poles of order p at T∞ , with leading term in Eq∞ . The growth
 Closing the circle 63

condition that E(0,1,0)(0,1,∞) corresponds to is that of growth bounded (up to

polynomial factors) by exp(−x3 µ1 ) as x3 → ∞, and by exp(x3 µ1 ) as x3 → −∞;
for E(0,0,0)(1,0,∞) , it is growth bounded (up to polynomial factors) by exp(x3 µ1 )
as x3 → ∞, and by exp(x3 (µ0 − µ1 )) as x3 → −∞. Let us consider the case of
E(0,1,0)(0,1,∞) . Our bounds for the section defined by (4.45) tell us that it is a
global (holomorphic) section of E(0,1,0)(0,1,∞) on the P1 above (η, ζ) = (w, 0).
The sheaves E(p,q,0)(p ,q ,∞) have degree 2(p + p − 2) + q + q  on the fibres
of the projections to T P1, from our identifications of them given above. In
particular, the degree of E(0,1,0)(0,1,∞) is −2. Now, starting from a point of
the Nahm spectral curve, we have produced a global section of E(0,1,0)(0,1,∞)
over the corresponding Riemann sphere; the Riemann–Roch theorem then tells
us that the first cohomology is also non-zero, so that we are in the support of
R1 π̃∗ (E(0,1,0)(0,1,∞) ), which is exactly the caloron spectral curve S1 . As both S1 ’s
are of the same degree and unreduced, they are identical.
Proceeding similarly for the spectral Nahm spectral curve S0 , we obtain a
global section of E(0,0,0)(1,0,∞) over the corresponding twistor line, and so show
that the line lies in the support of R1 π̃∗ (E(0,0,0)(1,0,∞) ), which is exactly the
caloron spectral curve S0 .
We have now identified the spectral curves for the solutions to Nahm’s
equations and the caloron that it produces. We now note that, in addition to
the bundle V over S 1 × R3 that Nahm’s construction produces, there is another
bundle, V̂ , obtained by taking the spectral data associated to the solution of
Nahm’s equations, and feeding it in to the reconstruction of a caloron from its
spectral data. Indeed, from a solution to Nahm’s equations, we have seen that
we can define sheaves Lz ; the fact that they satisfy the boundary conditions by
Proposition 4.5 tells us we can reconstruct a bundle F of infinite rank over T P1
using the sequence (4.10). We now identify the bundles V and V̂ .
Starting from a caloron, we obtained F as the push-down from the twistor
space T of a rank 2 bundle E. If w is a fibre coordinate on T → T P1 , we saw that
it induced an automorphism W of F , such that, at w = w0 , Ew0  F/Im(W −
w0 ). In the sequence (4.10), the shift map W identifies the nth entry in the middle
column with the (n + 2)th (e.g. L(p−1)µ0 +µ1 (2k + j) ⊗ IS01 with Lpµ0 −µ1 (2k +
j) ⊗ IS10 ) and similarly in the right-hand column. If w = w(ζ) is the equation
of a twistor line L in T , then quotienting by W − w(ζ) in the sequence (4.10)
gives for E over the line (recalling that Lµ0 is trivial over the line)

Lµ1 (2k + j) ⊗ IS01 Lµ1 (2k + j)[−S01 ]|S0

E → ⊕ → ⊕ → 0. (4.46)
L−µ1 (2k + j) ⊗ IS10 L−µ1 (2k + j)[−S10 ]|S1

From the twistor point of view, the space of global sections of E along the
real line Lx in T correspond to the fibre of the caloron bundle V̂ over the
corresponding x ∈ S 1 × R3 . On the other hand, from the Nahm point of view,
64 The Nahm transform for calorons

Vx correspond to ker Dx∗ , and so to the cohomology ker(D2 )/Im(D1 ) in the

complex (4.40).
To identify V and V̂ , we follow closely Hurtubise and Murray (1989, pp. 80–84),
so we simply summarize the ideas. The identification is made for the x whose
lines Lx which do not intersect S0 ∩ S1 ; to do it on this dense set is sufficient,
as one is also identifying the connections.
Let us denote the sequence of sections corresponding to (4.46) by

0 → H 0 (Lx , E) → Uµ1 ⊕ U−µ1 → W0 ⊕ W1 . (4.47)

What Hurtubise and Murray (1989) do is to identify W0 and W1 with solutions

to Dx∗ (s) = 0 on the interval (µ1 , µ0 − µ1 ) and (−µ1 , µ1 ), respectively, and Uµ1
and U−µ1 with L2 solutions on a neighborhood of µ1 and −µ1 , respectively.
The kernel H 0 (Lx , E) then gets identified with global L2 solutions on the circle,
which are elements of Vx . Since from the twistor point of view V̂x  H 0 (Lx , E),
we are done. The case j = 0 is similar.
Having identified the bundles, one then wants to ensure that the connections
defined in both case are the same, and we do so again according to Hurtubise and
Murray (1989). It suffices to do this identification along one null plane through
the point x, as changing coordinates will do the rest, and we can suppose that
x is the origin, so η = 0 in T P1 . Let us choose the plane corresponding to ζ = 0.
From the twistor point of view, parallel sections along this null plane correspond
to sections on the corresponding family of lines with fixed values at ζ = 0. From
the Nahm point of view, a parallel section along this null plane is a family of
cohomology classes represented by cocycles s = (s1 , s2 ) in ker(D2 ) that satisfy
∂
( ∂x 1
∂
+ i ∂x 2
∂
)s = 0, ( ∂x 0
− i ∂x
∂
3
)s = 0, modulo coboundaries. In particular, the
derivatives of s2 are of the form (T1 + iT2 )φ = A0 φ for a suitable φ depending
on x. Following through the identifications of Hurtubise and Murray (1989
pp. 80–84) tells us that the derivatives ∇ŝ in suitable trivializations of the
corresponding sections ŝ on the family of lines are of the form ∇ŝ = A0 ψ
for the corresponding section ψ. However, the defining relation for A(ζ) is
(ηI − A0 − ζA1 − ζ 2 A2 )φ = 0 over the spectral curve. At the boundaries of the
inverval, it extends to all of T . In particular, at (ζ, η) = 0, A0 φ vanishes as a
function, and so ŝ is parallel, as desired.

4.5.3 From Nahm to caloron to Nahm to caloron

Let us first place ourselves in the generic situation. We have seen first that the
family of generic calorons maps continuously and injectively into the family of
generic spectral data; secondly, that generic spectral data and generic solutions
to Nahm’s equations are equivalent; thirdly, that starting from a caloron and
producing a solution to Nahm’s equations then taking the spectral data of both
objects gives the same result, and fourthly, that the generic solution to Nahm’s
equation gives the same caloron, whether you pass through the Nahm transform
or through the spectral data via the twistor construction. Taken together, those
 Closing the circle 65

facts tell us that the map from caloron to spectral data is a bijection, that all
three sets of data are equivalent, and that the six transforms between them are
pairwise inverses of each other.
Let us now leave the generic set. If we now take an arbitrary solution T
to Nahm’s equations, again satisfying all the conditions, we can fit it into a
continuous family T (t) with T (0) = T with T (t) generic for t = 0; that it can
be done follows from our description of moduli given in Section 4.6. The Nahm
transform of this family is a continuous family C(t) of calorons, which in turn
produces a continuous family of solutions T̃ (t) to Nahm’s equations. We do not
know a priori whether the boundary and symmetry conditions are satisfied at
t = 0. However, for t = 0, T̃ (t) = T (t), and so T̃ (0) = T (0), and we are done; the
transform Nahm to caloron to Nahm is the identity.
If the caloron C lies in the closure of the set of generic calorons (presumably
all calorons do, but it needs to be proved, perhaps using the methods of Taubes
(1984) to show that the moduli space is connected), we again fit it as C(0) into
a family C(t) of calorons, with C(t) generic for t = 0, all of same charge. The
Nahm transform produces a family T (t) of solutions to Nahm’s equations, with
T (t) satisfying all the conditions for t = 0. We also get a family S(t) of spectral
data, which for all t corresponds to both C(t) and T (t). We need to show that
T (0) satisfy all the conditions.
Taking a limit, it is fairly clear that the symmetry condition is satisfied also
for t = 0. The boundary conditions are less obvious. From the small side, there
is no problem, as the vanishing theorem (proposition 4.3) holds at C(0), and
so the solutions to Nahm’s equations are continuous at the boundary of the
interval. From the large side, what saves us is the rigidity of representations
into SU (2); indeed, the polar parts of the solutions are given by representations,
and so are fixed, in a suitable family of unitary gauges, hence preserved in a
limit. The process, given in Section 4.4.2, of passing to a constant gauge by
solving ds/dz + A+ (z)s = 0 applies in the limit near the singular points. We
have in the limit the same type of transformations relating the basis of K in
which one has the solution to Nahm’s equations to the continuous basis obtained
from lifting up and pushing down as to obtain the sequence (4.16), with the
same process of producing endomorphisms A(z, ζ)(t), giving us again in the
limit a continuous endomorphism of a bundle V over T P1 × S 1 , with V in fact
lifted from P1 × S 1 . The restrictions V (±µ1 ) (on the large side) are of fixed
type Vµ11 = Ok ⊕ Vj−1,−j+1 for all t. The summand Vj−1,−j+1 corresponds to
∞
the subsheaf O(j) of the sheaf F/(Fp,1 0
+ F−p,1 ) ⊗ L±µ1 , which exists in the
limit. The polar part C(p(ζ, t)) of (4.26), mapping Vj−1,−j+1 to itself, has limit
C(p(ζ, 0)). By (4.27), this latter limit is determined by the spectral curves, and
so is well defined. The other summand Ok is also well defined in the limit, as it
is the piece transferred from the small side, which is still Ok in the limit because
of the vanishing theorem. The off-diagonal vanishing as one goes back to the
trivialization for Nahm’s equations is simply a consequence of the polar behaviour
of solutions to ds/dz + A+ (z)s = 0, and of the normal form for A(z, ζ)(t) as in
66 The Nahm transform for calorons

(4.25). In short, the limit has exactly the same normal form, and so the same
boundary behaviour.
Finally, there is the question of irreducibility. The reducible solutions corre-
spond to calorons for which charge has bubbled off; as our family has constant
charge, this is precluded. The limit solutions satisfies all the conditions necessary
to produce a caloron by the opposite Nahm transform; as the transform is
involutive on the generic member of the family, it is also involutive in the limit,
and so the circle closes.
In short one has the following theorem:

Theorem 4.16

1. There is an equivalence between

a. Generic calorons of charge (k, j)
b. Generic solutions to Nahm’s equations on the circle satisfying the
conditions of Section 4.2.1.2
c. Generic spectral data
The equivalences of (a) and (b) are given in both directions by the Nahm
transform.
2. The Nahm transforms give equivalences between
a. Calorons of charge (k, j), in the closure of the generic set
b. Solutions to Nahm’s equations on the circle satisfying the conditions of
Section 4.2.1

4.6 Moduli
The equivalence exhibited above allows us to classify calorons by classifying
appropriate solutions to Nahm’s equations. One has to guide us through the
example of monopoles, as classified in Donaldson (1984) and Hurtubise (1989);
the (framed) monopoles for gauge group G, with symmetry breaking to a torus
at infinity, and of charge k are classified by the space Ratk (P1 , GC /B) of based
degree k rational maps from the Riemann sphere into the flag manifold GC /B.
As our calorons are Kač–Moody monopoles, we should have the same theorem,
with framed calorons equivalent to rational maps from P1 into the loop group.
Following an idea developed in Atiyah (1984), one thinks of these as bundles on
P1 × P1 , with some extra data of a flag along a line, and some framing.
The rank 2 bundle corresponding to an SU (2) caloron, again following the
example of monopoles, should be the restriction of the bundle E on T̃ corre-
sponding to the caloron, to the inverse image of a point, say ζ = 0, in P1 . This
inverse image is P1 × C; along the divisor {0} × C, the bundle has the flag E0 ,
and along {∞} × C, the bundle has the flag E∞ . We extend this bundle to
infinity, using a framing, in such a way that E∞ is a trivial subbundle, and E0
has degree −j. This bundle will be corresponding to the caloron.
 Moduli 67

The theorem, however, is proven in terms of solutions to Nahm’s equations;

indeed, we show in Charbonneau and Hurtubise (2007) that both bundles
and solutions to Nahm’s equations satisfying the appropriate conditions are
describable in terms of a geometric quotient of a family of matrices.

Theorem 4.17 Let k  1, j  0 be integers. There is an equivalence between

1. Vector bundles E of rank 2 on P1 × P1 , with c1 (E) = 0, c2 (E) = k, trivi-
alized along P1 × {∞} ∪ {∞} × P1 , with a flag φ : O(−j) → E of degree j
along P1 × {0} such that φ(∞)(O(−j)) = span(0, 1).
2. Framed, irreducible solutions to Nahm’s equations on the circle, with rank k
over (µ1 , µ0 − µ1 ), rank k + j over (−µ1 , µ1 ), with the boundary conditions
defined above, modulo the action of the unitary gauge group.
3. Complex matrices A, B (k × k), C (k × 2), D (2 × k), A (j × k), B  (1 ×
k), C  (j × 2), satisfying the monad equations
[A, B] + CD = 0,


B
A + S(j, 0)A − A B − C  D = 0,
0
 
−e+ A + 1 0 D = 0,
and the genericity conditions
⎞ ⎛
A−y
⎝B − x⎠ is injective for all x, y ∈ C,
D
 
x−B A−y C is surjective for all x, y ∈ C,
⎛ ⎞

x − B A C 0
⎜ −B  ⎟
⎜ A C  x − S(j, 0)⎟ is surjective for all x ∈ C,
⎝ 0 ⎠
 
0 0 10 −e+






A C2 C2 j−1 C2
M · · · M is an isomorphism,
A C2 C2 C2
modulo the action of Gl(k, C) given by
(A, B, C, D, A , B  , C  ) → (gAg −1 , gBg −1 , gC, Dg −1 , A g −1 , B  g −1 , C  ).
Here Ci , Ci are the ith column 
 of C, C , and Di the ith row of D, (4.32) gives
S(j, 0), and e+ := 0 · · · 0 1 ,
⎛ ⎞

B  −C1 e+
M := ⎝ B ⎠.
S(j, 0) − C1 e+
0
68 The Nahm transform for calorons

We see in Charbonneau and Hurtubise (2007) that B and M are conjugate

to the matrix T1 + iT2 on the intervals (µ1 , µ0 − µ1 ) and (−µ1 , µ1 ), respectively.
Choosing B diagonal, and B and M with distinct and disjoint eigenvalues, it
is not difficult to build solutions to the various matrix constraints. Hence, the
spectral curves for the obtained solutions to Nahm’s equations are reduced and
have no common components, as the spectral curve over ζ = 0 is the spectrum
of T1 + iT2 . Thus, generic generic solutions to Nahm’s equations, and so generic
calorons, exist.

Acknowledgement
This research was conducted while the first author was at McGill University
and was supported by an NSERC PDF. The second author is supported by
NSERC and FQRNT grants. The diagrams in this chapter were created using
Paul Taylor’s Commutative Diagram package.

References
Adams, M. R., Harnad, J., and Hurtubise, J. (1990). Isospectral Hamiltonian flows in
finite and infinite dimensions. II. Integration of flows. Commun. Math. Phys. 134(3),
555–85.
Atiyah, M. F. (1984). Instantons in two and four dimensions. Commun. Math.
Phys. 93(4), 437–51.
Atiyah, M. F., Drinfel
d, V. G., Hitchin, N. J., and Manin, Yu. I. (1978). Construction
of instantons. Phys. Lett. A 65(3), 185–7.
Boyer, C. P., Mann, B. M., Hurtubise, J. C., and Milgram, R. J. (1994). The topology
of the space of rational maps into generalized flag manifolds. Acta Math. 173(1),
61–101.
Braam, P. J. and van Baal, P. (1989). Nahm’s transformation for instantons. Commun.
Math. Phys. 122(2), 267–80.
Bruckmann, F. and van Baal, P. (2002). Multi-caloron solutions. Nuclear Phys.
B 645(1–2), 105–33.
Bruckmann, F., Nógrádi, D., and van Baal, P. (2003). Constituent monopoles through
the eyes of fermion zero-modes. Nuclear Phys. B 666(1–2), 197–229.
Bruckmann, F., Nógrádi, D., and van Baal, P. (2004). Higher charge calorons with
non-trivial holonomy. Nuclear Phys. B 698(1–2), 233–54.
Chakrabarti, A. (1987). Periodic generalizations of static, self-dual SU(2) gauge fields.
Phys. Rev. D (3) 35(2), 696–706.
Charbonneau, B. (2004). Analytic aspects of periodic instantons. PhD thesis, Sep-
tember, Massachusetts Institute of Technology, Cambridge, MA. Available at
http://hdl.handle.net/1721.1/26746.
Charbonneau, B. (2006). From spatially periodic instantons to singular monopoles.
Commun. Anal. Geom. 14(1), 183–214.
Charbonneau, B. and Hurtubise, J. (2008).Calorons, Nahm’s equations on S 1 and
bundles over P1 × P1 . Commun. Math. Phys. 280(2), 315–49.
Cherkis, S. and Kapustin, A. (2001). Nahm transform for periodic monopoles and
N = 2 super Yang–Mills theory. Commun. Math. Phys. 218(2), 333–71.
 References 69

Corrigan, E. and Goddard, P. (1984). Construction of instanton and monopole solutions

and reciprocity. Ann. Phys. 154(1), 253–79.
Donaldson, S. K. (1984). Nahm’s equations and the classification of monopoles. Com-
mun. Math. Phys. 96(3), 387–407.
Donaldson, S. K. and Kronheimer, P. B. (1990). The geometry of four-manifolds. Oxford
Mathematical Monographs. The Clarendon Press Oxford University Press, New York.
Garland, H. and Murray, M. K. (1988). Kac–Moody monopoles and periodic instantons.
Commun. Math. Phys. 120(2), 335–51.
Gravesen, J. (1989). On the topology of spaces of holomorphic maps. Acta Math.
162(3–4), 247–86.
Hitchin, N. J. (1982). Monopoles and geodesics. Commun. Math. Phys. 83(4), 579–602.
Hitchin, N. J. (1983). On the construction of monopoles. Commun. Math. Phys. 89(2),
145–90.
Hurtubise, J. (1989). The classification of monopoles for the classical groups. Commun.
Math. Phys. 120(4), 613–41.
Hurtubise, J. and Murray, M. K. (1989). On the construction of monopoles for the
classical groups. Commun. Math. Phys. 122(1), 35–89.
Hurtubise, J. and Murray, M. K. (1990). Monopoles and their spectral data. Commun.
Math. Phys. 133(3), 487–508.
Jardim, M. (2001). Construction of doubly-periodic instantons. Commun. Math.
Phys. 216(1), 1–15.
Jardim, M. (2002a). Classification and existence of doubly-periodic instantons. Q. J.
Math. 53(4), 431–42.
Jardim, M. (2002b). Nahm transform and spectral curves for doubly-periodic instan-
tons. Commun. Math. Phys. 225(3), 639–68.
Jardim, M. (2004). A survey on Nahm transform. J. Geom. Phys. 52(3), 313–27.
Kraan, T. C. (2000). Instantons, monopoles and toric hyperKähler manifolds. Commun.
Math. Phys. 212(3), 503–33.
Kraan, T. C. and van Baal, P. (1998a). Exact T -duality between calorons and Taub-
NUT spaces. Phys. Lett. B 428(3–4), 268–76.
Kraan, T. C. and van Baal, P. (1998b). Monopole constituents inside SU (n) calorons.
Phys. Lett. B 435, 389–95.
Kraan, T. C. and van Baal, P. (1998c). Periodic instantons with non-trivial holonomy.
Nuclear Phys. B 533(1–3), 627–59.
Lee, K. (1998). Instantons and magnetic monopoles on R3 × S 1 with arbitrary simple
gauge groups. Phys. Lett. B 426(3–4), 323–8.
Lee, K. and Lu, C. (1998). Su(2) calorons and magnetic monopoles. Phys. Rev. D
(2) 58, 025011.
Lee, K. and Yi, S.-H. (2003). 1/4 BPS dyonic calorons. Phys. Rev. D (3) 67(2),
025012, 9.
Mazzeo, R. and Melrose, R. B. (1998). Pseudodifferential operators on manifolds with
fibred boundaries. Asian J. Math. 2(4), 833–66.
Nahm, W. (1983). All self-dual multimonopoles for arbitrary gauge groups. In Structural
Elements in Particle Physics and Statistical Mechanics (Freiburg, 1981), Vol. 82 of
NATO Advanced Study Institute Series B: Physics, pp. 301–10. Plenum, New York.
Nahm, W. (1984). Self-dual monopoles and calorons. In Group Theoretical Methods in
Physics (Trieste, 1983), Vol. 201 of Lecture Notes in Physics, pp. 189–200. Springer,
Berlin.
70 The Nahm transform for calorons

Nógrádi, D. (2005). Multi-calorons and their moduli. PhD thesis, Institute Lorentz
for Theoretical Physics, University of Leiden, Leiden, the Netherlands. arXiv:hep-
th/0511125.
Norbury, P. (2000). Periodic instantons and the loop group. Commun. Math.
Phys. 212(3), 557–69.
Nye, T. M. W. (2001). The geometry of calorons. PhD thesis, University of Edinburgh,
Edinburgh, Scotland. arXiv:hep-th/0311215.
Nye, T. M. W. and Singer, M. A. (2000). An L2 -index theorem for Dirac operators on
S 1 × R3 . J. Funct. Anal. 177(1), 203–18.
Schenk, H. (1988). On a generalised Fourier transform of instantons over flat tori.
Commun. Math. Phys. 116(2), 177–83.
Szabó, S. (2007). Nahm transform for integrable connections on the Riemann sphere.
Mém. Soc. Math. Fr. (N.S.) No. 110.
Taubes, C. H. (1984). Path-connected Yang–Mills moduli spaces. J. Differential Geom.
19(2), 337–92.
Ward, R. S. (2004). Symmetric calorons. Phys. Lett. B 582(3–4), 203–10.
 V
NAHM’S EQUATIONS AND FREE-BOUNDARY PROBLEMS

Simon K. Donaldson
Dedicated to Nigel Hitchin, with gratitude and affection

5.1 Introduction
In Donaldson (1984), following up work of Hitchin (1983), the author found it
useful to express Nahm’s equations, for a matrix group, in terms of the motion
of a particle in a Riemannian symmetric space, subject to a potential field. This
point of view lead readily to an elementary existence theorem for solutions of
Nahm’s equation, corresponding to particle paths with prescribed end points.
The original motivation for this chapter is the question of formulating an
analogous theory for the Nahm equations associated to the infinite-dimensional
Lie group of area-preserving diffeomorphisms of a surface – in the spirit of
Donaldson (1993). We will see that this can be done, and that a form of the
appropriate existence theorem holds – essentially a special case of a result of
Chen. However the main focus of the chapter is not on existence proofs but on
the various formulations of the problem, and connections between them. In these
developments, one finds that the natural context is rather more general than the
original question, so we will start out of a different tack, and return to Nahm’s
equations in Section 5.5.
Consider the following set-up in Euclidean space R3 , in which we take coordi-
nates (x1 , x2 , z) – thinking of z as the vertical direction. (We will use the notation
∂ ∂
∂xi = ∂i , ∂z = ∂z .) Suppose we have a strictly positive function H(x1 , x2 ). This
defines a domain

ΩH = {(x1 , x2 , z) : 0 < z < H(x1 , x2 )},

whose boundary has two components {z = 0} and {z = H}. We consider the

Dirichlet problem for the standard Laplacian: to find a harmonic function θ
on ΩH with θ = 0 on {z = 0} and θ = 1 on {z = H}. To set up this problem
precisely, let us assume that the data H is Z2 periodic on R2 , and seek a Z2 -
periodic solution θ. Now we have a unique solution θ to our Dirichlet problem.
Consider the flux of the gradient of θ through the boundary {z = H}. This
defines another function ρ on R2 . To be precise, if ιH is the obvious map from
R2 to the boundary {z = H} then the flux is defined by

ι∗H (∗dθ) = ρdx1 dx2 . (5.1)

72 Nahm’s equations and free-boundary problems

Explicitly
ρ = ∂z θ − (∂1 θ∂1 H + ∂2 θ∂2 H),
with the right-hand side evaluated at (x1 , x2 , H(x2 , x2 )). By the maximum
principle, ρ is a positive function, since the normal derivative of θ is positive
in the positive z direction on {z = H}. We consider the following free-boundary
problem: Given a positive periodic function ρ does it arise from some periodic
H, and is H unique?
One can gain some physical intuition for this question by supposing that the
lower half-space {z ≤ 0} represents a body with an infinite specific heat capacity
fixed at temperature 0 and ΩH corresponds to a layer of ice covering this body.
We choose units so that the melting temperature of the ice is 1. Sunlight shines
vertically downwards onto the upper surface {z = H} of the ice, but with a
variable intensity so that heat is transmitted to the surface according to the
density function ρ. We suppose that the surface of the ice is sprinkled by rain,
which will instantly freeze if the surface temperature of the ice is less than 1. We
also suppose that a wind blows across the surface, instantly removing any surface
water. Then we see that the solution to our free-boundary problem represents
a static physical state, in which the upper surface of the ice is just at freezing
point, the lower surface is at the imposed sub-freezing temperature, and the
heat generated by the given sunlight flows through the ice without changing the
temperature. Physical intuition suggests that there should indeed be a unique
solution.
We can express the free-boundary problem considered above as a special case
of another question. Suppose now that we have a pair of periodic functions
H0 , H1 on R2 with H0 < H1 . Then we have a domain ΩH0 ,H1 = {H0 (x1 , x2 ) <
z < H1 (x1 , x2 )}, with two boundary components. Let θ be the harmonic function
in this domain equal to 0, 1 on {z = H0 }, {z = H1 }, respectively. Then we obtain
a pair of flux functions ρ0 , ρ1 as before. By Gauss’ theorem, these satisfy a
constraint
 
ρ0 dx = ρ1 dx, (5.2)
[0,1]2 [0,1]2

since [0, 1]2 is a fundamental domain for the Z2 action. Obviously, if we replace
H0 , H1 by H0 + c, H1 + c for any constant c we get the same pair ρ0 , ρ1 . We ask:
given ρ0 , ρ1 satisfying the integral constraint (5.2), is there a corresponding pair
(H0 , H1 ), and if so is the solution unique up to the addition of a constant? A
positive answer to this question implies a positive answer to the previous one, by a
simple reflection argument. (Given ρ, as in the first problem, take ρ0 = ρ1 = ρ/2.
Then uniqueness implies that the solution has reflection symmetry about θ = 1/2
and we get a solution to the first problem by changing θ to 2θ − 1.)
Of course we can also imagine a physical problem corresponding to this second
question: for example, a layer of ice in the region ΩH0 ,H1 . We can now vary
the problem by supposing that in place of ice we have a horizontally stratified
 Introduction 73

material in which heat can only flow in the horizontal directions. Thus the
steady-state condition, for a temperature distribution θ(x1 , x2 , z), is
 2 
∂1 + ∂22 θ = 0. (5.3)

We define flux functions ρ0 , ρ1 by pulling back the two-form

∂1 θ dx2 dz − ∂2 θ dx1 dz,

and the integral constraint (5.2) still holds. So we ask: given ρ0 , ρ1 satisfying
(5.2), is there a pair H0 , H1 and a function θ on ΩH0 ,H1 , equal to 0, 1 on the
two boundary components, which has these fluxes, and is the solution essentially
unique? (In this case, one has to relax the condition on the domain to H0 ≤ H1 .)
It is natural to extend these questions to a general compact-oriented Rie-
mannian manifold X (which would be the flat torus R2 /Z2 in the discussion
above). Write dµ for the Riemannian volume form on X. We fix a real parameter
≥ 0 and define a map ∗ from T ∗ (X × R) to Λn T ∗ (X × R) by

∗ dz = dµ, ∗ α = (∗X α)dz,

for α ∈ T ∗ X, where ∗X is the usual Hodge ∗ operator on X. Then for a function

θ on a domain in X × R

d ∗ dθ = (∆ θ)dzdµ,

where
 
∆ θ = − ∂z2 + ∆X θ,

with ∆X the standard Laplace operator on X. (We use the sign convention that
∆X is a positive operator, so when = 1 our ∆ is the standard Laplace operator
on X × R.) If θ is defined on a domain ΩH0 ,H1 , as above, we define the flux ρi
on the boundary {z = Hi } by pulling back ∗ dθ just as before. We consider a
pair of functions ρ0 , ρ1 > 0 with
  
ρ0 dµ = ρ1 dµ = dµ
X X X

and we ask

Question 5.1 Is there a pair H0 ≤ H1 and a function θ on the set ΩH0 ,H1 ⊂
X × R with θ = 0, 1 on the hypersurfaces {z = H0 }, {z = H1 }, with fluxes ρi and
with ∆ θ = 0? If so, is the solution essentially unique?

For any > 0 the equation ∆ θ = 0 can be transformed into the standard
Laplace equation on the product, by rescaling the z variable. When = 0 the
equation has a very different character: it is not elliptic and we obviously do not
have automatic interior regularity with respect to z.
74 Nahm’s equations and free-boundary problems

5.2 An infinite-dimensional Riemannian manifold

We now start in a different direction. Given our compact Riemannian manifold
X we let H be the set of functions φ on X such that 1 − ∆X φ > 0. We make H
into a Riemannian manifold, defining the norm of a tangent vector δφ at a point
φ by

δφφ =
2
(δφ)2 (1 − ∆X φ)dµ.
X

Thus a path φ(t) in H, parametrized by t ∈ [0, 1] say, is simply a function on

X × [0, 1] and the “energy” of the path is
 
2
1 1 ∂φ
(1 − ∆X φ)dµdt. (5.4)
2 0 X ∂t
When X is two-dimensional and orientable, this definition coincides with the
metric on the space of “Kähler potentials” discussed by Mabuchi, Semmes, and
the author (Mabuchi 1987; Semmes 1992; Donaldson 1999). The general context
in those references is a compact Kähler manifold: here we are considering a
different extension of the two-dimensional case, and we will see that some new
features emerge. The account below follows the approach in Donaldson (1999)
closely.
It is straightforward to find the Euler–Lagrange equations associated to the
energy (5.4). These are

|∇X φ̇|2
φ̈ = .
1 − ∆X φ
These equations define the geodesics in H. We can read off the Levi-Civita
connection of the metric from this geodesic equation, as follows. Let φ(t) be
any path in H and ψ(t) be another function on X × [0, 1], which we regard as
a vector field along the path φ(t). Then the covariant derivative of φ along the
path is given by
dφ
Dt φ = + (Wt , ∇X ψ), (5.5)
dt
where
−1
Wt = ∇X φ̇
1 − ∆X φ
and ( , ) is the Riemannian inner product on tangent vectors to X. (We write
∇X , or sometimes just ∇, for the gradient operator on X, so Wt is a vector
field on X.) This has an important consequence for the holonomy group of the
manifold H. Observe that the tangent space to H at a point φ is the space of
functions on X endowed with the standard L2 inner product associated to the
 An infinite-dimensional Riemannian manifold 75

measure

dµφ = (1 − ∆φ)dµ0 .

So, in a general way, the parallel transport along a path from φ0 to φ1 should
be an isometry from L2 (X, dµφ0 ) to L2 (X, dµφ1 ). (Here we are ignoring the
distinction between, e.g., smooth functions and L2 functions.) What we see from
(5.5) is that this isometry is induced by a diffeomorphism f : X → X with

f ∗ (dµφ1 ) = dµφ0 . (5.6)

The diffeomorphism is obtained by integrating the time-dependent vector field

Wt and (5.6) follows from the identity


1 d
LWt dµφ = d ∗ dφ̇ ∗X dµφ = ∆φ̇ = − µφ .
1 − ∆X dt
(Here L denotes the Lie derivative on X.) We conclude that the holonomy
group of H is contained in the group G of volume-preserving diffeomorphisms of
(X, dµ0 ), regarded as a subgroup of the orthogonal group of L2 (X, dµ0 ). (This
can also be expressed by saying that there is an obvious principal G bundle over
H with the tangent bundle as an associated vector bundle, and the Levi-Civita
connection is induced by a connection on this G bundle.)
We now move on to discuss the curvature tensor of H. Let φ be a point of
H and let α, β be tangent vectors to H at φ—so α and β are just functions on
X. The curvature Rα,β should be a linear map from tangent vectors to tangent
vectors: that is, from functions on X to functions on X. The discussion of the
holonomy above tells us that this map must have the form

Rα,β (ψ) = (να,β , ∇ψ), (5.7)

for some vector field να,β on X, determined by φ, α, and β. Moreover we know

that we must have

Lνα,β (dµφ ) = 0.

To identify this vector field we introduce some notation. For vector fields v, w on
X we write v × w for the exterior product: a section of the bundle Λ2 T X. We
define a differential operator

curl : Γ(Λ2 T X) → Γ(T X)

to be the composite of the standard identification:

Λ2 T X ∼
= Λn−2 T ∗ X
(using the Riemannian volume form dµ), the exterior derivative

d : Γ(Λn−2 T ∗ X) → Γ(Λn−1 T ∗ X),

76 Nahm’s equations and free-boundary problems

and the standard identification

Λn−1 T ∗ X ∼
= TX
(using the volume form dµ again). Then we have
Theorem 5.1 The curvature of H is given by (5.7) and the vector field


1 1
να,β = curl ∇α × ∇β .
1 − ∆φ 1 − ∆φ
Corollary 5.1 The manifold H has non-positive sectional curvature.
The sectional curvature corresponding to a pair of tangent vectors α, β at a point
φ is
Kα,β = Rα,β (α), β.
In our case this is

Kα,β = (να,β , ∇α)β(1 − ∆X φ)dµ.
X

Unwinding the algebraic identifications we used above, the integrand can be

written in terms of differential forms as


1 1
dα ∧ d ∗ (dα ∧ dβ) β(1 − ∆X φ).
1 − ∆X φ 1 − ∆X φ
So


1
Kα,β = dα ∧ d ∗ (dα ∧ dβ) β.
X 1 − ∆X φ
Applying Stokes’ theorem this is
 
1 1
Kα,β = − dα ∧ dβ ∧ ∗(dα ∧ dβ) = − |dα ∧ dβ|2 dµ ≤ 0.
X 1 − ∆X φ X 1 − ∆X φ

In the proof of Theorem 5.1 we will make use of two identities. For any pair
of vector fields v, w and function f
curl (v × w) = [v, w] + (div v)w − (div w)v (5.8)

curl (f (v × w)) = f curl (v × w) + (v, ∇f )w − (w, ∇f )v. (5.9)

We leave the verification as an exercise. (Considering geodesic coordinates we
see that it suffices to treat the case of Euclidean space. Our notation has been
chosen to agree with standard notation in the case of vector fields in R3 .)
To calculate the curvature we consider a two-parameter family φ(s, t) in H,
with a corresponding vector field ψ(s, t) along the family. Then we will compute
the commutator (Dt Ds − Ds Dt )ψ(s, t). Evaluating at φ = φ(0, 0) this is Rα,β (ψ)
where ψ = ψ(0, 0), α = ∂s φ, β = ∂t φ.
 An infinite-dimensional Riemannian manifold 77

Now we write
∂ ∂
Ds = + W s , Dt = + Wt ,
∂s ∂t
where the vector fields Ws , Wt are regarded as operators on the functions on X.
So Dt Ds − Ds Dt is the operator given by the vector field
∂Ws ∂Wt
ν= − − [Ws , Wt ],
∂t ∂s
and ν is exactly the vector field να,β we need to identify. Recall that
−∇∂s φ −∇∂t φ
Ws = , Wt = .
1 − ∆φ 1 − ∆φ
So


∂Ws −1 ∂2φ 1
= ∇ + ∆∂s φ∇∂t φ.
∂t 1 − ∆φ ∂s∂t (1 − ∆φ)2
Evaluating at s = t = 0 where ∂s φ = α, ∂t φ = β we have
∂Ws ∂Wt 1
− = (∆α∇β − ∆β∇α).
∂t ∂s (1 − ∆φ)2
Write g for the function (1 − ∆φ)−1 . Combining with the Lie bracket term we
obtain
να,β = [g∇α, g∇β] + g 2 (∆α∇β − ∆β∇α).

Now applying (5.8) we have

[g∇α, g∇β] = curl (g 2 ∇α × ∇β) − div (g∇α)g∇β + div (g∇β)g∇α.

Applying (5.9) we have

curl (g 2 ∇α × ∇β) = g curl (g∇α × ∇β) + g((∇g, ∇α)∇β − (∇g, ∇β)∇α).

Since

div (g∇α) = g∆α + (∇g, ∇α), div (g∇β) = g∆β + (∇g, ∇β)

we see that

να,β = g curl (g∇α × ∇β),

as required.
In the case when X has dimension 2—as discussed in Donaldson (1999),
Mabuchi (1987), and Semmes (1992)—the space H is formally a symmetric space.
This is not true in general, since the curvature tensor is not preserved by the
action of the group G.
78 Nahm’s equations and free-boundary problems

We define a functional on H by

V (φ) = φ dµ.
X

This function is convex along geodesics in H, since the geodesic equation implies
φ̈ ≥ 0. Now introduce a real parameter ≥ 0 and consider the functional on paths
in H:

1 2
E= |φ̇| + V (φ) dt, (5.10)
2 φ
corresponding to the motion of a particle in the potential − V . The Euler–
Lagrange equations are
|∇X φ̇|2 +
φ̈ = . (5.11)
1 − ∆X φ

5.3 Three equivalent problems

In this section we will explain that there are three equivalent formulations of the
same PDE problem associated to a compact Riemannian manifold X. We have
essentially encountered two of these already.
r θ equation. This is the problem we set up in Section 5.1. We are given positive
functions ρ0 , ρ1 on X, with
 
ρi dµ = dµ. (5.12)
X X

We seek a domain ΩH0 ,H1 ⊂ X × R defined by H0 , H1 : X → R and a func-

tion θ on ΩH0 ,H1 , equal to 0, 1 on the two boundary components, with fluxes
ρ0 , ρ1 and satisfying the equation

∆ θ = 0.
r Φ equation. Here we are given φ0 , φ1 on X, with 1 − ∆φi > 0. We seek a
function Φ on X × [0, 1], equal to φ0 , φ1 on the two boundary components,
with 1 − ∆Φ > 0 for all t and satisfying the non-linear equation


∂2Φ ∂Φ 2
(1 − ∆X Φ) − |∇ | = . (5.13)
∂t2 ∂t
As we have explained in Section 5.2, this is the same as finding a path in the
space H, with end points φ0 , φ1 , corresponding to the motion of a particle
in the potential − V .

Now we introduce the third problem.

 Three equivalent problems 79

r U equation. We are given positive functions φ0 , φ1 , with 1 − ∆φi > 0, as

above. Define a function L on X × R by
L(x, z) = max(φ0 (x) − φ1 (x) + z, 0).
We seek a C function U (x, z) on X × R with U ≥ L everywhere and
1

satisfying the equation

∆ U = (1 − ∆φ0 ) (5.14)
on the open set Ω where U > L.
The equivalence of these three problems (assuming suitable regularity for
the solutions in each case) arises from elementary, but not completely obvious,
transformations. We describe these now.

5.3.1 θ equation =⇒ φ equation

∂θ
Suppose we have a solution θ on a domain ΩH0 ,H1 . Then ∂z θ = ∂z is positive
on the boundary components of ΩH0 ,H1 . The function ∂z θ satisfies the equation
∆ (∂z θ) = 0 and it follows from this that ∂z θ is positive throughout the domain.
This implies that, for any t ∈ [0, 1], the set θ−1 (t) is the graph of a smooth
function ht on X. By definition h0 = H0 and h1 = H1 . We also write this function
as h(t, x) where convenient. For each fixed t we can define a function ρt on X
by the flux of ∗ dθ, just as before.
We claim that
∂ρt
= ∆X ht (5.15)
∂t
We show this by direct calculation (there are more conceptual, geometric
arguments). For simplicity
 we treat the ∂case when the metric on X is locally
Euclidean, so ∆X = − ∂i2 where ∂i = ∂x i
, for local coordinates xi . The identity
θ(x, ht (x)) = t
implies that
∂i θ + ∂z θ ∂i h = 0 (5.16)
and
∂z θ ∂t h = 1. (5.17)
Now

∆X ht = − (∂i + (∂i h)∂z )∂i h,
i

and this is

∂i θ


∂i θ

∆X ht = − ∂i − ∂z −
i
∂z θ ∂z θ
80 Nahm’s equations and free-boundary problems

which is

∂ 2θ ∂i θ∂i ∂z θ ∂i θ∂i θ∂z ∂z θ

− i
−2 + .
i
∂z θ (∂z θ)2 (∂z θ)3

On the other hand the flux ρt is given by pulling back the differential form ∗ dθ
on the product by the map x → (x, ht (x)) and this gives
1 
ρt = ∂z θ + (∂i θ)2 .
∂z θ i

So

 2

∂ρ 1 i (∂i θ)
= ∂z ∂z θ + .
∂t ∂z θ ∂z θ
This is
  2
∂z ∂z θ ∂i θ∂i ∂z θ i (∂i θ) ∂z ∂z θ
∂t ρt = +2 i
− .
∂z θ (∂z θ)2 (∂z θ)3

So we see that ∂t ρt = −∆X ht , since ∂z ∂z θ = − i ∂i ∂i θ.
Now the normalization (5.13) implies that there is a function φ0 on X such
that ρ0 = 1 − ∆X φ0 . For t > 0 we define φt by
 t
φt = φ0 + hτ dτ.
0
We can also regard this family of functions as a single function Φ on X × [0, 1],
Then (5.15) implies that ρt = 1 − ∆X φt for each t. We have
∂2Φ 1
= ∂t h =
∂t2 ∂z θ
and
1 
1 − ∆X Φ = ∂z θ + (∂i θ)2 .
∂z θ i
So

2
∂2Φ ∂i θ
2
(1 − ∆X Φ) = + .
∂t ∂z θ
i

Now since
−1
∂i ∂t Φ = ∂i ht = ∂i θ
∂z θ
we can write the above as
2
∂2Φ ∂
(1 − ∆X Φ) = + ∇X Φ ,
∂t2 ∂t
as required.
 Three equivalent problems 81

5.3.2 Φ equation =⇒ U equation

Here we suppose we have a solution Φ(x, t) of the Φ equation and we write
Φ(x, 0) = φ0 , Φ(x, 1) = φ1 . We essentially take the Legendre transform in the t-
variable. The discussion is slightly more complicated when = 0, so for simplicity
we treat the case when > 0 and ∂t2 Φ is strictly positive. Write H1 (x), H2 (x)
for the derivatives ∂t Φ evaluated at (x, 0), (x, 1), respectively, so H0 < H1 . We
calculate first in the open set ΩH0 ,H1 . For each fixed x ∈ X and each z in the
interval (H0 (x), H1 (x)) there is a t = t(x, z) such that z = ∂t Φ. We set
U (x, z) = Φ(x, 0) − Φ(x, t) + zt.
This defines a function U in ΩH0 ,H1 . We define U outside this set by setting
U (x, z) = 0 if z ≤ H0 (x) and U (x, z) = L(x, z) = φ0 − φ1 − z if z ≥ H1 (x). It
follows from the definitions that U is C 1 , that U ≥ L, and that the set where
U > L is exactly ΩH0 ,H1 . We calculate on this set. Then ∂z U = t and
∂z2 U = (∂t2 Φ)−1 . (5.18)
Differentiating with respect to the parameters xi we have
∂i U = ∂i φ0 − ∂i Φ,
and
∂t ∂ 2 Φ
∂i2 U = ∂i2 φ0 − ∂i2 Φ − .
∂xi ∂t∂xi
Differentiating the identity z = ∂t Φ gives
∂2Φ ∂ 2 Φ ∂t
0= + 2 ,
∂t∂xi ∂t ∂xi
so we can write
1
∂i2 U = ∂i2 φ0 − ∂i2 Φ + (∂t ∂i Φ)2 .
∂t2 Φ
Summing over i and using (5.18) for ∂z2 U we obtain
1  
∂z2 U − ∆X U = + |∇Xt Φ|2 − ∆X φ0 + 1,
∂t2 Φ
and so
∆ U = 1 − ∆X φ0 .

5.3.3 U equation =⇒ θ equation

Now suppose we have a solution U of ∆ U = ρ0 in a domain ΩH0 ,H1 , satisfying
the appropriate boundary conditions, where ρ0 = 1 − ∆X φ0 . We set
∂U
θ= .
∂z
82 Nahm’s equations and free-boundary problems

Then ∆ θ = 0 and θ = 0, 1 on the two boundary components. We have to check

that the fluxes of ∗ dθ on the boundary components are ρi = 1 − ∆X φi . Consider
first the boundary component where z = H0 . The flux is
|∇X θ|2 1 
∂z θ + = ∂z2 F + 2 (∂z ∂i F )2 .
∂z θ ∂z F
Now we have identities
(∂i F )(x, H0 (x)) = 0, (∂z F )(x, H0 (x)) = 0.
Differentiating the first of these with respect to xi we get
∂i2 F + ∂i H0 ∂i ∂z F = 0
on the boundary. Differentiating the second gives
∂i ∂z F + ∂i H0 ∂z2 F = 0
on the boundary. Combining these we have
(∂z ∂i F )2 = (∂z2 F )(∂i2 F ).
Hence the flux is

∂z2 F + ∂i2 F = ρ0 .
i

The argument for the other boundary component {z = H1 (x)} is similar.

5.4 Existence results and discussion

We have set up a class of PDE problems associated to any compact Riemannian
manifold, and seen that these have three equivalent formulations. In this section
we will make some remarks about existence results, and comparison with the free-
boundary literature. This discussion is unfortunately rather incomplete, mainly
due to the author’s limited grasp of the background.

5.4.1 Monge–Ampère and the results of Chen

For a function Φ on X × (0, 1) write q(Φ) for the non-linear differential operator
∂ 2
q(Φ) = ∂t2 Φ(1 − ∆X Φ) − |∇X Φ| .
∂t
So our “Φ equation” is q(Φ) = . When X has dimension 1—a circle with local
coordinate x—we can write ∆X = −∂x2 and the equation is the real Monge–
Ampère operator

2 
∂t Φ ∂x ∂t Φ
q(Φ)= det
∂x ∂t Φ 1 + ∂x2 Φ
When X has dimension 2 the operator can be expressed as a complex Monge–
Ampère operator. That is, we regard X as a Riemann surface and identify
 Existence results and discussion 83

the Laplace operator on X with i∂∂. We take the product with a circle, with
coordinate α, and let τ = t + iα be a complex coordinate on the Riemannn
surface S 1 × (0, 1). Then, in differential form notation, our non-linear operator
is given by

(ω0 + i∂∂Φ)2 = q(Φ)ω0 dτ dτ ,

where ω0 is the Riemannian area form of X lifted to X × S 1 × (0, 1). Our

Dirichlet problem becomes a Dirichlet problem for S 1 -invariant solutions of this
complex Monge–Ampère equation on X × S 1 × (0, 1). This was studied by Chen
(2000), and it follows from his results that, for any > 0 there is a unique
solution to our problem, and hence an affirmative answer to Question 5.1 in
this case. (Chen does not state this result explicitly, but it follows from the
continuity method developed in Chen 2000, section 3, that for any strictly
positive smooth function ν on X × [0, 1] there is a solution of the equation
q(Φ) = ν with prescribed boundary values φ0 , φ1 .)
It seems quite likely that the techniques used by Chen can be extended to
the higher dimensional case. The foundation for this should be provided by a
convexity property of the non-linear operator which we will now derive. Let A
be a symmetric (n + 1) × (n + 1) matrix with entries Aij 0 ≤ i, j ≤ n. Define


n 
n
Q(A) = A00 Aii − A2i0 .
i=1 i=1

Thus Q is a quadratic function on the vector space of symmetric (n + 1) × (n + 1)

matrices.

Lemma 5.1

1. If A > 0 then Q(A) > 0 and if A ≥ 0 then Q(A) ≥ 0.

2. If A, B are matrices with Q(A) = Q(B) > 0 and if the entries A00 , B00 are
positive then for each s ∈ [0, 1]

Q(sA + (1 − s)B) ≥ Q(A), Q(A − B) < 0.

Moreover, if A = B then strict inequality holds.

To see the first item, observe that we can change basis in Rn ⊂ Rn+1 to reduce
to the case when the block Aij , 1 ≤ i, j ≤ n is diagonal, say, with entries bi . Then
if A ≥ 0 we have A00 bi ≥ A20i and so
  
Q(A) = A00 bi − A20i ≥ 0,

with strict inequlality if A > 0.

84 Nahm’s equations and free-boundary problems

For the second item, we just have to observe that Q is induced from a quadratic
form of Lorentzian signature on Rn+2 by the linear map

n
π : A → (A00 , Aii , A0i ).
i=1

The hypotheses imply that π(A) and π(B) are in the same component of a hyper-
boloid defined by this Lorentzian form and the statements follow immediately
from elementary geometry of Lorentz space.
Using this lemma we can deduce the uniqueness of the solution to our Dirichlet
problem, in any dimension.
Proposition 5.1 If φ0 , φ1 ∈ H then there is at most one solution Φ of the
equation Q(Φ) = on X × [0, 1] with 1 − ∆X Φ > 0 for all t and with Φ(x, 0) =
φ0 (x), Φ(x, 1) = φ1 (x).
We show that the functional E(Φ) given by (5.10) is convex with respect to the
obvious linear structure. Thus we consider a one-parameter family Φs = Φ + sψ,
with the fixed end points. We have
 1
d
E(Φs ) = 2Φ̇s ψ̇(1 − ∆X Φs ) − Φ̇2 ∆X ψ.
ds 0 X

Integrating by parts (just as in the derivation of the geodesic equation) we

obtain
 1
d
E(Φs ) = (q(Φs ) − ) ψ dµ.
ds 0 X

Suppose that Φ0 , Φ1 are two different solutions, so when s = 0, 1 the term

q(Φs ) − in the above expression vanishes pointwise. Item (2) in Lemma 5.1
implies that for s ∈ (0, 1) we have q(Φs ) − ≥ 0, with strict inequality some-
where. This means that E(Φ1 ) > E(Φ0 ). Interchanging the roles of Φ0 , Φ1 we
obtain the reverse inequality, and hence a contradiction.
One can also prove this uniqueness using the maximum principle. Note too that
the uniqueness is what one would expect, formally, from the negative curvature
of the space H and the convexity of the functional V .

5.4.2 Comparison with the free-boundary literature

The author is not at all competent to make this comparison properly. Suffice it
to say, first, that the problem we are considering is very close to those which
have been studied extensively in the applied literature. For example, in the θ-
formulation, the condition of prescribing the pull-back of the flux on the free
boundary is the same as that in the classical problem of the “porous dam” (Elliot
and Ockendon 1982, chapter 4.4; Baiocchi and Capelo 1984, chapter 8), but with
the difference that in that case ρ is constant and there are additional boundary
conditions on other boundary components. Second, the constructions we have
 Existence results and discussion 85

introduced in Section 5.3 all appear in this literature. The transformation from
θ to Φ taking the harmonic function θ as a new independent variable is called in
Crank (1984, chapter 5) the “isothermal migration method.” The transformation
from the formulation in terms of θ to that in terms of U is known as the Baiocchi
transformation (Elliot and Ockendon 1982; Baiocchi and Capelo 1984; Crank
1984). The transformation of the free-boundary problem for a linear equation to
a non-linear Dirichlet problem is used in Kinderlehrer and Nirenberg (1979) to
derive fundamental regularity results.
An important feature of the U -formulation is that it admits a variational
description. Recall that we are given a function L = max(φ0 − φ1 + z, 0) on X ×
R and we seek a C 1 function U with U ≥ L satisfying the equation ∆ U = ρ0 on
the set where U > L. This can be formulated as follows. We fix a large positive
M and consider the functional

1
EM (U ) = |∇X U |2 + |∂z U |2 − ρ0 U dµdz
2
over the space of functions satisfying the constraint U ≥ L, where the integral
is taken over X × [−M, M ] in X × R (which, a posteriori, should contain the
set ΩH0 ,H1 on which U > L). Then the solution minimizes EM over all functions
U ≥ L. This can be used to give another proof of the uniqueness of the solution
to our problem. It seems likely that it could also be made the basis of an existence
proof, following standard techniques in the free-boundary literature. Now recall
that our Φ-formulation was based on a variational principle, with Lagrangian
(5.10). To relate the two, we consider any function Φ on X × [0, 1] with ∂t ∂t Φ ≥ 0
and define U by the recipe of Section 5.3. We suppose that −M < ∂t Φ(x, 0) and
∂t Φ(x, 1) < M for all x ∈ X. Then we have

Proposition 5.2 The functional EM (U ) is


1
E(Φ) + M (1 − ∆X φ0 )(φ0 − φ1 ) + |∇(φ1 − φ0 )|2 dµ
X 2

2  
M
+ + M dµ − φ1 dµ.
2 X X

Thus if we fix M and the end points φ0 , φ1 the two functionals differ by a

constant. The central step in the proof is the fact that the integrals
 1
∂t2 Φ|∇X Φ|2 dµ dt,
0 X

 1 
∆X Φ(∂t Φ)2 dµ dt
0 X

are equal modulo boundary terms. We leave the full calculation as an exercise
for the reader.
86 Nahm’s equations and free-boundary problems

5.4.3 Degenerate case

So far, in this section, we have discussed the case when > 0. In that case the
equations we are studying are elliptic. The degenerate case, when = 0, is much
more delicate. In fact Chen’s main concern in Chen (2000) was to obtain results
about this case, taking the limit as tends to 0. Chen shows that the Dirichlet
problem for Φ, with = 0, has a C 1,1 solution but the question of smoothness is
open. The formulation of the problem in terms of the function U has particular
advantages here, because the problem is set-up as a family of elliptic problems,
and the issue becomes one of smooth dependence on parameters. (This is
related to another approach, involving families of holomorphic maps, discussed
in Semmes 1992 and Donaldson 1999.) We can express the central question as
follows. Suppose we have a smooth function λ on a compact Riemannian manifold
X and fix a smooth positive function ρ. Let J be the functional

1
J(u) = |∇u|2 − ρu.
X 2
For each z ∈ R we set λz = max(λ, z) and minimize the functional J over the
set of functions u ≥ λz . Suppose we know that there is a minimizer uz which
is smooth on the open set Ωz ⊂ X where uz > λz . Let Ω = {(x, z) : x ∈ Ωz } ⊂
X × R.
Question 5.2 In this situation, does uz vary smoothly with z in Ω?
The interesting case here seems to be when z is a critical value of g.

5.5 Relation with Nahm’s equations

We recall that Nahm’s equations are a system of ODE for three functions T1 , T2 ,
and T3 taking values in a fixed Lie algebra:
dTi
= [Tj , Tk ], (5.19)
dt
where i, j, k run over cyclic permutations of 1, 2, 3. To simplify notation, let us
fix on the Lie algebra u(n). It is equivalent (at least in the finite-dimensional
case) to introduce a fourth function T0 and consider the equations
dTi
+ [T0 , Ti ] = [Tj , Tk ], (5.20)
dt
with the action of the “gauge group” of U (n)-valued functions u(t):
du −1
Ti → uTi u−1 , T0 → uT0 u−1 − u
dt
which preserved solutions to (5.20) (i.e. using the gauge group we can transform
T0 to 0). The equations imply that
d
(T2 + iT3 ) = [T0 + iT1 , T2 + iT3 ], (5.21)
dt
 Relation with Nahm’s equations 87

so T2 + iT3 moves in a single adjoint orbit in the Lie algebra of GL(n, C).
Conversely if we fix some B in this complex Lie algebra, introduce a function
g(t) taking values in GL(n, C) and define skew-Hermitian matrices Ti (t) by
dg −1
T0 + iT1 = g ,
ds

T2 + iT3 = gBg −1 ,

then two of the three Nahm equations are satisfied identically. The remaining
equation can be expressed in terms of the function h(t) = g ∗ (t)g(t), taking values
in the space H of positive definite Hermitian matrices, which we can also regard
as the quotient space GL(n, C)/U (n). This equation for h(t) is a second-order
ODE which is the Euler–Langrange equation for the Lagrangian

dh
E(h) = | |2H + VB (h)dt.
dt
Here | |H denotes the standard Riemannian metric on H. The function V on H
is

VB (h) = Tr(hBh−1 B ∗ ).

If g is any element of GL(n, C) with g ∗ = h then

VB (h) = |gBg−1 |2 ,

so VB is determined by the norm of matrices in the adjoint orbit of B. (See

Donaldson (1984) for details of the manipulations involved in all the above.)
The result in Donaldson (1984), mentioned in the introduction of this chapter, is
that for any two points h0 , h1 ∈ H there is a unique solution h(t) to the Euler–
Lagrange equations for t ∈ [0, 1] with h(0) = h0 , h(1) = h1 .
These constructions go over immediately to the case when U (n) is replaced
by any compact Lie group and GL(n, C) by the complexified group. We want
to extend them to the situation where U (n) is replaced by the group G of
Hamiltionian diffeomorphisms of a surface Σ with a fixed area form (or more
precisely, the extension of this group given by a choice of Hamiltionian). The
essential difficulty is that this group does not have a complexification. However,
as explained in Donaldson (1999), Mabuchi (1987), and Semmes (1993), the
space H of Kähler potentials behaves formally like the quotient space G c /G for a
fictitious group G c . Thus the problem we have formulated in Section 5.2 can be
viewed as an analogue of the desired kind provided that our potential function
V can be seen as an analogue of VB in the finite-dimensional case.
If we have a path φt in H with φ0 = 0 and a function β : Σ → C we can write
down a differential equation for a one-parameter family βt which corresponds,
formally, to the adjoint action of the complexified group G c , with the initial
88 Nahm’s equations and free-boundary problems

condition β0 = β. The equation has the shape

∂βt
= ∇φ̇∂βt .
∂t
The problem is that this evolution equation will not have solutions, even for a
short time, in general. But if we suspend for a moment our assumption that we
are working over a compact Riemann surface and suppose that β is a holomorphic
function then there is a trivial solution βt = β. So, formally, the functional Vβ
on H is given by the L2 norm of β with respect to the measure dµφ :

(1 − ∆X φ)|β|2 .

Even if this integral is divergent, the variation with respect to compactly

supported variations in φ is well defined, and this is what corresponds to the
gradient of VB appearing in the equations of motion. Moreover, we can integrate
by parts to get another formal representation of a functional with the same
variation
 
− φ∆X |β|2 = φ|∇β|2 .

Now take the compact Riemann surface Σ to be a two-torus, and identify the
space H with periodic Kähler potentials on the universal cover C. On this cover
the identity function β is holomorphic, and we see from the above that the formal
expression

Vβ = φ
C

is analogous to the function VB in the finite-dimensional case. Of course the

integrand is periodic and so the integral will be divergent but we can return to
the compact surface Σ and consider the well-defined functional

Vβ (φ) = φ
Σ

which will generate the same equations of motion. So we see that, modulo some
blurring of the distinction between Σ and its universal cover, the functional we
have been considering is indeed analogous to that in the finite-dimensional case.
Using the transformation from the Φ equation to the θ equation, we obtain
a relation between Nahm’s equations for the Hamiltonian diffeomorphisms of a
surface and harmonic functions on R3 . This can be seen in other ways. Most
directly, we consider three one-parameter families of functions hi (t) on a surface
Σ with an area form which satisfy
dhi
= {hj , hk }, (5.22)
dt
 Relation with Nahm’s equations 89

where { , } is the Poisson bracket. We think of these as a one-parameter family

of maps ht : Σ → R3 , and assume for simplicity that these are embeddings, with
disjoint images. Then it is a simple exercise to show that (5.22) imply that
the images ht (Σ) are the level sets of a harmonic function on a domain in R3 .
From another point of view, the geometric structure defined by a solution to the
Φ equation is an S 1 invariant Kähler metric Ω = ω0 + i∂∂Φ on Σ × S 1 × (0, 1)
with volume form

Ω2 = dτ dβdτ dβ.

Since dτ dβ is an S 1 -invariant holomorphic two-form, what we have is an S 1 -

invariant hyperkähler structure. Then the relation with harmonic functions
appears as the Gibbons–Hawking construction for hyperkähler metrics.
The development above is rather limited, since we have only been able to
formulate an analogue of our Nahm’s equation problem for a single function
β. One can go further, and arrive at other interesting free-boundary problems.
Consider for example the case when the surface Σ is the two-sphere with the
standard area form, and the orientation-reversing map σ : Σ → Σ given by
reflection in the x1 , x2 plane. Now consider maps β : Σ → C with β = β ◦ σ
which are diffeomorphisms on each hemisphere. Then the push-forward of the
area form on the upper hemisphere defines a two-form ρβ on C with support
in a topological disc β(Σ) ⊂ C. (The form ρβ will not usually be smooth, but
will behave like d−1/2 where d is the distance to the boundary of β(Σ).) Clearly
the form ρβ determines β up to the action of the σ-equivariant Hamiltionian
diffeomorphisms of Σ. Suppose that h is a σ-invariant function on Σ. We can
regard this as an element of the Lie algebra of G c and consider its action on
β. This is given by ∆C h where h is thought of as a function on C, vanishing
outside β(Σ). So a reasonable candidate for a model of the quotient of the
space of maps β by the action of G c is given by the following. We consider
two-forms ρ supported on topological discs in C, with singularities at the
boundary of the kind arising above, and impose the equivalence relation that
ρ0 ∼ ρ1 if there is a compactly supported harmonic function F on C with
∆F = ρ0 − ρ1 .
Now let θ(x1 , x2 , z) be a harmonic function on an open set Ω ⊂ R3 , with
θ(x1 , x2 , z) = θ(x1 , x2 , −z). Suppose that Ω is diffeomorphic to S 2 × (0, 1), and
that θ = 0 on the inner boundary component Σ0 and θ = 1 on the outer
boundary component Σ1 . Suppose also that the projections of Σ0 , Σ1 to the
(x1 , x2 ) plane are diffeomorphisms on each upper hemisphere, mapping to a
pair of topological disc D0 ⊂ D1 . Then the flux of ∇θ on each boundary
component pushes forward to define a pair of compactly supported two-forms
ρ0 , ρ1 on C. These are equivalent in the sense above, since ρ0 − ρ1 = ∆C F for the
function

∂θ
F (x1 , x2 ) = z dz,
∂z
90 Nahm’s equations and free-boundary problems

where the integral is taken over the intersection of the vertical line through
(x1 , x2 , 0) with Ω. Our hypotheses imply that F ≥ 0, and F is supported on the
larger disc D1 .
The question we are lead to is the following:
Question 5.3 Suppose that D0 ⊂ D1 are topological discs in C, that ρi are
two-forms supported on Di , and that there is a non-negative function F on C,
supported on D1 , with ρ0 − ρ1 = ∆C F (where the Laplacian is defined in the
distributional sense). Do ρ0 , ρ1 arise from a unique harmonic function θ on a
domain in R3 , by the construction above?
(For simplicity we have not specified precisely what singularities should be
allowed in the forms ρi : this specification should be a part of the question.)
Hitchin showed in Hitchin (1983) that Nahm’s equations form an integrable
system. The root of this is the invariance of the conjugacy class given by
(5.21), together with the family of similar statements that arise from the SO(3)
action on the set-up. In this vein, we can write down infinitely many conserved
quantities for the solutions of (5.11) on the Riemannian manifold H. Let fλ be
an eigenfunction of the Laplacian ∆X , with eigenvalue λ > 0. Then we have
Proposition 5.3 For any > 0, if φt satisfies (5.11) then the quantity
 " 
λ
exp φ̇ fλ (1 − ∆φ) dµ,
X

does not vary with t.

This becomes rather transparent in the θ-formulation, using the fact that the
function
" 
λ
Kλ (x, z) = fλ (x) exp z ,

satisfies ∆ Kλ = 0.

Acknowledgements
The author is grateful to Professors Colin Atkinson, Xiuxiong Chen, Darryl
Holm, and John Ockendon for helpful discussions.

References
Baiocchi, C. and Capelo, A. (1984). Variational and Quasivariational Inequalities.
Wiley, New York.
Chen, X.-X.(2000). The space of Kahler metrics. Journal of Differential Geometry 56,
189–234.
Crank, J. (1984). Free and Moving Boundary Problems. Oxford University Press,
Oxford.
 References 91

Donaldson, S. (1984). Nahm’s equations and the classification of monopoles. Commu-

nications in Mathematical Physics 96, 387–407.
Donaldson, S. (1993). Complex cobordism, Ashtekar’s equations and diffeomorphisms.
In: Symplectic Geometry, D. Salamon (ed.). Cambridge University Press, Cambridge,
pp. 45–55.
Donaldson, S. (1999). Symmetric spaces, Kahler geometry and Hamiltonian dynamics.
In: Northern California Symplectic Geometry Seminar, Y. Eliashberg et al. (eds.).
American Mathematical Society, Providence, RI, pp.13–33.
Donaldson, S. (2002). Holomorphic discs and the complex Monge-Ampère equation.
Journal of Symplectic Geometry 1, 171–96.
Elliot, C. and Ockendon, J. (1982). Weak and Variational Methods for Moving Bound-
ary Problems. Pitman, Boston, MA.
Hitchin, N. (1983). On the construction of monopoles. Communications in Mathemat-
ical Physics 89, 145–90.
Kinderlehrer, D. and Nirenberg, L. (1979). Regularity in free boundary problems.
Annali della Scuola Normale Superiore di Pisa 4(2), 373–91.
Mabuchi, T. (1987). Some symplectic geometry on Kahler manifolds, I. Osaka Journal
of Mathematics 24, 227–52.
Semmes, S. (1992). Complex Monge-Ampère and symplectic manifolds. American
Journal of Mathematics 114, 495–550.
 VI
SOME ASPECTS OF THE THEORY OF HIGGS PAIRS

S. Ramanan
Dedicated to Nigel Hitchin

This is an exposition of certain aspects of the theory of Higgs pairs (also called
‘Higgs bundles’ by some authors), naturally with some emphasis on topics in
which I played a part. This is an area to which fundamental contributions have
been made by Nigel Hitchin, and I take pleasure in presenting this summary on
the occasion of his sixtieth birthday.

6.1 Moduli of vector bundles

Let C be a projective nonsingular curve over C, or what is the same, a compact
Riemann surface. We will denote its genus by g. We will assume that g ≥ 2
throughout. It is well known that isomorphism classes of line bundles (or divisor
classes) of degree 0 on C are in one-to-one correspondence with unitary characters
of the fundamental group π. Indeed every character of π gives rise to a line
bundle, namely, the one associated to the universal covering which is a principal
π-bundle. It is this association which gives rise to the above bijection.
Since the abelianization of π is free of rank 2g, the space of unitary characters
of π can be identified with the 2g-fold product of S 1 . Thus our remark above
implies that there is a natural bijection of the moduli of line bundles of degree
0 with the 2g-dimensional (real) torus.
The moduli space of line bundles of fixed degree d comes with a natural
complex structure. Indeed, it is a nonsingular projective variety J d . When d = 0,
what we get is a natural structure of a projective variety on the torus, in other
words, an abelian variety. This is the classical Jacobian variety J of C.
If we think of a line bundle as a (fibred) family of one-dimensional spaces,
parametrized by points of C, the ambiguity in identifying the fibres with a
standard one-dimensional vector space, say, C, is given by the action of C∗ on C.
We can then ask for a higher rank, or non-abelian analogue of this construction.
In other words, one considers vector bundles of rank n, instead of line bundles
and replaces the abelian group C∗ by the non-abelian group GL(n, C).
Again, as before, any unitary representation ρ of π gives rise to a vector bundle
Eρ . This was considered by A. Weil in 1939, who observed that vector bundles
associated to unitary representations ρ and ρ are isomorphic if and only if ρ is
equivalent to ρ (Weil 1938). Moreover, any infinitesimal deformation of a unitary
representation ρ clearly gives rise to one of Eρ . If ρ is irreducible, this gives an
injective map from the infinitesimal deformation space of ρ into that of Eρ . This
is just the tangent space level analogue of the above statement.
 Moduli of vector bundles 93

Weil also observed that actually, every infinitesimal deformation of Eρ arises

from an infinitesimal deformation of ρ. In other words, if the set of isomorphism
classes of all vector bundles of a given rank and degree 0 had the structure of
a variety, the map from (the set of equivalence classes of) irreducible unitary
representations to the space of vector bundles would be bijective at the infin-
itesimal level. Finally he also remarked that although not every vector bundle
arises from a unitary representation, those that so arise ‘play, without doubt, an
important part’.
There things stood for a quarter century, till Mumford, thanks to his newly
developed tool of ‘Geometric Invariant Theory’, based on ideas of Hilbert,
came up with a new notion. He showed that the bundles that are classifiable
in the sense of allowing a natural structure of a variety in the set of their
isomorphism classes can be characterized purely from the algebraic geometric
point of view. He called them ‘stable’ bundles. This is defined in terms of the slope
of a vector bundle. Indeed, to every vector bundle E of rank n, one associates a
line bundle, its determinant, namely, the nth exterior power Λn (E). The degree
of the latter is called the degree of E as well. The slope is the rational number
µ(E) = degree(E)/rank(E).

Definition 6.1 A vector bundle E is stable if for every proper subbundle F , we

have µ(F ) < µ(E). A vector bundle E is polystable if E is isomorphic to a direct
sum ⊕Ei of stable subbundles Ei all of whose slopes are the same (as that of E).

The fundamental theorem of Narasimhan and Seshadri (1965) resolved the

puzzle completely.

Theorem 6.1 (Narasimhan and Seshadri 1965) A vector bundle of rank n

is isomorphic to one associated to a unitary representation π → U (n) if and only
if it is polystable of degree 0. This gives a bijection between equivalence classes
of unitary representations of π, and isomorphism classes of polystable vector
bundles of degree 0. Under this bijection, irreducible representations correspond
to stable bundles.

If in the definition of stable bundles, we replace the strict inequality < by

the non-strict inequality ≤, we get the notion of semi-stability. Every semistable
bundle E admits a filtration by subbundles such that the associated graded
bundle is polystable. Although the filtration itself is not unique, the associated
graded polystable bundles given by any two such filtrations are isomorphic. Hence
one can associate to any semistable bundle E, a polystable bundle Gr(E), well
defined up to isomorphism.

Definition 6.2 A vector bundle E is said to be S-equivalent to E  if Gr(E) is

isomorphic to Gr(E  ).

It turns out that from the moduli point of view, it is better to think of the
moduli space as the set of S-equivalence classes of semi-stable bundles, rather
than as isomorphism classes of polystable bundles.
94 Some aspects of the theory of Higgs pairs

Theorem 6.2 (Seshadri 1967) The set of S-equivalence classes of semi-

stable bundles of a given rank n and degree d has a natural structure of a normal
(irreducible) projective variety of dimension n2 (g − 1) + 1.
We will denote the above variety corresponding to rank n and degree d by
UC (n, d) or simply U (n, d) where C is understood. If the rank is 1 and the
degree 0, we get back the Jacobian J. If n = 1 but the degree d is arbitrary, the
corresponding moduli space is denoted J d . Clearly J is a group variety under
tensor product, and it acts simply transitively on J d for all d.
There is a natural morphism of U (n, d) into J d namely the association of
det(E) = Λn (E) to E. Notice that ξ → ξ n gives a morphism J → J which is a
Galois covering with Galois group Jn , the group {ξ ∈ J : ξ n = 1} of n-division
points of the Jacobian. Under tensor product operation, J (and in particular,
Jn ) acts on U (n, d). If SU (n, d) denotes the subspace of U (n, d) consisting of
bundles with some fixed determinant, then the action of Jn leaves SU (n, d)
invariant. Thus the Jn -Galois covering J → J gives rise to the associated bundle
over J with SU (n, d) as fibre. This can be identified with U (n, d). Thus, for the
most part, the study of U (n, d) can be reduced to one of SU (n, d) (together with
the action of Jn on it).
In the 1960s and 1970s, these varieties were studied intensely, especially by
Narasimhan, Newstead, Seshadri, and myself. I will recall here some of the results
obtained, in order to facilitate the rest of this survey.
Firstly, when n and d are relatively prime, the varieties U (n, d) and SU (n, d)
are smooth projective. Under this assumption, semi-stability and stability of
vector bundles are equivalent, and indeed it is true that points of U (n, d)
(respectively, SU (n, d)) corresponding to stable bundles are always smooth. The
tangent space at a stable point E to U (n, d) (respectively, SU (n, d)) can be
identified with H 1 (C, End(E)) (respectively, H 1 (C, ad(E)), where ad(E) denotes
the bundle of trace-free endomorphisms). It is easy to verify that H 0 (C, End(E))
is one-dimensional (respectively, H 0 (C, ad(E)) = 0). Hence the dimension of
the tangent space at E to U (n, d) (respectively, SU (n, d)) is n2 (g − 1) + 1
(respectively, (n2 − 1)(g − 1)). Moreover, SU (n, d) is simply connected and its
second Betti number is 1. Let us denote by Θ the unique ample line bundle whose
Chern class generates H 2 (SU (n, d), Z). The determinant of the tangent bundle
is ample and is an even multiple of Θ.

6.2 Hecke correspondence

One of the tools which Narasimhan and I came up with (Narasimhan and
Ramanan 1975) in the study of U (n, d) is the Hecke correspondence. This sets
up a correspondence between U (n, d) (respectively, SU (n, d)) and U (n, d − 1)
(respectively, SU (n, d − 1)). The idea is very simple. Let E be a vector bundle
of rank n and degree d. Fix a point a of C and take a nonzero linear form
on the fibre Ea . It can be considered as a homomorphism of the locally free
sheaf defined by E onto the structure sheaf of the point a. If E is sufficiently
 Moduli of Higgs pairs 95

stable, which is generically the case, the kernel is also stable and therefore defines
a point of SU (n, d − 1). Incidentally, the kernels corresponding to two linear
forms are isomorphic if they are scalar multiples of each other. Hence the kernel
is determined by an element of P (Ea∗ ).
Two points E, E  of SU (n, d) are Hecke modifications of each other if there
exist elements l ∈ P (Ea∗ ) and l ∈ P (Ea∗ ) such that ker(l) is isomorphic to ker(l ).
Thus we have a closed subvariety HC of SU (n, d) × SU (n, d) which is invariant
under the transposition of coordinates.
For a generic E ∈ SU (n, d + 1) with determinant ξ ⊗ O(a) we can thus see that
P (Ea∗ ) is contained in SU (n, d) with determinant ξ. These projective spaces are
called Hecke cycles. In particular, if we consider a projective line in P (Ea∗ ) it
gives a projective line in SU (n, d) which we call a Hecke curve. Thus SU (n, d)
contains a lot of rational curves. This is not surprising, since it is easily seen that
SU (n, d) are unirational, and indeed (not so easily) even rational if n and d are
coprime (King and Schofield 1999).

6.3 Moduli of Higgs pairs

One may ask whether the set of all (i.e. not necessarily unitary) representations
of π can also be described purely algebraically. Donaldson (1993) in reproving
the result of Narasimhan and Seshadri, started this process of thinking. The
upshot was the following.
Let L be a fixed line bundle. Consider pairs (E, α) where E is a vector bundle
and α is an L-twisted endomorphism, namely, a linear map E → E ⊗ L. A
subbundle F of E is invariant if α maps F into F ⊗ L. One requires the condition
of stability now only for invariant subbundles and arrives at the notion of stable,
semi-stable, and polystable pairs. Also one can make sense of S equivalence and
construct the moduli of such pairs. We will call it the moduli space of L-twisted
Higgs pairs.
We can define the direct sum (respectively, tensor product) of L-twisted Higgs
pairs (E, α) and (E, α ) to be (E ⊕ E  , α ⊕ α ) (respectively, (E ⊗ E  , α ⊗ 1E  +
1E ⊗ α )). The dual of (E, α) is defined to be (E ∗ , −αt ⊗ 1L ).
The case when L is the canonical line bundle K, namely, the cotangent bundle
of C, is of particular interest. In this case, we will omit the phrase ‘K-twisted’.
The component α in (E, α) can then be treated as a one-differential form with
values in the bundle of endomorphisms of E and is called the Higgs field.

Remark 6.1 A Higgs pair (E, α) may be stable or semi-stable even if E is

not stable as a vector bundle. However if (E, α) is stable, E cannot be terribly
nonstable. For example, if E is of rank 2 and degree 0, and ξ is a line subbundle
with deg(ξ) > g − 1, then deg(E/ξ) < 1 − g and so any homomorphism of ξ into
E/ξ ⊗ K is zero. In particular, the restriction of α to ξ followed by the projection
E ⊗ K → E/ξ ⊗ K is zero. Hence ξ is invariant, contradicting the stability of
(E, α).
96 Some aspects of the theory of Higgs pairs

Example 6.1 Let ξ be a ‘theta characteristic’, namely, a line bundle such that
ξ 2 is isomorphic to K. In particular the degree of ξ is g − 1. Take for E the vector
bundle ξ ⊕ ξ −1 . It is of rank 2 and degree 0. Since it has ξ as a subbundle, it
is not semi-stable. Now a
homomorphism
 α : E → E ⊗ K can be written in the
α 1 α2
form of a (2 × 2) matrix where α1 and α4 are sections of K, α2 is a
α3 α4
section of K 2 , and α3 is a section of O. If α3 is nonzero, then ξ is not invariant
under α. Any line subbundle of E other than ξ admits a nonzero homomorphism
into ξ −1 and hence has degree at most −(g − 1). Hence for any such α, the pair
(E, α) is stable.

Remark 6.2 On the other hand, if E is stable, then any α (including 0) gives
rise to a Higgs pair (E, α).

I believe that the following theorem may be attributed to Donaldson (1993)

and Hitchin (1987) in the case of rank 2 and to Simpson (1992) and Corlette
(1988), who generalized it to higher rank and also to bundles on higher dimen-
sional spaces:

Theorem 6.3 There is a canonical bijection between the set of equivalence

classes of completely reducible representations of π in GL(n) and the set of
isomorphism classes of polystable Higgs pairs of rank n and degree 0.

Remark 6.3 We regain the theorem of Narasimhan and Seshadri by noting

that under the above bijection, unitarisable representations correspond to those
Higgs pairs in which the Higgs field vanishes.

We observed that every representation ρ of π gives rise to a vector bundle

Eρ , associated to the principal π-bundle, namely the universal covering. This
is in general not the bundle in the Higgs pair associated to ρ in the above
correspondence.
The bundle Eρ comes with a natural flat (holomorphic) connection ∇ρ . If the
given representation is unitarizable, then Eρ admits a metric along the fibres
which is invariant (or flat) for the connection ∇ρ . If ρ is not unitarizable, such
a metric of course does not exist. Nevertheless one seeks ‘the most favourable’
unitary metric along the fibres of Eρ .
Using the flat connection, one can trivialize the bundle and the connection
locally. Any two such trivializations are mediated by a locally constant map into
GL(n, C). The metric we seek can then be viewed locally as a differentiable map
of a domain in C into the space P of positive definite Hermitian metrics on
Cn . Since P is a (negatively curved) Hermitian manifold, it makes sense to say
that the map is harmonic. Clearly harmonicity does not depend on the particular
local trivialization that we use. The favourable metric we referred to is one which
is harmonic locally. Using ideas of Eells and Sampson (1964), Donaldson (1987)
proved the following theorem:
 Higgs pairs and the fundamental group 97

Theorem 6.4 Complete reducibility of ρ is a necessary and sufficient condition

for the existence of a global harmonic metric along the fibres on Eρ .

Remark 6.4 Note that when the representation is unitary, the flat metric
gives, on local trivialization on a domain, a constant map, which is of course
harmonic.

Any unitary metric m on a holomorphic bundle gives rise to a connection ∇m ,

called the Chern connection. The difference of the flat connection on Eρ , and
∇m , is a one-form β on C with values in the bundle End(Eρ ). The harmonic
equation is essentially equivalent to the following. If the complex structure on
Eρ is modified by the (0, 1) part of β, we get a holomorphic bundle with respect
to which the (1, 0) part of β is holomorphic.
As in the case of vector bundles, one can also construct (Nitsure 1991;
Simpson 1992) the moduli space Hg(n, d) of Higgs pairs corresponding to a
given rank n and degree d. These are normal, irreducible, and quasi-projective
(but not projective) varieties. If n and d are coprime, they are even smooth. If E
is a stable vector bundle, the entire vector space Γ(C, End(E) ⊗ K) is contained
in Hg(n, d). Since this space is dual to H 1 (C, End(E)), the tangent space to
the moduli space U (n, d) at E, we conclude that the cotangent bundle over the
stable locus in U (n, d) is contained in Hg(n, d). In any family, stable bundles
form a Zariski open set and so, the cotangent bundle is in fact a dense, open
subset of Hg.

6.4 Higgs pairs and the fundamental group

There is another kind of structure that Higgs pairs constitute which has profound
significance. Polystable Higgs pairs of degree 0 form an additive category Hg.
(In fact, morphisms form vector spaces, and compositions are bilinear.) Using
semi-stability, one checks that it is actually an abelian category.
On the other hand, completely reducible π-modules also form such a category
Rep. The above bijection can be elevated to a categorical equivalence of Rep with
Hg. In fact, this functor respects the structures of direct sum, tensor product,
duals, etc. which make sense in Hg as well as in Rep.
One axiomatizes these structures and calls a category so equipped, a tensor
category.
Finally, one also notices that if we fix a base point x0 in C, then the functor
that associates to the Higgs pair (E, α) the fibre Ex0 is an additive (indeed
linear) functor of Hg into the category of vector spaces, which respects all the
structures we mentioned above. This functor is actually faithful in the sense
that a morphism (E, α) → (E  , α ) is zero if and only if the induced morphism
Ex0 → Ex 0 at the fibre level is 0. The faithful functor Φ taking (E, α) to Ex0 of
the Higgs-tensor category into the tensor category of vector spaces, is called a
fibre functor. A tensor category provided with a fibre functor is called a Tannaka
category.
98 Some aspects of the theory of Higgs pairs

Note that the category Rep of completely reducible representations of a group

G is a tensor category in a natural way. Furthermore, the forgetful functor, which
associates to a representation ρ of G the underlying vector space Vρ , is faithful
and makes Rep a Tannaka category as well. The above correspondence is an
equivalence of the two Tannaka categories.
What is its importance? To every Tannaka category C associate the group of
automorphisms of its fibre functor. It is called the Tannaka dual T n(C) of the
category. If we start with a Lie group G, and consider Rep, then there is a natural
homomorphism of G into T n(Rep). We get, by definition, for any g ∈ G, and for
each representation ρ, an automorphism of the vector space Vρ , namely, ρ(g).
This can actually be regarded as an automorphism of the fibre functor. The map
which takes g to this automorphism gives a homomorphism of G into T n(Rep).
It has the universal property that any completely reducible representation of G
‘extends’ to a unique representation of T n(Rep).
We refer to it therefore as the Tannakian completion T n(G) of G. Tannaka
considered this in the case when G is a compact Lie group. Chevalley interpreted
Tannaka’s results as showing that Tannakian completion T n(G) of a compact
Lie group G is a reductive algebraic group which complexifies G.
The Tannakian completion T n(Γ) of a finitely generated group Γ is a pro-
reductive algebraic group. The homomorphism Γ → T n(Γ) is indeed universal
for homomorphisms of Γ into reductive algebraic groups.
The category Hg has been defined purely in terms of the algebraic structure
unlike the fundamental group which is a priori a highly transcendental construct.
We have therefore accessed the Tannaka completion T n(Hg) = T n(π) of the
fundamental group purely by algebraic geometric methods.
The profinite completion of π is a group P f (π), together with a homomorphism
of π into it which is universal for homomorphisms into finite groups. It is called
the Grothendieck fundamental group.
Since each homomorphism of π into a finite group also gives rise to a unique
homomorphism of T n(π) into that group, we get a homomorphism of T n(π)
into P f (π). Thus the group T n(π) is much richer than P f (π). Incidentally, Nori
(1976) recovered the group P f (π) as the Tannaka dual of the category of what
he called ‘finite’ bundles. These are bundles E that satisfy an isomorphism of
the type

⊕mi (⊗i E)  ⊕nj (⊗j E), mi , nj ∈ N.

Here we interpret mE as the m-fold direct sum of E and assume that the i’s
and j’s are disjoint. He showed these form a Tannaka category N and that its
Tannaka dual is P f (π).

6.5 Non-abelian Hodge theory

Note that for any λ ∈ C∗ and polystable Higgs pair (E, α), the pair (E, λα) is
also polystable. Since this action is compatible with all the structures of the
 Hitchin morphism 99

Tannaka category Hg, one easily concludes that there is a natural action of C∗
on T n(π) as well.
The group of homomorphisms of π into C is isomorphic to H 1 (C, C) and the
Hodge decomposition can be described as the action of C∗ on it, acting by the
natural (identity) character on H 1,0 and by its inverse (namely, the character
λ → λ−1 on H 0,1 ).
From this point of view, the above action of C∗ on T n(π), which mimics the
Hodge action, can be thought of as the non-abelian version of the Hodge action.
It should be pointed out that this action depends on the complex structure of
the curve.
Actually, the considerations above can be generalized in two directions. Firstly,
we can take a reductive group G and consider principal G-bundles. Stability of a
principal G-bundle E will then mean that for any reduction of its structure
group G to a parabolic subgroup P (which can be interpreted to mean a
section s of the bundle E/P ), the pull-back of the tangent bundle along the
fibres is of positive degree. As for a Higgs pair, it now consists of a principal
G-bundle E and a differential form with values in the associated adjoint bundle
ad(E), namely the vector bundle associated to E via the adjoint representation
of G in its Lie algebra. We will discuss this generalization in some detail
below.
Secondly, we may take, instead of C, an arbitrary smooth, polarized projective
variety X of dimension n. This essentially means that we are given an ample line
bundle ξ on X. Then one can define the degree of a line bundle L to be the
intersection number c1 (L).ξ n−1 , and proceed as before. A Higgs field is, in the
general case, a principal G-bundle E, and a one-form α with values in ad(E)
such that the two-form [α, α] vanishes. Thus we can get an algebraic description
of the pro-reductive algebraic completion of the fundamental group, but more
importantly, an action of C∗ as well.
This would show that many (non-uniform) discrete subgroups of real semi-
simple groups cannot be the fundamental group (nor even split quotients of the
fundamental group) of smooth projective varieties. This is because one can check,
thanks to rigidity theorems on uniform discrete subgroups of Lie groups, that
these do not admit an action of C∗ on the above lines. For example, SL(n, Z),
n ≥ 3 are not.

6.6 Hitchin morphism

We have already mentioned that there is a natural inclusion of the cotangent
bundle of the moduli space of stable vector bundles as an open dense set in the
Higgs moduli space.
As is well known, there is a natural sympectic structure on the cotangent
bundle of any smooth variety. Actually this structure extends to the whole of
the locus of stable Higgs pairs.
100 Some aspects of the theory of Higgs pairs

Let (E, α) be a stable Higgs pair. Consider the map

End(E) → End(E) ⊗ K,

given by bracketing with α. We may regard this as a complex of sheaves having

only two terms C 0 and C 1 . It is easy to check (Biswas and Ramanan 1994)
that the first hypercohomology space of this complex is the tangent space at E
to the Higgs moduli space. The dual of this complex tensored with K, namely,
the Serre dual of the complex, can be identified with the complex itself. In this
way, we get a symplectic form on the first hypercohomology of the complex,
that is, the tangent space at (E, α) to Hg. It is this that gives the canonical
(Hamiltonian) symplectic form at those points of the Higgs moduli which belong
to the cotangent bundle of the moduli of vector bundles, that is, when E is itself
stable.
Now Hitchin defined a morphism of the Higgs moduli space into the affine
space Ht = ⊕i=r i
i=1 Γ(C, K ) as follows. To any Higgs field α one can associate the
trace as an element of Γ(K). In a similar manner one can lift α to a map Λi (α)
of Λi (E) into Λi (E) ⊗ K i and take its trace, interpreted as a section of K i . In
other words, the coefficients of the characteristic polynomial of α make sense as
elements of Γ(K i ). This gives rise to the Hitchin morphism of Hg mentioned
above into the Hitchin space Ht.
From the above description of the tangent space and the symplectic form, one
can easily conclude that the form vanishes identically on the i=rfibres of the Hitchin
morphism. On the other hand, the dimension of Ht is g + i=2 (2i − 1)(g − 1) =
r 2 (g − 1) + 1. In other words, the generic fibres are Lagrangian submanifolds.
It was shown by Nitsure (1991) that the Hitchin morphism is actually a proper
map, so that the fibres are projective varieties. By its very nature one can
prove that if there is a completely integrable system, the Lagrangian fibres are,
generally speaking, (open sets in) tori. In our case, the generic fibres in Hg should
perhaps be abelian varieties. If so, what are these?
Fix a point s = (s1 , . . . , sr ) ∈ Ht. Then one can construct a curve S and a
(finite) morphism π : S → C such that π∗ (O) = V = O ⊕ K −1 ⊕ · · · ⊕ K −(r−1)
as vector bundles. But the left side comes with an algebra structure over OC .
We only have to identify this algebra structure on V in order to construct S. In
fact, S is then Spec(V ).
Now we have natural isomorphisms K −i ⊗ K −j → K −(i+j) ⊂ V , if i + j ≤
r − 1. If i + j ≥ r, we have a homomorphism K −i−j → K −k for each k, 0 ≤ k ≤
r − 1 given by tensoring with si+j−k . All these add up to a morphism V ⊗ V →
V , giving the required structure. If s is sufficiently general, it can be checked that
S is actually smooth. It is called the spectral curve defined by s. The pull-back
π ∗ (K) comes with a canonical section which can be described as follows. We
have π∗ (π ∗ (K)) = K ⊗ π∗ (O) = K ⊕ O ⊕ K −1 ⊕ · · · ⊕ K −(r−2) . The canonical
 Quantization 101

section of the second summand, namely, O, gives a section of π∗ (π ∗ (K)) and

hence that of π ∗ (K) as claimed.
If ξ is any line bundle on S, then the direct image π∗ (ξ) is a vector bundle of
rank r on C. Also the canonical section of π ∗ (K) gives rise to a homomorphism
ξ → ξ ⊗ π ∗ (K) and therefore of E = π∗ ξ into π∗ (ξ ⊗ (π ∗ (K)) = E ⊗ K. It is
therefore a Higgs pair. It is straightforward to check that its image under the
Hitchin map is s. Equally one can check that the Higgs pair thus obtained is
semi-stable and that the Hitchin morphism takes this Higgs pair to (s). Thus
the fibre over s of the Hitchin map can be identified with the Jacobian (of suitable
degree) of S. The intersection of this fibre with the cotangent bundle over the
stable locus of the moduli space of vector bundles is a Zariski open subset of the
above Jacobian, and is Lagrangian for the Hamiltonian symplectic form.
The above result can be restated slightly differently. Let us choose a base
for Ht. Then the Hitchin map can be considered as a map into Cl given by
l functions on the Higgs moduli space and hence on the cotangent bundle of
the vector bundle moduli space as well. The key point is that these functions
commute in the Poisson algebra. It is rarely that one finds the maximum number
of Poisson commuting functions on the cotangent bundle. These functions are not
only algebraic, but actually homogenous polynomials on the cotangent spaces.
One talks, in this case, of algebraic complete integrability.
Beauville, Narasimhan, and I (1989) considered the map of the Jacobian of S
into U (r, 0) as well as the corresponding Prym variety (the kernel of the natural
‘norm map’ from the Jacobian of S onto the Jacobian of C) into SU (r, 0). Note
that it is defined only on a big Zariski open subset but it is dominant, that
is, has dense image. By looking into the induced map from Γ(SU (r, 0), Θ) into
the space of sections of the pull-back of Θ on the Prym variety, it was shown
that dim Γ(SU (r, 0), Θ) = r g . Thus the theory of Higgs pairs has provided many
other interesting results on linear systems on U (n, d) as well.

6.7 Quantization
If D is a (linear) differential operator of order ≤ k, then its (kth order) symbol is
a function on the cotangent bundle which is a homogeneous polynomial of degree
k on the fibres. In the above context therefore one may ask if one can associate
in a canonical fashion a basis of commuting differential operators whose symbols
are the Poisson commuting functions defined above.
Some motivation or analogy may be in order here. The Hitchin space is the
analogue of (actually the dual of) the space of invariant polynomials for the
conjugacy action of GL(n) on the vector space of (n, n) matrices. The Hitchin
map itself can be thought of as the analogue of associating to any matrix A, the
linear form on invariant polynomials which maps any f to the evaluated complex
number f (A). One can associate canonically to any homogeneous invariant
polynomial f , a bi-invariant differential operator Df on the Lie group, with
102 Some aspects of the theory of Higgs pairs

symbol f . Moreover, the map f → Df is an isomorphism of the algebra of

invariant polynomials onto the algebra of bi-invariant differential operators on
GL(n, C).
So one might expect a homomorphism of the dual Ht∗ of the Hitchin space
into the space of differential operators on the moduli space of vector bundles.
This is called the quantization of the Hitchin map.
However, it is immediately seen that this is too naive. Indeed, the moduli space
has no global differential operators of higher order at all! Now the canonical
bundle of SU (n, d) is divisible by 2. In other words, there is a (unique) line
bundle η on it whose square is the canonical bundle. It turns out that these
symbols can be lifted to a commutative subalgebra of differential operators from
η to itself. We will be more precise below, where we deal with the general case
of a semi-simple group. This is related to the so-called geometric Langlands
programme.

6.8 Hecke transformation and Hitchin discriminant

We may ask if one can start with a stable Higgs pair (E, α) and modify it by
a Hecke transformation. Let a ∈ C and l ∈ Ea∗ \ {0}. We then get the kernel E 
from the exact sequence
0 → E  → E → Oa → 0.
The question is if α also gives rise to a Higgs field on E  . By composing the
map E  → E with α : E → K ⊗ E, we get a map E  → K ⊗ E. In order for this
to factor through a map E  → K ⊗ E  , the induced map Ea → Ka ⊗ Ea has to
have its image contained in the kernel of l. This happens for every l if and only
if α vanishes at a. In particular, it does not happen for a generic (E, α). Indeed,
notice first that even if we start with a Hecke line, and lift it to Hg, the Hitchin
morphism (being a map into an affine space) has to be a constant on it. On the
other hand, it cannot be contained in a generic Hitchin fibre, since we know that
the fibre is an abelian variety and so does not contain any rational curve. So it
is clear that Hecke curves do not in general lift to Hg.
Now we will look into the question if any Hecke curve at all lifts to Hg. Again
such a lift has to be contained in a Hitchin fibre. So let us assume that this fibre
is not generic, that is to say, the spectral curve S is not smooth. We will assume
for our analysis that the spectral curve has only one nodal singularity. Let S 
be its normalization. Denote by b the node in S over a ∈ C and a1 , a2 the two
points in S  over b. A line bundle on S is given by a line bundle L on S  together
with an isomorphism η of La1 with La2 . Thus the space of line bundles on S
can be regarded as a C∗ -bundle over the Jacobian of S  . Consider the associated
P1 bundle over the Jacobian. It comes with two canonical sections. Identify the
 Hitchin component 103

first section with the second section after translating by the divisor class a2 − a1 .
This is the compactified Jacobian of C. This can be identified with the Hitchin
fibre.
Let E  be the direct image of L on C. Its fibre at a is of the form V ⊕ La1 ⊕ La2 ,
where V is a vector space of dimension n − 2. It comes with a one-dimensional
subspace, namely, the graph of the isomorphism η. The corresponding Hecke
transform is the bundle E, obtained as the direct image on C of the line bundle
on S given by (L, η).
In other words, the Hitchin fibre corresponding to the spectral curve S is the
compactified Jacobian J˜ described above. The explicit Higgs pair given rise to
by a point of J˜ is also given above. If we deal with SU (n, d) instead of U (n, d),
one has to replace the Jacobian by the Prym variety. The point is of course that
we get on this fibre, projective lines parametrized by the Jacobian, or the Prym
variety of S  . These lines map to Hecke curves on SU (n, d).
Hwang and I (2004) considered the tangents to Hecke curves passing through
the generic point E of SU (n, d) and studied their lifts to the cotangent bundle.
The unions of all these are, as a simple consequence of our remark above, dense in
the Hitchin discriminant. The intersection of this discriminant with the cotangent
space at E is a cone and defines a projective subvariety of P (TE∗ ). We showed
from the above considerations that the dual of this variety is the variety of all
the tangents to Hecke curves through E.
This implies that from SU (n, d) as a variety one can geometrically extract the
Hitchin discriminant and hence the variety of Hecke tangents. The latter has as
its Albanese image, the curve C, so that we have a geometric description of the
curve – a Torelli kind of result.
Indeed, one can work a little more and show the following (Hwang and
Ramanan 2004):
Theorem 6.5 If C and C  are two curves, then there exists a nonconstant
morphism from SUC (n, d) into SUC (n, d) if and only if C is isomorphic to C  .
Moreover any isomorphism is actually induced by an isomorphism of C with C  ,
tensoring by line bundles of order n and possibly dualizing.

6.9 Hitchin component

We will now investigate stable pairs (E, α) in which E itself is highly non-
stable. Let us first analyse the case of rank 2 bundles with trivial deter-
minant. We observed in Remark 6.1 that every line subbundle of E is of
degree ≤ g − 1.
We also gave an example where E admits a line subbundle of degree
g − 1, namely, E = ξ ⊕ ξ −1 , with ξ a theta characteristic and a suitable α.
104 Some aspects of the theory of Higgs pairs

The (K-twisted) automorphisms of E can be represented as matrices ( λ0 λω−1 )

with ω ∈ Γ(K). Also α can be expressed as ( α α2
α3 −α1 ) where α1 is a section
1

of K, α2 and α3 are sections of K and O, respectively. Here α3 has to be

2

nonzero. Taking λ2 = α3 and ω = −α1 /λ, we see that (E, α) is equivalent to

(E, β), where β is a matrix of the form ( 01 −γ 0 ). Thus fixing the above bundle
E and varying γ, we get an embedding of Γ(C, K 2 ) in Hg(2, 0). This is called
the Hitchin component. One should bear in mind though that it depends on the
choice of theta characteristic ξ. Thus there are 22g Hitchin components and the
group J2 acts transitively on the set of these components.
Note that the Hitchin map in this case is simply the map (E, β) → γ ∈
Γ(C, K 2 ) and that it maps the Hitchin component isomorphically onto Ht. Also
this component stays in the complement of the cotangent bundle of the stable
locus in SUC (2).
We will now explain why we call it the Hitchin component. It is a natural
question to ask whenever one sees a geometrically defined subvariety in Hg if it
has a good meaning in the space of representations. We will now discuss this in
the case of the Hitchin component.
Consider those completely reducible representations of π in SL(2, C) which
come from representations into SL(2, R). We then of course get oriented real
bundles. Even topologically, these have an invariant, namely, the Euler class.
Since we are discussing the case of compact, oriented (real) manifolds, this
invariant is an integer. It was observed by Milnor (1957) and Wood (1976) that
the Euler class does not take arbitrary values.
Indeed, we have the Milnor–Wood inequality for the Euler class c, namely,
−(g − 1) ≤ c ≤ g − 1. Once we fix the Euler class, the bundle is topologically
determined. There is an involution acting on SL(2, R) given by conjugation by
the matrix ( 0i 0i ). If two representations into SL(2, R) are taken to each other
by the above involution, then obviously they are equivalent as representations
into SL(2, C). It is easy to see that two completely reducible representations into
SL(2, R) are equivalent as representations into SL(2, C) if and if only they are
either conjugate in SL(2, R), or conjugate after composing one of them with the
involution above. The Euler invariant of a representation is the negative of that
of its transform by this involution.
The Hitchin components correspond to those representations which come from
a representation into SL(2, R) with Euler invariant g − 1 (or equally −(g − 1)).
Hitchin himself called it the Teichmüller component since it is the component
consisting of representations for which the image of π in SL(2, R) are such that
the quotients are compact. Thus these give again curves of genus g.
The representations with a fixed Euler invariant other than ±(g − 1) form
however an irreducible variety.
Hitchin wrote down for each non-negative integer c ≤ g − 1, the subset of
Higgs pairs in Hg(2, 0) that come from representations into SL(2, R) and has
Euler invariant c.
 Reductive groups and principal bundles 105

6.10 Reductive groups and principal bundles

As we have already remarked, many of the above considerations can be extended
to any principal bundle E with a reductive structure group G. We will quickly
run through these analogues.
For simplicity, we will confine ourselves to semi-simple groups here. Generally
speaking, the corresponding notions for a reductive group G can be reduced to
this case.
Let then G be a complex semi-simple algebraic group and g its Lie algebra.
We are interested in principal G-bundles over C.

Definition 6.3 Let E be a principal G-bundle. Then it is said to be stable

(respectively, semi-stable) if for every reduction τ of the structure group of E
from G to any parabolic subgroup P , namely, a section of the associated bundle
with G/P as fibre, the pull-back of the tangent bundle along the fibres has positive
(respectively, non-negative) degree.

Using this definition, A. Ramanathan (1996) showed that the isomorphism

classes of stable principal G-bundles form a quasi-projective variety of dimension
(dimg)(g − 1). In the case of reductive groups, the dimension is (dimg)(g − 1) +
dimz, where z is the centre of g.
A. Ramanathan also showed that if E is semi-stable, then there exists a
reduction to a parabolic subgroup P with the following property. Let R = P/U
be the reductive group obtained by passing to the quotient by the unipotent
radical U of P . Then the R-bundle associated to the reduced P -bundle is stable.
Also any character on P gives a line bundle of degree 0. Moreover, the R-bundle
depends up to isomorphism only on E.
Two semi-stable bundles are defined to be S-equivalent if the corresponding
R-bundles are isomorphic. A polystable bundle is one which can be reduced to
the reductive part of a parabolic subgroup as a stable bundle of degree 0.
The S-equivalence classes of semi-stable G-bundles form a projective variety
MC G or M G in which stable bundles form a dense open set. However, a point
corresponding to a stable point can be singular, unlike the case when G = GL(n)
or SL(n). One can only say these are orbifold singularities. This is because, in
the general case, a stable bundle may admit non-trivial automorphisms. The
automorphism group Aut(E), which is finite, acts on H 1 (C, ad(E)), where ad(E)
is the vector bundle associated to the adjoint representation of G in its Lie
algebra. The singularity is like this quotient. Of course generically, we have
Aut(G) is isomorphic to the centre of G.
Let U be a maximal compact subgroup of G. A. Ramanathan (1975) proved
that a bundle is polystable if and only if it comes from a homomorphism of the
fundamental group into U , and that two such homomophisms are conjugate in
U if and only if the corresponding polystable bundles are isomorphic.
106 Some aspects of the theory of Higgs pairs

6.11 Reductive groups and Higgs pairs

We now turn to the corresponding considerations for Higgs pairs. Firstly, a Higgs
pair consists of a G-bundle E and a section α of ad(E) ⊗ K. Given a reduction F
of the structure group of E to a subgroup H of G, we have an inclusion of ad(F ) in
ad(E). If α comes from a section of ad(F ) ⊗ K, we say the Higgs pair is reduced
to H. Then we define stability, semi-stability, polystability, and S-equivalence
as above, but now in terms of reductions of the Higgs pair to parabolic
subgroups.
One can show (Nitsure 1991) that S-equivalence classes of polystable Higgs
pairs form a quasi-projective variety Hg(G). It contains a dense open set which
is isomorphic to the cotangent bundle over the locus of those stable bundles in
the moduli M (G) of G-bundles, which do not admit non-trivial automorphisms.
We now come to the Hitchin morphism in the general case. It is a well-
known theorem of Chevalley that the algebra I(G) of polynomial functions
on g invariant under the adjoint action is isomorphic to a polynomial ring
generated by l homogeneous polynomials of degrees mi , i = 1, . . . , l, where l
is the rank of G. Indeed, let h be a Cartan subalgebra of g. Then one can show
that by restricting polynomial functions on g to h, one gets an isomorphism of
the algebra of adjoint invariant polynomial functions on g with the algebra of
Weyl group invariant polynomial functions on h. Using the fact that the Weyl
group is generated by reflections in hyperplanes, Chevalley proved the above
theorem.
To any x ∈ g, one associates the graded algebra homomorphism from P (g) →
C[t] which takes any f ∈ P to the polynomial t → P (tx). By restriction, we also
get a map g → GrAlgHom(I(G), C[t]). This can be elevated to the level of Higgs
pairs.
Let (E, α) be a Higgs pair. At every point a ∈ C, the section α ∈ Γ(ad(E) ⊗ K)
gives rise to an element of ad(E)a ⊗ Ka and hence a graded algebra homomor-
phism from I(G) to Sym(Ka ). This in turn induces a map of Γ((ad(E) ⊗ K) →
GrAlgHom(I(G), Sym(K)). The right side is defined to be the Hitchin space
Ht(G) and the above morphism from Hg(G) to Ht(G), the Hitchin mor-
phism. More concretely, if I(G) is the polynomial algebra on {f1 , . . . , fl }, with
fi ’s, homogeneous of respective degrees mi , then Ht(G)  is the affine space
⊕i=l
i=1 (Γ(C, K mi
)). Note
 that the Hitchin space has dimension (2mi − 1)(g − 1),
and one knows that (2mi − 1) = dimg.
As we mentioned above, the Higgs moduli space contains, as a Zariski open
set, the cotangent bundle of the smooth stable points of M (G). The standard
symplectic structure on the cotangent bundle extends to the whole of the smooth
locus of Hg(G).
The tangent space to the Hitchin space at (E, α) is the first hypercohomology
of the two-term complex

ad(E) → ad(E) ⊗ K
 Reductive groups and Higgs pairs 107

where the map is given by Lie bracketing with α. Using this, one can identify
the symplectic form and show that the fibres of the Hitchin morphism are
Lagrangian. This would prove the complete integrability in general (Biswas and
Ramanan 1994).
Nitsure showed in Nitsure (1991) that the Hitchin map is proper. One can seek
therefore to identify the abelian varieties which are generic fibres of the Hitchin
morphism. This has been done in a beautiful paper of Donagi (1993).
The Hitchin map has also been quantized in the general case (Frenkel 2004;
Beilinson and Drinfeld, to appear). I should point out here that the moduli space
may be gainfully replaced by the moduli ‘stack’ and it is in this set-up that the
subject is dealt with in the above references. Since the moduli stack is smooth,
many of the circumlocutions in the statements can thereby be avoided. Moreover,
in the context of the Langlands programme, the moduli stack is more natural,
but I prefer to leave things as they are, rather than launch into the theory of
algebraic stacks here.
The algebra D of differential operators on the moduli space M (G) from the
square root of the canonical bundle to itself turns out to be commutative and
their symbols give the space of functions on Ht(G). In other words, the algebra
of their symbols has as its spectrum the space Ht(G).
One may then ask for the spectrum of D itself. What is it natually isomorphic
to? This turns out to be an affine space, called the space of opers.
Here there is another subtlety. There is a canonical isomorphism (or rather
duality) between the algebra of adjoint invariant polynomials on the Lie algebra
of a semi-simple group G and those of its Langlands dual L G. The Langlands
dual is a group associated to the root system, which is dual to that of G. If G
is simply connected, L G is the adjoint group with the dual root system. Also
the centre of G and the fundamental group of L G are canonically dual to each
other. In general, the Langlands dual of G/A, where A is a central subgroup of
G, is defined to be the covering group of L G corresponding to the quotient A∗
of π1 (L G).
For simplicity, let us assume here that G is simply connected. We have
remarked that the algebra of adjoint invariant polynomials on g and L g are
isomorphic to their respective Weyl group invariant polynomials on their Cartan
algebras. The Cartan algebras are canonically dual, compatible with the Weyl
group actions.
A G-oper consists of a principal G-bundle E provided with a connection ∇, and
a reduction F of E to a Borel subgroup B of G. The connection is not supposed
to be compatible with the reduction. On the contrary, one can associate to the
connection its second fundamental form ψ which measures the extent to which
∇ does not come from an F -connection. This is defined for example, locally, by
taking a connection ∇loc on F and taking the difference ψ = ∇ − ∇loc . It is a
one-form with values in ad(E). Using the natural map g → g/b, we get a section
ϕ of K ⊗ Fρ , where Fρ is the vector bundle associated to F via the representation
of B in g/b. This form is obviously independent of ∇loc and therefore glues into
108 Some aspects of the theory of Higgs pairs

a global form. It is called the second fundamental form of ∇. We will assume

actually that it is highly nondegenerate in the following sense.
Consider the set in g/b which gives an open orbit O for B. Since by definition
B acts on this set, the bundle associated to F with O as fibre makes sense and
is a subbundle of ad(E). We will denote this by FO . We assume that ϕ takes
values in FO .
It is fairly straightforward to check that if a connection ∇ of the type required
exists, then the corresponding B-reduction is unique. Indeed, it is the Harder–
Narasimhan reduction (Ramanathan 1996) of the bundle E. Moreover, it is
also easy to check that the bundle E itself is uniquely determined under our
assumptions (if G is the adjoint group).
We will start with an example. Let ξ be a theta characteristic. There is then
a canonical non-trivial extension

0 → ξ → P → ξ −1 → 0
given by the generator of H 1 (C, K). The bundle P can be considered an
SL(2, C) bundle with a reduction to a Borel subgroup B0 . It is clear that it
is indecomposable of degree 0. Hence by Weil’s theorem, it admits a connection.
Although P depends on the choice of the theta characteristic, the corresponding
P SL(2)-bundle is entirely canonical.
In general, that is to say, for any semi-simple Lie group G, Kostant considered
all non-trivial homomorphisms of SL(2) into it, up to conjugacy. He isolated one
such conjugacy class as particularly interesting, calling it the principal TDS. TDS
stands for three-dimensional simple Lie group. If we fix one such homomorphism
κ, we get the associated principal G-bundle κ(P ). It comes with a reduction to
a Borel subgroup B of G. Thus, we have a principal G-bundle, a B-reduction
which admits a G-connection. Its Harder–Narasimhan filtration is essentially the
reduction to B.
The set of (equivalence classes of) opers is an affine space with the torsor
vector space Ht(G). The conclusion is that the sought for spectrum of D is
the set of opers for the Langlands dual group. There does not yet seem to be
a simple understanding of this nice relationship. For further details one may
consult Beilinson and Drinfeld (to appear).

6.12 Real forms

Again let G be any semi-simple group and we will stick to the above notation.
Let B be a Borel subgroup in G and U = [B, B] its unipotent radical. Consider
a principal TDS G0 in G whose Borel subgroup B0 is contained in B. Restrict
the adjoint representation of G in its Lie algebra g to G0 , and consider the
U0 -fixed subspace L in it. If f generates the nilpotent radical of the opposite
Borel subalgebra in g0 , then the coset f + L containing f maps isomorphically
on the space GrAlgHom0 (I(G), C[t]). Also, the Kostant subspace Cf ⊕ L splits
into a direct sum of one-dimensional spaces under the diagonal matrices of G0 .
 Real forms 109

These are acted upon via the character λ → λ−2 on Cf and as characters λ →
λ2mi −2 on these one-dimensional subspaces in L.
We considered the SL(2, C)-bundle E = ξ −1 ⊕ ξ and determined all the Higgs
pairs with that as the underlying bundle. Take the Kostant homomorphism of
SL(2, C) in G and consider the associated bundle EG . We will now consider Higgs
fields on the bundle EG . Now ad(EG ) can be identified with the bundle associated
to E via the restriction of the adjoint representation of G to SL(2). Since the
bundle E comes from the line bundle ξ by the inclusion of C∗ as diagonal matrices
in SL(2), the Kostant subbundle also makes sense in ad(EG ). Clearly this bundle
is isomorphic to the direct sum ξ −2 ⊕i=l
i=1 (ξ)
2mi −2
G ) ⊗ K contains
. Thus, ad(E
the trivial bundle, and Higgs fields on EG of the form α = 1 + βi , βi ∈ Γ(K mi )
therefore make sense. Then one can show that the set of all Higgs pairs (EG , α)
with α, as above, are mapped by the Hitchin morphism isomorphically on the
Hitchin space. This is ‘the’ generalized Hitchin component. It of course depends
on the choice of ξ.
Note that there is a canonical conjugacy class of real forms Gr of G, namely,
the Chevalley or split form. Reductive representations of π in Gr give a closed
subspace of Rep(π, G). As we have seen this space is not connected. In some cases,
such as SL(2, R), there is a topological invariant (called the Toledo invariant)
which takes different values on different components. The invariants which serve
to nail down all components is, to my knowledge, not fully understood yet.
One of the components in this case is the Hitchin component we have described
above.
We have thus some understanding of the Higgs pairs corresponding to the
compact and the split forms. It is natural to consider the other real forms as
well. A real form is defined as the fixed point set of a complex anti-holomorphic
involution of G, but the fixed points in Hg(G) constitute a complex subvariety.
The relation between Higgs pairs and representations can then be understood
in the following way. Let σ be a complex conjugation defining a real form Gσ of
G. Then there is a suitable compact real form H of G which is invariant under
σ. Now extend σ|H as a complex analytic involution of the complexification
H c of H in G. Let m be the intersection of h and gσ . Then there is a Cartan
decomposition of the vector space gσ as m ⊕ p. Thus we have g = hc = mc ⊕ pc .
Then we have

Theorem 6.6 There is a bijection between the set of equivalence classes of

polystable G-Higgs pairs (E, α) where E comes from a principal M c -bundle V and
α from a section of K ⊗ Vpc on the one hand, and the set of conjugacy classes of
reductive homomorphisms of the fundamental group into G which factor through
Gσ , on the other.

Remark 6.5 Since the normalizer of Gσ in G may contain Gσ properly, the

map Rep(Gσ ) → Rep(G) is not injective in general. This happens already for
SL(2, R) → SL(2, C), since the matrix ( 01 10 ) normalizes SL(2, R).
110 Some aspects of the theory of Higgs pairs

The connected components of the space of representations into classical real

forms have been studied in Bradlow, Garcia-Prada, and Gothen (2003). Let
us take a look at the example of SL(2)-Higgs pairs. Notice that the complex
analytic involution (E, α) → (E, −α) of the moduli of Higgs pairs fixes the
moduli of vector bundles embedded in Hg by taking α = 0. This variety of course
corresponds to representations into SU (2). But the above involution fixes other
varieties including the Hitchin component! On the representation side, we note
that representations into SU (2) as well as those into SL(2, R) are fixed under
this involution. On the Higgs side, recall that the Hitchin component is defined
by E = ξ ⊕ ξ −1 (with ξ 2  K) with α of the form ( −γ 0 1
0 ) with γ ∈ Γ(C, K ).
2

It is easy to see that the diagonal automorphism with i and −i on the diagonal
takes α to −α, so that (E, α) and (E, −α) are indeed equivalent.
The involution (E, α) → (E, −α) obviously fixes the Higgs pairs given by
representations into SU (n) for all n, since α is 0 for these pairs. On the other
hand, it is easily seen that points in the Hitchin component are not fixed by this
involution if n ≥ 3. Why is this so?
A little reflection immediately gives us the clue. The two anti-holomorphic
involutions of SL(n, C) given by A → A and A → (A )−1 give rise to the two real
forms SL(n, R) and SU (n, C) as fixed points, respectively. Consider two complex
conjugations as equivalent if they differ by an inner automorphism. Then the
maps induced on the set of equivalence classes of representations coming from
either of these two real forms are clearly the same. In the case of SL(2), the two
anti-holomorphic involutions are indeed equivalent, but in the case of SL(n)’s
for n ≥ 3, they are not!
One may also consider other involutions and determine their fixed points, and
this gives rise to a nice interplay between the geometry of etale coverings of C,
the equivalence among real forms, etc. These are being studied in Garcia-Prada
and Ramanan (in preparation).

Acknowledgement
I would like finally to thank the referee for reading the typescript carefully,
pointing out many typographical errors, and catching at least one misleading
statement.

References
Beauville, A., Narasimhan, M. S. and Ramanan, S. (1989). Spectral curves and the
generalized theta divisor. J. Reine Angew. Math. 398, 801–28.
Beilinson, A. and Drinfeld, V. (2005). Opers—Quantization of Hitchin integrable
systems and Hecke eigensheaves. math. AG/0501398.
Biswas, I. and Ramanan, S. (1994). An infinitesimal study of the moduli of Hitchin
pairs. J. Lond. Math. Soc. 49, 219–31.
Bradlow, S. B., Garcia-Prada, O. and Gothen, P. B. (2003). Surface group representa-
tions and U (p, q)-Higgs bundles. J. Diff. Geom. 64, 111–70.
 References 111

Corlette, K. (1988). Flat G-bundles with canonical metrics. J. Diff. Geom. 28, 361–82.
Donagi, R. (1993). Decompositions of spectral covers. Astérisque 218, 145–75.
Donaldson, S. K. (1987). Twisted harmonic maps and self-duality equations. Proc.
Lond. Math. Soc. 55, 127–31.
Donaldson, S. K. (1993). A new proof of a theorem of Narasimhan and Seshadri. J.
Diff. Geom. 18, 269–77.
Drezet M. and Narasimhan, M. S. (1989). Groupe de Picard des variétés de modules
de fibré semi-stable sur les courbes algèbriques. Inv. Math. 97, 53–94.
Eells, J. and Sampson, J. H. (1964). Harmonic mappings of Riemannian manifolds.
Am. J. Math. 86, 109–60.
Frenkel, E. (2004). Recent advances in the Langlands program. Bull. Am. Math. Soc.
41, 151–84.
Garcia-Prada, O. and Ramanan, S. Involutions on Higgs moduli space on curves, in
preparation.
Goldman, W. M. (1988). Topological components of spaces of representations. Inv.
Math. 93, 557–607.
Hitchin, N. J. (1987). The self-duality equations on a Riemann surface. Proc. Lond.
Math. Soc. 55, 59–126.
Hitchin, N. J. (1992). Lie groups and Teichmüller space. Topolgy 31, 449–73.
Hwang, J.-M. and Ramanan, S. (2004). Hecke curves and Hitchin discriminant. Ann.
Sci. Ecole Norm. Sup. 37, 801–17.
King, A. and Schofield, A. (1999). Rationality of moduli of vector bundles on curves.
Indagationes Mathematicae 10, 519–35.
Kostant, B. (1959). The principal three-dimensional subgroup and the Betti numbers
of a complex simple group. Am. J. Math. 81, 973–1032.
Kostant, B. (1963). Lie group representations on polynomial rings. Am. J. Math. 82,
327–404.
Milnor, J. W. (1957). On the existence of a connection with curvature zero. Commun.
Math. Helv. 32, 216–23.
Mumford, D. (1962). Projective invariants of projective structures and applications.
Proceedings of the International Congress on Mathematics, Stockholm, Sweden,
pp. 526–30.
Narasimhan, M. S. and Ramanan, S. (1975). Deformations of the moduli space of vector
bundles over an algebraic curve. Ann. Math. 101, 391–417.
Narasimhan, M. S. and Seshadri, (1965). C. S. Stable and unitary vector bundles on
compact Riemann surfaces. Ann. Math. 82, 540–64.
Nitsure, N. (1991). Moduli spaces of semistable pairs on a curve. Proc. Lond. Math.
Soc. 62, 275–300.
Nori, M. V. (1976). On the representations of the fundamental group. Compositio
Mathematica 33, 29–41.
Ramanan, S. (1973). The moduli space of vector bundles over an algebraic curve. Math.
Ann. 200, 69–84.
Ramanathan, A. (1975). Stable and principal bundles on a compact Riemann surface.
Math. Ann. 213, 129–52.
Ramanathan, A. (1996). Moduli for principal bundles over algebraic curves. Proc.
Indian Acad. Sci. 106, 301–28.
Seshadri, C. S. (1967). Space of unitary vector bundles on a compact Riemann surface.
Ann. Math. 85, 303–36.
112 Some aspects of the theory of Higgs pairs

Simpson, C. T. (1992). Higgs bundles and local systems. Inst. Hautes Études Sci. Publ.
Mat. 75, 5–95.
Toledo, D. (1989). Representations of surface groups in complex hyperbolic space. J.
Diff. Geom. 29, 125–33.
Uhlenbeck, K. and Yau, S. T. (1986). On the existence of Hermitian-Yang-Mills
connections in stable vector bundles. Commun. Pure Appl. Math. 39, 5257–93.
Weil, A. (1938). Généralisation de fonctions abéliennes. J. Math. Pure Appl. 17, 47–87.
Wood, J. W. (1976). Bundles with totally disconnected structure group. Commun.
Math. Helv. 51, 183–99.
 VII
MIRROR SYMMETRY, HITCHIN’S EQUATIONS,
AND LANGLANDS DUALITY

Edward Witten
Dedicated to Nigel Hitchin on the occasion of his 60th birthday

7.1 A-model and B-model

Let G be a compact Lie group and let GC be its complexification. And let C
be a compact-oriented two-manifold without boundary. We write Y(G, C) (or
simply Y(G) or Y if the context is clear) for the moduli space of flat GC bun-
dles E → C, modulo gauge transformations. Equivalently, Y(G, C) parametrizes
homomorphisms 1 of the fundamental group of C to GC .
Y(G, C) is in a natural way a complex symplectic manifold, that is, a complex
manifold with a nondegenerate holomorphic two-form. The complex structure
comes simply from the complex structure of GC , and the symplectic form, which
we call Ω, comes from the intersection pairing 2 on H 1 (C, ad(E)), where ad(E)
is the adjoint bundle associated to a flat bundle E. Since Y(G, C) is a complex
symplectic manifold, in particular it follows that its canonical line bundle is
naturally trivial.
Geometric Langlands duality is concerned with certain topological field theo-
ries associated with Y(G, C). The most basic of these are the B-model that is
defined by viewing Y(G, C) as a complex manifold with trivial canonical bundle,
and the A-model that is defined by viewing it as a real symplectic manifold with
symplectic form 3 ω = Im Ω.

1 Actually, it is best to define Y(G, C) as a geometric invariant theory quotient that para-

metrizes stable homomorphisms plus equivalence classes of semi-stable ones. This refinement
will not concern us here (see Section 7.6.1).
2 The definition of this intersection pairing depends on the choice of an invariant quadratic

form on the Lie algebra of G. It can be shown using Hitchin’s C∗ action on the moduli space
of Higgs bundles that the A-model that we define shortly is independent of this choice, up to
a natural isomorphism. The geometric Langlands duality that one ultimately defines likewise
does not depend on this choice.
3 The usual definition of Ω is such that Im Ω is cohomologically trivial, while Re Ω is not.

The fact that ω = Im Ω is cohomologically trivial is a partial explanation of the fact, mentioned
in the last footnote, that the A-model of Y is invariant under scaling of ω.
114 Mirror symmetry, Hitchin’s equations, and Langlands duality

These are the topological field theories that are relevant to the most basic
form of geometric Langlands duality. However, there is also a generalization that
is relevant to what is sometimes called quantum geometric Langlands. From the
A-model side, it is obvious that a generalization is possible, since we could use a
more general linear combination of Re Ω and Im Ω in defining the A-model. What
is less evident is that the B-model can actually be deformed, as a topological
field theory, into this family of A-models. This rather surprising fact is natural
from the point of view of generalized complex geometry (see Hitchin 2003) and
has been explained from that point of view in section 4.6 of Gualtieri (2003), as a
general statement about complex symplectic manifolds. In sections 5.2 and 11.3
of Kapustin and Witten (2007), it was shown that quantum geometric Langlands
is naturally understood in precisely this setting.
Here, however, to keep things simple, we will focus on the most basic B-model
and A-model that were just described.

7.2 Mirror symmetry and Hitchin’s equations

The next ingredient we need is Langlands or Goddard–Nuyts–Olive duality.
To every compact Lie group G is naturally associated to its dual group L G.
For example, the dual of SU (N ) is P SU (N ) = SU (N )/ZN , the dual of E8
is E8 , and so on. And we must also recall the concept of mirror symmetry
between A-models and B-models (e.g. see Hori et al. 2003). This is a quantum
symmetry of two-dimensional non-linear sigma models whose most basic role is to
transform questions of complex geometry into questions of symplectic geometry.
The geometric Langlands correspondence does not appear at first sight to be an
example of mirror symmetry, but it turns out that it is.
With a little bit of hindsight (the question was first addressed in Hausel and
Thaddeus 2002, following earlier works by Bershadsky et al. 1995, and Harvey
et al. 1995), we may ask whether the B-model of Y(L G, C) may be mirror to the
A-model of Y(G, C). Even once this question is asked, it is difficult to answer
it without some additional structure. The additional structure that comes in
handy is provided by Hitchin’s equations (see Hitchin 1987a). Until this point,
C has simply been an oriented two-manifold (compact and without boundary).
But now we pick a complex structure and view C as a complex Riemann surface.
Hitchin’s equations with gauge group G are equations for a pair (A, φ). Here A
is a connection on a G-bundle E → C (we stress that the structure group of E
is now the compact group G), and φ is a one-form on C with values in ad(E).
Hitchin’s equations, which are elliptic modulo the gauge group, are the system
F −φ∧φ=0
Dφ = D  φ = 0. (7.1)
Here  is the Hodge  operator determined by the complex structure on C. The
role of the complex structure of C is that it enables us to write the last of these
equations.
 Mirror symmetry and Hitchin’s equations 115

A solution of Hitchin’s equations has two interpretations. On the one hand,

given such a solution, we can define the complex-valued connection A =
A + iφ. Hitchin’s equations imply that the corresponding curvature F = dA +
A ∧ A vanishes, so a solution of Hitchin’s equations defines a complex-valued flat
connection, and thus a point in Y(G, C).
On the other hand, the (0, 1) part of the connection A determines a ∂¯ operator
on the bundle E (or rather its complexification, which we also call E). There is no
integrability condition on ∂¯ operators in complex dimension 1, so this ∂¯ operator
endows E with a complex structure; it becomes a holomorphic GC bundle over
C. Moreover, let us write φ = ϕ + ϕ̄, where ϕ and ϕ̄ are the (1, 0) and (0, 1) parts
of φ, respectively. Then Hitchin’s equations imply that ϕ, regarded as a section
of K ⊗ ad(E) (with K the canonical line bundle of C), is holomorphic. The pair
(E, ϕ), where E → C is a holomorphic GC bundle and ϕ ∈ H 0 (C, K ⊗ ad(E)),
is known as a Higgs bundle.
We write MH for the moduli space of solutions of Hitchin’s equations,
modulo a gauge transformation. The fact that a solution of these equations
can be interpreted in two different ways means that MH is endowed with two
different natural complex structures. In one complex structure, which has been
called I, MH parametrizes isomorphism classes of semi-stable Higgs bundles
(E, ϕ). In another complex structure, J, it parametrizes equivalence classes of
flat GC -bundles or in other words homomorphisms ρ : π1 (C) → GC . I, J, and
K = IJ fit together to a natural hyper-Kähler structure on MH , as described in
Hitchin (1987a). In particular, there are holomophic two-forms ΩI , ΩJ , ΩK and
Kähler forms ωI , ωJ , ωK . These are all related by ΩI = ωJ + iωK , and cyclic
permutations of this statement, as is usual in hyper-Kähler geometry.
In complex structure J, MH is the same as the variety Y that we described
earlier. The natural holomorphic symplectic form Ω of Y is the same as iΩJ . And
the real symplectic form ω = Im Ω used in defining the A-model coincides with
ωK . Complex structure J and the holomorphic symplectic form ΩJ = ωK + iωI
do not depend on the chosen complex structure on C, in contrast to the rest of
the hyper-Kähler structure of MH .

Remark 7.1 As an aside, one may ask how closely related φ, known in the
present context as the Higgs field, is to the Higgs fields of particle physics. Thus,
to what extent is the terminology that was introduced in Hitchin (1987a) actually
justified? The main difference is that Higgs fields in particle physics are scalar
fields, while φ is a one-form on C (valued in each case in some representation
of the gauge group). However, although Hitchin’s equations were first written
down and studied directly, they can be obtained from N = 4 supersymmetric
gauge theory via a sort of twisting procedure (similar to the procedure that
leads from N = 2 supersymmetric gauge theory to Donaldson theory). In this
twisting procedure, some of the Higgs-like scalar fields of N = 4 super Yang–
Mills theory are indeed converted into the Higgs field that enters in Hitchin’s
equations. This gives a reasonable justification for the terminology.
116 Mirror symmetry, Hitchin’s equations, and Langlands duality

As we will explain next, it is possible, with the aid of Hitchin’s equations, to

answer the question of whether the B-model of Y(L G, C) is mirror to the A-model
of Y(G, C). This in fact was first pointed out in Hausel and Thaddeus (2002),
and used in Kapustin and Witten (2007) as a key ingredient in understanding
the geometric Langlands correspondence.

7.3 Hitchin fibration

We will have to use the Hitchin fibration. This is the map, holomorphic in
complex structure I, that takes a Higgs bundle (E, ϕ) to the characteristic
polynomial of ϕ. For example, for G = SU (2), (E, ϕ) is mapped simply to the
quadratic differential det ϕ. The target of the Hitchin fibration is thus in this
case the space B = H 0 (C, K 2 ) that parametrizes quadratic differentials. This
has a natural analog for any G. From the standpoint of complex structure I,
the generic fiber of the map π : MH → B is a complex abelian variety (or to
be slightly more precise, in general a torsor for one). The fibers are Lagrangian
from the standpoint of the holomorphic symplectic form ΩI . Such a fibration by
complex Lagrangian tori turns MH into a completely integrable Hamiltonian
system in the complex sense (Hitchin 1987b).
There is, however, another way to look at the Hitchin fibration, as first
described in Hausel and Thaddeus (2002). Let us go back to the A-model
defined with the real symplectic structure ω. Since the fibers of π : MH → B are
Lagrangian for ΩI = ωJ + iωK , they are in particular Lagrangian for ω = ωK .
Moreover, being holomorphic in complex structure I, these fibers are actually
area-minimizing in their homology class—here areas are computed using the
hyper-Kähler metric on MH . So the Hitchin fibration, from the standpoint of
the A-model, is actually a fibration of MH by special Lagrangian tori.
Mirror symmetry is believed to arise from T -duality on the fibers of a special
Lagrangian fibration (see Strominger et al. 1996). Generally, it is very difficult to
explicitly exhibit a non-trivial special Lagrangian fibration. The present example
is one of the few instances in which this can actually be done, with the aid of
the hyper-Kähler structure of MH and its integrable nature. Non-trivial special
Lagrangian fibrations are hard to understand because it is difficult to elucidate
the structure of the singularities. In the hyper-Kähler context, the fact that the
fibers are holomorphic in a different complex structure makes everything far
more accessible.
Once we actually find a special Lagrangian fibration, what we are supposed
to do with it, in order to give an example of mirror symmetry, is to construct
the dual special Lagrangian fibration, which should be mirror to the original
one. The mirror map exchanges the symplectic structure on one side with the
complex structure on the other side.
In the present context, there is a very beautiful description of the dual
fibration: it is, as first shown in Hausel and Thaddeus (2002), simply the Hitchin
fibration of the dual group. Thus, one considers MH (L G, C), the moduli space
 Hitchin fibration 117

of solutions of Hitchin’s equation for the dual group L G. It turns out that the
bases of the Hitchin fibrations for G and L G can be identified in a natural way.
The resulting picture is something like this:
MH (L G, C) MH (G, C)
 
B

In complex structure I, the fibers over a generic point b ∈ B are, roughly

speaking, dual abelian varieties (more precisely, they are torsors for dual abelian
varieties).
Alternatively, the fibers are special Lagrangian submanifolds in the symplectic
structure ω = ωK . From this second point of view, the same picture leads to a
mirror symmetry between the B-model of MH (L G, C) in complex structure J
and the A-model of MH (G, C) with symplectic structure ωK .
As we have just explained, the tools that make this mirror symmetry visible
are the hyper-Kähler structure of MH and its Hitchin fibration. Those structures
depend on the choice of a complex structure on C, but in fact, the resulting mirror
symmetry does not really depend on this choice. This was shown in Kapustin
and Witten (2007) in the process of deriving this example of mirror symmetry
from a four-dimensional topological field theory. The topological field theory in
question is obtained by twisting of N = 4 super Yang–Mills theory.

7.3.1 A few hints

There are a few obstacles to overcome to go from this instance of mirror sym-
metry to the usual formulation of geometric Langlands duality. Unfortunately,
it will not be practical here to give more than a few hints.
One key point is that in the usual formulation, the dual of a B-brane on
MH (L G, C) is supposed to be not an A-brane on MH (G, C)—which is what
we most naturally get from the above construction—but a sheaf of D-modules
on M(G, C), the moduli space of G-bundles over C (a sheaf of D-modules is
by definition a sheaf of modules for the sheaf D of differential operators on
M(G, C)). The link between the two statements is explained in section 11
of Kapustin and Witten (2007), using the existence of a special A-brane on
MH (G, C) that is intimately related to differential operators on M(G, C). This
relation is possible because, as explained in Hitchin (1987a), MH (G, C) contains
T ∗ Mst (G, C) as a Zariski open set; here Mst (G, C) is the subspace of M(G, C)
parametrizing strictly stable bundles.
Another key point is the following. A central role in the usual formulation is
played by the geometric Hecke operators, which act on holomorphic G-bundles
over C and therefore also on D-modules on M(G, C). They have a natural role in
the present story, but this is one place that one misses something if one attempts
to express this subject just in terms of two-dimensional sigma models and mirror
symmetry. This particular instance of mirror symmetry actually originates from a
118 Mirror symmetry, Hitchin’s equations, and Langlands duality

duality in an underlying four-dimensional gauge theory. Once this is understood,

basic facts about the Wilson and ’t Hooft line operators of gauge theory lead to
the usual statements about Hecke eigensheaves, as explained in sections 9 and
10 of Kapustin and Witten (2007). The geometric Hecke operators are naturally
reinterpreted in this context in terms of the Bogomolny equations of three-
dimensional gauge theory, which are of great geometrical as well as physical
interest and have been much studied, for example in Atiyah and Hitchin (1988).
A proper formulation of some of these statements leads to another important
role for four dimensions. The usual formulation of geometric Langlands involves
D-modules not on the moduli space of semi-stable G-bundles but on the moduli
stack of all G-bundles. The main reason for this is that one cannot see the
action of the Hecke operators if one considers only semi-stable bundles. As we
will explain in Section 7.6, the role of stacks in the standard description can
be understood as a strong clue for an alternative approach that starts in four-
dimensional gauge theory.

7.4 Ramification
Before getting back to stacks, however, I want to give an idea of what is called
“ramification” in the context of geometric Langlands.
A simple generalization of what we have said so far is to consider flat bundles
not on a closed oriented two-manifold C but on a punctured two-manifold C  =
C\p; that is, C  is C with a point p omitted.
We pick a conjugacy class C ⊂ GC , and we let Y(G, C  ; C) denote the moduli
space of homomorphisms ρ : π1 (C  ) → GC , up to conjugation, such that the
monodromy around p is in the conjugacy class C.
Many statements that we made before have natural analogs in this punctured
case. In particular, Y(G, C  ; C) has a natural structure of a complex symplectic
manifold. It has a natural complex structure and holomorphic symplectic form
Ω. Just as in the unpunctured case, we can define a B-model of Y(G, C  ; C).
Also, viewing Y(G, C  ; C) as a real symplectic manifold with symplectic form
ω = Im Ω, we can define an A-model. The B-model and the A-model are both
completely independent of the complex structure of C  .
Next, introduce the dual group L G and let L C denote a conjugacy class in
its complexification. Again, the space Y(L G, C  ; L C) has a natural B-model and
A-model.
Based on what we have said so far, one may wonder if, for some map
between C and L C, there might be a mirror symmetry between Y(G, C  ; C)
and Y(L G, C  ; L C). The answer to this question is “not quite,” for a number
of reasons. One problem is that there is no natural correspondence between
conjugacy classes in GC and in L GC . A more fundamental problem is that the
B-model of Y(G, C  ; C) varies holomorphically with the conjugacy class C, but
the A-model of the same space does not. To find a version of the statement
 Ramification 119

that has a chance of being right, we have to add additional parameters to find a
mirror-symmetric set.
In any event, regardless of what parameters one adds, it is very difficult to
answer the question about mirror symmetry if C  is viewed simply as an oriented
two-manifold with a puncture. To make progress, just as in the unramified case
(i.e. the case without punctures), it is very helpful to endow C  with a complex
structure and to use Hitchin’s equations. This actually also helps us in finding the
right parameters, because an improved set of parameters appears just in trying
to give a natural formulation of Hitchin’s equations on a punctured surface. Let
z be a local parameter near the puncture and write z = reiθ . In the punctured
case, it is natural (see Simpson 1990) to introduce variables α, β, γ taking values
in the Lie algebra t of a maximal torus T ⊂ G, and consider solutions of Hitchin’s
equations on C  whose behavior near z = 0 is as follows:

A = α dθ + . . . (7.2)
dr
φ=β − γ dθ + . . . . (7.3)
r
The ellipses refer to terms that are less singular near z = 0.
All the usual statements about Hitchin’s equations have close analogs in this
situation. The moduli space of solutions of Hitchin’s equations with this sort
of singularity is a hyper-Kähler manifold MH (G, C; α, β, γ). In one complex
structure, usually called J, it coincides with Y(G, C; C), where C is the conju-
gacy class that contains 4 U = exp(−2π(α − iγ)). In another complex structure,
often called I, MH (G, C; α, β, γ) parametrizes Higgs bundles (E, ϕ), where
ϕ ∈ H 0 (C  , K ⊗ ad(E)) has a pole at z = 0, with ϕ ∼ 12 (β + iγ)(dz/z). More-
over, there is a Hitchin fibration, and most of the usual statements about the
unramified case—those that we have explained and those that we have omitted
here—have close analogs. For a much more detailed explanation, and references
to the original literature, see Gukov and Witten (2006).
The variables α, β, γ are a natural set of parameters for the classical geometry.
However, quantum mechanically, there is one more natural variable η (analogous
to the usual θ angles of gauge theory), as described in section 2.3 of Gukov
and Witten (2006). With the complete set of parameters (α, β, γ, η) at hand,
it is possible to formulate a natural duality statement, according to which
MH (L G, C; L α, L β, L γ, L η) is mirror to MH (G, C; α, β, γ, η), with a certain map
between the parameters, described in section 2.4 of Gukov and Witten (2006).
The main point of this map is that (α, η) = (L η, −L α). Since the monodromy U
depends on L α, this shows that the dual of the monodromy involves the quantum
parameter η that is invisible in the classical geometry. In the A-model, η becomes
the imaginary part of the complexified Kähler class.

4 For simplicity, we assume that U is regular. The more involved statement that holds in

general is explained in Gukov and Witten (2006).

120 Mirror symmetry, Hitchin’s equations, and Langlands duality

This duality statement leads, after again mapping A-branes to D-modules, to

a statement of geometric Langlands duality for this situation similar to what
has been obtained via algebraic geometry and two-dimensional conformal field
theory in Frenkel and Gaitsgory (2005).

Remark 7.2 We pause here to explain one very elementary fact about the
classical geometry that will be helpful as background for Section 7.5. In complex
structure J, a solution of Hitchin’s equations with the singularity of (7.2)
describes a flat GC bundle E → C  with monodromy around the puncture p.
E can be extended over p as a holomorphic bundle, though of course not as a
flat one, and moreover from a holomorphic point of view, E can be trivialized
near p. The flat connection on E → C  is then represented, in this gauge, by a
holomorphic (1, 0)-form on C  (valued in the Lie algebra of GC ) with a simple
pole at p:


α − iγ
A = dz + ... , (7.4)
iz

where the omitted terms are regular at z = 0. The singularity of the connection
at z = 0 is a simple pole because the ansatz (7.2) for Hitchin’s equations only
allows a singularity of order 1/|z|. A holomorphic connection with such a simple
pole is said to have a regular singularity.

In geometric Langlands, what is usually called tame ramification is, roughly

speaking, the case that we have just arrived at: a holomorphic bundle E → C
that has a holomorphic connection form with a regular singularity. Actually,
the phrase “tame ramification” is sometimes taken to refer to the case that
the residue of the simple pole is nilpotent, while in (7.4) we seem to be in the
opposite case of semi-simple residue. In Gukov and Witten (2006), it is explained
that, with some care, mirror symmetry for MH (G, C  ; α, β, γ, η) is actually a
sufficient framework to understand geometric Langlands for a connection with a
simple pole of any residue. For example, the case of a nilpotent residue can be
understood by setting L α = L γ = 0 (or γ = η = 0 in the dual description).

7.5 Wild ramification

Based on an analogy with number theory, geometric Langlands is usually formu-
lated not only for the case of tame ramification. One goes on to inquire about
an analogous duality statement involving a holomorphic bundle E → C with
a holomorphic connection that has a pole of any order. In other words, after
trivializing the holomorphic structure of E near a point p ∈ C, the connection
looks like


Tn Tn−1 T1
A = dz + + · · · + + · · · , (7.5)
zn z n−1 z
 Wild ramification 121

where regular terms are omitted. A meromorphic connection with a pole of degree
greater than 1 is said to have an irregular singularity.
Trying to formulate a duality statement for this situation poses, at first sight,
a severe challenge for the approach to geometric Langlands described here. Our
basic point of view is that the fundamental duality statements depend on C only
as an oriented two-manifold. A complex structure on C is introduced only as a
tool to answer certain natural questions that can be asked without introducing
the complex structure.
From this point of view, tame ramification is natural because a simple pole
in this sense has a clear topological meaning. A meromorphic connection with a
simple pole at a point p ∈ C is a natural way to encode the monodromy about
p of a flat connection on C  = C\p. And this monodromy, of course, is a purely
topological notion. But what could possibly be the topological meaning of a
connection with a pole of degree greater than 1?
A closely related observation is that T1 is the residue of the pole in A at
z = 0, and so is independent of the choice of local coordinate z. However, the
coefficients T2 , . . . , Tn of the higher order poles most definitely do depend on the
choice of a local coordinate. How can we hope to include them in a theory that
is supposed to depend on C only as an oriented two-manifold?
Moreover, if the plan is to formulate a duality conjecture of a topological nature
and then prove it using Hitchin’s equations, we face the question of whether
Hitchin’s equations are compatible with an irregular singularity. Hitchin’s equa-
tions for a pair Φ = (A, φ) are schematically of the form dΦ + Φ2 = 0. If near
z = 0, we have a singularity with |Φ| ∼ 1/|z|n , then |dΦ| ∼ 1/|z|n+1 and |Φ|2 ∼
1/|z|2n . For n = 1, dΦ and |Φ|2 are comparable in magnitude, and therefore
Hitchin’s equations look reasonable. However, for n > 1, we have |Φ|2 >> |dΦ|,
and it looks like the non-linear term in Hitchin’s equations will be too strong.
Both questions, however, have natural answers. The answer to the first ques-
tion is that, despite appearances, one actually can associate to a connection
with irregular singularity something that goes beyond the ordinary monodromy
and has a purely topological meaning. The appropriate concept is an extended
monodromy that includes Stokes matrices as well as the ordinary monodromy.
Stokes matrices are part of the classical theory of ordinary differential equations
with irregular singularity (e.g. see Wasow 1965).
Assuming for brevity that the leading coefficient Tn of the singular part of
the connection is regular semi-simple, one can make a gauge transformation to
conjugate T1 , . . . , Tn to the maximal torus. Then one defines a moduli space
Y(G, C; T1 , . . . , Tn ) that parametrizes, up to a gauge transformation, pairs con-
sisting of a holomorphic GC -bundle over C and a connection with an irregular
singularity of the form described in (7.5). As shown in Boalch (2001), it turns
out that this space Y(G, C; T1 , . . . , Tn ) is in a natural way a complex symplectic
manifold, with a complex symplectic structure that depends on C only as an
oriented two-manifold. This can be proved by adapting to the present setting
the gauge theory definition of the symplectic structure, formulated in Atiyah and
122 Mirror symmetry, Hitchin’s equations, and Langlands duality

Bott (1982). Moreover, the complex symplectic structure of Y(G, C; T1 , . . . , Tn )

is independent of T2 , . . . , Tn (as long as Tn remains semi-simple).
At this point, the important concept of isomonodromic deformation, intro-
duced by Jimbo et al. (1981), comes into play. There is a natural way to vary the
parameters T2 , . . . , Tn , without changing the generalized monodromy data that is
parametrized by Y(G, C; T1 , . . . , Tn ). Moreover, as has been proved quite recently
in Boalch (2001), the complex symplectic structure of the space of generalized
monodromy data is invariant under isomonodromic deformation. Thus, roughly
speaking, one can define a complex symplectic manifold Y(G, C; T1 , n) that
depends only on T1 and the integer n ≥ 1.
The fact that the parameters T2 , . . . , Tn turn out to be, in the sense just
described, inessential is certainly welcome, since as we have already observed,
they have no evident topological meaning. Now we are in a situation very similar
to what we had in the unramified and tamely ramified cases. Given Y(G, C; T1 , n)
as a complex symplectic manifold, with complex symplectic form Ω, we can
define its B-model or its A-model using the real symplectic form ω = Im Ω.
Of course, we can do the same for the dual group, defining another complex
symplectic manifold Y(L G, C; L T1 , n), with its own B-model and A-model. And,
just as in the unramified case, we can ask if these two models are mirror to each
other.
Even before trying to answer this question, we should refine it slightly. Because
of the constraint that Tn should be regular semi-simple, it is not quite correct to
simply forget about Tn . There can be monodromies as Tn varies. We think of Tn
as taking values in treg C ⊗ Kp
n−1
, with notation as follows: tC is the Lie algebra of
reg
a maximal torus in GC , tC is its subspace consisting of regular elements, and Kp
is the fiber at p of the cotangent bundle to C. The fundamental group of treg C is
known as the braid group of G; we call it B(G). Because of the monodromies, one
really needs to choose a basepoint ∗ ∈ treg C to define Y( G, C; T1 , n); to be more
L L

precise, we can denote this space as Y( G, C; T1 , n, ∗). The group B(G) acts via
L L

monodromies on both the B-model and the A-model of Y(G, C; T1 , n, ∗). Dually,
the corresponding braid group B(L G) acts on the B-model and the A-model
of Y(L G, C; L T1 , n, ∗). However, the two groups B(G) and B(L G) are naturally
isomorphic; indeed, modulo a choice of an invariant quadratic form, there is
a natural map from treg L reg
C to tC , so the two spaces have the same fundamental
group and a choice of basepoint in one determines a basepoint in the other, up to
homotopy. A better (but still not yet precise) question is whether there is a mirror
symmetry between Y(G, C; T1 , n, ∗) and Y(L G, C; L T1 , n, ∗) that commutes with
the braid group.
We expect as well that this mirror symmetry depends on C only as an oriented
two-manifold, and so commutes with the mapping class group. We can think of
the mapping class group of C and the braid group as playing quite parallel roles.
In fact, because of the appearance of the fiber Kp of the canonical bundle in the
last paragraph, these two groups do not simply commute with each other; the
group that acts is an extension of the mapping class group by B(G).
 Four-dimensional gauge theory and stacks 123

Just as in the tamely ramified case, to get the right mirror symmetry con-
jecture, we need to extend the parameters slightly to get a mirror-symmetric
set. But we also face the fundamental question of whether Hitchin’s equations
are compatible with wild ramification. As explained above, the non-linearity of
Hitchin’s equations makes this appear doubtful at first sight. But happily, it
turns out that all is well, as shown in Biquard and Boalch (2004). The key point
is that, again with Tn assumed to be regular semi-simple, we can assume that
the singular part of the connection is abelian. Though Hitchin’s equations are
non-linear, they become linear in the abelian case, and once abelianized, they are
compatible with a singularity of any order. Using this as a starting point, it turns
out that, for any n, one can develop a theory of Hitchin’s equations with irregular
singularity that is quite parallel to the more familiar story in the unramified case.
For example, the moduli space MH of solutions of the equations is hyper-Kähler.
In one complex structure, MH parametrizes flat connections with a singularity
similar to that in (7.5); in another complex structure, it parametrizes Higgs
bundles (E, ϕ) in which ϕ has an analogous pole of order n. There is a Hitchin
fibration, and all the usual properties have close analogs.
All this gives precisely the right ingredients to use Hitchin’s equations to
establish the desired mirror symmetry between the two moduli spaces. See
Witten (2007) for a detailed explanation in which this classical geometry is
embedded in four-dimensional gauge theory. Many of the arguments are quite
similar to those given in the tame case in Gukov and Witten (2006). The
construction makes it apparent that the duality commutes with isomonodromic
deformation.
Finally, one might worry that the assumption that Tn is regular semi-simple
may have simplified things in some unrealistic way. This is actually not the
case. For one thing, the analysis in Biquard and Boalch (2004) requires only
that T2 , . . . , Tn should be simultaneously diagonalizable (in some gauge), and in
particular semi-simple, but not that Tn is regular. But even if these coefficients
are not semi-simple, there is no essential problem. In the classical theory of
ordinary differential equations, it is shown that given any such equation with
an irregular singularity at z = 0, after possibly passing to a finite cover of the
punctured z-plane and changing the extension of a holomorphic bundle over
the puncture at z = 0, one can reduce to the case that the irregular part of
the singularity has the properties assumed in Biquard and Boalch (2004). Given
this, one can adapt all the relevant arguments concerning geometric Langlands
duality to the more general case, as is explained in section 6 of Witten (2007).

7.6 Four-dimensional gauge theory and stacks

To a physicist, it is natural, in studying dualities involving gauge theory, to begin
in four dimensions, which is often found to be the natural setting for gauge theory
duality. There is a simple reason for this. The curvature, which is one of the most
fundamental concepts in gauge theory, is a two-form. In d dimensions, the dual
124 Mirror symmetry, Hitchin’s equations, and Langlands duality

of a two-form is a (d − 2)-form, so it is only a two-form if d = 4. This suggests

that d = 4 is the most natural dimension in which the dual of a gauge theory
might be another gauge theory.
Moreover, N = 4 supersymmetric Yang–Mills theory, originally constructed in
Brink et al. (1977), is a natural place to start, as it has the maximal possible
supersymmetry, and has the celebrated duality whose origins go back to the
early work of Montonen and Olive (1977). It indeed turns out that geometric
Langlands has a natural origin in a twisted version of N = 4 super Yang–Mills
theory in four dimensions. The twisting is quite analogous to the twisting of
N = 2 super Yang–Mills theory that leads to Donaldson theory.
That particular motivation may seem opaque to some, and instead I will
adopt here a different approach in explaining why it is natural to begin in four
dimensions for understanding geometric Langlands, instead of relying only on
the B-model and the A-model of MH (G, C).
First of all, the B-model and the A-model of any space X are both twisted
versions of a quantum sigma model that governs maps Φ : Σ → X, where Σ is
a two-manifold (or better, a supermanifold of bosonic dimension 2). Since the
A-model involves in its most elementary form a counting of holomorphic maps
Φ : Σ → X that obey appropriate conditions, the roles of Σ and Φ are clear in the
A-model. Mirror symmetry indicates that it must be correct to also formulate the
B-model in terms of maps Φ : Σ → X, and this is done in the usual formulation
by physicists.
In the present case, we are interested, roughly speaking, in the B-model and
A-model of MH (G, C), for some compact Lie group G and two-manifold C.
Therefore, roughly speaking, we want to study a sigma model of maps Φ : Σ →
MH (G, C), where as before Σ is an auxiliary two-manifold.
The reason that this description is rough is that MH (G, C) has singularities, 5
and the sigma model of target MH (G, C) is therefore not really well defined.
Therefore, a complete description cannot be made purely in terms of a sigma
model in which the target space is MH (G, C), viewed as an abstract manifold.
We need a more complete description that will tell us how to treat the singular-
ities. What might this be?
By definition, a point in MH (G, C) determines up to gauge-equivalence a pair
(A, φ) obeying Hitchin’s equations. A and φ are fields defined on C, so let us
write them more explicitly as (A(y), φ(y)), where y is a coordinate on C.
Now suppose that we have a map Φ : Σ → MH (G, C), where Σ is a Riemann
surface with a local coordinate x. Such a map is described by a pair (A(y), φ(y))
that also depends on x. So we can describe the map Φ via fields (A(x, y), φ(x, y))
that depend on both x and y. We would like to interpret these fields as fields on
the four-manifold M = Σ × C. The pair (A(x, y), φ(x, y)) is not quite a natural

5 Moreover, these singularities are worse than orbifold singularities. Orbifold singularities

would cause no difficulty. See Frenkel and Witten (2007) for a discussion of orbifold singularities
in geometric Langlands.
 Four-dimensional gauge theory and stacks 125

set of fields on M but can be naturally completed to one. For example, A(x, y) is
locally a one-form tangent to the second factor in M = Σ × C; to get a four-
dimensional gauge field, we should relax the condition that A is tangent to
the second factor. Similarly, we can extend φ to an adjoint-valued one-form
on Σ × C. N = 4 super Yang–Mills theory, or rather its twisted version that is
related to geometric Langlands, is obtained by completing this set of fields to a
supersymmetric combination in a minimal fashion.
In N = 4 super Yang–Mills theory, there are no singularities analogous to
the singularities of MH (G, C). The space of gauge fields, for example, is an
affine space, and the other fields (such as φ) take values in linear spaces. The
problems with singularities that make it difficult to define a sigma model of
maps Φ : Σ → MH (G, C) have no analog in defining gauge theory on M = Σ × C
(or any other four-manifold). The relation between the two is that the two-
dimensional sigma model is an approximation to the four-dimensional gauge
theory. The approximation breaks down when one runs into the singularities of
MH (G, C). Any question that involves those singularities should be addressed
in the underlying four-dimensional gauge theory.
But away from singularities, it suffices to consider only the smaller set of fields
that describe a map Φ : Σ → MH (G, C). Many questions do not depend on the
singularities and for these questions the description via two-dimensional sigma
models and mirror symmetry is adequate.

7.6.1 Stacks
To conclude, we will make contact with the counterpart of this discussion in the
usual mathematical theory. We start with bundles rather than Higgs bundles
because this case will be easier to explain.
In the usual mathematical theory, the right-hand side of the geometric Lang-
lands correspondence is described in terms of D-modules on, roughly speaking,
the moduli space of all holomorphic GC bundles on the Riemann surface C.
However, instead of the moduli space M(G, C) of semi-stable holomorphic GC
bundles E → C, one considers D-modules on the “stack” BunG (C) of all such
bundles. The main reason for this is that to define the action of Hecke operators,
it is necessary to allow unstable bundles. Unstable bundles are related to the
nonorbifold singularities of M(G, C).
What is a stack? Roughly, it is a space that can everywhere be locally described
as a quotient. The trivial case is a stack that can actually be described globally as
a quotient. Interpreting BunG (C) as a global quotient would mean finding a pair
(Y, WC ), consisting of a smooth algebraic variety Y and a complex Lie group WC
acting on Y , with the following properties. Isomorphism classes of holomorphic
GC bundles E → C should be in one-to-one correspondence with WC orbits on
Y , and for every E → C, its automorphism group should be isomorphic to the
subgroup of WC leaving fixed the corresponding point in Y .
126 Mirror symmetry, Hitchin’s equations, and Langlands duality

A pair (Y, WC ) representing in this way the stack BunG (C) does not exist
if Y and WC are supposed to be finite-dimensional. Indeed, the GC -bundle
E → C can be arbitrarily unstable, so there is no upper bound on the dimension
of its automorphism group. So no finite-dimensional WC can contain all such
automorphism groups as subgroups.
However as shown in Atiyah and Bott (1982), taking G to be of adjoint type
for simplicity, there is a natural infinite-dimensional pair (Y, WC ). One simply
takes Y to be the space of all connections on a given G-bundle E → C which
initially is defined only topologically. One defines W to be the group of all gauge
transformations of the bundle E; thus, if E is topologically trivial, we can identify
W as the group Maps(C, G). Then we take WC to be the complexification of W ,
or in other words Maps(C, GC ). (This complexification acts on Y as follows. We
associate to a connection A the corresponding ∂¯ operator ∂¯A . Then a complex-
valued gauge transformation acts by ∂¯A → g ∂¯A g −1 .)
Suppose then that we were presented with the problem of making sense of the
supersymmetric sigma model of maps Φ : Σ → M(G, C), given the singularities
of M(G, C). (This is a practice case for our actual problem, which involves
MH (G, C) rather than M(G, C).) Our friends in algebraic geometry would tell
us to replace M(G, C) by the stack BunG (C). We interpret this stack as the pair
(Y, WC ), where Y is the space of all connections on a G-bundle E → C and WC
is the complexified group of gauge transformations. The connected components
of the stack correspond to the topological choices for E.
By a supersymmetric sigma model with target a pair (Y, WC ), with WC a
complex Lie group acting on a complex manifold Y , we mean 6 in the finite-
dimensional case a gauge-invariant supersymmetric sigma model in which the
gauge group is W (a maximal compact subgroup of WC ) and the target is Y .
Actually, to define this sigma model, we want Y to be a Kähler manifold with a
W -invariant (but of course not WC -invariant) Kähler structure. The sigma model
action contains a term which is the square of the moment map for the action
of W . This term is minimized precisely when the moment map vanishes. The
combined operation of setting the moment map to zero and dividing by W is
equivalent classically to dividing by WC .
To write down the term in the action that involves the square of the moment
map (and in fact, to write down the kinetic energy of the gauge fields) one needs
an invariant and positive definite quadratic form on the Lie algebra of W . If W
is finite-dimensional, existence of such a form is equivalent to W being compact.
However, the appropriate quadratic form also exists in the infinite-dimensional
case that W = Maps(C, G) for some space C. (An element of the Lie algebra of
W is a g-valued function on C, and the quadratic form is defined by C dµ ( , ),
where (, ) is an invariant positive-definite quadratic form on g, and µ is a suitable
measure on C.)

6 For a discussion of this construction in relation to stacks, see Pantev and Sharpe (2006).
 References 127

Now, suppose we construct the two-dimensional sigma model of maps from

a Riemann surface Σ to the stack BunG (C), understood as above. What is the
group of gauge transformations of the sigma model? In general, in a gauge theory
on any space Σ with gauge group W , the group of gauge transformations (of a
topologically trivial W -bundle, for simplicity) is the group of maps from Σ to
W , or Maps(Σ, W ). In our case, W is in turn Maps(C, G). So Maps(Σ, W ) is the
same as Maps(M, G), where M = Σ × C. But this is simply the group of gauge
transformations in gauge theory on M , with gauge group G. In the present case,
Σ and C are two-manifolds and M is a four-manifold. We have arrived at four-
dimensional gauge theory. If we chase through the definitions a little more, we
learn that the supersymmetric sigma model of maps Φ : Σ → BunG (C) should
be understood as four-dimensional N = 2 supersymmetric gauge theory, with
gauge group G, on the four-manifold M = Σ × C. (This is the theory that after
twisting is related to Donaldson theory.)
Now let us return to the original problem. Geometric Langlands duality is a
statement about the B-model and A-model not of M(G, C) but of MH (G, C),
the corresponding moduli space of Higgs bundles, and its analog for the dual
group L G. To deal with the singularities, we want to “stackify” this situation.
We are now in a hyper-Kähler context and the appropriate concept of a stack
should incorporate this. (What algebraic geometers would call the stack of Higgs
bundles does not quite do justice to the situation, since it emphasizes one complex
structure too much.) Since quaternionic Lie groups do not exist, we cannot ask
to construct MH (G, C) as the quotient of a smooth space by a quaternionic
Lie group. However, the notion of a symplectic quotient does have a good
analog in the hyper-Kähler world, namely, the hyper-Kähler quotient, described
in Hitchin et al. (1987). The analog of what we explained for M(G, C) is to
realize MH (G, C) as the hyper-Kähler quotient of a smooth space Y by a group
W . It may be impossible to do this with finite-dimensional Y and W , but in
the infinite-dimensional world, this problem has a natural solution described in
Hitchin’s original paper (1987a) on the Hitchin equations. (Y is the space of pairs
(A, φ) on C, and W = Maps(C, G).) Taking this as input and interpreting what
it should mean to have a sigma model whose target is the hyper-Kähler stack
corresponding to MH (G, C), one arrives at the twisted version of N = 4 super
Yang–Mills theory that was the starting point in Kapustin and Witten (2007).

Acknowledgement
This chapter was supported in part by NSF Grant Phy-0503584.

References
Atiyah, M. F. and Bott, R. (1982). The Yang-Mills equations over Riemann surfaces.
Phil. Trans. R. Soc. London A308, 166–86.
Atiyah, M. F. and Hitchin, N. J. (1988). The Physics and Geometry of Monopoles.
Princeton University Press, Princeton, NJ.
128 Mirror symmetry, Hitchin’s equations, and Langlands duality

Bershadsky, M., Johansen, A., Sadov, V. and Vafa C. (1995). Topological reduction of
4 − D SYM to 2 − D sigma models. Nucl. Phys. B448, 166–86.
Biquard, O. and Boalch, P. (2004). Wild non-abelian Hodge theory on curves. Compo-
sitio Mathematica 140, 179–204.
Boalch, P. (2001). Symplectic manifolds and isomonodromic deformations. Adv. Math.
163, 137–205.
Brink, L., Schwarz, J. H. and Scherk, J. (1977). Supersymmetric Yang-Mills theories.
Nucl. Phys. B121, 77.
Frenkel, E. and Gaitsgory, D. (2005). Local geometric Langlands correspondence and
affine Kac-Moody algebras. math.RT/0508382.
Frenkel, E. and Witten, E. (2007). Geometric endoscopy and mirror symmetry.
arXiv.0710.5939.
Gualtieri, M. (2003). Generalized complex geometry. DPhil thesis, Oxford University,
Oxford, math.DG/0401221.
Gukov, S. and Witten, E. (2006). Gauge theory, ramification, and the geometric
Langlands program. hep-th/0612073.
Harvey, J. A., Moore, G. W. and Strominger A. (1995). Reducing S duality to T duality.
Phys. Rev. D52, 7161–7.
Hausel, T. and Thaddeus, M. (2002). Mirror symmetry, Langlands duality, and the
Hitchin system. math.AG/0205236.
Hitchin, N. J. (1987a). The self-duality equations on a Riemann surface. Proc. London
Math. Soc. (3) 55, 67.
Hitchin, N. J. (1987b). Stable bundles and integrable systems. Duke Math. J. 54, 91–
114.
Hitchin, N. J. (2003). Generalized Calabi-Yau manifolds. Q. J. Math. 54, 281–308,
math.DG/0209099.
Hitchin, N. J., Karlhede, A., Lindstrom, U. and Rocek M. (1987). Hyperkähler metrics
and supersymmetry. Commun. Math. Phys. 108, 535–89.
Hori, K. et al., (2003). editors. Mirror Symmetry. American Mathematical Society,
Providence, RI.
Jimbo, M., Miwa, T. and Ueno, K. (1981). Monodromy preserving deformations of
linear differential equations with rational coefficients. Physica 2D, 407–48.
Kapustin, A. and Witten, E. (2007). Electric-magnetic duality and the geometric
Langlands program. Commun. Number Theory Phys. 1, 1–236, hep-th/0604151.
Montonen, C. and Olive, D. (1977). Magnetic monopoles as gauge particles?. Phys.
Lett. B72, 117.
Pantev, T. and Sharpe, E. (2006). GLSM’s for gerbes (and other toric stacks). Adv.
Theor. Math. Phys. 10, 77–121, hep-th/0502053.
Simpson, C. (1990). Harmonic bundles on noncompact curves. J. Am. Math. Soc. 3,
713–70.
Strominger, A., Yau, S.T. and Zaslow, E. (1996). Mirror symmetry is T duality. Nucl.
Phys. B479, 243–59, hep-th/9606040.
Wasow, W. (1965). Asymptotic Expansions for Ordinary Differential Equations. Wiley,
New York.
Witten, E. (2007). Gauge theory and wild ramification. arXiv:0710.0631.
 VIII
HIGGS BUNDLES AND GEOMETRIC STRUCTURES
ON SURFACES

William M. Goldman
Dedicated to Nigel Hitchin on the occasion of his 60th birthday

8.1 Introduction
In the late 1980s Hitchin (1987) and Simpson (1988) discovered deep connec-
tions between representations of fundamental groups of surfaces and algebraic
geometry. The fundamental group π = π1 (Σ) of a closed orientable surface Σ of
genus g > 1 is an algebraic object governing the topology of Σ. For a Lie group
G, the space of conjugacy classes of representations π → G is a natural algebraic
object Hom(π, G)/G whose geometry, topology, and dynamics intimately relates
to the topology of Σ and the various geometries associated with G. In partic-
ular Hom(π, G)/G arises as a moduli space of locally homogeneous geometric
structures as well as flat connections on bundles over Σ.
Giving Σ a conformal structure profoundly affects π and its representations.
This additional structure induces further geometric and analytic structure on the
deformation space Hom(π, G)/G. Furthermore this analytic interpretation allows
Morse-theoretic methods to compute the algebraic topology of these non-linear
finite-dimensional spaces.
For example, when G = U(1), the space of representations is a torus of dimen-
sion 2g. Giving Σ a conformal structure denotes the resulting Riemann surface
by X. The classical Abel–Jacobi theory identifies representations π1 (X) −→ U(1)
with topologically trivial holomorphic line bundles over X. The resulting Jacobi
variety is an abelian variety, whose structure strongly depends on the Riemann
surface X. However the underlying symplectic manifold depends only on the
topology of Σ, and indeed just the fundamental group π. (Compare Goldman
1984).
Another important class of representations of π arises from introducing the
local structure of hyperbolic geometry to Σ. Giving Σ a Riemannian metric
of curvature −1 determines a representation ρ in the group G = Isom+ (H2 ) ∼ =
PSL(2, R). These representations, which we call Fuchsian, are characterized
as embeddings of π onto discrete subgroups of G. Equivalence classes of
130 Higgs bundles and geometric structures on surfaces

Fuchsian representations comprise the Fricke–Teichmüller space F(Σ) of marked

hyperbolic structures on Σ, which embeds in Hom(π, G)/G as a connected com-
ponent. This component is a cell of dimension 6g − 6 upon which the mapping
class group acts properly.
The theory of Higgs bundles, pioneered by Hitchin and Simpson, provides
an analytic approach to studying surface group representations and their defor-
mation space. The purpose of this chapter is to describe the basic examples
of this theory, emphasizing relations to deformation and rigidity of geometric
structures. In particular we report on some very recent developments when G is
a real Lie group, either a split real semisimple group or an automorphism group
of a Hermitian symmetric space of noncompact type.
In the 20 years since the appearance of Hitchin’s and Simpson’s work, many
other developments directly arose from their works. These relate to variations
of Hodge structures, spectral curves, integrable systems, Higgs bundles over
noncompact Riemann surfaces and higher dimensional Kähler manifolds, and
the finer topology of the deformation spaces. None of these topics are discussed
here. It is an indication of the power and the depth of these ideas that so many
mathematical subjects have been profoundly influenced by the pioneering work
of Hitchin and Simpson.

8.2 Representations of the fundamental group

8.2.1 Closed surface groups
Let Σ = Σg be a closed orientable surface of genus g > 1. Orient Σ, and choose a
smooth structure on Σ. Ignoring basepoints, denote the fundamental group π1 (Σ)
of Σ by π. The familiar decomposition of Σ as a 4g-gon with 2g identifications
(depicted in Figures 8.1 and 8.2) of its sides leads to a presentation
π = A1 , B1 , . . . , Ag , Bg | [A1 , B1 ] · · · [Ag , Bg ] = 1 (8.1)
where [A, B] := ABA−1 B −1 .

8.2.2 Representation variety

ρ
Denote the set of representations π − → G by Hom(π, G). Evaluation on a collec-
tion γ1 , . . . , γN ∈ π defines a map
Hom(π, G) −→ GN
⎡ ⎤
ρ(γ1 )
⎢ ⎥
ρ −→ ⎣ ... ⎦ (8.2)
ρ(γN )
 Representations of the fundamental group 131

b2

a2 a1

b1

Figure 8.1 The pattern of identifications for a genus 2 surface. The sides of
an octagon are pairwise identified to construct a surface of genus 2. The 8
vertices identify to a single 0-cell in the quotient, and the 8 sides identify to
four 1-cells, which correspond to the 4 generators in the standard presentation
of the fundatmental group.

which is an embedding if γ1 , . . . , γN generate π. Its image consists of N -tuples

(g1 , . . . , gN ) ∈ GN
satisfying equations
R(g1 , . . . , gN ) = 1 ∈ G
where R(γ1 , . . . , γN ) are defining relations in π satisfied by γ1 , . . . , γN . If G is a
linear algebraic group, these equations are polynomial equations in the matrix
entries of gi . Thus the evaluation map (8.2) identifies Hom(π, G) as an algebraic
subset of GN . The resulting algebraic structure is independent of the generating

a2
b2

b1 a1

Figure 8.2 The genus 2 surface as an identification space.

132 Higgs bundles and geometric structures on surfaces

set. In particular Hom(π, G) inherits both the Zariski and the classical topology.
We consider the classical topology unless otherwise noted.
In terms of the standard presentation (8.1), Hom(π, G) identifies with the
subset of G2g consisting of

(α1 , β1 , . . . αg , βg )

satisfying the single G-valued equation

[α1 , β1 ] · · · [αg , βg ] = 1.

8.2.3 Symmetries
The product Aut(π) × Aut(G) acts naturally by left and right compositions, on
Hom(π, G): An element

(φ, α) ∈ Aut(π) × Aut(G)

transforms ρ ∈ Hom(π, G) to the composition

φ−1 ρ α
π −−→ π −
→G−
→ G.

The resulting action preserves the algebraic structure on Hom(π, G).

8.2.4 Deformation space

For any group H, let Inn(H) denote the normal subgroup of Aut(H) comprising
inner autormorphisms. The quotient group Aut(H)/Inn(H) is the outer auto-
morphism group, denoted Out(H).
We will mainly be concerned with the quotient
 
Hom(π, G)/G := Hom(π, G)/ {1} × Inn(G) ,

which we call the deformation space. For applications to differential geometry,

such as moduli spaces of flat connections (gauge theory) or locally homogeneous
geometric structures, it plays a more prominent role than the representation vari-
ety Hom(π, G). Although Inn(G) preserves the algebraic structure, Hom(π, G)/G
will generally not admit the structure of an algebraic set.
Since the Inn(G) action on Hom(π, G) absorbs the Inn(π) action on Hom(π, G),
the outer automorphism group Out(π) acts on Hom(π, G)/G. By a theorem of M.
Dehn and J. Nielsen (compare Nielsen 1927 and Stillwell 1987), Out(π) identifies
with the mapping class group
 
Mod(Σ) := π0 Diff(Σ) .

One motivation for this study is that the deformation spaces provide natural
objects upon which mapping class groups act (Goldman 2006).
 Abelian groups and rank 1 Higgs bundles 133

8.3 Abelian groups and rank 1 Higgs bundles

The simplest groups are commutative. When G is abelian, then the commmuta-
tors [α, β] = 1 and the defining relation in (8.1) is vacuous. Thus
Hom(π, G) ←→ G2g .
Furthermore Inn(G) is trivial so
Hom(π, G)/G ←→ G2g
as well.

8.3.1 Symplectic vector spaces

Homological machinery applies. By the Hurewicz theorem and the universal
coefficient theorem,
Hom(π, G) ∼ = Hom(H1 (Σ), G) ∼
= Hom(π/[π, π], G) ∼ = H 1 (Σ, G)
(or H 1 (π, G) if you prefer group cohomology). In particular when G = R, then
Hom(π, G)/G is the real vector space
H 1 (Σ, R) ∼
= R2g
which is naturally a symplectic vector space under the cup-product pairing
H 1 (Σ, R) × H 1 (Σ, R) −→ H 2 (Σ, R) ∼
= R,
the last isomorphism corresponding to the orientation of Σ.
Similarly when G = C, the representation variety and the deformation space
Hom(π, G)/G = Hom(π, G) ←→ H 1 (π, C) ∼
= H 1 (Σ, C)
is a complex-symplectic vector space, that is, a complex vector space with a
complex-bilinear symplectic form.
The mapping class group action factors through the action on homology of Σ,
or equivalently the abelianization of π, which is the homomorphism
Mod(Σ) −→ Sp(2g, Z).

8.3.2 Multiplicative characters: G = C∗

Representations π −→ C∗ correspond to multiplicative characters, and are easily
understood using the universal covering
C −→ C∗
z −→ exp(2πiz)
with kernel Z ⊂ C. Such a representation corresponds to a flat complex line
bundle over Σ. The deformation space Hom(π, G) identifies with the quotient
H 1 (Σ, C)/H 1 (Σ, Z).
134 Higgs bundles and geometric structures on surfaces

Restricting to unit complex numbers G = U(1) ⊂ C∗ identifies Hom(π, G) with

the 2g-dimensional torus
H 1 (Σ, R)/H 1 (Σ, Z),
the quotient of a real symplectic vector space by an integer lattice, Mod(Σ) acts
on this torus by symplectomorphisms.

8.3.3 Jacobi variety of a Riemann surface

The classical Abel–Jacobi theory (compare, e.g., Farkas and Kra 1980) identifies
unitary characters π1 (X) −→ U(1) of the fundamental group of a Riemann sur-
face X with topologically trivial holomorphic line bundles over X. In particular
Hom(π, G) identifies with the Jacobi variety Jac(X).
While the basic structure of Hom(π, G) is a 2g-dimensional compact real torus
with a parallel symplectic structure, the conformal structure on X provides much
stronger structure. Namely, Jac(X) is a principally polarized abelian variety,
a projective variety with the structure of an abelian group. Indeed this extra
structure, by Torelli’s theorem, is enough to recover the Riemann surface X.
In particular the analytic/algebraic structure on Jac(X) is definitely not invari-
ant under the mapping class group Mod(Σ). However the symplectic structure
on Hom(π, G) is independent of the conformal structure X and is invariant under
Mod(Σ).
The complex structure on Jac(X) is the effect of the complex structure on the
tangent bundle T X (equivalent to the Hodge -operator). The Hodge theory of
harmonic differential forms finds unique harmonic representatives for cohomology
classes, which uniquely extend to holomorphic differential forms. Higgs bundle
theory is nonabelian Hodge theory (Simpson 1992) in that it extends this basic
technique from ordinary one-dimensional cohomology classes to flat connections.
When G = C∗ , then Hom(π, G) acquires a complex structure J coming from
the complex structure on C∗ . This depends only on the topology Σ, in fact
just its fundamental group π. Cup product provides a holomorphic symplectic
structure Ω on this complex manifold, giving the moduli space the structure of
a complex-symplectic manifold.
As for the U(1)-case above, Hodge theory on the Riemann surface X deter-
mines another complex structure by I; then these two complex structures anti-
commute:
IJ + JI = 0,
generating a quaternionic action on the tangent bundle with K := IJ. The
symplectic structure arising from cup product is not holomorphic with respect to
I; instead it is Hermitian (of Hodge type (1, 1)) with respect to I, extending the
Kähler structure on Jac(X). Indeed with the structure I, Hom(π, C∗ ) identifies
with the cotangent bundle T ∗ Jac(X) with Kähler metric defined by
g(X, Y ) := Ω(X, IY ).
 Stable vector bundles and Higgs bundles 135

The triple (Ω, I, J) defines a hyper-Kähler structure refining the complex-

symplectic structure. If one thinks of a complex-symplectic structure as a G-
structure where G = Sp(2g, C), then a hyper-Kähler refinement is a reduction
of the structure group to the maximal compact Sp(2g, C) ⊃ Sp(2g). The more
common definition of a hyper-Kähler structure involves the Riemannian met-
ric g which is Kählerian with respect to all three complex structures I, J, K;
alternatively it is characterized as a Riemannian manifold of dimension 4g with
holonomy reduced to Sp(2g) ⊂ SO(4g).
For a detailed exposition of the theory of rank 1 Higgs bundles on Riemann
surfaces, compare Goldman and Xia (2008).

8.4 Stable vector bundles and Higgs bundles

Narasimhan and Seshadri (1965) generalized the Abel–Jacobi theory above
to identify Hom(π, G)/G with a moduli space of holomorphic objects over a
Riemann surface X, when G = U(n). (This was later extended by Ramanathan
1975 to general compact Lie groups G.)
A notable new feature is that, unlike line bundles, not every topologically
trivial holomorphic rank n vector bundle arises from a representation π −→ U(n).
Furthermore equivalence classes of all holomorphic Cn bundles does not form an
algebraic set.
Narasimhan and Seshadri define a degree 0 holomorphic Cn -bundle V over X
to be stable (respectively, semistable) if and only if every holomorphic vector
subbundle of V has negative (respectively, nonpositive) degree. Then a holomor-
phic vector bundle arising from a unitary representation ρ is semistable, and
the bundle is stable if and only if the representation is irreducible. Furthermore
every such semistable bundle arises from a unitary representation. Narasimhan
and Seshadri show the moduli space Mn,0 (X) of semistable bundles of degree 0
and rank n over X is naturally a projective variety, thus defining such a structure
on Hom(π, G)/G. The Kähler structure depends heavily on the Riemann surface
X, although the symplectic structure depends only on the topology Σ.
It is useful to extend the notions of stability to bundles which may not have
degree 0. In particular we would like stability to be preserved by tensor product
with holomorphic line bundles. Define a holomorphic vector bundle V to be stable
if every holomorphic subbundle W ⊂ V satisfies the inequality
deg(W ) deg(V )
< .
rank(W ) rank(V )
Semistability is defined similarly by replacing the strict inequality by a weak
inequality.
In trying to extend this correspondence to the complexification G =
GL(n, C) of U(n), one might consider the cotangent bundle T ∗ Mn,0 (X) of
the Narasimhan–Seshadri moduli space, and relate it to representations π →
GL(n, C). In particular since cotangent bundles of Kähler manifolds tend to be
136 Higgs bundles and geometric structures on surfaces

M
hyper-Kähler, relating Hom(π, G)/G to Tn,0 (X) might lead to a hyper-Kähler
geometry on Hom(π, G)/G.
Thus a neighborhood of the U(n) representations in the space of GL(n, C)
corresponds to a neighborhood of the zero-section of T ∗ Mn,0 (X). In turn,
elements in this neighborhood identify with pairs (V, Φ) where V is a semistable
holomorphic vector bundle and Φ is a tangent covector to V in the space of
holomorphic vector bundles. Such a tangent covector is with a Higgs field, by
definition, an End(V )-valued holomorphic one-form on X.
Although one can define a hyper-Kähler structure on the moduli space of
such pairs, the hyper-Kähler metric is incomplete and not all irreducible linear
representations arise. To rectify this problem, one must consider Higgs fields on
possibly unstable vector bundles.
Following Hitchin (1987) and Simpson (1988), define a Higgs pair to be a pair
(V, Φ) where V is a (not ncessarily semistable) holomorphic vector bundle and
the Higgs field Φ a End(V )-valued holomorphic one-form. Define (V, Φ) to be
stable if and only if for all Φ-invariant holomorphic subbundles W ⊂ V ,
deg(W ) deg(V )
< .
rank(W ) rank(V )
l
The Higgs bundle (V, Φ) is polystable if and only if (V, Φ) = i=1 (Vi , Φi ) where
each summand (Vi , Φi ) is stable and
deg(Vi ) deg(V )
=
rank(Vi ) rank(V )
for i = 1, . . . , l.
The following basic result follows from Hitchin (1987) and Simpson (1988),
with a key ingredient (the harmonic metric) supplied by Corlette (1988) and
Donaldson (1987):
Theorem 8.1 The following natural bijections exist between equivalences
classes:
 )  )
Stable Higgs pairs Irreducible representations
←→ ρ
(V, Φ) over Σ π1 (Σ) −
→ GL(n, C)

 )  )
Polystable Higgs pairs Reductive representations
←→ ρ
(V, Φ) over Σ π1 (Σ) −
→ GL(n, C)
When the Higgs field Φ = 0, this is just the Narasimhan–Seshadri theorem,
identifying stable holomorphic vector bundles with irreducible U(n) represen-
tations. Allowing the Higgs field Φ to be nonzero, even when V is unstable, leads
to a rich new class of examples, which can now be treated using the techniques
of geometric invariant theory.
 Hyperbolic geometry 137

8.5 Hyperbolic geometry: G = PSL(2, R)

Another important class of surface group representations are Fuchsian repre-
sentations, which arise by endowing Σ with the local geometry of hyperbolic
space H2 . Here G is the group of orientation-preserving isometries Isom+ (H2 ),
which, using Poincaré’s upper half-space model, identifies with PSL(2, R). Fuch-
sian representations are characterized in many different equivalent ways; in
ρ
particular a representation π −→ G = PSL(2, R) is Fuchsian if and only if it is
a discrete embedding, that is, ρ embeds π isomorphicly onto a discrete subgroup
of G.

8.5.1 Geometric structures

Let H2 be the hyperbolic plane with a fixed orientation and G ∼ = Isom+ (H2 ) ∼
=
PSL(2, R) its group of orientation-preserving isometries. A hyperbolic structure
on a topological surface Σ is defined by a coordinate atlas {(Uα , ψα )}α∈A where
r The collection {Uα }α∈A of coordinate patches covers Σ (for some index
set A).
r Each coordinate chart ψα is an orientation-preserving homeomorphism of
the coordinate patch Uα onto an open subset ψα (Uα ) ⊂ H2 .
r For each connected component C ⊂ Uα ∩ Uβ , there is (necessarily unique)
gC,α,β ∈ G such that

ψα |C = gC,α,β ◦ ψβ |C .

The resulting local hyperbolic geometry defined on the patches by the coordinate
charts is independent of the charts, and extends to a global structure on Σ. The
surface Σ with this refined structure of local hyperbolic geometry will be called
a hyperbolic surface and denoted by M . Such a structure is equivalent to a
Riemannian metric of constant curvature −1. The equivalence follows from two
basic facts:
r Any two Riemannian manifolds of curvature −1 are locally isometric.
r A local isometry from a connected subdomain of H2 extends globally to an
isometry of H2 .
φ
Suppose M1 , M2 are two hyperbolic surfaces. Define a morphism M1 − → M2 as
a map φ, which, in the preferred local coordinates of M1 and M2 , is defined
by isometries in G. Necessarily a morphism is a local isometry of Riemannian
f
manifolds. Furthermore, if M is a hyperbolic surface and Σ − → M is a local
homeomorphism, there exists a hyperbolic structure on Σ for which f is a mor-
phism. In particular every covering space of a hyperbolic surface is a hyperbolic
surface.
In more traditional terms, a morphism of hyperbolic surfaces is just a local
isometry.
138 Higgs bundles and geometric structures on surfaces

8.5.2 Relation to the fundamental group

While the definitions involving coordinate atlases or Riemannian metrics have
certain advantages, another point of view underscores the role of the fundamental
group.
Let M be a hyperbolic surface. Choose a universal covering space M̃ → M and
give M̃ the unique hyperbolic structure for which M̃ → M is a local isometry.
dev
Then there exists a developing map M̃ −−−M → H2 , a local isometry, which induces
the hyperbolic structure on M̃ from that of H2 . The group π1 (M ) of deck
transformations of M̃ → M acts on H2 by isometries and dev is equivariant
respecting this action: for all γ ∈ π1 (M ), the figure
dev
M̃ −−−−
M
→ H2
⏐ ⏐
⏐
γ+
⏐ρ(γ)
+
M̃ −−−−→ H2 .
devM

commutes. The correspondence γ −→ ρ(γ) is a homomorphism,

hol
π1 (M ) −−−
→ Isom+ (H2 ),
M

the holonomy representation of the hyperbolic surface M . The pair (devM , holM )
is unique up to the G-action defined by
g
(devM , holM ) −→ (g ◦ devM , Inn(g) ◦ holM )
for g ∈ Isom+ (H2 ).
If the hyperbolic structure is complete, that is, the Riemannian metric is
geodesically complete, then the developing map is a global isometry M̃ ≈ H2 .
In that case the π-action on H2 defined by the holonomy representation ρ is
equivalent to the action by deck transformations. Thus ρ defines a proper free
action of π on H2 by isometries. Conversely if ρ defines a proper free isometric
π-action, then the quotient
M := H2 /ρ(π)
is a complete hyperbolic manifold with a preferred isomorphism
ρ
π1 (Σ) −
→ ρ(π) ⊂ G.
This isomorphism (called a marking) determines a preferred homotopy class of
homotopy equivalences
Σ −→ M.

8.5.3 Examples of hyperbolic structures

We now give three examples of surface group representations in PSL(2, R). The
first example is Fuchsian and corresponds to a hyperbolic structure on a surface
of genus 2. The second example is not Fuchsian, but corresponds to a hyperbolic
 Hyperbolic geometry 139

structure with a single branch point, that is, a point with local coordinate
given by a branched conformal mapping z −→ z k where k  1. (The nonsingular
case corresponds to k = 1.) In our example k = 2 and the singular point has a
neighborhood isometric to a hyperbolic cone of cone angle 4π.

8.5.3.1 A Fuchsian example

Here is a simple example of a hyperbolic surface of genus 2. Figure 8.1 depicts
a topological construction for a genus 2 surface Σ. Realizing this topological
construction in hyperbolic geometry gives Σ a local hyperbolic geometry as
follows. Take a regular octagon P with angles π/4. Label the sides as

A− − + + − − + +
1 , B1 , A1 , B1 , A2 , B2 , A2 , B2 ,

ai pairs Bi− to Bi+ and bi pairs A− +

i to Ai , respectively.
Pair the sides by

a1 , b1 , a2 , b2 ∈ PSL(2, R)

according to the pattern described in Figure 8.1. Given any two oriented geo-
desic segments in H2 of equal length, a unique orientation-preserving isometry

Figure 8.3 A regular octagon with vertex angles π/4 can be realized in the
tiling of H2 by triangles with angles π/2, π/4, and π/8. The identifications
depicted in Figure 8.1 are realized by orientation-preserving isometries. The eight
angles of π/4 fit together to form a cone of angle 2π, forming a coordinate chart
for a hyperbolic structure around that point.
140 Higgs bundles and geometric structures on surfaces

maps one to the other. Since the polygon is regular, one can realize all four
identifications in Isom+ (H2 ).
The quotient (compare Figure 8.2) contains three types of points:
r A point in the open 2-cell has a coordinate chart which is the embedding
P → H2 .
r A point on the interior of an edge has a half-disc neighborhood, which
together with the half-disc neighborhood of its part, gives a coordinate chart
for the corresponding point in the quotient.
r Around the single vertex in the quotient is a cone of angle

8(π/4) = 2π,

a disc in the hyperbolic plane.

The resulting identification space is a hyperbolic surface of genus g = 2. The

above isometries satisfying the defining relation for π1 (Σ):

a1 b1 a−1 −1 −1 −1
1 b1 a2 b2 a2 b2 = 1

Figure 8.4 A regular right-angled octagon can also be realized in the tiling of
H2 by triangles with angles π/2, π/4, and π/8. The identifications depicted in
Figure 8.1 are realized by orientation-preserving isometries. The eight angles of
π/2 fit together to form a cone of angle 4π, forming a coordinate chart for a
singular hyperbolic structure, branched at one point.
 Hyperbolic geometry 141

and define a Fuchsian representation

ρ
π1 (Σ) −
→ PSL(2, R).

Compare Figure 8.3.

8.5.3.2 A branched hyperbolic structure
We can modify the preceding example to include a singular structure, again on
a surface of genus 2. Take a regular right-angled octagon. Again, labeling the
sides as before, side pairings a1 , b1 , a2 , b2 exist. Now eight right angles compose
a neighborhood of the vertex in the quotient space. The quotient space is a
hyperbolic structure with one singularity of cone angle 4π = 8(π/2). Since the
product of the identification mappings
a1 b1 a−1 −1 −1 −1
1 b1 a2 b2 a2 b2

is rotation through 4π (the identity), the holonomy representation ρ̂ of the

nonsingular hyperbolic surface Σ \ {p} extends:
 
π1 S \ {p}
MMM ρ̂
MMM
 MMM
 M&
π1 (Σ) _ _ _ _/ PSL(2, R)
ρ

Although ρ(π) is discrete, ρ is not injective. Compare Figure 8.4.

Figure 8.5 A degree 1 map from a genus 3 surface to a genus 2 surface which
collapses a handle. Such a map is not homotopic to a smooth map with branch
point singularities (such as a holomorphic map).
142 Higgs bundles and geometric structures on surfaces

8.5.3.3 A representation with no branched structures

Consider a degree 1 map f from a genus 3 surface Σ to a genus 2 hyperbolic
µ
surface M , depicted in Figure 8.5. Let π1 (M ) −
→ G denote the holonomy repre-
sentation of M and consider the composition
f∗ µ
π = π1 (Σ) −→ π1 (M ) → G.
Then a branched hyperbolic structure with holonomy µ ◦ f∗ corresponds to a
mapping with branch singularities
F
Σ−
→ H2 /Image(µ ◦ f∗ ) = M,
inducing the homomorphism
f∗
π = π1 (Σ) −→ π1 (M ).
In particular F  f . However, since deg(f ) = 1, any mapping with only branch
point singularities of degree 1 must be a homeomorphism, a contradiction.

8.6 Moduli of hyperbolic structures and representations

To understand “different” geometric structures on the “same” surface, one
introduces markings. Fix a topological type Σ and let the geometry M vary.
The fundamental group π = π1 (Σ) is also fixed, and each marked structure
determines a well-defined equivalence class in Hom(π, G)/G. Changing the mark-
ing corressponds to the action of the mapping class group Mod(Σ) = Out(π)
on Hom(π, G)/G. Unmarked structures correspond to the orbits of the Mod(Σ)
action.

8.6.1 Deformation spaces of geometric structures

A marked hyperbolic structure on Σ is defined as a pair (M, f ) where M is
a hyperbolic surface and f is a homotopy equivalence Σ → M . Two marked
hyperbolic structures
f f
Σ−
→ M, Σ −→ M 
φ
→ M  such that
are equivalent if and only if there exists an isometry M −

M
}> 
f } 
}}
}}} 
φ
}
Σ / M
f

homotopy commutes, that is, φ ◦ f  f  . The Fricke space F(Σ) of Σ is the

space of all such equivalence classes of marked hyperbolic structures on Σ (Bers
and Gardiner 1986). The Fricke space is diffeomorphic to R6g−6 . The theory of
 Moduli of hyperbolic structures and representations 143

moduli of hyperbolic structures on surfaces goes back at least to Fricke and Klein
(1897/1912).
The Teichmüller space T(Σ) of Σ is defined similarly, as the space of equiva-
lence classes of marked conformal structures on Σ, that is, pairs (X, f ) where X
f
is a Riemann surface and Σ − → X is a homotopy equivalence. Teichmüller used
quasiconformal mappings to parametrize T(Σ) by elements of a vector space,
define a metric on T(Σ), and prove analytically that T(Σ) is a cell. Using these
ideas, Ahlfors (1960) proved T(Σ) is naturally a complex manifold.
Since a hyperbolic structure is a Riemannian metric, every hyperbolic struc-
ture has an underlying conformal structure. The uniformization theorem asserts
that if χ(Σ) < 0, then every conformal structure on Σ underlies a unique
hyperbolic structure. The resulting identification of conformal and hyperbolic
structures identifies T(Σ) with F(Σ). As discussed below, F(Σ) identifies with
an open subset of Hom(π, PSL(2, R))/PSL(2, R) which has no apparent complex
structure. Thus the complex structure on T(Σ) is more mysterious when T(Σ)
is viewed as a space of hyperbolic structures. For a readable survey of classical
Teichmüller theory, see Bers (1972), Abikoff (1980), Nag (1988), Imayoshi and
Taniguchi (1992), Gardiner-Lakic (2000) or Hubbard (2006).

8.6.2 Fuchsian components of Hom(π, G)/G

To every equivalence class of marked hyperbolic structures is associated a well-
defined element
[ρ] ∈ Hom(π, G)/G.
ρ
A representation π −→ G is Fuchsian if and only if it arises as the holonomy of
a hyperbolic structure on Σ. Equivalently, it satisfies the three conditions:
r ρ is injective.
r Its image ρ(π) is a discrete subgroup of G.
r The quotient G/ρ(π) is compact.
The first condition asserts that ρ is an embedding, and the second two conditions
assert that ρ(π) is a cocompact lattice. Under our assumption ∂Σ = ∅, the third
condition (compactness of G/ρ(π)) follows from the first two. In general, we say
that ρ is a discrete embedding (or discrete and faithful) if ρ is an embedding with
discrete image (the first two conditions).
Theorem 8.2 Let G = Isom(H2 ) = PGL(2, R) and Σ a closed connected sur-
face with χ(Σ) < 0. Fricke space, the subset of Hom(π, G)/G consisting of G-
equivalence classes of Fuchsian representations, is a connected component of
Hom(π, G)/G.
This result follows from three facts:
r Openness of Fricke space (Weil 1962)
r Closedness of Fricke space (Chuckrow 1968)
r Connectedness of Fricke space
144 Higgs bundles and geometric structures on surfaces

Chuckrow’s theorem is a special case of a consequence of the Kazhdan–

Margulis uniform discreteness (compare Raghunathan 1972 and Goldman and
Millson 1987). These ideas go back to Bieberbach and Zassenhaus in connection
with the classification of Euclidan crystallographic groups. Uniform discretness
applies under very general hypotheses, to show that discrete embeddings form a
closed subset of the representation variety. For the proof of connectedness, see,
for example, Jost (2006, section 4.3), Buser (2002), Thurston (1979) or Ratcliffe
(2006) for elementary proofs using Fenchel–Nielsen coordinates. Connectedness
also follows from the uniformization theorem, together with the identification of
Teichmüller space T(Σ) as a cell.
When G = Isom+ (H2 ) = PSL(2, R), the situation slightly complicates, due to
the choice of orientation. Assume Σ is orientable, and orient it. Orient H2
f
as well. Let Σ − → M be a marked hyperbolic structure on Σ. The orientation
of M induces an orientation of M̃ which is invariant under π1 (M ). However,
the developing map devM may or may not preserve the (arbitrarily chosen)
orientations of M̃ and H2 . Accordingly Isom+ (H2 )-equivalence classes of Fuchsian
representations in G fall into two classes, which we call orientation-preserving
and orientation-reversing, respectively. These two classes are interchanged by
inner automorphisms of orientation-reversing isometries of H2 .
Theorem 8.3 Let G = Isom+ (H2 ) = PSL(2, R) and Σ a closed connected ori-
ented surface with χ(Σ) < 0. The set of G-equivalence classes of Fuchsian rep-
resentations forms two connected components of Hom(π, G)/G. One component
corresponds to orientation-preserving Fuchsian representations and the other to
orientation-reversing Fuchsian representations.

8.6.3 Characteristic classes and maximal representations

Characteristic classes of flat bundles determine invariants of representations. In
the simplest cases (when G is compact or reductive complex), these determine
the connected components of Hom(π, G).
8.6.3.1 Euler class and components of Hom(π, PSL(2, R))
The components of Hom(π, G) were determined in Goldman (1988) using an
invariant derived from the Euler class of the oriented H2 -bundle

H2 / (H2 )ρ


Σ
ρ
associated to a representation π −
→ PSL(2, R) as follows. The total space is the
quotient
(H2 )ρ := (Σ̃ × H2 )/π
 Moduli of hyperbolic structures and representations 145

where π acts diagonally on Σ̃ × H2 by deck transformations on Σ̃ and via ρ on H2 .

Isomorphism classes of oriented H2 -bundles over Σ are determined by the Euler
class, which lives in H 2 (Σ, Z). The orientation of Σ identifies this cohomology
group with Z. The resulting map

Hom(π, G) −−−→ H 2 (S; Z) ∼

Euler
=Z

satisfies

|Euler(ρ)|  |χ(S)| = 2g − 2 (8.3)

(Milnor 1958 and Wood 1976). Call a representation maximal if equality holds
in (8.3), that is, Euler(ρ) = ±χ(Σ).
The following converse was proved in Goldman (1980) (compare also Hitchin
1987 and Goldman 1988).

Theorem 8.4 ρ is maximal if and only if ρ is Fuchsian.

Suppose M is a branched hyperbolic surface with branch points p1 , . . . , pl where

pi is branched of order ki , where each ki is a positive integer. In other words, each
pi has a neighborhood which is a hyperbolic cone of cone angle 2πki . Consider
a marking Σ → M , determining a holonomy representation ρ. Then


l
Euler(ρ) = χ(Σ) + ki .
i=1

Consider the two examples for genus 2 surfaces:

r The first (Fuchsian) example (Section 8.5.3.1) arising from a regular octagon
with π/4 angles has Euler class −2 = χ(σ).
r In the second example (Section 8.5.3.2), the structure is branched at one
point, so that l = k1 = 1 and the Euler class equals −1 = χ(Σ) + 1.

8.6.4 Quasi-Fuchsian representations: G = PSL(2, C)

When the representation

π −→ PSL(2, R) → PSL(2, C)

is deformed inside PSL(2, C), the action on CP1 is topologically conjugated to

the original Fuchsian action (Figure 8.6). Furthermore there exists a Hölder ρ-
equivariant embedding S 1 → CP1 , whose image Λ has Hausdorff dimension >1,
unless the deformation is still Fuchsian. The space of such representations is
the quasi-Fuchsian space QF(Σ). By Bers (1960), QF(Σ) naturally identifies
with

T(Σ) × T(Σ) ≈ R12g−12 .

146 Higgs bundles and geometric structures on surfaces

Figure 8.6 A quasi-Fuchsian subgroup of PSL(2, C) obtained by deforming the

genus 2 surface with a fundamental domain, the regular octagon with π/4 angles
in CP1 . The limit set is a nonrectifiable Jordan curve, but the new action of
π1 (Σ) is topologically conjugated to the original Fuchsian action.

Bers’s correspondence is the following. The action of ρ on the complement CP1 \

Λ is properly discontinuous, and the quotient
 
CP1 \ Λ /ρ(π)

consists of two Riemann surfaces, each with a canonical marking determined

by ρ. Furthermore these surfaces possess opposite orientations, so the pair of
marked conformal structures determine a point in T(Σ) × T(Σ). Bonahon (1986)
and Thurston (1979) proved that the closure of QF(Σ) in Hom(π, G)/G equals
the set of equivalence classes of discrete embeddings. The frontier ∂QF(Σ) ⊂
Hom(π, G)/G is nonrectifiable, and is near non-discrete representations.
However, the two connected components of Hom(π, G)/G are distinguished
by the characteristic class (related to the second Stiefel–Whitney class w2 )
which detects whether a representation in PSL(2, C) lifts to the double covering
SL(2, C) → PSL(2, C) (Goldman 1988). Contrast this situation with PSL(2, R)
where the discrete embeddings form connected components, characterized by
maximality.

8.6.4.1 Higher rank Hermitian spaces: the Toledo invariant

Vladimir Turaev (1984, 1985) and Domingo Toledo (1989) generalized the Euler
class of flat PSL(2, R)-bundles to flat G-bundles, where G is the autormorphism
group of a Hermitian symmetric space X of noncompact type.
 Moduli of hyperbolic structures and representations 147

ρ
Let π −
→ G be a representation and let
X / (X)ρ


Σ
be the corresponding flat (G, X)-bundle over Σ. Then the G-invariant Kähler
form ω on X defines a closed exterior two-form ωρ on the total space (X)ρ . Let
s
Σ− → (X)ρ be a smooth section. Then the integral

s ∗ ωρ
Σ

is independent of s, depends continuously on ρ, and, after suitable normalization,

assumes integer values. The resulting Turaev-Toledo invariant
τ
Hom(π, G) −
→Z
satisfies
|τ (ρ)|  (2g − 2)rankR (G)
(Domic and Toledo 1987, and Clerc and Ørsted 2003). Define ρ to be maximal
if and only if
|τ (ρ)| = (2g − 2)rankR (G).
ρ
Theorem 8.5 (Toledo 1989) π − → U(n, 1) is maximal if and only if ρ is a
discrete embedding preserving a complex geodesic, that is, ρ is conjugate to a
representation with
ρ(π) ⊂ U(1, 1) × U(n − 1).
This rigidity has a curious consequence for the local geometry of the deforma-
tion space. Let G := U(n, 1) and
G0 = U(1, 1) × U(n − 1) ⊂ G.
Then, in an appropriate sense,
 
dim Hom(π, G)/G = 2g + (2g − 2) (n + 1)2 − 1 = (2g − 2)(n + 1)2 + 2,
but Toledo’s rigidity result implies that the component of maximal representa-
tions has strictly lower dimension:
dim Hom(π, G0 )/G0 = 4g + (2g − 2)3 + (2g − 2)((n − 1)2 − 1)
with codimension
8(n − 1)(g − 1) − 2.
Compare Goldman and Millson (1988).
148 Higgs bundles and geometric structures on surfaces

8.6.5 Teichmüller space: marked conformal structures

The Teichmüller space T(Σ) of Σ is the deformation space of marked conformal
structures on Σ.
A marked conformal structure on Σ is a pair (X, f ) where X is a Riemann
surface and f is a homotopy equivalence Σ → X. Marked conformal structures
f f
Σ−
→ X, Σ −→ X 
φ
→ X  such that
are equivalent if and only if there exists a biholomorphism X −

X
f }}> 
}}  φ
}}} 
} 
Σ / X
f

homotopy commutes.
Theorem 8.6 (Uniformization) Let X be a Riemann surface with χ(X) < 0.
Then there exists a unique hyperbolic metric whose underlying conformal struc-
ture agrees with X.
Since every hyperbolic structure possesses an underlying conformal structure,
Fricke space F(Σ) maps to Teichmüller space T(Σ). By the uniformization
theorem, F(Σ) → T(Σ) is an isomorphism. It is both common and tempting
to confuse these two deformation spaces. In the present context, however, it
seems best to distinguish between the representation/hyperbolic structure and
the conformal structure.
For example, each Fuchsian representation determines a marked hyperbolic
structure, and hence an underlying marked conformal structure. An equivalence
class of Fuchsian representations thus determines a special point in Teichmüller
space. This contrasts sharply with other representations which do not generally
pick out a preferred point in T(Σ). This preferred point can be characterized as
the unique minimum of an energy function on Teichmüller space.
The construction, due to Tromba (1992), is as follows. Given a hyperbolic
f
surface M and a homotopy equivalence X − → M , then by Eels and Sampson
F
(1964) a unique harmonic map X − → M exists homotopic to f . The harmonic
map is conformal if and only if M is the uniformization of X. In general the
nonconformality is detected by the Hopf differential Hopf(F ) ∈ H 0 (X, KX 2
),
defined as the (2, 0) part of the pull-back by F of the complexified Riemannian
metric on M . The resulting mapping
F(X) −→ H 0 (X, KX
2
)
(f, M ) −→ Hopf(F )
is a diffeomorphism.
 Moduli of hyperbolic structures and representations 149

Fixing M and letting the marked complex structure (f, X) vary over T(Σ)
yields an interesting invariant discussed in Tromba (1992), and extended in
Goldman and Wentworth (2005) and Labourie (to appear). The energy of the
harmonic map F = F (f, X, M ) is a real-valued function on T(Σ). In the present
context it is the square of the L2 -norm of Hopf(F ).

Theorem 8.7 (Tromba 1992) The resulting function T(Σ) → R is proper,

convex, and possesses a unique minimum at the uniformaization structure X.

For more applications of this energy function to surface group representations,

compare Goldman and Wentworth (2007), where properness is proved for convex
cocompact discrete embeddings, and Labourie (2008), where the above result is
extended to quasi-isometric embeddings π → G.

8.6.6 Holomorphic vector bundles and uniformization

ρ
Let π −
→ PSL(2, R) be a Fuchsian representation corresponding to a marked
f
hyperbolic structure Σ −
→ M . A spin structure on Σ determines a lifting of
ρ to
ρ̃
π−
→ SL(2, R) ⊂ SL(2, C)

and hence a flat C2 -bundle (C2 )ρ over Σ.

Choose a marked Riemann surface X corresponding to a point in Teichmüller
space T(Σ). Since locally constant maps are holomorphic for any complex
structure on Σ, the flat bundle (C2 )ρ has a natural holomorphic structure; denote
the corresponding holomorphic rank 2 vector bundle over X by Eρ → X.
In trying to fit such a structure into a moduli problem over X, the first problem
is that this holomorphic vector bundle is unstable and does not seem susceptible
to geometric invariant theory techniques. Indeed, its instability intimately relates
to its role in uniformization. Namely, the developing map
dev
M̃ −−→ CP1

determines a holomorphic line bundle L ⊂ Eρ̃ . Since deg(Eρ̃ ) = 0, and dev is

nonsingular, the well-known isomorphism
 
T CP1 ∼ = Hom(γ, γ −1 )

where γ → CP1 is the tautological line bundle implies that

L2 ∼
= KX

and deg(L) = g − 1 > 0. Therefore Eρ̃ is unstable. In fact, Eρ̃ is a non-trivial

extension

L −→ Eρ̃ −→ Eρ̃ /L ∼
= L−1
150 Higgs bundles and geometric structures on surfaces

determined by the fundamental cohomology class ε in

H 1 (X, Hom(L−1 , L) ∼
= H 1 (X, K) ∼
=C
defining Serre duality (compare Gunning (1967)).
One resolves this difficulty by changing the question. Replace the extension
class ε by an auxiliary holomorphic object, a Higgs field

Φ ∈ H 0 (X; KX ⊗ End(E)),

for the vector bundle E := L ⊕ L−1 so that the Higgs pair (E, Φ) is stable in the
appropriate sense. In our setting the Higgs field corresponds to the everywhere
nonzero holomorphic section of the trivial holomorphic line bundle

C∼
= KX ⊗ Hom(L, L−1 ) ⊂ KX ⊗ End(E).
Now the only Φ-invariant holomorphic subbundle of E is L−1 which is negative,
and the pair (E, Φ) is stable.

8.7 Rank 2 Higgs bundles

Now we impose a conformal structure on the surface to obtain extra structure
on the deformation space Hom(π, G)/G. As before Σ denotes a fixed oriented
smooth surface, and X a Riemann surface with a fixed marking Σ → X.

8.7.1 Harmonic metrics

Going from ρ to (V, Φ) involves finding a harmonic metric, which may be
regarded as a ρ-equivariant harmonic map
,
−
M
h
→ GL(n, C)/U(n)

into the symmetric space GL(n, C)/U(n). The metric h determines a reduction of
structure group of Eρ̃ from GL(n, C) to U(n), giving Eρ̃ a Hermitian structure.
Let A denote the unique connection on Eρ̃ which is unitary with respect to h.
The harmonic metric determines the Higgs pair (V, ∂¯V , Φ) as follows:
r The Higgs field Φ is the holomorphic (1, 0)-form ∂h ∈ Ω1 End(V ), where the
tangent space to GL(n, C)/U(n) is identified with a subspace of h∗ End(V ).
r The holomorphic structure d on V arises from conformal structure Σ and
A
the Hermitian connection A.

The Higgs pair satisfies the self-duality equations with respect to the Hermitian
metric h:

dA (Φ) = 0
F (A) + [Φ, Φ∗ ] = 0 (8.4)
 Rank 2 Higgs bundles 151

Here F (A) denotes the curvature of A, and Φ∗ denotes the adjoint of Φ with
respect to h. Conversely, Hitchin and Simpson show that every stable Higgs pair
determines a Hermitian metric satisfying (8.4).

8.7.2 Higgs pairs and branched hyperbolic structures

Choose an integer d satisfying
0  d < 2g − 2.
Hitchin identifies the component Euler−1 (2 − 2g + d) with Higgs pairs (V, Φ)
where
V = L1 ⊕ L2
is a direct sum of line bundles L1 and L2 defined as follows. Choose a square
1/2 −1/2
root KX of the canonical bundle KX and let KX be its inverse. Let D  0
be an effective divisor of degree d. Define line bundles
−1/2
L1 := KX ⊗D
1/2
L2 := KX .
Define a Higgs field
0 sD
Φ=
Q 0
where
r sD is a holomorphic section of the line bundle corresponding to D, which
determines the component of Φ in
 
KX ⊗ Hom(L2 , L1 ) ∼
= D ⊂ Ω1 Σ, End(V ) .
r Q ∈ H 0 (Σ, K 2 ) is a holomorphic quadratic differential with div(Q)  D,
X
which determines the component of Φ in
 
KX ⊗ Hom(L1 , L2 ) ∼ 2
= KX ⊂ Ω1 Σ, End(V ) .
Then (V, Φ) is a stable Higgs pair.
When Q = 0, this Higgs bundle corresponds to the uniformization representa-
tion. In general, when d = 0, the harmonic metric is a diffeomorphism (Schoen
and Yau 1978) Q is its Hopf differential.
The Euler class of the corresponding representation equals
deg(L2 ) − deg(L1 ) = 2 − 2g + d.

Theorem 8.8 (Hitchin 1987) The component Euler−1 (2 − 2g + d) identifies

with a holomorphic vector bundle over the symmetric power Symd (X). The fiber
152 Higgs bundles and geometric structures on surfaces

over D ∈ Symd (X) is the vector space

{Q ∈ H 0 (X, KX
2
) | div(Q)  D} ∼
= C3(g−1)−d .

The quadratic differential Q corresponds to the Hopf differential of the harmonic

metric h. When Q = 0, the harmonic metric is holomorphic, and defines a
developing map for a branched conformal structure, with branching defined by D.
When e = 2 − 2g, then d = 0 and the space
 F(X) of Fuchsian representations
2 ∼ 3(g−1)
identifies with the vector space H 0 X, KX =C .

8.7.3 Uniformization with singularities

McOwen 1993 and Troyanov 1991 proved a general uniformization theorem
for hyperbolic structures with conical singularities. Specifically, let D = (p1 ) +
· · · + (pk ) be an effective divisor, with pi ∈ X. Choose real numbers θi > 0 and
introduce singularities in the conformal structure on X by replacing a coordinate
chart at pi with a chart mapping to a cone with cone angle θi . The following
uniformization theorem describes when there is a singular hyperbolic metric in
this singular conformal structure.

Theorem 8.9 (McOwen 1993, Troyanov 1991) If


k
2 − 2g + (θi − 2π) > 0,
i=1

there exists a unique singular hyperbolic surface conformal to X with cone angle
θi at pi .

When the θi are multiples of 2π, then this structure is a branched structure
(and the above theorem follows from Hitchin (1987)). The moduli space of such
branched conformal structures forms a bundle Sd over T(Σ) where the fiber over
a marked Riemann surface Σ → X is the symmetric power Symd (X) where

1 
k
d= (θi − 2π).
2π i=1

The resulting uniformization map

 
→ Euler−1 2 − 2g + d ⊂ Hom(π, G)/G
U
Sd −
is homotopy equivalence, which is not surjective, by the example in Section
8.5.3.3.

Conjecture 8.1 Every representation with non-discrete image lines in the

image of U
 Split R-forms and Hitchin’s Teichmüller component 153

8.8 Split R-forms and Hitchin’s Teichmüller component

When G is a split real form of a semisimple Lie group, Hitchin (1992) used
Higgs bundle techniques to determine an interesting connected component of
Hom(π, G)/G, which is not detected by characteristic classes. A Hitchin compo-
nent of Hom(π, G) is the connected component containing a composition

ρ0 K
π → SL(2, R) −
→G

where ρ0 is Fuchsian and K is the representation corresponding to the three-

dimensional principal subgroup discovered by Kostant (1959). When G =
SL(n, R), then Kostant’s representation K is the irreducible n-dimensional rep-
resentation corresponding to the symmetric power Symn−1 (R2 ).
The compositions K ◦ ρ0 above determine a subset of Hom(π, G)/G which
identifies with the Fricke–Teichmüller space, and Hitchin’s main result is that
each Hitchin component is a cell of (the expected) dimension (2g − 2) dim G.
For example, if G = SL(n, R), then Hitchin identifies this component with the
2(g − 1)(n2 − 1)-cell:

   
H 0 X; KX
2
⊕H 0 X; KX
3
⊕ · · · ⊕ H 0 (X; KX
n
)
∼
= C3(g−1) ⊕ C5(g−1) ⊕ · · · ⊕ C(2n−1)(g−1) .

When n is odd, Hitchin proves there are exactly three components. The second
Stiefel–Whitney characteristic class is nonzero on exactly one component; it is
zero on two components, one of which is the Hitchin–Teichmüller component.

8.8.1 Convex RP2 -structures: G = SL(3, R)

When G ∼ = PGL(3, R) ∼= SL(3, R), Hitchin (1992) conjectured that his component
corresponded to the deformation space C(Σ) of marked convex RP2 -structures,
proved in Goldman (1990) to be a cell of dimension 16(g − 1). In Choi and
Goldman (1993) proves this conjecture. A convex RP2 -manifold is a quotient Ω/Γ
where Ω ⊂ RP2 is a convex domain and Γ a discrete group of collineations acting
properly and freely on Ω (Figures 8.7 and 8.8). If χ(M ) < 0, then necessarily Ω is
properly convex (contains no complete affine line), and its boundary ∂Ω is a C 1+α
strictly convex curve, for some 0 < α  1. Furthermore α = 1 if and only if ∂Ω is
a conic and the RP2 -structure arises from a hyperbolic structure. These facts are
due to Kuiper (1954) and Benzécri (1960) and have recently been extended and
amplified to compact quotients of convex domains in RPn−1 by Benoist (2000,
2004).
154 Higgs bundles and geometric structures on surfaces

Figure 8.7 A triangle tesselation in the hyperbolic plane, drawn in the

Beltrami–Klein projective model. Its holonomy representation is obtained by
composing a Fuchsian representation in SL(2, R) with the irreducible represen-
tation SL(2, R) −→ SL(3, R).

Figure 8.8 A deformation of a Fuchsian representation preserving an exotic

convex domain. The boundary is a strictly convex C 1 curve which is not C 2 .

8.8.2 Higgs bundles and affine spheres

The Higgs bundle theory of Hitchin (1992) identifies, for an arbitrary Riemann
surface X, the Hitchin component C(Σ) with the complex vector space
   
3 ∼ 8g−8
H 0 X, KX
2
⊕ H 0 X, KX =C
 
and the component in H 0 X, KX 2
of the Higgs field corresponds to the Hopf
differential of the harmonic metric. Using the theory of hyperbolic affine spheres
developed by Calabi, Loewner–Nirenberg, Cheng–Yau, Gigena, Sasaki, Li, Wang,
and Labourie (1997, to appear) and Loftin (2001) proved
 Split R-forms and Hitchin’s Teichmüller component 155

Theorem 8.10 The deformation space C(Σ) naturally identifies with the holo-
morphic vector bundle over T(Σ) whose fiber over a marked Riemann surface
Σ → X is H 0 X, KX 3
.
For every such representation, there exists a unique conformal structure so
that
h̃
Σ̃ −
→ SL(3, R)/SO(3)
 
is a conformal map, that is, the component of the Higgs field in H 0 Σ, KX 2
—the
Hopf differential Hopf(h)—vanishes. This defines the projection C(Σ) → T(Σ).
The zero section corresponds to the Fuchsian RP2 structures, that is, the RP2
structures arising from hyperbolic structures on Σ.
It is natural to attempt to generalize this as follows. For any split real form
G, and Riemann surface X with π1 (X) ∼ = π, Hitchin (1992) identifies a certain
direct sum of holomorphic line bundles VX naturally associated to X so that a
Hitchin component of Hom(π, G)/G identifies with the complex vector space
 
H 0 X, KX 2
⊕ H 0 (X, VX ).
However, this identification depends crucially on the Riemann surface X and
fails to be Mod(Σ)-invariant. Generalizing the Labourie–Loftin Theorem 8.10,
we conjecture that each Hitchin component of Hom(π, G)/G identifies naturally
with the total space of a holomorphic vector bundle E(Σ) over T(Σ), whose fiber
over a marked Riemann surface X equals H 0 (X, VX ).

8.8.3 Hyperconvex curves

Labourie (2006) discovered an important property of the Hitchin component:
Theorem 8.11 (Labourie) A representation in the Hitchin component for
G = SL(n, R) is a discrete quasi-isometric embedding
ρ
π → SL(n, R)
with reductive image.
f
A crucial ingredient in his proof is the following notion. A curve S 1 −
→ RPn−1 is
hyperconvex if and only if for all x1 , . . . , xn ∈ S distinct,
1

f (x1 ) + · · · + f (xn ) = Rn .
Theorem 8.12 (Guichard 2005, Guichard 2008, Labourie 2006) ρ is
Hitchin if and only if ρ preserves hyperconvex curve.
Recently Fock and Goncharov (2006, 2007) have studied this component of rep-
resentations, using global coordinates generalizing Thurston and Penner’s shear-
ing coordinates. In these coordinates the Poisson structure admits a particularly
simple expression, leading to a quantization. Furthermore they find a positive
156 Higgs bundles and geometric structures on surfaces

structure which leads to an intrinsic characterization of these semi-algebraic

subsets of Hom(π, G)/G. Their work has close and suggestive connections with
cluster algebras and K-theory.

8.9 Hermitian symmetric spaces: maximal representations

We return now to the maximal representations into groups of Hermitian
type, concentrating on the unitary groups U(p, q) and the symplectic groups
Sp(n, R).

8.9.1 Unitary groups U(p, q)

The Milnor–Wood inequality (8.3) may be the first example of the boundedness
of a cohomology class. In a series of papers (Burger and Monod 2002; Burger
and Iozzi 2002, 2007; Iozzi 2002; Burger et al. 2003, preprint), Burger, Monod,
Iozzi, and Wienhard place the local and global rigidities in the context of the
Turaev-Toledo invariant being a bounded cohomology class. A consequence of
these powerful methods for surface groups is the following, announced in Burger
et al. (2003):
Theorem 8.13 (Burger et al. 2003) Let X be a Hermitian symmetric space,
and maximal representation
ρ
π−
→ G.
r The Zariski closure L of ρ(π) is reductive.
r The symmetric space associatied to L is a Hermitian symmetric tube domain,
totally geodesicly embedded in the symmetric space of G.
r ρ is a discrete embedding.
Conversely, if X is a tube domain, then there exists a maximal ρ with ρ(π)
Zariski-dense.
For example, if G = U(p, q), where p  q, then ρ is conjugate to the normalizer
U(p, p) × U(q − p) of U(p, p) in U(p, q). As in the rank 1 case (compare Sec-
tion 8.6.4.1), the components of maximal representations have strictly smaller
dimension. (In earlier work Hernández Lamoneda 1991 considered the case of
U(2, q).)
Furthermore every maximal representation deforms into the composition of a
ρ
Fuchsian representation π −→ SU(1, 1) with the diagonal embedding
p
∆
- ./ 0
SU(1, 1) ⊂ U(1, 1) → U(1, 1) × · · · × U(1, 1) ⊂ U(p, p) ⊂ U(p, q).
At roughly the same time, Bradlow, Garcı́a-Prada, and Gothen (2003) inves-
tigated the space of Higgs bundles using infinite-dimensional Morse theory, in
a similar way to Hitchin (1987). Their critical point analysis also showed that
maximal representations formed components of strictly smaller dimension. They
 Hermitian symmetric spaces: maximal representations 157

found that the number of connected components of Hom(π, U(p, q)) equals
2(p + q) min(p, q) (g − 1) + gcd(p, q).
(For a survey of these techniques and other results, compare Bradlow et al. 2006
as well as their recent column Bradlow et al. 2007.)

8.9.2 Symplectic groups Sp(2n, R)

The case G = Sp(2n, R) is particularly interesting, since G is both R-split
and of Hermitian type. Gothen (2001) showed there are 3 · 22g + 2g − 4 com-
ponents of maximal representations when n = 2. For n > 2, there are 3 · 22g
components of maximal representations (Garcı́a-Prada et al., 2008). For n = 2,
the components of the nonmaximal representations are just the preimages of the
Turaev–Toledo invariant, comprising 1 + 2(2g − 3)= 4g − 5 components.
 Thus
the total number of connected components of Hom π, Sp(4, R) equals
 
2 3 · 22g + 2g − 4 + 4g − 5 = 6 · 4g + 10g − 13.
The Hitchin representations are maximal and comprise 22g+1 of these compo-
nents. They correspond to deformations of compositions of Fuchsian representa-
ρ0
tions π −→ SL(2, R) with the irreducible representation
 
SL(2, R) −→ Aut Sym2n−1 (R2 ) → Sp(2n, R)

where R2n ∼ = Sym2n−1 (R2 ) with the symplectic structure induced from R2 .
Another class of maximal representations arises from deformations of compo-
ρ0
sitions of a Fuchsian representation π −→ SL(2, R) with the diagonal embedding
n
∆
- ./ 0
SL(2, R) → SL(2, R) × · · · × SL(2, R) → Sp(2n, R).
More generally, the diagonal embedding extends to a representation
,
∆
SL(2, R) × O(n) → Sp(2n, R)
corresponding to the SL(2, R) × O(n)-equivariant decomposition of the symplec-
tic vector space
R2n = R2 ⊗ Rn
as a tensor product of the symplectic vector space R2 and the Euclidean inner
product space Rn . Deformations of compositions of Fuchsian representations
into SL(2, R) × O(2) with ∆, provide 22g more components of maximal represen-
tations.
For n > 2, these account for all the maximal components. This situation is
more complicated when n = 2. In that case, 4g − 5 components of maximal
representations into Sp(4, R) do not contain representations into smaller compact
extensions of embedded subgroups isomorphic to SL(2, R). In particular the
158 Higgs bundles and geometric structures on surfaces

image of every representation in such a maximal component is Zariski dense

in Sp(4, R), in contrast to the situation for U(p, q) and Sp(2n, R) for n > 2 (see
Guichard and Wienhard in preparation, and also Bradlow, Garcı́a-Prada, and
Gothen 2009, for more details).

8.9.3 Geometric structures associated to Hitchin representations

Fuchsian representations into SL(2, R) correspond to hyperbolic structures on Σ,
and Hitchin representations into SL(3, R) correspond to convex RP2 structures
on Σ. What geometric structures correspond to other classes of surface group
representations?
Guichard and Wienhard (to appear) associate to a Hitchin representation in
SL(4, R), an RP3 structure on the unit tangent bundle T1 (Σ) of a rather special
type. The trajectories of the geodesic flow on T1 (Σ) (for any hyperbolic metric
on Σ) develop into projective lines. The leaves of the weak-stable foliations of
this structure develop into convex subdomains of projective planes in RP3 . The
construction of this structure uses the hyperconvex curve in RP3 . This convex-
foliated structure is a geometric structure corresponding to Hitchin representa-
tions in SL(4, R).
For the special case of Hitchin representations into Sp(4, R) (which are readily
Hitchin representations into SL(4, R)), the convex-foliated structures are char-
acterized by a duality. Furthermore the symplectic structure on R4 induces a
contact structure on T 1 (Σ) which is compatible with the convex-foliated RP3
structure. In addition, another geometric structure on another circle bundle over
Σ arises naturally, related to the local isomorphism Sp(4, R) −→ O(3, 2) and the
identfication of the Grassmannian of Lagrangian subspaces of the symplectic
vector spaceR4 with the conformal compactification of Minkowski (2 + 1)-space,
the (2 + 1)-Einstein universe (compare Borbot et al., to appear for an exposition
of this geometry). The interplay between the contact RP3 geometry, flat confor-
mal Lorentzian structures, the dynamics of geodesics on hyperbolic surfaces, and
the resuting deformation theory promises to be a fascinating extension of ideas
rooted in the work of Nigel Hitchin.

Acknowledgments
Goldman gratefully acknowledges partial support from National Science Foun-
dation grants DMS-070781, DMS-0103889, and DMS-0405605. This chapter was
presented at the Geometry Conference honoring Nigel Hitchin, Consejo Superior
de Investigaciones Cientı́ficas, Madrid, Spain, 5 September 2006. We gratefully
acknowledge support from the Oswald Veblen Fund at the Institute for Advanced
Study where this work was completed.
I am grateful to Nigel Hitchin for the inspiration of these ideas, and to him and
Simon Donaldson for the opportunity to study with them at the Maths Institute
in Oxford in 1989. Over the years my knowledge of the subject has benefitted
enormously from conversations with Yves Benoist, Steve Bradlow, Marc Burger,
 References 159

Suhyoung Choi, Kevin Corlette, Oscar Garcı́a-Prada, Peter Gothen, Olivier

Guichard, Alessandra Iozzi, Misha Kapovich, François Labourie, John Loftin,
John Millson, Ignasi Mundet i Riera, Walter Neumann, Carlos Simpson, Ser-
Peow Tan, Richard Wentworth, Anna Wienhard, Graeme Wilkin, Mike Wolf,
Scott Wolpert, and Eugene Xia. I also thank the students in the Experimental
Geometry Lab at the University of Maryland, in particular Ryan Hoban, Rachel
Kirsch, and Anton Lukyanenko for help producing the illustrations for this
chapter. I am grateful to the anonymous referee for several helpful suggestions.
Finally I wish to thank the Mathematical Sciences Research Institute, the
Institute for Advanced Study, and the Mathematics Department at Princeton
University for their hospitality in 2007–2008.

References
Abikoff, W. (1980). The real-analytic theory of Teichmüller spaces. Lecture Notes in
Mathematics Vol. 820, Springer-Verlag, Berlin, Germany.
Ahlfors, L. (1960). The complex analytic struture of the space of closed Rie-
mann surfaces. in Analytic Functions. Princeton University Press Princeton, NJ,
pp. 45–66.
Barbot, T., Charette, V., Drumm, T., Goldman, W. and Melnick, K. (2008). A primer
on the (2 + 1) Einstein universe in Recent developments in pseudo-Riemannian
geometry. ESI Lect. Math. Phys., Math. Soc., Zurich, 179–229.
Benoist, Y. (2000). Automorphismes des cônes convexes. Inv. Math. 141(1), 149–93.
Benoist, Y. Convexes divisibles. I, Algebraic groups and arithmetic, Tata Inst. Fund.
Res., Mumbai (2004) 339–74. II, Duke Math. J. 120(1) (2003), 97–120. III, Ann. Sci.
École Norm. Sup. (4) 38(5) (2005), 793–832. IV, Structure du bord en dimension 3,
Inv. Math. 164(2) (2006), 249–78.
Benzécri, J. P. (1960). Sur les variétés localement affines et projectives. Bull. Soc. Math.
France 88, 229–332.
Bers, L. (1960). Simultaneous uniformization. Bull. Am. Math. Soc. 66, 94–7.
Bers, L. (1972). Uniformization, moduli, and Kleinian groups. Bull. Lond. Math. Soc.
4, 257–300.
Bers, L. and Gardiner, F. (1986). Fricke spaces. Adv. Math. 62, 249–84.
Bonahon, F. (1986). Bouts des varit́és hyperboliques de dimension 3. Ann. Math. (2)
124(1), 71–158.
Bradlow, S., Garcı́a-Prada, O. and Gothen, P. (2003). Surface group representations
and U (p, q)-Higgs bundles. J. Diff. Geom. 64(1), 111–70.
Bradlow, S., Garcı́a-Prada, O. and Gothen, P. (2006). Maximal surface group repre-
sentations in isometry groups of classical Hermitian symmetric spaces. Geom. Ded.
122, 185–213.
Bradlow, S., Garcı́a-Prada, O. and Gothen, P. (2007). What is a Higgs bundle?. Notices
Am. Math. Soc. 54(8), 980–1.
Bradlow, S., Garcı́a-Prada, O. and Gothen, P. (2009). Deformations of maximal repre-
sentations in Sp(4, R). arXiv:0903.5496.
Burger, M. and Iozzi, A. (2002). Boundary maps in bounded cohomology. Geom. Funct.
Anal. 12, 281–92.
160 Higgs bundles and geometric structures on surfaces

Burger, M., Iozzi, A., Labourie, F. and Wienhard, A. (2005). Maximal representations
of surface groups: symplectic Anosov structures, Pure Appl. Math. Q., Special Issue:
In Memory of Armand Borel Part 2 of 3, 1(3), 555–601.
Burger, M. and Iozzi, A. (2007). Bounded differential forms, generalized Milnor–Wood
inequality and applications to rigidity theory. Geom. Ded. 125, 1–23.
Burger, M., Iozzi, A. and Wienhard, A. (2003). Surface group representations with
maximal Toledo invariant. C. R. Acad. Sci. Paris, Sér. I. 336, 387–90.
Burger, M., Iozzi, A. and Wienhard, A. Representations of surface groups with maximal
Toledo invariant. Preprint. math.DG/0605656.
Burger, M. and Monod, N. (2002). Continuous bounded cohomology and applications
to rigidity theory. Geom. Funct. Anal. 12, 219–80.
Buser, P. (1992). Geometry and spectra of compact Riemann surfaces. Progress in
Mathematics. Vol. 106, Birkhuser Boston, Inc., Boston, MA.
Choi, S. and Goldman, W. (1993). Convex real projective structures on closed surfaces
are closed. Proc. Am. Math. Soc. 118(2), 657–61.
Chuckrow, V. (1968). On Schottky groups with applications to Kleinian groups. Ann.
Math. (2) 88, 47–61.
Clerc, J. L. and Ørsted, B. (2003). The Gromov norm of the Kähler class and the
Maslov index. Asian J. Math. 7, 269–96.
Corlette, K. (1988). Flat G-bundles with canonical metrics. J. Diff. Geom. 28, 361–82.
Domic, A. and Toledo, D. (1987). The Gromov norm of the Kaehler class of symmetric
domains. Math. Ann. 276(3), 425–32.
Donaldson, S. (1987). Twisted harmonic maps and the self-duality equations. Proc.
Lond. Math. Soc. (3) 55(1), 127–31. MR MR 887285 (88g:58040).
Eells, J. and Sampson, J. (1964). Harmonic mappings of Riemannian manifolds. Am.
J. Math. 85, 109–60.
Farkas, H. and Kra, I. (1980). Riemann surfaces. Graduate Texts in Mathematics. Vol.
71, Springer-Verlag, Berlin/Heidelberg/New York.
Fock, V. and Goncharov, A. (2006). Moduli spaces of local systems and
higher Teichmüller theory. Publ. Math. Inst. Hautes Études Sci. 103, 1–211.
math.AG/0311149, 2003.
Fock, V. and Goncharov, A. (2007). Moduli spaces of convex projective structures on
surfaces. Adv. Math. 208(1), 249–273. math.DG/0405348, 2004.
Fricke, R. and Klein, F. Vorlesungen der Automorphen Funktionen, Teubner, Leipzig,
Germany, Vol. I (1897), Vol. II (1912).
Gardiner, F. P. and Lakic, N. (2000). Quasiconformal Teichmüller theory. Mathematical
Surveys and Monographs. Vol. 76, American Mathematical Society Providence, RI.
Garcı́a-Prada, O. and Mundet i Riera, I. (2003). Representations of the fundamental
group of a closed orientable surface in Sp(4, R). Topology 43, 831–55.
Garcı́a-Prada, O., Gothen, P. and Mundet i Riera, I. (2008). Higgs bundles and surface
group representations in the real symplectic group. Preprint arXiv:0809.0576v3.
Goldman, W. (1980). Discontinuous groups and the Euler class, Doctoral thesis.
University of California, Berkeley, CA.
Goldman, W. (1984). The symplectic nature of fundamental groups of surfaces. Adv.
Math. 54, 200–25.
Goldman, W. (1988). Topological components of spaces of representations. Inv. Math.
93(3), 557–607.
 References 161

Goldman, W. (1990). Convex real projective structures on compact surfaces. J. Diff.

Geom. 31(3), 791–845.
Goldman, W. (2006). Mapping class group dynamics on surface group representations,
in Problems on Mapping Class Groups and Related Topics. B. Farb (ed.), Proceedings
of the Symposium on Pure Mathematics, Vol. 74, American Mathematical Society,
Providence, RI, pp. 189–214. math.GT/0509114.
Goldman, W. and Millson, J. J. (1987). Local rigidity of discrete groups acting on
complex hyperbolic space. Inv. Math. 88, 495–520.
Goldman, W. and Millson, J. J. (1988). The deformation theory of representations of
fundamental groups of Kähler manifolds. Publ. Math. IHES. 67, 43–96.
Goldman, W. and Wentworth, R. (2007). Energy of twisted harmonic maps of Rie-
mann surfaces. In the tradition of Ahlfors-Bers, IV H. Masur and R. Canary
(eds.), Contemp. Math., 432, Amer. Math. Soc., Providence, RI, 45–61. math.DG/
0506212.
Goldman, W. and Xia, E. (2008). Rank one Higgs bundles and fundamental groups of
Riemann surfaces. Memoirs of the Am. Math. Soc. 904, 1–69. math.DG/0402429.
Gothen, P. (2001). Components of representations and stable triples. Topology 40,
823–50.
Guichard, O. (2005). Une dualité pour les courbes hyperconvexes. Geom. Ded. 112,
141–64. MR MR2163895.
Guichard, O. (2008). Composantes de Hitchin et representations hyperconvexes de
groupes de surface. J. Diff. Geom. 80(3), 391–431.
Guichard, O. and Wienhard, A. (2008). Convex foliated projective structures and
the Hitchin component for SL(4, R). Duke Math. J. 144(3), 381–445. math/
0702184.
Guichard, O. and Wienhard, A. Topological invariants for Anosov representations,
(Submitted).
Gunning, R. (1967). Special coordinate coverings of Riemann surfaces. Math. Ann. 170,
67–86.
Hernández Lamoneda, L. (1991). Maximal representations of surface groups in bounded
symmetric domains. Trans. Am. Math. Soc. 324, 405–20.
Hitchin, N. (1987). The self-duality equations on Riemann surfaces. Proc. Lond. Math.
Soc. 55, 59–126.
Hitchin, N. (1992). Lie groups and Teichmüller space. Topology 31(3), 449–73.
Hubbard, J. H. (2006). Teichmüller Theory, and Applications to Geometry, Topology
and Dynamics, Volume 1, Matrix Editions, Ithaca, NY.
Imayoshi, Y. and Taniguchi, M. (1992). An Introduction to Teichmüller Spaces,
Springer-Verlag, Tokyo, Japan.
Iozzi, A. (2002). Bounded cohomology, boundary maps, and representations into
Homeo+ (S 1 ) and SU(1, n). in Rigidity in Dynamics and Geometry, Cambridge, UK,
2000, Springer Verlag, Berlin, Germany, pp. 237–60.
Jost, J. (2006). Compact Riemann surfaces. in An Introduction to Contemporary
Mathematics, 3rd edition, Universitext, Springer-Verlag, Berlin, Germany.
Kostant, B. (1959). The principal three-dimensional subgroup and the Betti numbers
of a complex simple Lie group. Am. J. Math. 81, 973–1032.
Kuiper, N. (1954). On convex locally projective spaces. Convegno Internazionale di
Geometria Differenziale, Italy, pp. 200–213.
162 Higgs bundles and geometric structures on surfaces

Labourie, F. (1991). Existence d’applications harmoniques tordues valeurs dans les

variétés à courbure négative. Proc. Am. Math. Soc. 111(3), 877–82.
Labourie, F. (1997). RP2 -structures et differentielles cubiques holomorphes, in Proceed-
ings of the GARC Conference in Differential Geometry, Seoul National University.
Labourie, F. (2006). Anosov flows, surface group representations and curves in projec-
tive space. Inv. Math. 165(1), 51–114. math.DG/0401230.
Labourie, F. (2007). Flat projective structures on surfaces and cubic holomorphic
differentials, Special Issue: Volume in honor of the 60th birthday of Gregory Margulis,
Part 1 Pure Appl. Math. Q. 3(4), 1057–1099. math.DG/0611250.
Labourie, F. (2008). Cross ratios, Anosov representations and the energy func-
tional on Teichmüller space. Ann. Sci. École Norm. Sup. (4), 41(3), 437–469.
math.DG/0512070.
Loftin, J. (2001). Affine spheres and convex RPn−1 -manifolds. Am. J. Math. 123, 255–
74.
McOwen, R. (1993). Prescribed curvature and singularities of conformal metrics on
Riemann surfaces. J. Math. Anal. Appl. 177(1), 287–98.
Milnor, J. (1958). On the existence of a connection with curvature zero. Commun.
Math. Helv. 32, 215–23.
Nag, S. (1988). The Complex Analytic Theory of Teichmüller Spaces. Wiley-
Interscience, New York.
Narasimhan, M. S. and Seshadri, C. S. (1965). Stable and unitary vector bundles on a
compact Riemann surface. Ann. Math. 82, 540–67.
Nielsen, J. (1927). Untersuchungen zur Topologie der geschlossenen zweiseitigen
Flächen I. Acta Math. 50, 189–358.
Ramanathan, A. (1975). Stable principal bundles on a compact Riemann surface. Math.
Ann. 213, 129–52.
Raghunathan, M. (1972). Discrete Subgroups of Lie Groups, Springer Verlag, New York.
Ratcliffe, J. (2006). Foundations of hyperbolic manifolds. Graduate Texts in Mathemat-
ics, Vol. 149, Springer, New York.
Schoen, R. and Yau, S. T. (1978). Univalent harmonic maps between surfaces. Invent.
Math. 44, 265–78.
Simpson, C. (1988). Constructing variations of Hodge structure using Yang-Mills theory
and applications to uniformization. J. Am. Math. Soc. 1(4), 867–918.
Simpson, C. (1992). Higgs bundles and local systems. Inst. Hautes Études Sci. Publ.
Math. 75, 5–95.
Stillwell, J. (1987). The Dehn-Nielsen theorem, Appendix to Papers on Group Theory
and Topology, by Max Dehn, Springer Verlag, Berlin/Heidelberg/New York.
Thurston, W. (1979). The geometry and topology of 3-manifolds, Princeton University
Lecture Notes (unpublished). Available at http://www.msri.org/communications/
books/gt3m.
Toledo, D. (1989). Representations of surface groups in complex hyperbolic space. J.
Diff. Geom. 29(1), 125–33.
Tromba, A. (1992). Teichmüller theory in Riemannian geometry. ETHZ Lectures in
Mathematics, Birkhäuser, Basel, Switzerland.
Troyanov, M. (1991). Prescribing curvature on compact surfaces with conical singular-
ities. Trans. Am. Math. Soc. 324(2), 793–821.
 References 163

Turaev, V. G. (1985). The first symplectic Chern class and Maslov indices. (Russian)
Studies in topology, V. Zap. Nauchn. Sem. Leningrad. Otdel. Mat. Inst. Steklov.
(LOMI) 143, 110–129, 178.
Turaev, V. G. (1984). A cocycle of the symplectic first Chern class and Maslov indices.
(Russian) Funktsional. Anal. i Prilozhen. 18 (1), 43–48.
Weil, A. (1962). On discrete subgroups of Lie groups II. Ann. Math. 75, 578–602.
Wood, J. (1976). Bundles with totally disconnected structure group. Commun. Math.
Helv. 51, 183–99.
 IX
LOCALITY OF HOLOMORPHIC BUNDLES,
AND LOCALITY IN QUANTUM FIELD THEORY

Graeme Segal
Dedicated to Nigel Hitchin, on his 60th birthday

Nigel Hitchin has been a close colleague of mine for most of my mathematical life,
and I have profited enormously from my contact with him. His wonderful skill
in asking the right questions and in obtaining deep results without sacrificing
concreteness or simplicity has both inspired me and filled me with envy. Having
seen, however, how many others at this meeting are better qualified than I am
to talk about Nigel’s work, I decided it would be best to keep to my own terrain,
and talk about locality in quantum field theory. All the same, many of the ideas
involved are well exemplified in the study of bundles on Riemann surfaces which
Nigel is famous for, and I shall begin there, especially as the question of locality
relates to an aspect of his work that has not been talked about so far at this
meeting, namely, its role in so-called ‘geometric Langlands theory’.
We can approach the subject by contrasting two opposite ways of looking
at holomorphic vector bundles E on a compact Riemann surface Σ. At one
extreme, if we remove any finite set σ of points from Σ then E is trivial on the
remaining surface Σ − σ, so we can think of all the ‘twisting’ of the bundle as
being concentrated into tiny neighbourhoods of the points of σ. At the other
extreme, we can try to spread the twist as evenly as possible over all of Σ.
The classical case of line bundles is very simple – perhaps misleadingly so.
On the one hand, any line bundle L can be constructed from a divisor z =
n1 z1 + · · · + nk zk – an element of the free abelian group on the set of points
of Σ – by attaching the trivial bundle L0 on Σ − {z1 , . . . , zk } to trivial bundles
Li in the neighbourhood of each zi by means of the clutching function ζini ,
where ζi is a local parameter at zi . (The resulting bundle L(z) depends up to
canonical isomorphism only on the divisor z.) On the other hand, for any choice
of Riemannian metric on Σ, any line bundle can be given a unique unitary
connection with constant curvature, so that it looks exactly the same in the
neighbourhood of any point of Σ. The isomorphism classes of holomorphic line
bundles on Σ naturally form a commutative complex Lie group Pic(Σ), and the
correspondence between the two ways of looking at a bundle amounts to the
classical theorem that Pic(Σ) is – in the holomorphic category – the free abelian
group on Σ, traditionally called its ‘Albanese variety’, that is, that the map
 Noncommutative geometry 165

Σ → Pic(Σ) given by z → L(z) is universal 1 among holomorphic maps from Σ

to a commutative group.
We encounter two main obstacles when we try to formulate an analogue of
this attractive picture for higher dimensional bundles. The first is that the
isomorphism classes of n-dimensional holomorphic bundles on a Riemann surface
Σ do not form a nice space; and the second is that the bundles trivialized in the
complement of a given point do not form an abelian group. For brevity I shall
refer to these as the problems of ‘noncommutative geometry’ and of ‘algebraic
structures up to homotopy’, respectively, and shall say a little about each in
turn.

9.1 Noncommutative geometry

If Σ is the Riemann sphere then every n-dimensional holomorphic bundle E
on Σ is a sum of line bundles, of degrees k1 ≥ k2 ≥ · · · ≥ kn say, and E is
determined up to isomorphism by the n-tuple of degrees. Equivalently, E can be
obtained by attaching trivial bundles on Σ − {∞} and Σ − {0} by a holomorphic
attaching function which is a homomorphism λ : C× → GLn (C), unique up to
conjugation in GLn (C). Now the conjugacy classes of homomorphisms from C×
to a complex reductive group G – and hence the isomorphism classes of G-bundles
on the sphere – are in one-to-one correspondence with the finite-dimensional
holomorphic irreducible representations of the Langlands dual group L G. (If G
is GLn (C) then L G = G.) This is the starting point of geometric Langlands
theory, but it is not my subject here. Let us notice, however, that whereas
the isomorphism classes of representations form a countable discrete set the
behaviour of the bundles is quite different. When a holomorphic bundle E (t) on
the sphere depends holomorphically on a parameter t ∈ T we find that – if T is
connected – the isomorphism class of E (t) is constant on a dense open subset of
the parameter space T , but ‘jumps’ when t belongs to certain submanifolds: the
space T is stratified by the isomorphism class of E (t) . Thus the countable set BΣ
of isomorphism classes of bundles consists of a sequence of connected components

1 This may seem strange. The set-theoretical free abelian group F generated by the points
Σ
of Σ fits into an exact sequence

×
KΣ → FΣ → PicΣ → 1,

where KΣ is the field of rational functions on Σ. The group FΣ is the disjoint union of
(n) (n)
a sequence of finite-dimensional algebraic varieties FΣ , where FΣ consists of all Σnk xk
(n)
such that Σ|nk | = n. Furthermore, FΣ has a natural topology in which the closure of FΣ is
(m)
compact, and is the union of the FΣ for m ≤ n. But we cannot say that FΣ is any kind of
algebraic variety: in fact it is easy to see that if U is a neighbourhood of the identity element
(n)
of FΣ then any continuous f : U → C for which f |U ∩ FΣ is holomorphic for each n has to
×
be constant along the orbits of KΣ .
166 Locality of holomorphic bundles, and locality in quantum field theory

indexed by the first Chern class of E, but each connected component, though
infinite, is the closure of a single point.
To make better sense of unpromising ‘spaces’ such as BΣ a number of different
approaches are commonly used. In algebraic geometry the main candidates are
1. To regard BΣ as a stack, that is, to work with the category – actually a
groupoid – of bundles and their isomorphisms rather than just with the set
π0 (BΣ ) of isomorphism classes of objects
2. the approach of geometric invariant theory, which picks out a class of
‘stable’ bundles whose isomorphism classes do form a nice space, in fact
an algebraic variety
Here I am going to talk about the first approach, which is more obviously
related to quantum field theory, especially in the treatment of the Langlands
theory by Kapustin and Witten (2007). Nigel Hitchin’s own main tool, however,
was geometric invariant theory.
Stepping back a little from algebraic geometry, one can say that the category
of bundles on a space resembles the category of representations of a group. For
example, the Hitchin moduli space associated to a surface Σ and a compact group
G is – among other things – the space of conjugacy classes of homomorphisms
from the fundamental group π1 (Σ) to the complexified group GC . The ‘space’
of irreducible unitary representations of a group Γ is an archetypal example
in Connes’s theory of noncommutative geometry (Connes 1994). He observes
that the set of irreducible representations is the set Spec(AΓ ) of irreducible
∗-representations of a C∗ -algebra AΓ associated to Γ. If A is a commutative
algebra its irreducible representations are the algebra-homomorphisms A → C,
which form a topological space Spec(A) on which A is an algebra of continuous
complex-valued functions. Connes’s idea is to think of the geometry of the set of
irreducible representations of the group Γ as defined by the noncommutative
algebra AΓ rather than by the – often too small – commutative algebra of
continuous functions on Spec(AΓ ), which in fact is the centre of AΓ .
To relate this picture to the stacks or groupoids of algebraic geometry we
think of a groupoid as halfway between a group and a space. More precisely,
a groupoid B consists of two sets together with some maps between them: a
set B0 of objects, and a set B1 of morphisms which is the disjoint union of
the sets B(x, y) of morphisms from x to y, where x and y run through all the
objects in B0 . We are concerned here, however, with topological groupoids, for
which the sets B0 and B1 have topologies and the structural maps between them
are continuous. At one extreme, if B0 is a point then we have a topological
group; at the other, if B0 = B1 , we have no morphisms except identities, and
the groupoid is simply a space. As Connes has emphasized, a groupoid B has a
groupoid-algebra AB , which interpolates between the group-algebra of a group
and the commutative algebra of functions on a space. (The algebra AB is devised
so that an AB -module is the same thing as a functor from B to vector spaces.
For a discrete groupoid, AB has a vector space basis ef indexed by the set
 Noncommutative geometry 167

B1 of morphisms, and composition is given by ef eg = ef ◦g when f and g are

composable, and ef eg = 0 otherwise. For the actual groupoids at hand, we must
make a choice of what we mean by the algebra, just as for a Lie group we can
define different group algebras depending on the geometric context in which we
are working.)
The groupoids B we are concerned with all arise from the action of a group
G on a space B0 , so that B1 = G × B0 , and the set of morphisms from x to y is
{g ∈ G : gx = y}. I shall denote this groupoid by B0 //G. The algebra AB in this
case is the twisted group algebra G
A0 , where A0 is the commutative algebra
of functions on B0 .
The essential features of a groupoid B are the (often not very nice) space
π0 (B) of isomorphism classes of its objects, and the groups Aut(x) of automor-
phisms of the objects, which are the obstructions to B’s being a space. These
features depend only on the equivalence class of the groupoid (in the sense of
category theory), and we are interested in groupoids only up to equivalence.
In fact we are content (cf. Segal 1973, 2.8) even with what I shall call ‘weak
equivalences’: functors T : B → B  which induce bijections π0 (B) → π0 (B ) and
Aut(x) → Aut(T (x)) for each object x, and which are ‘covering maps’ (in the
sense that they have local cross-sections) on the space of objects and morphisms.
Thus if G acts freely on B0 , making it a locally trivial principal G-bundle on the
space B0 /G, then we are not interested in the difference between the groupoid
B0 //G and the space B0 /G. Similarly, whenever we decompose a closed surface
Σ into two pieces Σ1 and Σ2 intersecting in a curve S we can identify the set
of isomorphism classes of n-dimensional bundles on Σ with the (non-Hausdorff)
double-coset space

GΣ1 \GS /GΣ2 ,

where GΣi is the group of holomorphic maps from Σi to G = GLn (C), and GS
is the group of smooth maps S → G, for any bundle can be trivialized on a
non-closed surface. We then have three groupoids:

1. B (1) formed by the action of GΣ1 on the homogeneous space GS /GΣ2

2. B (2) formed by the action of GΣ2 on GΣ1 \GS
3. B (12) formed by the action of GΣ1 × GΣ2 on GS

All three are weakly equivalent, and are weakly equivalent to the groupoids
obtained from any other way of decomposing Σ into pieces. According to the
‘stack’ point of view, any of them can be taken as the ‘space’ of bundles
on Σ.
The relation of equivalence of groupoids translates into Morita equivalence of
algebras. Two algebras are Morita equivalent if their categories of left modules
are equivalent. Recall that if A and B are algebras then an (A, B)-bimodule
F defines an additive functor F∗ : MB → MA , where MA and MB are the
categories of left A and B modules, by F∗ (M ) = F ⊗B M . In particular, for the
168 Locality of holomorphic bundles, and locality in quantum field theory

(A, A)-bimodule A we have A∗ = idMA , and if G is a (B, C)-bimodule we have

F∗ ◦ G∗ = (F ⊗B G)∗ : MC → MA .
On the level of discrete groupoids, at least, there is a dictionary
groupoid B −→ algebra AB
functor T : B → B  −→ (AB , AB )-bimodule FT
natural transformation Φ : T → T  −→ isomorphism Φ∗ : FT → FT  .
In the light of the previous discussion we see that a ‘noncommutative space’ is
defined not by an algebra but by a Morita equivalence class of algebras, and so
it is more naturally described by a linear category – the category of left modules
for the algebra. In holomorphic geometry, however, the algebras that arise, even
when they are commutative, are not semisimple, and their module categories
do not have very convenient properties. It is therefore better to go further, and
replace the module categories by the categories of cochain complexes of modules.
That is why the geometric Langlands correspondence is stated in terms of the
linear categories of complexes of coherent sheaves 2 – or of D-modules – on the
moduli spaces of bundles.
Yet another version – slightly more general still – of the notion of noncommu-
tative space arises from quantum field theory, and it is perhaps the most natural
one in the geometric Langlands theory. I shall sketch it below. Before leaving the
present discussion, however, it may be worth making another remark.
A generalized space defined by a topological groupoid has a homotopy type,
just like an ordinary space. For – like any topological category (cf. Segal 1968) –
the groupoid B0 //G has a ‘realization’ |B0 //G| as a space, and equivalent
groupoids have homotopy-equivalent realizations. A generalized space defined
by a noncommutative ring – or, better, by a linear category C – has no such
homotopy type. The best one can do is consider the stable homotopy type (or
spectrum) KC defined by applying the usual K-theory construction to C: this
is (cf. Segal 1977) the ‘space’ whose homotopy groups are the K-groups of the
category.
In the case of the stack BΣ of holomorphic G-bundles on a Riemann surface Σ
the space |BΣ | has the same homotopy type as the realization of the correspond-
ing topological groupoid BΣ sm
of smooth bundles, which in turn has the homotopy
type of the space Map(Σ; BG) of continuous maps from Σ to the classifying space
of G. (This follows at once from the fact that for a non-closed Riemann surface
Σi the space of holomorphic maps from Σi to G has the homotopy type of the
space of continuous maps.) On the other hand, we know from the beautiful work
of Atiyah and Bott 1982 that the moduli space of holomorphic bundles in the
sense of geometric invariant theory can be identified with the minimum level of
the Yang–Mills functional on the space of smooth bundles.
2 Sheaves rather than modules, because the algebraic varieties are not affine. But that is a

standard technicality.
 Algebraic structures up to homotopy 169

9.2 Algebraic structures up to homotopy

Let us consider the space B̂Σ of pairs (E, e) consisting of a holomorphic vector
bundle E on Σ equipped with a meromorphic trivialization e – that is, E is
trivialized, in the complement of a finite subset σ of Σ, by n holomorphic
sections e1 , . . . , en which extend to meromorphic sections on Σ. A pair (E, e)
is completely determined by giving σ and, for each z ∈ σ, the local information
consisting of a bundle in a neighbourhood of z trivialized away from z. The items
of local information can be prescribed independently, so B̂Σ can be regarded
as a labelled configuration space. The group GLn (KΣ ) acts on B̂Σ by changing
the meromorphic trivialization, and the orbit space is the space of isomorphism
classes of holomorphic bundles. (The isotropy group of any pair (E, e) is just the
group of holomorphic automorphisms of E.) Let us now recall a few aspects of
the theory of labelled configuation spaces.
For an d-dimensional manifold M let Č(M ) denote the manifold of all finite
unordered
1 subsets of M , with its natural topology in which it is the disconnected
union k≥0 Ck (M ) of the spaces Ck (M ) of subsets with exactly k elements. If P
1
is an arbitrary auxiliary space, we can also form Č(M ; P ) = Ck (M ; P ), whose
points are the finite subsets σ of M with each point z ∈ σ ‘labelled’ by a point
pz of P . The space Ck (M ; P ) is fibred over Ck (M ) with fibre P k .
The configuration spaces we are interested in, however, have a topology which
allows the points of a configuration σ to move continuously into coincidence
and the labels at the same time to amalgamate – or ‘add’ – in some sense. For
example, the free abelian group FΣ is the configuration space of Σ labelled by
the group Z, with the usual addition when points merge. We can define such
a ‘configuration space with amalgamation’ whenever the labelling space P has
a composition-law which is sufficiently associative and commutative to make it
what is called in homotopy theory an d-fold loop space – in fact the existence of
the amalgamated configuration space C(M ; P ) for all d-manifolds M is a good
way of defining an d-fold loop space. 3 To be precise, I shall say, in the style of
Beilinson and Drinfeld, that an amalgamated configuration space is any space
C(M ; P ) to each of whose points is associated a finite subset σ of M called its
support, and which is equipped with a map iz : P k → C(M ; P ), for each sequence
z = {z1 , . . . , zk } of distinct points of M , with the properties:

(i) the image of iz consists of configurations with support z, and

(ii) if U is the disjoint union of k small open balls U1 , . . . , Uk such that Ui
contains zi , and CU (M ; P ) denotes the part of C(M ; P ) with supports in
U , then iz : P k → CU (M ; P ) is a homotopy equivalence.
3 The conventional way to define the structure of an d-fold loop space on a space P is to

give the amalgamation maps

Ck (Rd ; P ) → P

for each k > 0; but these must of course satisfy various compatibilities.
170 Locality of holomorphic bundles, and locality in quantum field theory

Although it is not strictly necessary, we may as well assume that CU (M ; P ) is

an open subset of C(M ; P ) when U is an open subset of M , and that it can be
identified with C(U ; P ). That is certainly true in our examples, where C(M ; P )
is simply Č(M ; P ) with a coarser topology.
In fact we need a slight generalization to include bundles of d-fold loop spaces.
We shall allow the label of a point z ∈ M to lie not in a fixed space P but rather
in the fibre at z of a bundle P on the manifold M . It is clear how the definition
of C(M ; P ) should be adapted to this case.
When we have a single d-fold loop space P we define its d-fold classifying
space – or dth ‘delooping’ – as the space
B d P = C(U ; P )/CV (U ; P ),
where U is an open ball – say U = {z ∈ Rd :  z  < 1} – and V is the annular
region {z ∈ U : 1/2 ≤  z  < 1}, and the notation means that the subspace
CV (U ; P ) of C(U ; P ) is collapsed to a single point, which is a natural base
point in B d P .
Similarly, when P is a bundle of d-fold loop spaces on M we can define a
bundle B d P whose fibre at z ∈ M is constructed using a neighbourhood U of z
in M .
The main theorem about labelled configuration spaces is the following.

Theorem 9.1 There is a natural map

C(M ; P ) −→ Γcpt (M ; B d P ),
where Γcpt denotes the space of cross-sections which are equal to the base point
outside of a compact subset of M . If the composition law of P makes the set of
components π0 (P ) into a group then the map is a homotopy equivalence.

Notice that applying the theorem when M is an open ball tells us that P is
homotopy-equivalent to the d-fold based loop space of B d P .
The equivalences are defined by the scanning map (see McDuff 1975, 1977
and Segal 1979) which associates to an element c ∈ C(M ; P ), the section of
B d P whose value at z ∈ M is the image of c in C(M ; P )/CM −W (M ; P ) =
C(U ; P )/CV (U ; P ), where W ⊂ U are two concentric open balls around z, and
V = U − W . The theorem is very easy to prove, and not at all deep, in the form
I have stated it here: in applications the main difficulty may be to verify the
hypotheses.
In the application to holomorphic vector bundles on a surface, at a point z on
the surface, with neighbourhood U , the labelling space P is the quotient GǓ /GU ,
where GǓ is the group of holomorphic maps U − {z} → GLn (C) which extend
meromorphically over U , and GU is the group of holomorphic maps U → GLn (C).
This can be identified with GLn (C(t))/GLn (C[t]). It is (see Pressley and Segal
1986) the union of a sequence of compact finite-dimensional algebraic varieties,
and it has the homotopy type of the based loop space of GLn (C). To define the
 Quantum field theory 171

topology of B̂Σ we first define a topology on the part B̂Σ,U with support in a
disjoint union U of open discs by identifying it with G∂U /GU , where G∂U is the
group of smooth maps ∂U → GLn (C) which are boundary values of meromorphic
maps U → GLn (C), and then we give B̂Σ the finest topology compatible with
these. Having checked the hypotheses of the theorem, it tells us that B̂Σ has the
homotopy type of the space of continuous maps from Σ to B 2 P  BGLn (C),
that is, the homotopy type of the space of smooth bundles, as we expect from
the stack picture.
Before leaving this topic, let us mention the converse question to the one
answered by the theorem. If we are given a space Q, can we model the mapping
spaces Map(M ; Q) for varying d-manifolds M – at least up to homotopy – by
labelled configuration spaces C(M ; P )? The answer, clearly, is: if and only if the
space Q is d-connected; for the delooping B d P of any d-fold loop space P is d-
connected. If we want to model spaces of maps into less highly connected spaces
Q we would have to allow not just ‘particles’ but also configurations of higher
dimensional submanifolds – presumably, up to dimension m if Q is (d − m)-
connected. (One way to see this is to realize Q as an open manifold with a Morse
function with critical points of indices ≥ d − m, and to make maps M → Q flow
downwards along the gradient flow.)

9.3 Quantum field theory

All the mathematical phenomena I have been discussing play an important
role in quantum field theory. In particular, noncommutative geometry enters
in two somewhat opposite ways, first because the moduli spaces of theories are
noncommutative spaces, and also because field theories can be used to give a
new formulation of noncommutative geometry. It is only the second aspect that
I am going to talk about. But let us begin at the beginning. . . .
In quantum mechanics a system is described at any time by giving an algebra
A of ‘observables’ and a linear map θ : A → C called the ‘state’. 4 In quantum
field theory, in space-time of dimension d, this picture is enriched by supposing
that the observables are spread out over a given space-time manifold M . More
precisely, for each x ∈ M there is given a sub-vector space Ox of A formed by the
observables which can be measured in the neighbourhood of x. Quantum field
theory assumes that – up to some global topological effects 5 to which I shall
return at the end of this talk – the complete information about the system is
contained in the maps
Θk : Ox1 ⊗ · · · ⊗ Oxk −→ C,
4 It is perhaps more usual to say that A is an algebra of operators in a Hilbert space H,

and that the state is a unit vector ψ in H related to θ by θ(a) = ψ, a ψ. But the equivalent
description by (A, θ) seems more satisfactory to me.
5 The essential example is the Bohm–Aharonov phenomenon, when an electromagnetic field

in a non-simply connected region is described by a flat connection which is undetectable in any

simply connected subregion.
172 Locality of holomorphic bundles, and locality in quantum field theory

for each finite set {x1 , . . . , xk } of distinct points of M , which are got by
multiplying in A and composing with θ. The Θk are traditionally called vacuum
expectation values.
It is not easy to say what properties the vector spaces Ox and the functions
Θk must have for them to constitute a quantum field theory. A first attempt at
an answer can be given by defining a d-dimensional theory as a rule which
1. associates a complex topological vector space HY to each compact-oriented
Riemannian manifold Y of dimension d − 1, functorially with respect to
diffeomorphisms Y → Y  , and
2. associates a trace-class operator UX : HY0 → HY1 to each oriented
Riemannian cobordism X from Y0 to Y1 .
These data are constrained to satisfy two axioms:
(a) Concatenation:
UX  ◦X = UX  ◦ UX
when X  ◦ X is the cobordism from Y0 to Y2 obtained by concatenating
X from Y0 to Y1 with X  from Y1 to Y2 .
(b) Tensoring: We are given associative natural isomorphisms
∼
=
HY ⊗ HY  → HY Y

UX ⊗ U X  = U X X

when we have disjoint unions Y ! Y  or X ! X  of (d − 1)-manifolds or

cobordisms.
Notice that it follows from property (a) that HY = C if Y is the empty (d − 1)-
manifold.
When we have a theory in this sense we can reconstruct the local observables
and their expectation values. We define the vector space Ox of observables for
each point x in a closed d-manifold M by
Ox = lim H∂D ,
←

where the inverse limit is over the ordered set of all closed balls D in M which
are neighbourhoods of x, ordered by
◦
D  > D ⇐⇒ D ⊂D,
in which case we have a canonical map U ◦ : H∂D  → H∂D defined by the
D−D 
annular cobordism.
If x1 , . . . , xk are distinct points of M , and D1 , . . . , Dk are disjoint discs with
xi in the interior of Di , let M0 denote the manifold with boundary obtained by
 Quantum field theory 173

deleting from M the interiors of the discs Di . We regard M0 as a cobordism from

1
∂Di to the empty manifold, and define
Θk : Ox1 ⊗ · · · ⊗ Oxk −→ C
as the inverse limit of the maps
UM0 : H∂D1 ⊗ · · · ⊗ H∂Dk → C
as the discs Di shrink to points around the points xi .
If the points xi are all contained in the interior of a disc D then the map Θk
clearly factorizes through H∂D . We can interpret this as saying that the local
observables have the structure of an algebra which is associative and commutative
up to homotopy, for if D is a small neighbourhood of a point x ∈ M then we can
regard H∂D as a completion Ôx of Ox (for the maps H∂D → H∂D in the system
defining Ox are always injective with dense image), and we have a map
Ox1 ⊗ · · · ⊗ Oxk → Ôx
for any family x1 , . . . , xk of distinct points sufficiently close to x. This is an exact
linear analogue of the d-fold loop space structures discussed above, and it is what
in traditional quantum field theory is called the operator product expansion.
It is still not clear, however, that we have made our definition of a quantum
field theory sufficiently rigid. One feels that the vector space HY associated
to a (d − 1)-manifold Y should be constructed locally from Y , and perhaps
more axioms are needed to ensure this. A natural second approximation to the
definition is the notion of a three-tier theory, which gives an additional layer
of structure to allow (d − 1)-manifolds to be cut into pieces. In a three-tier
theory
1. to each compact oriented Riemannian (d − 2)-manifold Z there is associated
a linear category CZ ,
2. to each (d − 1)-dimensional Riemannian cobordism Y from Z0 to Z1 there
is associated an additive functor HY : CZ0 → CZ1 , and
3. to each d-dimensional Riemannian cobordism X from Y to Y  , where Y
and Y  are cobordisms from Z0 to Z1 , there is associated a transformation
of functors UX : HY → HY  .
As with the earlier definition, the data are required to satisfy two axioms of
concatenation and tensoring. A theory in the earlier two-tier sense is obtained
from the three-tier structure by restricting to closed (d − 1)-manifolds, which
can be regarded as cobordisms from the empty (d − 2)-manifold ∅ to itself. The
tensoring axiom implies that C∅ is the category of vector spaces, and since any
additive functor C∅ → C∅ is given by tensoring with a vector space, we can identify
HY with a vector space when Y is closed.
In the form I have just stated, the three-tier definition is too vague to be of
much use. In this talk I shall not try to elaborate it, as my purpose is to make
174 Locality of holomorphic bundles, and locality in quantum field theory

just two points. The first is that a three-tier two-dimensional field theory seems
to have a good claim as a candidate definition of a ‘noncommutative manifold’.
For, schematically at least, a natural way to give the data of a three-tier theory
is to associate an algebra AZ to each (d − 2)-manifold, and to take CZ to be the
category of left AZ -modules; then to a cobordism Y is associated an (AZ1 , AZ0 )-
bimodule HY , which defines a functor CZ0 → CZ1 by
M → HY ⊗AZ0 M ;
and to a cobordism between cobordisms is associated a homomorphism of
bimodules. When d = 2 this simply means that we have a dual pair of linear
categories – the left and right modules for an algebra – associated to a point
with its two orientations, while the one-dimensional data expresses the cate-
gorical duality, and the two-dimensional data gives us ‘trace’ or ‘integration’
maps. The field theory even leads one naturally from categories of modules to
categories of cochain complexes of modules, if we assume that the field theory is
supersymmetric.
The idea that two-dimensional theories should replace manifolds is, of course,
the central proposal of string theory, which models space-time by a two-
dimensional conformal field theory, with the category of D-branes in space-time
as the category which the field theory associates to a point. It is also what arises
in the Kapustin–Witten treatment of geometric Langlands duality. There, one
begins from the maximally supersymmetric four-dimensional Yang–Mills theory
associated to a compact group G, and observes that, for any compact surface
Σ, a four-dimensional theory gives a two-dimensional theory by dimensional
reduction along Σ – that is, by composing with the functor M → Σ × M from i-
manifolds to (i + 2)-manifolds. The two-dimensional theory obtained from Yang–
Mills theory for G by reducing along Σ is supposed to associate to a point
the category of D-modules on the moduli space of holomorphic GC -bundles
on Σ.
My second point is even vaguer. The obvious fear, if one starts to study
three-tier theories, is that one will be impelled to believe that a d-dimensional
theory should really mean a (d + 1)-tier d-dimensional theory, which associates
a two-category to a manifold of dimension d − 3, and even worse things to
lower dimensional manifolds, until one gets to a (d − 1)-category associated to a
point.
This is not completely mad, as it works well in the one famous example afforded
by three-dimensional Chern–Simons theory for a compact group G at a given
‘level’. There, the category associated to a circle S is the category of positive
energy representations of the loop group of maps S → G at the specified level,
and the two-category associated to a point is the same thing, but remembering
its tensor structure coming from the fusion of loop group representations. (We
think of this as a two-category with just one object, whose linear category of
endomorphisms is the category of loop group representations, with fusion as its
composition law.)
 References 175

I cannot believe, however, that genuine – non-topological – quantum field

theory will be advanced by higher categories. The algebras-up-to-homotopy
formed by the field operators look much more promising. It is interesting that,
in the context of homotopy theory, a d-fold loop space P does indeed give rise
to a precise analogue of the structure of a (d + 1)-tier field theory, as follows.

1. To a closed d-manifold X we associate the space QX = C(X; P ).

2. To a closed (d − 1)-manifold Y we associate the ‘group’ – that is, one-fold
loop space – GY = C(I × Y ; P ), where I is an open interval.
3. To a closed d-manifold X with boundary we associate the G∂X -space
QX = C(X ◦ ; P ), where X ◦ is the interior of X, noticing that if X =
X1 ∪Y X2 then QX  QX1 ×GY QX2 . And, more generally, to a cobordism
from Y0 to Y1 we can associate a (GY0 × GY1 )-space, and hence a functor
from GY0 spaces to GY1 spaces.
4. To a closed (d − 2)-manifold Z we associate the two-fold loop space GZ =
C(I × I × Z; P ), noticing that if Y is a (d − 1)-manifold with boundary
then GY = C(I × Y ◦ ; P ) is a one-fold loop space on which the two-fold
loop space G∂Y acts by maps of one-fold loop spaces.

And so on – I shall not spell out all the details.

I hope that a linear and analytical version of this picture is the model for
quantum field theory. One point where the analogy may help is in considering
whether one should expect that the d-fold algebra of field operators defined in a
small d-dimensional ball should determine the theory on an arbitrary space-time,
in the way that the d-fold loop space P determines all the spaces C(X; P ). We
saw that a mapping space Map(X; Q) is of the form C(X; P ) only when the target
space Q is d-connected, and I would guess that an analogous distinction applies
to field theories: just as the mapping space cannot be modelled by particles, but
requires m-dimensional ‘objects’, if Q is only (d − m)-connected, so we know
that a field theory in general has non-local observables that can be seen only
in topologically non-trivial regions of space-time. It seems, however, that the
non-local observables are ‘topological’ in the sense that they contribute only a
finite number of degrees of freedom to the infinite-dimensional physical system
that the field theory describes.

References
Atiyah, M. F. and Bott, R. (1982). The Yang-Mills equations over Riemann surfaces.
Phil. Trans. Roy. Soc. London 308A, 523–615.
Connes, A. (1994). Noncommutative Geometry. Academic Press, San Diego, CA.
Kapustin, A. and Witten, E. (2007). Electro-magnetic duality and the geometric
Langlands program. Commun. Number Theory Phys. 1, 1–236. hep-th/0610210.
McDuff, D. (1975). Configuration spaces of positive and negative particles. Topology
14, 91–107.
176 Locality of holomorphic bundles, and locality in quantum field theory

McDuff, D. (1977). Configuration spaces. In K-Theory and Operator Algebras. Proceed-

ings Georgia 1975, Springer Lecture Notes 575, pp. 88–95.
Pressley, A. and Segal, G. (1986). Loop Groups. Oxford University Press, Oxford.
Segal, G. (1968). Classifying spaces and spectral sequences. Proc. Math. IHES Paris
34, 105–12.
Segal, G. (1973). Configuration spaces and iterated loop-spaces. Invent. Math., 21,
213–21.
Segal, G. (1977). K-homology theory and algebraic K-theory. In K-Theory and Operator
Algebras. Proceedings Georgia 1975, Springer Lecture Notes 575, pp. 113–27.
Segal, G. (1979). The topology of spaces of rational functions. Acta Math 143, 39–72.
 X
TOEPLITZ OPERATORS AND HITCHIN’S
PROJECTIVELY FLAT CONNECTION

Jørgen Ellegaard Andersen

Dedicated to Nigel Hitchin on the occasion of his 60th birthday

10.1 Introduction
Let Z be a (2 + 1)-dimensional topological quantum field theory (TQFT). That
is, the two dimensional part of Z is a modular functor with finite label set 1 Λ:
⎧ ⎫
⎪ Category of ⎪
⎪
⎪ ⎪
⎪
⎪ (extended) closed ⎪ ⎪
⎪
⎪
⎪ ⎪  )
⎨oriented surfaces ⎪ ⎬ Category of finite
Z: with Λ-labeled → dimensional vector
⎪
⎪ ⎪
⎪
⎪
⎪ marked points with ⎪
⎪ spaces over C
⎪
⎪ ⎪
⎪
⎪
⎩ projective tangent ⎪
⎭
vectors
The three-dimensional part of the TQFT is an association of a vector
Z(M, L, λ) ∈ Z(∂M, ∂L, ∂λ)
to any compact oriented framed three-manifold M together with an oriented
framed link (L, ∂L) ⊆ (M, ∂M ) and a Λ-labeling λ : π0 (L) → Λ:

λ2

λ2
λ2
λ1

This association has to satisfy the Atiyah–Segal–Witten TQFT axioms (see, e.g.,
Witten 1989, Atiyah 1990, Segal 1992 and for a more comprehensive presentation
of the axioms see Turaev 1994).

1 The set is also equipped with an involution † and a trivial element which is preserved by

the involution.
178 Toeplitz operators and Hitchin’s projectively flat connection

Witten constructed via path integral techniques a quantization of Chern–

Simons theory in (2 + 1) dimensions and argued in Witten (1989) that this
produced a TQFT index by a compact simple Lie group and an integer level
(n)
k. For the group SU (n) and level k let us denote this TQFT by Zk . Witten
(2)
(1989) shows that the theory Zk determines the Jones polynomial of a knot in
3
S . Combinatorially this theory was first constructed by Reshetikin and Turaev
(1990, 1991) using representation theory of Uq (sl(n, C)) at q = e(2πi)/(k+n) .
(n)
Subsequently the TQFTs Zk were constructed using skein theory by Blanchet,
Habegger, Masbaum, and Vogel in Blanchet et al. (1992), (1995) and Blanchet
(2000).
(n)
For this theory the label set Λk is a finite subset (depending on k) of the set of
finite-dimensional irreducible representations of SU (n). We use the usual labeling
(n)
of irreducible representations by Young diagrams, so in particular
∈ Λk is the
(d) (n)
defining representation for SU (n). Let further λ0 ∈ Λk be the Young diagram
consisting of d columns of length k.
Part of this TQFT consists of the quantum SU (n) representations of the
mapping class group. Namely, if Σ is a closed oriented surface of genus g and
Γ is the mapping class group of Σ. Let p be a point on Σ. Then the modular
functor induces a representation
(n,d)  (n) (d) 
Zk : Γ → P Aut Zk (Σ, p, λ0 ) . (10.1)

For a general label of p we would need to choose a projective tangent vector

vp ∈ Tp Σ/R+ and we would get a representation of the mapping class group of
(d)
(Σ, p, vp ), but for the special labels λ0 , the dependence on vp is trivial and in
fact we get a representation of Γ. Furthermore, the curve operators are also part
(n)
of any TQFT: For γ ⊆ Σ − {p} an oriented simple closed curve and any λ ∈ Λk
we have the operators
(n,d) (n) (d) (n) (d)
Zk (γ, λ) : Zk (Σ, p, λ0 ) → Zk (Σ, p, λ0 ). (10.2)

λ
(d)
λ0

For the curve operators, we can derive an explicit formula using factorization:
Let Σ be the surface obtained from cutting Σ along γ and identifying the
two boundary components to two points, say {p+ , p− }. Here p+ is the point
(n)
corresponding to the “left” side of γ. For any label µ ∈ Λk we get a labeling of
the ordered points (p+ , p− ) by the ordered pair of labels (µ, µ† ).
 Introduction 179

(n) (n) (d)

Since Zk is also a modular functor one can factor the space Zk (Σ, p, λ0 )
(n)
into a direct sum “along” γ as a direct sum over Λk . That is, we get an
isomorphism

9
Zk (Σ, p, λ0 ) ∼ Z (k) (Σ , p+ , p− , p, µ, µ† , λ0 ).
(n) (d) (d)
= (10.3)
(n)
µ∈Λk

Strictly speaking we need a base point on γ to induce tangent directions at p± .

(d)
However, the corresponding subspaces of Z (k) (Σ, p, λ0 ) do not depend on the
choice of base point. The isomorphism (10.3) induces an isomorphism
9
End(Z (k) (Σ, p, λ0 )) ∼ End(Z (k) (Σ , p+ , p− , p, µ, µ† , λ0 )).
(d) (d)
=
(n)
µ∈Λk

(d)
which also induces a direct sum decomposition of End(Z (k) (Σ, p, λ0 )) which is
independent of the choice of the base point.
The TQFT axioms imply that the curve operator Z (k) (γ, λ) is diagonal with
respect to this direct sum decomposition along γ. One has the formula
9
Z (k) (γ, λ) = Sλ,µ (S0,µ )−1 IdZ (k) (Σ ,p ,p ,p,µ,µ† ,λ(d) ) .
+ − 0
(n)
µ∈Λk

(n)
Here Sλ,µ is the S-matrix 2 of the theory Zk (see, e.g., Blanchet 2000 for a
derivation of this formula).
In this chapter we review some of the results we have established regarding
these representations of the mapping class group and the curve operators. For
that we need the geometric construction of these theories as discussed below. We
will here discuss the applications to TQFT of the general program of combining
the theory of Toeplitz operators with the Hitchin connection. In this introduction
we will cover the applications. The general setting will be discussed in the rest
of this chapter.

(n)
10.1.1 Geometric construction of Zk
The geometric construction of these TQFTs was proposed by Witten (1989),
where he derived via the Hamiltonian approach to quantum Chern–Simons
theory, that the geometric quantization of the moduli spaces of flat connec-
tions should give the two-dimensional part of the theory. Further, he pro-
posed an alternative construction of the two-dimensional part of the theory
2 The S-matrix is determined by the isomorphism a modular functor induces from two

different way of glueing an annulus to obtain a torus. For its definition see for example Moore
and Seiberg (1989), Walker (1991), Segal (1992), or Bakalov and Kirillov (2000) and references
therein. It is also discussed in Andersen and Ueno (2006).
180 Toeplitz operators and Hitchin’s projectively flat connection

via WZW-conformal field theory. This theory has been studied intensively.
In particular the work of Tsuchiya, Ueno, and Yamada (1989) provided the
major geometric constructions and results needed. In Bakalov and Kirillov
(2000) their results was used to show that the category of integrable highest
modules of level k for the affine Lie algebra associated to any simple Lie
algebra is a modular tensor category. Further in Bakalov and Kirillov (2000)
this result is combined with the works of Kazhdan and Lusztig (1993, 1994a,
1994b) and of Finkelberg (1996) to argue that this category is isomorphic to
the modular tensor category associated to the corresponding quantum group,
from which Reshetikhin and Turaev constructed their TQFT. Unfortunately,
these results do not allow one to conclude the validity of the geometric con-
structions of the two-dimensional part of the TQFT proposed by Witten.
However, in joint work with Ueno (Andersen and Ueno, 2006, 2007a, 2007b,
2008), we have given a proof, based mainly on the results of Tsuchiya, Ueno
and Yamada (1989), that the TUY construction of the WZW-conformal field
theory after twist by a fractional power of an abelian theory satisfies all the
axioms of a modular functor. Furthermore, we have proved that the full (2 + 1)-
dimensional TQFT that results from this is isomorphic to the one constructed
by BHMV via skein theory mentioned above. Combining this with the theo-
rem of Laszlo (Laszlo 1998), which identifies (projectively) the representations
of the mapping class groups one obtains from the geometric quantization of
the moduli space of flat connections with the ones obtained from the TUY
constructions, one gets a proof of the validity of the construction proposed
by Witten (1989).
Let us now briefly recall the geometric construction of the representations
(n,d)
Zk of the mapping class group, as proposed by Witten, using geometric
quantization of moduli spaces.
We assume from now on that the genus of the closed oriented surface Σ is
at least 2. Let M be the moduli space of flat SU (n) connections on Σ − p
with holonomy around p equal to exp(2πid/n) Id ∈ SU (n). In the case (n, d)
are coprime, the moduli space is smooth. In all cases the smooth part of the
moduli space has a natural symplectic structure ω. There is a natural smooth
symplectic action of the mapping class group Γ of Σ on M . Moreover there is a
unique prequantum line bundle (L, ∇, (·, ·)) over (M, ω). The Teichmüller space
T of complex structures on Σ parametrizes naturally and Γ-equivariantly Kähler
structures on (M, ω). For σ ∈ T , we denote (M, ω) with its corresponding Kähler
structure by Mσ .
By applying geometric quantization to the moduli space M , one gets a certain
finite rank bundle over Teichmüller space T , which we will call the Verlinde
bundle Vk at level k, where k is any positive integer. The fiber of this bundle
over a point σ ∈ T is Vk,σ = H 0 (Mσ , Lk ). We observe that there is a natural
Hermitian structure ·, · on H 0 (Mσ , Lk ) by restricting the L2 inner product on
global L2 sections of Lk to H 0 (Mσ , Lk ).
The main result pertaining to this bundle is
 Introduction 181

Theorem 10.1 (Axelrod, Della Pietra, and Witten 1991; Hitchin 1990)
The projectivization of the bundle Vk supports a natural flat Γ-invariant connec-
ˆ
tion ∇.

This is a result proved independently by Axelrod, Della Pietra, and Witten

(1991) and by Hitchin (1990). In Section 10.2 we review our differential geometric
construction of the connection ∇ ˆ in the general setting discussed in Andersen
(2006b). We obtain as a corollary that the connection constructed by Axelrod,
Della Pietra, and Witten projectively agrees with Hitchin’s.
(n,d)
Definition 10.1 We denote by Zk the representation
(n,d)  (n) (d) 
Zk : Γ → P Aut Zk (Σ, p, λ0 )

obtained from the action of the mapping class group on the covariant constant
sections of P(Vk ) over T .

The projectively flat connection ∇ ˆ induces a flat connection ∇ ˆ e in End(Vk ).

Let End0 (Vk ) be the subbundle consisting of traceless endomorphisms. The
connection ∇ˆ e also induces a connection in End0 (Vk ), which is invariant under
the action of Γ.

10.1.2 Asymptotic faithfulness

In Andersen (2006a) we proved the following. (See also Freedman, Walker, and
Wang for an alternative proof in the (n, d) = (2, 0) case.)

Theorem 10.2 Assume that n and d are coprime or that (n, d) = (2, 0) when
g = 2. Then we have that

∞
:   {1, H} g = 2, n = 2 and d = 0
(n,d)
ker Zk =
k=1
{1} otherwise.

where H is the hyperelliptic involution.

Large parts of the proof of this theorem, which we gave in Andersen (2006a),
apply to the general setting discussed in the following sections. In particular,
we get the general asymptotic faithfulness Theorem 10.19 discussed in the last
section of this chapter. Theorem 10.2 follows directly from Theorem 10.19 as
argued in the end of section 6 of Andersen (2006a). See also Andersen (2008a)
and Andersen and Christ (2008) for the singular case.
The main ingredient behind this is the theory of Toeplitz operators associated
to smooth functions on M . For each f ∈ C ∞ (M (d) ) and each point σ ∈ T , we
have the Toeplitz operator
(k)    
Tf,σ : H 0 Mσ , Lkσ → H 0 Mσ , Lkσ
182 Toeplitz operators and Hitchin’s projectively flat connection

which is given by
(k)
Tf,σ = πσ(k) (f s)
(k)
for all s ∈ H 0 (Mσ , Lkσ ), where πσ is the orthogonal projection onto H 0 (Mσ , Lkσ )
induced from the L2 inner product on C ∞ (M, Lk ). We get a smooth section of
End(V (k) ) over T

∈ C ∞ (T , End(V (k) ))
(k)
Tf
(k) (k)
by letting Tf (σ) = Tf,σ (see Andersen 2006a). See Section 10.3 for further
discussion of the Toeplitz operators and their connection to deformation quan-
tization.
(k)
The sections Tf of End(V (k) ) over T are not covariant constant with respect
to Hitchin’s connection ∇ˆ e . However, they are asymptotic as k goes to infinity.
The precise meaning of this is stated in Theorem 10.18 of Section 10.5, where
it is proved in the general setting. Theorem 10.19 then follows in a rather
straightforward way from this asymptotic covariant constantness of the Toeplitz
operators and Theorem 10.3 due to Bordeman, Meinrenken, and Schlichenmaier
(1994).
Theorem 10.3 (Bordeman, Meinrenken, and Schlichenmaier 1994)
For any f ∈ C ∞ (M ) we have that
; ;
; (k) ;
lim ;Tf,σ ; = sup |f (x)|.
k→∞ x∈M

(k)
Since the association of the sequence of Toeplitz operators Tf,σ , k ∈ Z+ is
linear in f , we see from this theorem that this association is faithful.

10.1.3 Kashdan’s property (T) and the mapping class group

As an application of TQFT to a problem in geometric group theory, which has
been around for some time (see e.g. problem (7.2) in chapter 7, “A short list
of open questions,” of Bekka, de la Harpe, and Valette 2007), we proved in
(Andersen, 2007) that
Theorem 10.4 The mapping class group of a closed oriented surface of genus
at least 2 does not have Kazhdan’s property (T).
In Kazhdan (1967) the author introduced his property (T) for topological
groups. A topological group has Kazhdan’s property (T) if the trivial represen-
tation is isolated in the Fell topology on the space of unitary representations of
the group. Alternatively, we can formula Kazhdan’s property (T) as follows.
A discrete countable group G has Kazhdan’s property (T) if and only if for
all unitary Hilbert space representations of G, the existence of an almost fixed
vector implies the existence of a non-trivial fixed vector.
 Introduction 183

Definition 10.2 Let ρ : G → U (R) be a unitary Hilbert space representation of

a discrete countable group G on the Hilbert space R. By an almost fixed vector
for ρ we mean a sequence of unit vectors (vk ) ⊂ R with the property that
lim | ρ(g)vk − vk |= 0
k→∞

for all g ∈ G.
The Hilbert space representation of the mapping class group we consider is
the following.
Let
(2,1)
Rk = End0 (Zk (Σ)),
where End0 means the traceless endomorphisms. Since the TQFTs are unitary,
there is a natural unitary structure which is mapping class group invariant on
Rk . Now define
∞
9
R̃ = Rk
k+2 prime

and let R be the Hilbert space completion of R̃.

Theorem 10.5 (Roberts 2001) The only Γ invariant vector in R is 0.
(2,1)
This theorem follows from the fact that the representations Zk (Σ) are
irreducible as Γ-representation for k, such that k + 2 is prime. This result was
established in the un-twisted case by Roberts (2001) and his proof can be applied
word for word also to this case.
The basic idea behind building the required almost fixed vector for R is to
consider coherent states on Mσ , σ ∈ T .
Fix a point x ∈ M . Evaluation of sections at x gives a section of Vk∗ up to
(k)
scale. Using the Hermitian structure ·, · we get induced a section ex of Vk up
(k)
to scale. For each σ ∈ T , ex (σ) is the coherent state associated to x on Mσ .
(k)
Let Ex be the section of End(Vk ) obtained as the orthogonal projection (with
(k)
respect to ·, ·) onto the one-dimensional subspace spanned by ex . We observe
(k)
that Ex only depends on x.
(k)
Theorem 10.6 The sections Ex of End(Vk ) over T are asymptotically covari-
ant constant. That is, for any pair of points σ0 , σ1 ∈ T there exists a constant C
such that
  C
Pσe0 ,σ1 Ex(k) (σ0 ) − Ex(k) (σ1 ) ≤ ,
k
where Pσ0 ,σ1 is the parallel transport from σ0 to σ1 in End(Vk ) and the norm | · |
e

is the one associated to the Hermitian structure on Vk∗ ⊗ Vk induced from ·, ·
on Vk .
184 Toeplitz operators and Hitchin’s projectively flat connection

In fact the proof of this theorem is valid in the general setting discussed in the
following sections. However, the techniques used in Andersen (2007) go beyond
the techniques reviewed in this chapter and we therefore refer to Andersen (2007)
for the proof of this theorem.
To produce the almost fixed vectors, we now pick a finite subgroup Λ of SU (2)
which contains −1 ∈ SU (2) and we consider the finite subset X of M , consisting
of connections which reduce to Λ. For the dihedral group Λ, we get a non-empty
finite subset X of M this way which is invariant under the action of the mapping
(k)
class group. Let now EX be the section of End(Vk ) given by
(k)

EX = Ex(k) .
x∈X

(k) (k)
Let EX,0 be the traceless part of EX . As we have argued in section 7 of Andersen
(k)
(2007), for large enough k, EX = 0. Hence for large enough k we have a unique
(k)
(up to unit scale) unit vector in EX,0 ∈ Rk , which at σ0 agrees projectively with
(k)
EX,0 (σ0 ).
(k)
Theorem 10.7 The sequence EX,0 is an almost fixed vector for the action
of Γ on R.

This theorem is proved in Andersen (2007). Theorem 10.4 is of course a

consequence of Theorems 10.5 and 10.7.

10.1.4 Geometric construction of the curve operators

In the gauge theory picture, the decomposition (10.3) is obtained as follows (see
Andersen 2008b for details).
One considers a one-parameter family of complex structures σt ∈ T , t ∈ R+ ,
such that the corresponding family in the moduli space of curves converges in the
Mumford–Deligne boundary to a nodal curve, which topologically corresponds
to shrinking γ to a point. By the results of Andersen (1998) the corresponding
sequence of complex structures on the moduli space M converges to a non-
negative polarization on M , whose isotropic foliation is spanned by the Hamil-
tonian vector fields associated to the holonomy functions of γ. The main result
of Andersen (2008b) is that the covariant constant sections of V (k) along the
family σt converge to distributions supported on the Bohr–Sommerfeld leaves
of the limiting non-negative polarization as t goes to infinity. The direct sum
of the geometric quantization of the level k Bohr–Sommerfeld levels of this
non-negative polarization is precisely the left-hand side of (10.3). A sewing-
construction inspired from conformal field theory (see Tsuchiya, Ueno, and
Yamada 1989) is then applied to show that the resulting linear map from the
right-hand side of (10.3) to the left-hand side is an isomorphism. This is described
in detail in Andersen (2008b).
 Introduction 185

In Andersen (2008b) we further prove the following important asymptotic

result. Let hγ,λ ∈ C ∞ (M ) be the holonomy function obtained by taking the
trace in the representation λ of the holonomy around γ.
Theorem 10.8 For any one-dimensional oriented submanifold γ and any
labeling λ of the components of γ, we have that
(n,d) (k)
lim Zk (γ, λ) − Thγ,λ  = 0.
k→∞

Let us here give the main idea behind the proof of Theorem 10.8 and refer
to Andersen (2008b) for details. One considers the explicit expression for the
S-matrix, as given in formula (13.8.9) in Kac (1995):
µ̌+ρ̌
Sλ,µ /S0,µ = λ(e−2πi k+n ), (10.4)
where ρ is half of the sum of the positive roots and ν̌ (ν any element of Λ) is
the unique element of the Cartan subalgebra of the Lie algebra of SU (n) which
is dual to ν with respect to the Cartan–Killing form (·, ·).
From (10.4) one sees that under the isomorphism µ̌ → µ, Sλ,µ /S0,µ makes
sense for any µ̌ in the Cartan subalgebra of the Lie algebra of SU (n). Furthermore
one finds that the values of this sequence of functions (depending on k) are
asymptotic to the values of the holonomy function hγ,λ at the level k Bohr–
Sommerfeld sets of the limiting non-negative polarizations discussed above (see
Andersen 1998). From this one can deduce Theorem 10.8 (see again Andersen
2008b for details). Please see Andersen (2005) for the corresponding result in the
abelian case.

10.1.5 The Nielsen–Thurston classification of mapping classes is determined

by TQFT
Let us now review the result from Andersen (2006c), where via the use of curve
operators we show that the Nielsen–Thurston classification of mapping classes is
determined by TQFT.
The Nielsen–Thurston classification of mapping classes of compact oriented
surfaces splits mapping classes into three disjoint types (Thurston 1988) (see
also Fathi, Laudenbach, and Poénaru 1979, and Bleiler and Casson 1988). We
shall only be interested in the closed surface case here, so we state the Nielsen–
Thurston theorem in this case.
Suppose Σ is a closed oriented surface of genus g ≥ 2 and let Γ be the mapping
class group of Σ.
Theorem 10.9 (Nielsen–Thurston) A mapping class φ ∈ Γ has exactly one
of the following three properties:
1. The mapping class φ is, finite-order, that is, φ is a finite-order element in
Γ. This is equivalent to φ having an automorphism of a Riemann surface
as representative.
186 Toeplitz operators and Hitchin’s projectively flat connection

2. The mapping class φ is not finite order, but it is reducible, meaning there
exists a simple closed curve on the surface, whose non-trivial homotopy
class is preserved by some power of φ.
3. The mapping class φ is pseudo-Anosov, meaning that there exists ζ > 1,
two transverse measured foliations F s and F u on Σ and a diffeomorphism
f of Σ, which represents φ, such that
f∗ (F s ) = ζ −1 F s and f∗ (F u ) = ζF u .
In the pseudo-Anosov case, ζ is uniquely determined by φ and it is called the
streching factor for φ.
In the reducible case one continues the analysis of φ, by cutting Σ along
the preserved simple closed curve, to get a mapping class of a surface with
boundary. This mapping class is then classified in terms of the Nielsen–Thurston
classification of mapping classes of surfaces with boundary. The upshot of this is
that there is a diffeomorphism f of Σ, which represents φ and which preserves a
system of simple closed curves, and when one cuts the surface along these curves,
f induces a diffeomorphism of the resulting cut surface. For each component of
the cut surface, there is a smallest power of f , which preserves the component,
and this power of f is a diffeomorphism of the component, which is either finite
order or pseudo-Anosov (see Fathi, Laudenbach, and Poénaru 1979 for further
details regarding this).
The asymptotic faithfulness property gives us immediately the following the-
orem:
Theorem 10.10 For any mapping class φ ∈ Γ we have that there exists an
integer M such that
 M
(n,d)
Zk (φ) ∈ C Id

for all k if and only if φM = 1 (or φ2M = 1, in case (n, d) = (2, 0)).
This separates the finite-order ones from the rest. In order to separate the
reducibles from the pseudo-Anosov ones, we consider the curve operators. Sup-
pose that γ is a simple closed curve on the surface Σ. When we choose an
orientation on γ, we have the holonomy function hγ,
. Note that hγ,
only
depends on the free homotopy class of γ. Further hγ,
is constant if γ is
nul-homotopic.
Theorem 10.11 For any mapping class φ ∈ Γ and any homotopy class γ of a
simple closed curve on Σ we have that φ is reducible along γ, that is,
φ(γ) = γ
if and only if
(n,d) (n,d)
[Zk (φ), Zk (γ,
)] = 0
 Introduction 187

for infinitely many k and any fixed pair (n, d) (and any choice of orientation of
γ if n > 2).
Using these two conditions, we see immediately how to determine the Nielsen–
Thurston classification of a mapping class using TQFT:

1. Given a mapping class φ, we first determine if it is finite order or not by

checking if there exists an N such that ρk (φ)N = 1 for all k.
2. If not, we check through the non-trivial homotopy classes γ of simple closed
(n,d)
curves to see if there exists an integer M such that Zk (φM ) commutes
(n,d)
with Zk (γ,
) for infinitely many k for some pair (n, d). If so, φ is
reducible.
3. If not, then φ is pseudo-Anosov.
(n,d)
Proof. It is clear that φ(γ) = γ implies that Zk (φ) commutes with
(n,d)
Zk (γ,
), since

(φ−1 ) = Zk
(n,d) (n,d) (n,d) (n,d)
Zk (φ)Zk (γ,
)Zk (φ(γ),
). (10.5)
Conversely, suppose that
 
(n,d) (n,d)
Zk (φ), Zk (γ,
) = 0.

Then
(n,d) (n,d)
Zk (γ,
) = Zk (φ(γ),
)
by (10.5). From Theorem 10.8 we conclude that
; ;
; (k) ;
lim ;Thγ,
− Zk (φ)Thγ,
Zk (φ−1 ); = 0.
(n,d) (k) (n,d)
k→∞

But then from Proposition 10.5 below and Theorem 10.3, we get that hγ,
=
hφ(γ),
. We now get that φ(γ) = γ by proposition 1 in Andersen (2006c). The
main idea behind proposition 1 from Andersen (2006c) is that the holonomy
functions extend to holomorphic functions on the SL(n, C)-moduli space M.
The restriction of the holonomy functions from M to the real slice M ⊂ M is
injective. Moreover if two holonomy functions for two simple closed curves agree
on M for some d, then they also agree on M for d = 0. But this space contains a
copy of the Teichmüller space of Σ where it is clear that the holonomy function
of a simple closed curve determines the curve up to homotopy on Σ. We refer to
Andersen (2006c) for the full details of this argument.

We consider it an interesting problem to provide a TQFT formula for the
streching factor ζ of any pseudo-Anosov mapping class (see Andersen, Masbaum,
and Ueno 2006 for the first steps in this direction). We expect that the asymptotic
(n,d)
expansions of the operators Zk (γ,
) should be related to work of Andersen,
Mattes and Reshetikhin (1996, 1998)
188 Toeplitz operators and Hitchin’s projectively flat connection

10.1.6 General setting

Throughout the rest of this chapter, we consider the general setting treated
in Andersen (2006b). That is, we consider a general prequantizable symplec-
tic manifold (M, ω) with a prequantum line bundle (L, (·, ·), ∇) (as opposed
to only considering the moduli spaces). We assume that T is a complex
manifold which holomorphically and rigidly (see Definition 10.4) parameter-
izes Kähler structures on (M, ω). Under a mild cohomological condition we
then establish the existence of the Hitchin connection in this setting as
stated in Theorem 10.12 in Section 10.2. In Section 10.3 we recall well-
known results about the relation between Toeplitz operators and deforma-
tion quantization. In Section 10.4 we define and construct the formal Hitchin
connection as stated in Theorem 10.16. We further discuss how this allows
us under certain cohomological conditions to construct symmetry invariant
star-products (Theorem 10.17). In Section 10.5, we apply these results to
prove Theorem 10.18, which gives the asymptotic flatness of the Toeplitz
operators with respect to the Hitchin connection. We have here given a
somewhat simpler proof of this theorem, than the one given in Andersen
(2006a). The proof makes use of our notion of formal trivializations of the
formal Hitchin connection. In Section 10.6 we establish a generalized asymp-
totic faithfulness, Theorem 10.19, and a generalized condition for reducibility,
Theorem 10.20.
These general results form the analytic basis for the TQFT applications
discussed above.
Finally, we would like to add that we have several other applications of
the theory of Toeplitz operators to the study of the Hitchin connection and
TQFT. In particular the Toeplitz operator theory gives the complete asymptotic
expansion of the quantum invariant of closed three-manifolds. As a corollary of
this we get that the colored Jones polynomials of a knot is an unknot detector.
Moreover we also get the corollary that if a closed three-manifold X satisfies
that
(n) (n)
Zk (X) = Zk (S 3 )
then the only group homomorphism from π1 (X) to SU (n) is the trivial
one. By Perelman’s work we know that π1 (X) is residually finite, thus it
follows that X is simply connected and therefore by Perelman’s proof of
the Poincaré conjecture, that X = S 3 . The writing up of these results is in
progress.

10.2 The Hitchin connection

In this section we review our construction of the Hitchin connection using
the global differential geometric setting of Andersen (2006b). This approach is
close in spirit to Axelrod, Della Pietra, and Witten (1991), however we do not
use any infinite-dimensional gauge theory. In fact the setting is more general
 The Hitchin connection 189

than the gauge theory setting in which Hitchin (1990) constructed his original
connection. But when applied to the gauge theory situation we get the corollary
that Hitchin’s connection agrees with Axelrod, Della Pietra, and Witten’s.
Hence, we start in the general setting and let (M, ω) be any compact symplectic
manifold.
Definition 10.3 A prequantum line bundle (L, (·, ·), ∇) over the symplectic
manifold (M, ω) consists of a complex line bundle L with a Hermitian structure
(·, ·) and a compatible connection ∇ whose curvature is
i
F∇ = ω.
2π
We say that the symplectic manifold (M, ω) is prequantizable if there exists a
prequantum line bundle over it.
The condition on the curvature is to be interpreted in terms of the induced
principal U (1)-connection in the circle bundle of L in the above definition. In
terms of covariant derivatives, this means
ω(X, Y ) = [∇X , ∇Y ] − ∇[X,Y ] .
Recall that the condition for the existence of a prequantum line bundle
is that [ω] ∈ Im(H 2 (M, Z) → H 2 (M, R)) and that the inequivalent choices of
prequantum line bundles (if they exist) are parametriced by H 1 (M, R) (see e.g.
Woodhouse 1992).
We shall assume that (M, ω) is prequantizable and fix a prequantum line
bundle (L, (·, ·), ∇) over (M, ω).
Assume that T is a smooth manifold which smoothly parametrizes
Kähler structures on (M, ω). This means we have a smooth 3 map I :
T → C ∞ (M, End(T M )) such that (M, ω, Iσ ) is a Kähler manifold for each σ ∈ T .
We will use the notation Mσ for the complex manifold (M, Iσ ). For each σ ∈ T
we use Iσ to split the complexified tangent bundle T MC into the holomorphic
and the anti-holomorphic parts, which we denote
Tσ = E(Iσ , i) = Im(Id −iIσ )
and
T̄σ = E(Iσ , −i) = Im(Id +iIσ ),
respectively.
3 Here a smooth map from T to C ∞ (M, W ) for any smooth vector bundle W over M means a
smooth section of πM ∗ (W ) over T × M , where π
M is the projection onto M . Likewise a smooth
p-form on T with values in C ∞ (M, W ) is by definition a smooth section of πT∗ Λp (T ) ⊗ πM∗ (W )

over T × M . We will also encounter the situation where we have a bundle W̃ over T × M and
then we will talk about a smooth p-form on T with values in C ∞ (M, W̃σ ) and mean a smooth
section of πT∗ Λp (T ) ⊗ W̃ over T × M .
190 Toeplitz operators and Hitchin’s projectively flat connection

The real Kähler metric gσ on (Mσ , ω) extended complex linear to T MC is by

definition
gσ (X, Y ) = ω(X, Iσ Y ),
where X, Y ∈ C ∞ (M, T M ⊗ C).
Suppose V is a vector field on T . Then we can differentiate I along V
and we denote this derivative V [I] : T → C ∞ (M, End(T MC )). Differentiating the
equation I 2 = − Id, we see that V [I] anti-commutes with I. Hence we get that
    
V [I]σ ∈ C ∞ M, Tσ∗ ⊗ T̄σ ⊕ T̄σ∗ ⊗ Tσ
for each σ ∈ T . Let
V [I]σ = V [I]σ + V [I]σ
 
be the corresponding
 decomposition
 such that V [I]σ ∈ C ∞ M, Tσ∗ ⊗ T̄σ and
V [I]σ ∈ C ∞ M, T̄σ∗ ⊗ Tσ .
Now we will further assume that T is a complex manifold and that I is a
holomorphic map from T to the space of all complex structures on M . Concretely,
this means that
V  [I]σ = V [I]σ
and
V  [I]σ = V [I]σ
for all σ ∈ T , where V  means the (1, 0) part of V and V  means the (0, 1) part
of V over T .
Let us now define G̃(V ) ∈ C ∞ (M, T MC ⊗ T MC ) by
V [I] = G̃(V )ω,
and define G(V ) ∈ C ∞ (M, Tσ ⊗ Tσ ) such that
G̃(V ) = G(V ) + G(V )
for all real vector fields V on T . We see that G̃ and G are one-forms on T with
values in C ∞ (M, T MC ⊗ T MC ) and C ∞ (M, Tσ ⊗ Tσ ), respectively. We observe
that
V  [I] = G(V )ω,
and G(V ) = G(V  ).
It is easy to check that G̃ takes values in C ∞ (M, S 2 (T MC )) and therefore that
G takes values in C ∞ (M, S 2 (Tσ )).
On Lk we have the smooth family of ∂¯ operators ∇0,1 defined at a σ ∈ T by
1
∇0,1
σ = (1 + iIσ )∇.
2
 The Hitchin connection 191

For every σ ∈ T we consider the finite-dimensional subspace of C ∞ (M, Lk )

given by
 
Hσ(k) = H 0 (Mσ , Lk ) = s ∈ C ∞ (M, Lk )|∇0,1
σ s=0 .

ˆ t denote the trivial connection in the trivial bundle T × C ∞ (M, Lk ). Let

Let ∇
D(M, Lk ) denote the vector space of differential operators on C ∞ (M, Lk ). For
any smooth one-form u on T with values in D(M, Lk ) we have a connection ∇ ˆ
in H given by
(k)

ˆV = ∇
∇ ˆ tV − u(V )

for any vector field V on T .

Lemma 10.1 The connection ∇ ˆ in H(k) preserves the subspaces Hσ(k) ⊂

∞
C (M, L ), for all σ ∈ T if and only if
k

i
V [I]∇1,0 s + ∇0,1 u(V )s = 0 (10.6)
2
for all vector fields V on T and all smooth sections s of H (k) .
This result is not surprising (see Andersen 2006b for a proof this lemma).
(k)
Observe that if this condition holds, we can conclude that the subspaces Hσ ⊂
C ∞ (M, Lk ), for all σ ∈ T , form a subbundle H (k) of H(k) .
We observe that
V  [I]∇1,0 s = 0,
so u(V  ) = 0 solves (10.6) along the anti-holomorphic directions on T . In other
words the (0, 1) part of the trivial connection ∇ ˆ t induces a ∂¯ operator on H (k)
and hence makes it a holomorphic vector bundle over T .
This is of course not in general the situation in the (1, 0) direction. Let us now
consider a particular u and prove that it solves (10.6) under certain conditions.
On the Kähler manifold (Mσ , ω) we have the Kähler metric and we have
the Levi-Civita connection ∇ in Tσ . We also have the Ricci potential Fσ ∈
C0∞ (M, R). Here
<  =
C0∞ (M, R) = f ∈ C ∞ (M, R) | f ωm = 0
M

and the Ricci potential is the element of Fσ ∈ C0∞ (M, R) which satisfies
1 ¯
Ricσ = RicH
σ + i∂σ ∂σ Fσ ,
2
where Ricσ ∈ Ω1,1 (Mσ ) is the Ricci form and RicHσ is its harmonic part. We see
that we get this way a smooth function F : T → C0∞ (M, R).
For any G ∈ C ∞ (M, S 2 (Tσ )) we get a linear bundle map
G : Tσ∗ → Tσ
192 Toeplitz operators and Hitchin’s projectively flat connection

and we have the operator

∇1,0   G⊗Id
∆G : C ∞ (M, Lk ) →
σ
C ∞ M, Tσ∗ ⊗ Lk → C ∞ (M, Tσ ⊗ Lk )
∇1,0
σ ⊗Id + Id ⊗∇σ
1,0   Tr
→ C ∞ M, Tσ∗ ⊗ Tσ ⊗ Lk → C ∞ (M, Lk ).
For any smooth function f on M , we get a vector field
Gdf ∈ C ∞ (M, Tσ ).
Implicit in this definition is the projection from T M ∼
= Tσ ⊕ T̄σ to Tσ , which
takes df to ∂σ f .
Putting these constructions together we consider the following operator for
some n ∈ Z such that 2k + n = 0:
1
u(V ) = o(V ) + V  [F ] (10.7)
2k + n
where
1
∆G(V ) − ∇G(V )dF − nV  [F ].
o(V ) = (10.8)
2
ˆ and we call it the Hitchin
The connection associated to this u is denoted ∇
connection in H .
(k)

Definition 10.4 We say that the complex family I of Kähler structures on

(M, ω) is rigid if
∂¯σ (G(V )σ ) = 0
for all vector fields V on T and all points σ ∈ T .
We will assume our holomorphic family I is rigid.
Theorem 10.12 Suppose that I is a rigid family of Kähler structures on the
compact symplectic prequantizable manifold (M, ω), which satisfies that there
exists n ∈ Z such that the first Chern class of (M, ω) is n[ω] ∈ H 2 (M, Z) and
H 1 (M, R) = 0. Then u given by (10.7) and (10.8) satisfies (10.6) for all k such
that 2k + n = 0.
Hence the Hitchin connection ∇ˆ preserves the subbundle H (k) under the stated
conditions.
Theorem 10.12 is established in Andersen (2006b) through the following three
lemmas.
Lemma 10.2 Assume that the first Chern class of (M, ω) is n[ω] ∈ H 2 (M, Z).
For any σ ∈ T and for any G ∈ H 0 (Mσ , S 2 (Tσ )) we have the following formula:
∇0,1
σ (∆G (s) − 2∇GdFσ (s)) = −i(2k + n)Gω∇σ (s)
1,0

 
−ik Tr −2G∂σ F ω + ∇1,0
σ (G)ω s,

for all s ∈ H 0 (Mσ , Lk ).

 The Hitchin connection 193

Lemma 10.3 We have the following relation

4i∂¯σ (V  [F ]σ ) = Tr(2G(V )∂(F )ω − ∇1,0 (G(V ))ω)σ
provided H 1 (M, R) = 0.
Lemma 10.4 For any smooth vector field V on T we have that
1
(V  [Ric])1,1 = − ∂ Tr(∇1,0 (G(V ))ω). (10.9)
2
Let us here recall how Lemma 10.3 is derived from Lemmas 10.2 and 10.4. By
the definition of the Ricci potential,
¯
Ric = RicH +2id∂F
where RicH = nω by the assumption. Hence
1 ¯ 
V  [Ric] = −dV  [I]dF + id∂V [F ]
2
and therefore
¯  [F ] = 2(V  [Ric])1,1 + 2∂V  [I]∂F.
i∂ ∂V
From the above we conclude that
Tr(2G(V )∂F ω − ∇1,0 (G(V ))ω)σ − i∂¯σ V  [F ]σ ∈ Ω0,1
σ (M )

is ∂σ -closed. By Lemma 10.2 it is also ∂¯σ -closed, hence it is a closed one-form on

M . But since we assume that H 1 (M, R) = 0, we see it is exact, but then it in
fact vanishes since it is also of type (0, 1) on Mσ .
From the above we conclude that
< =
1 1
u(V ) = ∆G(V ) − ∇G(V )dF + 2kV  [F ]
2k + n 2
solves (10.6) and hence we have established Theorem 10.12.
We arrive at the following theorem:
Theorem 10.13 Suppose that I is a rigid family of Kähler structures on the
symplectic prequantizable compact manifold (M, ω), which satisfies that there
exists n ∈ Z such that the first Chern class of (M, ω) is n[ω] ∈ H 2 (M, Z) and
ˆ in the bundle H(k) preserves the
H 1 (M, R) = 0. Then the Hitchin connection ∇
(k)
subbundle H . It is given by
ˆ tV − u(V )
ˆV = ∇
∇
where ∇ ˆ t is the trivial connection in H(k) , V is any smooth vector field on
T , and u(V ) is the second-order differential operator given by (10.7) and
(10.8).
In Andersen, Gammelgaard, and Lauridsen (2008) we use half-forms and the
metaplectic correction to prove the existence of a Hitchin connection in the
context of half-form quantization. The assumption that the first Chern class of
194 Toeplitz operators and Hitchin’s projectively flat connection

(M, ω) is n[ω] ∈ H 2 (M, Z) is then just replaced by the vanishing of the second
Stiefel–Whitney class of M (see Andersen, Gammelgaard, and Lauridsen 2008
for more details).
In Andersen (2006b) we prove the following lemma:
Lemma 10.5 There exist smooth one-forms Xr , Z and functions Yr , r =
1, . . . , R, on T with values in C ∞ (M, Tσ ) such that

1  R
∆G(V ) − ∇G(V )dF = ∇Xr (V ) ∇Yr + ∇Z(V ) (10.10)
2 r=1

for all vector fields V on T .

This gives us the expression
 R 
1 
u(V ) = ∇Xr (V ) ∇Yr + ∇Z(V ) − nV  [F ] + V  [F ]. (10.11)
2k + n r=1

Suppose Γ is a group which acts by bundle automorphisms of L over M

preserving both the Hermitian structure and the connection in L. Then there
is an induced action of Γ on (M, ω). We will further assume that Γ acts on
T and that I is Γ-equivariant. In this case we immediately get the following
invariance:
Lemma 10.6 The natural induced action of Γ on H(k) preserves the subbundle
H (k) and the Hitchin connection.
We are actually interested in the induced connection ∇ˆ e in the endomorphism
(k) (k)
bundle End(H ). Suppose Φ is a section of End(H ). Then for all sections s
of H (k) and all vector fields V on T , we have that
ˆ eV Φ)(s) = ∇
(∇ ˆ V Φ(s) − Φ(∇
ˆ V (s)).

Assume now that we have extended Φ to a section of Hom(H(k) , H (k) ) over T .

Then
∇ ˆ e,t Φ + [Φ, u(V )]
ˆ eV Φ = ∇ (10.12)
V

ˆ e,t is the trivial connection in the trivial bundle End(H(k) ) over T .

where ∇

10.3 Toeplitz operators and Berezin–Toeplitz deformation

quantization
We shall in this section discuss the Toeplitz operators and their asymptotics
as the level k goes to infinity. The properties we need can all be derived from
the fundamental work of Boutet de Monvel and Sjöstrand. In Boutet de Monvel
and Sjöstrand (1976) they did a microlocal analysis of the Szegö projection,
which can be applied to the asymptotic analysis in the situation at hand,
 Toeplitz operators and Berezin–Toeplitz deformation quantization 195

as was done by Boutet de Monvel and Guillemin (1981) (in fact in a much
more general situation than the one we consider here) and others following
them. In particular the applications developed by Schlichenmaier and further by
Karabegov and Schlichenmaier to the study of Toeplitz operators in the geomet-
ric quantization setting is what will interest us here. Let us first describe the basic
setting.
For each f ∈ C ∞ (M ) we consider the prequantum operator, namely, the
differential operator Mf : C ∞ (M, Lk ) → C ∞ (M, Lk ) given by
(k)

(k)
Mf (s) = f s

for all s ∈ H 0 (M, Lk ).

These operators act on C ∞ (M, Lk ) and therefore also on the bundle H(k) ,
however, they do not preserve the subbundle H (k) . In order to turn these
operators into operators which act on H (k) we need to consider the Hilbert
space structure.
Integrating the inner product of two sections against the volume form associ-
ated to the symplectic form gives the pre-Hilbert space structure on C ∞ (M, Lk ):

1
s1 , s2  = (s1 , s2 )ω m .
m! M
We think of this as a pre-Hilbert space structure on the trivial bundle H(k)
which of course is compatible with the trivial connection in this bundle. This
pre-Hilbert space structure induces a Hermitian structure ·, · on the finite rank
subbundle H (k) of H(k) . The Hermitian structure ·, · on H (k) also induces the
operator norm  ·  on End(H (k) ).
Since Hσ is a finite-dimensional subspace of C ∞ (M, Lk ) = Hσ and therefore
(k) (k)
(k) (k) (k)
closed, we have the orthogonal projection πσ : Hσ → Hσ . Since H (k) is a
(k)
smooth subbundle of H(k) the projections πσ form a smooth map π (k) from
T to the space of bounded operators on the L2 completion of C ∞ (M, Lk ). The
easiest way to see this is to consider a local frame for (s1 , . . . , sRank H (k) ) of H (k) .
Let hij = si , sj . Let h−1
ij be the inverse matrix of hij . Then
  
πσ(k) (s) = s, (si )σ  h−1
ij σ (sj )σ . (10.13)
i,j

From these projections we can construct the Toeplitz operators associated to

any smooth function f ∈ C ∞ (M ), Tf,σ : Hσ → Hσ , defined by
(k) (k) (k)

(k)
Tf,σ (s) = πσ(k) (f s)
(k)
for any element s in Hσ and any point σ ∈ T . We observe that the Toeplitz
(k)
operators are smooth sections Tf of the bundle Hom(H(k) , H (k) ) and restrict
to smooth sections of End(H (k) ).
196 Toeplitz operators and Hitchin’s projectively flat connection

Remark 10.1 Similarly for any pseudo-differential operator A on M with

coefficients in Lk (which may even depend on σ ∈ T ), we can consider the asso-
ciated Toeplitz operator π (k) A and think of it as a section of Hom(H(k) , H (k) ).
However, whenever we consider asymptotic expansions of such or consider their
operator norms, we implicitly restrict them to H (k) and consider them as
sections of End(H (k) ) or equivalently assume that they have been precomposed
with π (k) .

Suppose we have a smooth section X ∈ C ∞ (M, Tσ ) of the holomorphic tangent

bundle of Mσ . We then claim that the operator π(k) ∇X is a zero-order Toeplitz
operator. Suppose s1 ∈ C ∞ (M, Lk ) and s2 ∈ H 0 (Mσ , Lk ), then we have that

X(s1 , s2 ) = (∇X s1 , s2 ).

Now, calculating the Lie derivative along X of (s1 , s2 )ω m and using the above,
one obtains after integration that

∇X s1 , s2  = −Λd(iX ω)s1 , s2 ,

where Λ denotes contraction with ω. Thus

(k)
π (k) ∇X = TfX , (10.14)

as operators from C ∞ (N, Lk ) to H 0 (N, Lk ), where fX = −Λd(iX ω).

Iterating (10.14), we find for all X1 , X2 ∈ C ∞ (M, Tσ ) that
(k)
π (k) ∇X1 ∇X2 = TfX f −X2 (fX1 ) (10.15)
2 X1

again as operators from C ∞ (M, Lk ) to H 0 (Mσ , Lk ).

For X ∈ C ∞ (M, Tσ ), the complex conjugate vector field X̄ ∈ C ∞ (M, T̄σ ) is a
section of the anti-holomorphic tangent bundle, and for s1 , s2 ∈ C ∞ (M, Lk ), we
have that

X̄(s1 , s2 ) = (∇X̄ s1 , s2 ) + (s1 , ∇X s2 ).

Computing the Lie derivative along X̄ of (s1 , s2 )ω m and integrating, we get that

∇X̄ s1 , s2  + (∇X )∗ s1 , s2  = Λd(iX̄ ω)s1 , s2 .

Hence we see that

(∇X )∗ = − (∇X̄ + fX̄ )

as operators on C ∞ (M, Lk ). In particular, we see that

π (k) (∇X )∗ π (k) = −TfX̄ |H 0 (Mσ ,Lk ) : H 0 (Mσ , Lk ) → H 0 (Mσ , Lk ).

(k)
(10.16)
 Toeplitz operators and Berezin–Toeplitz deformation quantization 197

For two smooth sections X1 , X2 of the holomorphic tangent bundle Tσ and a

smooth function h ∈ C ∞ (M ), we deduce from the formula for (∇X )∗ that
π (k) (∇X1 )∗ (∇X2 )∗ hπ (k) = π (k) X̄1 X̄2 (h)π (k) (10.17)
(k) (k)
+π fX̄1 X̄2 (h)π + πfX̄2 X̄1 (h)π
+π (k) X̄1 (fX̄2 )hπ + πfX̄1 fX̄2 hπ (k)
as operators on H 0 (Mσ , Lk ).
The product of two Toeplitz operators associated to two smooth functions will
in general not be the Toeplitz operator associated to a smooth function again,
but there is an asymptotic expansion of the product in terms of such Toeplitz
operators on a compact Kähler manifold by the results of Schlichenmaier (1998,
2000 and 2001).
Theorem 10.14 (Schlichenmaier) For any pair of smooth functions f1 , f2 ∈
C ∞ (M ), we have an asymptotic expansion
∞

k −l ,
(k) (k) (k)
Tf1 ,σ Tf2 ,σ ∼ T (l)
cσ (f1 ,f2 ),σ
l=0
(l) ∞
where ∈ C (M ) are uniquely determined since ∼ means the follow-
cσ (f1 , f2 )
ing: For all L ∈ Z+ we have that

L
k −l  = O(k−(L+1) )
(k) (k) (k)
Tf1 ,σ Tf2 ,σ − T (l)
cσ (f1 ,f2 ),σ
l=0
(0)
uniformly over compact subsets of T . Moreover, cσ (f1 , f2 ) = f1 f2 .
(l)
Remark 10.2 It will be useful for us to define new coefficients c̃σ (f, g) ∈
C ∞ (M ) which correspond to the expansion of the product in 1/(2k + n) (where
n is some fixed integer):
∞

(2k + n)−l .
(k) (k) (k)
Tf1 ,σ Tf2 ,σ ∼ T (l)
c̃σ (f1 ,f2 ),σ
l=0

Theorem 10.14 is proved in Schlichenmaier (1998, 2000), where it is also proved

(l)
that the formal generating series for the cσ (f1 , f2 )’s gives a formal deformation
4
quantization of the symplectic manifold (M, ω).
We recall the definition of a formal deformation quantization. Introduce the
space of formal functions Ch∞ (M ) = C ∞ (M )[[h]] as the space of formal power
series in the variable h with coefficients in C ∞ (M ). Let Ch = C[[h]].
Definition 10.5 A deformation quantization of (M, ω) is an associative prod-
uct ∗ on Ch∞ (M ) which respects the Ch -module structure. It is determined by a
4 We have the opposite sign convention on the curvature, which means our c are (−1)l c in
l l
Schlichenmaier (1998, 2001).
198 Toeplitz operators and Hitchin’s projectively flat connection

sequence of bilinear operators

c(l) : C ∞ (M ) ⊗ C ∞ (M ) → C ∞ (M )
defined by
∞

f g = c(l) (f, g)hl ,
l=0

where f, g ∈ C ∞ (M ). The deformation quantization is said to be differential if

the operators c(l) are bidifferential operators. Considering the symplectic action
of Γ on (M, ω), we say that a ∗-product ∗ is Γ-invariant if
γ ∗ (f ∗ g) = γ ∗ (f ) ∗ γ ∗ (g)
for all f, g ∈ C ∞ (M ) and all γ ∈ Γ.
Theorem 10.15 (Karabegov and Schlichenmaier 2001) The product BT
σ
given by
∞

f BT
σ g = (−1)l c(l) l
σ (f, g)h ,
l=0

where f, g ∈ C ∞ (M ) and cσ (f, g) are determined by Theorem 10.14, is a differ-

(l)

entiable deformation quantization of (M, ω).

Definition 10.6 The Berezin–Toeplitz deformation quantization of the com-
pact Kähler manifold (Mσ , ω) is the product BT
σ .

Remark 10.3 Let Γσ be the σ-stabilizer subgroup of Γ. For any element γ ∈

Γσ , we have that
 
γ ∗ Tf,σ = Tγ ∗ f,σ .
(k) (k)

This implies the invariance of BT

σ under the σ-stabilizer Γσ .

Remark 10.4 We define a new ∗-product by

∞

f ˜BT
σ g = (−1)l c̃(l) l
σ (f, g)h .
l=0

Then
 −1 −1

σ g = (f ◦ φ
f ˜BT σ (g ◦ φ
) BT ) ◦φ
for all f, g ∈ Ch∞ (M ), where φ(h) = h
2+nh .

10.4 The formal Hitchin connection

In this section we review the construction of the formal Hitchin connection as
defined and constructed in Andersen (2006b).
 The formal Hitchin connection 199

We assume the conditions on (M, ω) and I of Theorem 10.12, thus providing

us with a Hitchin connection ∇ˆ in H (k) over T and the associated connection
ˆ in End(H ). Let D(M ) be the space of smooth differential operators on M
∇ e (k)

acting on smooth functions on M . Let Ch be the trivial Ch∞ (M )-bundle over T .

Definition 10.7 A formal connection D is a connection in Ch over T of the
form
DV f = V [f ] + D̃(V )(f ),
where D̃ is a smooth one-form on T with values in Dh (M ) = D(M )[[h]], f is
any smooth section of Ch , V is any smooth vector field on T , and V [f ] is the
derivative of f in the direction of V .
For a formal connection we get the series of differential operators D̃(l) given
by
∞

D̃(V ) = D̃(l) (V )hl .
l=0

From Hitchin’s connection in H (k) we get an induced connection ∇ ˆ e in the

endomorphism bundle End(H (k) ). The Teoplitz operators are not covariant
ˆ e . They are so asymptotically in k in the
constant sections with respect to ∇
following very precise sense.
Theorem 10.16 There is a unique formal Hitchin connection D which satisfies
that
ˆ eV T (k) ∼ T (k)
∇ (10.18)
f (DV f )(1/(2k+n))

for all smooth section f of Ch and all smooth vector fields on T . Moreover
D̃ = 0 mod h.
Here ∼ means the following: For all L ∈ Z+ we have that
;  ;
; L ;
; ˆ e (k) 1 ;
; = O(k −(L+1) )
(k) (k)
;∇V Tf − TV [f ] + T (l)
; D̃V f (2k + n)l ;
l=1

uniformly over compact subsets of T for all smooth maps f : T → C ∞ (M ).

We call this formal connection D the formal Hitchin connection. Since D̃ =
0 mod h we see that if we now fix f ∈ C ∞ (M ) then we get that
; ;
; ˆ e (k) ;
;∇V Tf ; = O(k−1 ).

The proof of this theorem is given in Andersen (2006b). Let us recall formula
(19) from Andersen (2006b):
D̃(V )(f ) = V  [F ]f − V  [F ]˜BT f + h(E(V )(f ) − H(V )˜
BT f ) (10.19)
200 Toeplitz operators and Hitchin’s projectively flat connection

where E is the one-form on T with values in D(M ) such that

TE(V )f = π (k) o(V )∗ f π (k) + π (k) f o(V )π (k)

(k)

and H is the one-form on T with values in C ∞ (M ) such that H(V ) = E(V )(1).
Lemma 10.7 σ for each σ ∈ T ,
The formal operator DV is a derivation for BT
that is,
DV (f ˜BT g) = DV (f )˜BT g + f ˜
BT DV (g)
for all f, g ∈ C ∞ (M ).
Again this lemma is proved in Andersen (2006b), and in fact it follows basically
from
     
ˆ eV T (k) Tg(k) = ∇
∇ ˆ eV T (k) Tg(k) + T (k) ∇
ˆ eV Tg(k) .
f f f

If the Hitchin connection is projectively flat, then the induced connection in

the endomorphism bundle is flat and hence so is the formal Hitchin connection
by proposition 3 of Andersen (2006b).
Definition 10.8 A formal trivialization of a formal connection D is a smooth
map P : T → Dh (M ) which modulo h is an isomorphism for all σ ∈ T and such
that
DV (P (f )) = 0
for all vector fields V on T and all f ∈ Ch∞ (M ).
Clearly if D is not flat, such a formal trivialization will not exist even locally
on T . However, if D is flat, then we have the following result.
Proposition 10.1 Assume that D is flat and that D̃ = 0 mod h. Then locally
around any point in T there exists a formal trivialization. If H 1 (T , R) =
0 then there exist a formal trivialization defined globally on T . If further
HΓ1 (T , D(M )) = 0 then we can construct P such that it is Γ-equivariant.
In this proposition HΓ1 (T , D(M )) simply refers to the Γ-equivariant first de
Rham cohomology of T with coefficients in the real Γ-vector space D(M ).
Now suppose we have a formal trivialization P of the formal Hitchin connection
D. Then P is constant mod h and we may and will assume that
P = Id mod h.
We can then define a new smooth family of -products parametrized by T as
follows
f σ g = Pσ−1 (Pσ (f )˜BT
σ Pσ (g))

for all f, g ∈ C ∞ (M ) and all σ ∈ T .

 Asymptotic flatness of Toeplitz operators 201

Proposition 10.2 The -products σ are independent of σ ∈ T .

Theorem 10.17 Assume that the formal Hitchin connection D is flat and

HΓ1 (T , D(M )) = 0,

then there is a Γ-invariant trivialization P of D and the ∗-product

f  g = Pσ−1 (Pσ (f ) BT

σ Pσ (g))

is independent of σ ∈ T and Γ-invariant. If HΓ1 (T , C ∞ (M )) = 0 and the com-

mutant of Γ in D(M ) is trivial, then a Γ-invariant differential ∗-product on M
is unique.

This theorem is proved in Andersen (2006a).

Let P (r) : T → D(M )[r] be defined by

P ≡ P (r) mod hr+1 .

Proposition 10.3 For any f ∈ C ∞ (M ) and any vector field V on T with

compact support we have that

 ∇V TP (r) (f ) 1 = O(k −r−1 ).

(k)
(k)
Proof. This follows directly from the definition of a formal trivialization
combined with Theorem 10.16.

10.5 Asymptotic flatness of Toeplitz operators

In this section we will establish that the Toeplitz operators associated to
functions on M are asymptotically (in k) covariant constant sections of
ˆ e ).
(End(V (k) ), ∇
The setting is as in Section 10.2, that means we are under the assumptions of
Theorem 10.13. Then we have the following theorem:

Theorem 10.18 Let σ0 and σ1 be two points in T and Pσ0 ,σ1 be the parallel
ˆ e ) from σ0 to σ1 with respect to ∇
transport in the bundle (End(V (k) ), ∇ ˆ e . Then

Pσ0 ,σ1 Tf,σ0 − Tf,σ1  = O(k −1 ),

(k) (k)

 
where  ·  is the operator norm on H 0 Mσ1 , Lkσ1 .

As preparation for the proof of this theorem, we need to discuss a certain

Hermitian structure on H (k) over T , which is asymptotically flat with respect to
Hitchin’s connection. Namely, we let

1
s1 , s2 F = (s1 , s2 )e−F ω m . (10.20)
m! M
202 Toeplitz operators and Hitchin’s projectively flat connection

Along any smooth one-parameter curve σt in T , we have that

> ?
d ˆ t  s2 F − ∂F s1 , s2
ˆ t  s1 , s2 F + s1 , ∇
s1 , s2 F = ∇ .
σt σt
dt ∂t F

Let
d 2 @ˆ A @ A
ˆ σ s .
E(s) = |s|F − ∇σt s, s − s, ∇
dt F t
F

Recall that ∇ˆv = ∇ˆ tv − u(v) and u(v) = 1 o(v) + v  [F ], hence we have for all
2k+n
sections s of Vk that
1  −F 
E(s) = πe o (σt ) s, s + s, πe−F o (σt ) s .
2k + n
Hence by combining Theorem 10.3, (10.10), (10.14), and (10.15) we have proved
that
Proposition 10.4 The Hermitian structure (10.20) is asymptotically flat with
ˆ that is, for any compact subset K of T , there exists
respect to the connections ∇,
a constant C such that for all sections s of Vk over K, we have that
C
|E(s)| ≤ |s|2
k+n F
over K.
We note that this proposition implies the same proposition for sections of
End(Vk ) with respect to the induced Hermitian structure on End(Vk ) = Vk∗ ⊗ Vk ,
which we also denote ·, ·F . We denote the analogous quantity to E for the
endomorphism bundle by Ee .
Lemma 10.8 The Hermitian structure on Hk

1
s1 , s2 F = (s1 , s2 )e−F ω m
m! M
and the constant L2 -Hermitian structure on Hk

1
s1 , s2  = (s1 , s2 )ω m
m! M
are equivalent uniformly in k when restricted to Vk over any compact subset K
of T .
The proof of Lemma 10.8 can be found in section 5 of Andersen (2006a).
Proof. (Of Theorem 10.12). Let σt , t ∈ J, be a smooth one-parameter family
of complex structures such that σt is a curve in T between the two points in
question. By Lemma 10.8 the Hermitian structures ·, · and ·, ·F on Vk are
equivalent uniformly in k over compact subsets of T . It follows that there exists
 Asymptotic flatness of Toeplitz operators 203

constants C1 and C2 , such that we have the following inequalities over the image
of σt in T for the operator norm  ·  and the norm | · |F on End(Vk ):
B
 ·  ≤ C1 | · |F ≤ C2 Pg,n (k) · , (10.21)

where Pg,n (k) is the rank of Vk given by the Verlinde formula. By the Riemann–
Roch theorem this is a polynomial in k of degree m.
Because of these  inequalities, we choose an integer r bigger than m/2 and
let fr = P (r) (f ) k1 where P is a trivialization of the formal Hitchin connection
along σt , t ∈ J, which clearly exists and is unique under the condition that
Pσ = Id by Proposition 10.1. Then let nk : J →[0, ∞) be given by
nk (t) = |Θk (t)|2F
where
Θk (t) : Vk,σt → Vk,σt
is given by
(k) (k)
Θk (t) = Pσ0 ,σt T(fr )0 ,σ0 − T(fr )t ,σt .

The functions nk are differentiable in t and we compute that

dnk ˆ e  (Θk (t)), Θk (t)F + Θk (t), ∇
ˆ e  (Θk (t))F + Ee (Θk (t))
= ∇ σt σt
dt
ˆ e  T (k)
= −∇ ˆ e (k)
σt (fr )t ,σt , Θk (t)F − Θk (t), ∇σt T(fr )t ,σt F + Ee (Θk (t)).

Using the above, we get the following estimate:

dnk ˆ e  T (k)
≤ 2|∇ σt (fr )t ,σt |F |Θk (t)|F + |Ee (Θk (t))|
dt
B
≤ 2C2 Pg,n (k)∇ ˆ e  T (k) 1/2
σt (fr )t ,σt nk + |Ee (Θk (t))|.

Consequently we can apply Propositions 10.3 and 10.4 to obtain that there exists
a constant C such that
dnk C 1/2
≤ (nk + nk ).
dt k
This estimate implies that


 2
Ct
nk (t) ≤ exp −1 .
2k
But by (10.21) we get that
; ;
; (k) (k) ;
;Pσ0 ,σ1 T(fr )0 ,σ0 − T(fr )1 ,σ1 ; = Θk (1) ≤ C1 nk (1)1/2 .
204 Toeplitz operators and Hitchin’s projectively flat connection

The theorem then follows from these two estimates, since (fr )0 = f and
; ;
; (k) (k) ;
;T(fr )1 ,σ1 − Tf,σ1 ; = O(k−1 ).

10.6 General asymptotic faithfulness

In this section we will review how the asymptotics of Toeplitz operators can
be used to prove asymptotic faithfulness of certain representations, as discussed
in the gauge theory setting in Andersen (2006a). However, we shall discuss the
most general setting in which the argument from Andersen (2006a) applies.
In addition to the assumptions of Theorem 10.12, we will further assume,
throughout this section, that the Hitchin connection ∇ ˆ is projectively flat and
that we have a symmetry group acting on our set-up as discussed in Section 10.2.
We thus get an induced flat connection in the bundle P(Vk ). This then gives
us a projective representation of the mapping class group

ρk : Γ → Aut(P(Vk )),

where P(Vk ) denotes the covariant constant sections of P(Vk ) over T .

First we establish the following proposition:

Proposition 10.5 For any φ ∈ Γ, f ∈ C ∞ (M ) and σ ∈ T , we have that

; ; ; ;
; (k) ; ; (k) ;
lim ;Tf,σ − ρk (φ)Tf,σ ρk (φ−1 ); = lim ;T(f −f ◦φ),σ ; .
(k)
k→∞ k→∞

Proof. Suppose we have a φ ∈ Γ. Then φ induces a symplectomorphism of M

which we also just denote φ and we get the following commutative diagram for
any f ∈ C ∞ (M ):
    P  
φ∗ φ(σ),σ
H 0 Mσ , Lkσ −−−−→ H 0 Mφ(σ) , Lkφ(σ) −−−−−→ H 0 Mσ , Lkσ
⏐ ⏐ ⏐
(k) ⏐ ⏐ ⏐P
Tf,σ + Tf ◦φ,φ(σ) + + φ(σ),σ Tf(k)
(k)
◦φ,φ(σ)

  φ∗   Pφ(σ),σ  
H 0 Mσ , Lkσ −−−−→ H 0 Mφ(σ) , Lkφ(σ) −−−−−→ H 0 Mσ , Lkσ ,
   
where Pφ(σ),σ : H 0 Mφ(σ) , Lkφ(σ) → H 0 Mσ , Lkσ on the horizontal arrows
refers to parallel transport in the bundle V (k) , whereas Pφ(σ),σ refers to the
parallel transport in the endomorphism bundle End(V (k) ) in the last vertical
arrow. By the definition of ρk , we see that

ρk (φ)Tf,σ ρk (φ−1 ) = Pφ(σ),σ Tf ◦φ,φ(σ) .

(k) (k)
 General asymptotic faithfulness 205

By Theorem 10.18 we get that

; ; ; ;
; (k) ; ; (k) (k) ;
lim ;T(f −f ◦φ),σ ; = lim ;Tf,σ − Tf ◦φ,σ ;
k→∞ k→∞
; ;
; (k) (k) ;
= lim ;Tf,σ − Pφ(σ),σ Tf ◦φ,φ(σ) ;
k→∞
; ;
; (k) ;
= lim ;Tf,σ − ρk (φ)Tf,σ ρk (φ−1 ); .
(k)
k→∞

From this proposition we get the following generalization of asymptotic faith-

fulness:

Theorem 10.19 For any φ ∈ Γ, we have that

∞
:
φ∈ ker ρk
k=1

if and only if φ induces the identity on M .

C∞
Proof. Suppose φ ∈ k=1 ker ρk , then by Proposition 10.5 and Bordemann,
Meinrenken, and Schlichenmaier’s Theorem 10.3 we see that

f ◦φ=f

for all f ∈ C ∞ (M ), hence we must have that φ acts by the identity on M .

We also get the following generalized reducibility property:

Theorem 10.20 For any φ ∈ Γ, f ∈ C ∞ (M ) and σ ∈ T , we have that

; ;
; (k) ;
lim ;[ρk (φ), Tf,σ ]; = 0
k→∞

if and only if

f ◦ φ = f.

Proof. Let φ ∈ Γ and assume that

; ;
; (k) ;
lim ;[ρk (φ), Thγ ,σ ]; = 0
k→∞

for some f ∈ C ∞ (M ). We have that

; ; ;  ;
; (k) ; ; ;
;Tf,σ − ρk (φ−1 )Tf,σ ρk (φ); = ;ρk (φ−1 )ρk (φ) Tf,σ − ρk (φ−1 )Tf,σ ρk (φ) ;
(k) (k) (k)

; ;;;
 ;
(k) ;
≤ ;ρk (φ−1 ); ; ρk (φ), Tf,σ ; .
206 Toeplitz operators and Hitchin’s projectively flat connection

Lemma 10.8 and Proposition 10.4 give a uniform bound on ρk (φ−1 ). Hence
Proposition 10.5 implies that
; ;
; (k) ;
lim ;T(f −f ◦φ−1 ),σ ; = 0.
k→∞

Then by Bordemann, Meinrenken, and Schlichenmaier’s Theorem 10.3, we must

have that f = f ◦ φ−1 . The converse follows since we have that
; ; ; ;
; (k) ; ; (k) ;
; ρk (φ), Thγ ,σ ; ≤ ρk (φ) ;Tf,σ − ρk (φ−1 )Tf,σ ρk (φ); .
(k)

Acknowledgement
I would like to take this opportunity to thank Professor Nigel Hitchin for
introducing me to the subject at hand and further for the many very helpful
discussions we have had over the years regarding this interesting and facinating
research field of mathematics bordering on Mathematical Physics.

References
Andersen, J. E. (1998). New polarizations on the moduli space and the Thurston
compactification of Teichmuller space. Int. J. Math. 9(1), 1–45.
Andersen, J. E. (2005). Geometric quantization and deformation quantization of abelian
moduli spaces. Commun. Math. Phys. 2005, 727–45.
Andersen, J. E. (2006a). Asymptotic faithfulness of the quantum SU (n) representations
of the mapping class groups. Ann. Math. 163, 347–68.
Andersen, J. E. (2006b). Hitchin’s connection, Toeplitz operators and symmetry invari-
ant deformation quantization. math.DG/0611126.
Andersen, J. E. (2006c). The Nielsen-Thurston classification of mapping classes is
determined by TQFT. math.QA/0605036.
Andersen, J. E. (2007). Mapping class groups do not have Kazhdan’s property (T),
math.QA/0706.2184.
Andersen, J. E. (2008a). Asymptotic faithfulness of the quantum SU (n) representations
of the mapping class groups in the singular case. In preparation.
Andersen, J. E. (2008b). Asymptotic in teichmuller space of the Hitchin connection. In
preparation.
Andersen, J. E. and Christ, M. (2008). Asymptotic expansion of the Szegö kernel on
singular algebraic varieties. In preparation.
Andersen, J. E., Gammelgaard, N. L., and Lauritsen, M. (2008). Hitchin’s connection
in half-form quantization. math.DG/0711.3995.
Andersen, J. E., Masbaum, G., and Ueno, K. (2006). Topological quantum field theory
and the Nielsen-Thurston classification of M (0, 4). Math. Proc. Camb. Philos. Soc.
141(3), 477–88.
 References 207

Andersen, J. E., Mattes, J., and Reshetikhin, N. (1996). The Poisson Structure
on the moduli space of flat connections and chord diagrams. Topology 35,
1069–83.
Andersen, J. E., Mattes, J., and Reshetikhin, N. (1998). Quantization of the algebra
of chord diagrams. Math. Proc. Camb. Phil. Soc. 124, 451–67.
Andersen, J. E. and Ueno, K. (2006). Modular functors are determined by their genus
zero data. math.QA/0611087.
Andersen, J. E. and Ueno, K. (2007a). Abelian conformal field theories and determinant
bundles. Int. J. Math. 18, 919–93.
Andersen, J. E. and Ueno, K. (2007b). Constructing modular functors from conformal
field theories. J. Knot Theory Ramifications 16(2), 127–202.
Andersen, J. E. and Ueno, K. (2008). Construction of the Reshetikhin-Turaev TQFT
from conformal field theory. In preparation.
Atiyah, M. (1990). The Jones-Witten invariants of knots. Astérisque (189–190), Exp.
No. 715, 7–16. Séminaire Bourbaki, Vol. 1989/90.
Axelrod, S., Della Pietra, S., and Witten, E. (1991). Geometric quantization of Chern-
Simons gauge theory. J. Diff. Geom. 33(3), 787–902.
Bakalov, B. and Kirillov, A. (2000). Lectures on Tensor Categories and Modular
Functors, Vol. 21. AMS University Lecture Series, Providence, RI.
Bekka, B., de la Harpe, P., and Valette, A. (2007). Kazhdan’s Proporty (T). Cambridge
University Press. In press.
Blanchet, C. (2000). Hecke algebras, modular categories and 3-manifolds quantum
invariants. Topology 39(1), 193–223.
Blanchet, C., Habegger, N., Masbaum, G., and Vogel, P. (1992). Three-manifold
invariants derived from the Kauffman bracket. Topology 31, 685–99.
Blanchet, C., Habegger, N., Masbaum, G., and Vogel, P. (1995). Topological quantum
field theories derived from the Kauffman bracket. Topology 34, 883–927.
Bleiler, S. and Casson, A. (1988). Automorphisms of Sufaces after Nielsen and
Thurston. Cambridge University Press, Cambridge, UK.
Bordeman, M., Meinrenken, E., and Schlichenmaier, M. (1994). Toeplitz quantiza-
tion of Kähler manifolds and gl(N ), N → ∞ limit. Commun. Math. Phys. 165,
281–96.
Boutet de Monvel, L. and Guillemin, V. (1981). The Spectral Theory of Toeplitz
Operators, Vol. 99 of Ann. Math. Stud. Princeton University Press, Princeton, NJ.
Boutet de Monvel, L. and Sjöstrand, J. (1976). Sur la singularité des noyaux de
Bergmann et de Szegö. Asterique 34–35, 123–64.
Faltings, G. (1993). Stable G-bundles and projective connections. J. Alg. Geom. 2,
507–68.
Fathi, A., Laudenbach, F., and Poénaru, V. (1991/1979). Travaux de Thurston sur les
surfaces, Astérisque 66–67.
Finkelberg, M. (1996). An equivalence of fusion categories. Geom. Funct. Anal. 6,
249–67.
208 Toeplitz operators and Hitchin’s projectively flat connection

Freedman, M. H., Walker, K., and Wang, Z. (2002). Quantum SU(2) faithfully detects
mapping class groups modulo center. Geom. Topol. 6, 523–39.
Hitchin, N. (1990). Flat connections and geometric quantization. Commun. Math. Phys.
131, 347–80.
Kac, V. G. (1995). Infinite Dimensional Lie Algebras, 3 edn. Cambridge University
Press, Cambridge, UK.
Karabegov, A. V. and Schlichenmaier, M. (2001). Identification of Berezin–Toeplitz
deformation quantization. J. Reine Angew. Math. 540, 49–76.
Kazhdan, D. (1967). Connection of the dual space of a group with the structure of its
closed subgroups. Funct. Anal. Appl. 1, 64–65.
Kazhdan, D. and Lusztig, G. (1993). Tensor structures arising from affine Lie algebras.
I, II. J. Am. Math. Soc. 6(4), 905–47, 949–1011.
Kazhdan, D. and Lusztig, G. (1994a). Tensor structures arising from affine Lie algebras.
III. J. Am. Math. Soc. 7(2), 335–81.
Kazhdan, D. and Lusztig, G. (1994b). Tensor structures arising from affine Lie algebras.
IV. J. Am. Math. Soc. 7(2), 383–453.
Laszlo, Y. (1998). Hitchin’s and WZW connections are the same. J. Diff. Geom. 49(3),
547–76.
Moore, G. and Seiberg, N. (1989). Classical and quantum conformal field theory.
Commun. Math. Phys. 123, 177–254.
Reshetikhin, N. and Turaev, V. (1990). Ribbon graphs and their invariants derived fron
quantum groups. Commun. Math. Phys. 127, 1–26.
Reshetikhin, N. and Turaev, V. (1991). Invariants of 3-manifolds via link polynomials
and quantum groups. Invent. Math. 103, 547–97.
Roberts, J. (2001). Irreducibility of some quantum representations of mapping class
groups. J. Knot Theory Ramifications 10, 763–7.
Schlichenmaier, M. (1998). Berezin-Toeplitz quantization and conformal field theory.
PhD thesis.
Schlichenmaier, M. (2000). Deformation quantization of compact Kähler manifolds by
Berezin-Toeplitz quantization. In Conférence Moshé Flato 1999, Vol. II (Dijon), Vol.
22 of Math. Phys. Stud. pp. 289–306. Kluwer Academic Publication, Dordrecht, the
Netherlands.
Schlichenmaier, M. (2001). Berezin-Toeplitz quantization and Berezin transform. In
Long Time Behaviour of Classical and Quantum Systems (Bologna, 1999), Vol. 1 of
Ser. Concr. Appl. Math., pp. 271–87. World Scientific Publication, River Edge, NJ.
Segal, G. (1992). The Definition of Conformal Field Theory. Oxford University
Preprint.
Thurston, W. (1988). On the geometry and dynamics of diffeomorphisms of surfaces.
Bull. Am. Math. Soc. 19, 417–31.
Tsuchiya, A., Ueno, K., and Yamada, Y. (1989). Conformal field theory on uni-
versal family of stable curves with gauge symmetries. Adv. Stud. Pure Math. 19,
459–566.
 References 209

Turaev, V. G. (1994). Quantum Invariants of Knots and 3-Manifolds, Vol. 18 of de

Gruyter Stud. Math. Walter de Gruyter & Co., Berlin, Germany.
Walker, K. (1991). On Witten’s 3-manifold invariants. Preliminary version #2, preprint.
Witten, E. (1989). Quantum field theory and the Jones polynomial. Commun. Math.
Phys. 121, 351–98.
Woodhouse, N. J. (1992). Geometric Quantization. Oxford University Press, Oxford.
 XI
TOWARDS A NON-LINEAR SCHWARZ’S LIST

Philip Boalch
Dedicated to Nigel Hitchin for his 60th birthday

11.1 Introduction
The main theme of this chapter is ‘icosahedral’ solutions of (ordinary) differential
equations, a topic that seems suitable for a 60th birthday conference. We will
however try to go beyond the icosahedron, to see what comes next, and consider
various symmetry groups each of which could be thought of as the next in a
sequence, following the icosahedral group.
To fix ideas let us give a classical example. Recall the icosahedral rotation
group of order 60:

= PSL2 (F5 ) ∼
A5 ∼ = ∆235 ∼
=  a, b, c a2 = b3 = c5 = abc = 1 .

This is described via three generators a, b, and c whose product is the identity,
and so it is natural to look for ordinary differential equations (ODEs) on the
three-punctured sphere P1 (C) \ {0, 1, ∞} with monodromy group A5 . Now A5
is a three-dimensional rotation group so naturally lives in SO3 (R) which is a
subgroup of SO3 (C) which is isomorphic to PSL2 (C). Thus we are led to search
for connections


A1 A2
∇=d− + dz, Ai ∈ sl2 (C) (11.1)
z z−1

on rank 2 holomorphic vector bundles over the three-punctured sphere with

projective monodromy group equal to A5 .
Such connections are essentially the same as Gauss hypergeometric equations,
and H. Schwarz (1873) classified all such equations having finite monodromy
groups. The list he produced has 15 rows, 1 for the family of dihedral groups,
2 rows for each of the tetrahedral and octahedral groups, and 10 rows for the
icosahedral group (see Table 11.1).

Note added in proof: Lisovyy and Tykhyy have recently announced (arXiv:0809.4873) that
the ‘Non-linear Schwarz’s list’ constructed here is in fact complete.
 Introduction 211

Table 11.1. Schwarz’s list:

Inhalt
No. λ

µ

ν

Polyeder
π
1 1
aa∗ I 2 2
ν ν Regelmässige Doppelpyramide
1 1 1 1
abb II 2 3 3 6
=A Tetraeder
2 1 1 1
bbb III 3 3 3 3
= 2A
1 1 1 1
abg IV 2 3 4 12
=B Würfel und Oktaeder
2 1 1 1
bgg V 3 4 4 6
= 2B
1 1 1 1
abc VI 2 3 5 30
= C Dodekaeder und Ikosaeder
2 1 1 1
bbd VII 5 3 3 15
= 2C
2 1 1 1
bcc VIII 3 5 5 15
= 2C
1 2 1 1
acd IX 2 5 5 10
= 3C
3 1 1 2
bcd X 5 3 5 15
= 4C
2 2 2 1
ddd XI 5 5 5 5
= 6C
2 1 1 1
bbc XII 3 3 5 5
= 6C
4 1 1 1
ccc XIII 5 5 5 5
= 6C
1 2 1 7
abd XIV 2 5 3 30
= 7C
3 2 1 1
bdd XV 5 5 3 3
= 10C

Source: From Schwarz (1873).

A key point here is that the Gauss hypergeometric equation is rigid so the full
monodromy representation (of the fundamental group of the three-punctured
sphere into PSL2 (C)) is determined by the conjugacy classes of the monodromy
around each of the punctures. Thus in Schwarz’s list it is sufficient to list these
local monodromy conjugacy classes in order to specify the possible monodromy
representations (and from this it is easy to find a hypergeometric equation with
given monodromy). To ease recognition, to the left of the table we have listed the
triples of conjugacy classes which occur, labelling the four non-trivial conjugacy
classes of A5 by a, b, c, and d, representing rotations by 12 , 13 , 15 , and 25 of a turn,
respectively. (In the octahedral case one may also have rotations by a quarter of
a turn, which we label by g.)

11.1.1 Naive generalizations

Our basic aim is to discuss three naive generalizations of Schwarz’s list, as follows.
The first two arise simply by looking for non-rigid connections that are natural
generalizations of the hypergeometric connections considered above, obtained by
adding an extra singularity – the two cases are generalizations of two ways one
may view the hypergeometric equation as a connection. First of all we can simply
212 Towards a non-linear Schwarz’s list

add another pole at some point t:


A1 A2 A3
(A) ∇=d− + + dz, Ai ∈ sl2 (C)
z z−t z−1
and keep the coefficients in sl2 (C).
Secondly, we recall that the connection one obtains immediately upon choos-
ing a cyclic vector for the hypergeometric equation is as in (11.1) but with
A1 , A2 both rank 1 matrices (in gl2 (C)). Then the monodromy group will be
a complex reflection group (generated by two two-dimensional complex reflec-
tions 1 ) and the natural generalization is then to consider connections of the
form


B1 B2 B3
(B) ∇=d− + + dz, Bi ∈ gl3 (C)
z z−t z−1
with each Bi having rank 1, so the monodromy group will be generated by three
three-dimensional complex reflections. This is a very natural condition as we will
see.
Questions A, B: Find the analogue of Schwarz’s list for connections (A)
or (B).

These questions can now be answered and lead to two ‘non-rigid Schwarz’s
lists’, that is, to classifications of possible monodromy representations with
finite image (up to equivalence) and the construction of connections realizing
such representations. We should emphasize that the main focus has been the
construction of such connections with given monodromy representation for any
value of t (which is a tricky business in this non-rigid case), rather than just the
classification.

Example (of type (B)) . The full symmetry group of the icosahedron is the
icosahedral reflection group of order 120:
D E
H = H3 ∼ = r1 , r2 , r3 ri2 = 1, (r1 r2 )2 = (r2 r3 )3 = (r3 r1 )5 = 1
⊂ O3 (R) ⊂ GL3 (C).
This is generated by three reflections (whose product is not the identity) and
so it is natural to look for connections on rank 3 bundles over a four-punctured
sphere with monodromy H (generated by three reflections about three of the
punctures – that is, connections of the form (B) with each of the three residues
Bi having trace 12 so the corresponding reflections are of order 2). There turn out
to be three inequivalent triples of generating reflections of H, two of which are
related by an outer automorphism. The problem is to write down connections

1 That is, arbitrary automorphisms of the form ‘one plus rank 1’, not necessarily of order 2

or orthogonal.
 Introduction 213

of the desired form for any value t of the final pole position. One triple of
generating reflections is intimately related to K. Saito’s flat structure (1993)
for H (or icosahedral Frobenius manifold) and appears in Dubrovin’s article
(1995, appendix E). The other two triples were dealt with around 1997: see
Dubrovin and Mazzocco (2000); one is similar to the first case (since related to
it by an outer automorphism) but the final triple turned out to be much trickier,
and writing out the family of connections in this case involved a specific elliptic
curve which took about 10 pages of 40 digits integers to write down (see the
preprint version of op. cit. on the mathematics arxiv). We will eventually see
below that this elliptic solution is in fact equivalent to a solution with a simple
parametrization, agreeing with Hitchin’s philosophy that ‘nice problems should
have nice solutions’.

Remark. Before moving on to the third generalization let us add some other
historical comments. The ‘non-naive’ generalizations of the Gauss hypergeomet-
ric equation are the equations satisfied by the n Fn−1 hypergeometric functions
(the Gauss case being that of n = 2). The corresponding Schwarz’s list appears
in Beukers and Heckman (1989). In terms of connections this amounts to consid-
ering connections (11.1) on rank n vector bundles, still with three singularities
on P1 , but with A1 of rank n − 1 and A2 of rank 1; these connections are still
rigid.
Some work in the non-rigid case has been done (besides that we will recall
below) by considering generalizations of the hypergeometric equation as an
equation (rather than as a connection); for example, the algebraic solutions of
the Lamé equation were studied in Beukers and van der Waall (2004) (Lamé
equations are basically the second order Fuchsian equations with four singular
points on P1 such that three of the local monodromies are of order 2). In general
connections of type (A) with such monodromy representations will not come
from a Lamé equation (since upon choosing a cyclic vector the corresponding
equations will in general have additional apparent singularities; this can also be
seen by counting dimensions). Indeed it turns out (op. cit.) that Lamé equations
only have finite monodromy for special configurations of the four poles.

11.1.2 Non-linear analogue: the Painlevé VI equation

One reason hypergeometric equations are interesting is that they provide the
simplest explicit examples of Gauss–Manin connections. Indeed this is one reason
Gauss was interested in them: he observed that the periods of a family of elliptic
curves satisfy a (Gauss) hypergeometric equation. (The modern interpretation
of this is as the explicit form of the natural flat connection on the vector bundle
of first cohomologies over the base of the family of elliptic curves, written with
respect to the basis given by the holomorphic one-forms – and their derivatives –
on the fibres.) Nowadays there is still much interest in such linear differential
equations ‘coming from geometry’.
214 Towards a non-linear Schwarz’s list

Thus the non-linear analogue of the Gauss hypergeometric equation should be

the explicit form of the simplest nonabelian Gauss–Manin connection (i.e. the
explicit form of the natural connection on the bundle of first nonabelian coho-
mologies of some family of varieties). The simplest interesting case corresponds
to taking the universal family of four-punctured spheres and taking cohomology
with coefficients in SL2 (C) (one needs a non-trivial family of varieties with
nonabelian fundamental groups). This leads to the Painlevé VI equation (PVI ),
which is a second-order non-linear differential equation whose solutions, like those
of the hypergeometric equation, branch only at 0, 1, ∞ ∈ P1 . In particular we
may study the (non-linear) monodromy of solutions of PVI , by examining how
solutions vary upon analytic continuation along paths in the three-punctured
sphere.
Thus, since Schwarz lists fundamental solutions of hypergeometric equations
having finite monodromy, our main question is to construct the analogue of
Schwarz’s list for PVI :
Question C: What are the solutions of Painlevé VI having finite monodromy?
This question is still open; there is as yet no full classification – the main
effort (at least of the present author) has been towards finding and constructing
interesting solutions. So far all known finite-branching solutions are actually
algebraic (cf. Iwasaki 2008). Currently we are at the reasonably happy state of
affairs that all such solutions known to exist have actually been constructed. In
what follows I will explain various aspects of the problem, and in particular show
how the non-rigid lists (A) and (B) map to the list of (C). Some key points,
demonstrating the richness and variety of solutions, are
r There are algebraic solutions of PVI not related to finite subgroups of the
coefficient group SL2 (C).
r There are ‘generic’ solutions of PVI with finite monodromy; that is, not
lying on any of the reflection hyperplanes of the affine F4 Weyl group of
symmetries of PVI .
r There are entries on the list of (C) which do not come from either (A)
or (B).
In particular we will see PVI solutions related to the groups A6 , PSL2 (F7 )
and ∆237 .

11.2 What is Painlevé VI?

There are various viewpoints, and simply giving the explicit equation is perhaps
the least helpful introduction to it. In brief, Painlevé VI is
r The explicit form of the simplest nonabelian Gauss–Manin connection
r The equation controlling the ‘isomonodromic deformations’ of certain loga-
rithmic connections/Fuchsian systems on P1
r The most general second-order ODE with the so-called ‘Painlevé property’
 What is Painlevé VI? 215

r A certain dimensional reduction of the anti-self-dual Yang–Mills equations

(see e.g. Mason and Woodhouse (1996)),
r An equation related to certain elliptic integrals with moving endpoints (cf.
Fuchs (1905) and Manin (1998))
r The second-order ODE for a complex function y(t)


2

d2 y 1 1 1 1 dy 1 1 1 dy
= + + − + +
dt 2 2 y y−1 y−t dt t t−1 y−t dt


y(y − 1)(y − t) t (t − 1) t(t − 1)
+ α+β +γ +δ
t2 (t − 1)2 y2 (y − 1)2 (y − t)2
where α, β, γ, δ ∈ C are constants

The Painlevé property means that any local solution y(t) defined in a disc
in the three-punctured sphere P1 \ {0, 1, ∞} extends to a meromorphic function
on the universal cover of P1 \ {0, 1, ∞}. It is this property that enables us to
speak of the monodromy of PVI solutions. Concerning solutions there is a basic
trichotomy (see Watanabe (1998)):
⎧
⎪
⎨a ‘new’ transcendental function, or
A solution of PVI is either a solution of a first-order Riccati equation, or
⎪
⎩
an algebraic function.

In particular if one is interested in constructing new explicit solutions of PVI

then, since the Riccati solutions are all well understood, the algebraic solutions
are the first place to look.
The standard approach to PVI is as isomonodromic deformations of rank 2
logarithmic connections with four poles on P1 , as the poles move (generic such
connections are of the form (A), and then t parametrizes the possible pole
configurations). In particular one can see the four constants in PVI directly
in terms of the eigenvalues of the residues of the connection: if we set θi to
be the difference of the eigenvalues (in some order) of the residue Ai (i =

1, 2, 3, 4, where A4 = − 31 Ai is the residue at infinity) then the relation to the
constants is

α = (θ4 − 1)2 /2, β = −θ12 /2, γ = θ32 /2, and δ = (1 − θ22 )/2.

Before going into more detail let us also mention one further property of PVI :
it admits a group of symmetries isomorphic to the affine Weyl group of type
F4 (see Okamoto (1987) or the exposition in Boalch (2006)). Indeed treating
θ = (θ1 , . . . , θ4 ) ∈ C4 as the set of parameters for PVI is useful since the affine
F4 Weyl group of symmetries acts in the standard way on this C4 . (We will see
below that these four parameters may also be interpreted as coordinates on the
moduli space of cubic surfaces.)
216 Towards a non-linear Schwarz’s list

11.2.1 Conceptual approach to Painlevé VI

Consider the universal family of smooth four-punctured rational curves with
labelled punctures. Write B := M0,4 ∼ = P1 \ {0, 1, ∞} for the base, F for the
standard fibre, and C for the total space:
C ←−−−− F ∼ = P1 \ 4 points
⏐
⏐
+
B
Now replace each fibre F by H1 (F, G) where G = SL2 (C). Here we will use
two viewpoints/realizations of this nonabelian cohomology set H1 :
1. Betti: Moduli of fundamental group representations
H1 (F, G) ∼
= Hom(π1 (F), G)/G
2. De Rham: Moduli of connections on holomorphic vector bundles over F
These two viewpoints are related by the Riemann–Hilbert correspondence
(the nonabelian De Rham functor), taking connections to their monodromy
representations. The point is that algebraically these realizations of H1 are very
different and the Riemann–Hilbert map is transcendental (things written in
algebraic coordinates on one side will look a lot more complicated from the
other side).
Thus we get two non-linear fibrations over the base B, with fibres the
De Rham or Betti realizations of H1 (F, G), respectively:
Riemann–Hilbert
MDe Rham −−−−−−−−−−−→ MBetti
⏐ ⏐
⏐ ⏐
+ +
B B
As in the case with abelian coefficients one still gets a natural connection on
these cohomology bundles. The surprising fact is that it is algebraic on both sides
(approximating the De Rham side in terms of logarithmic connections to give it
an algebraic structure Nitsure 1993). Thus when written explicitly we will get
non-linear algebraic differential equations ‘coming from geometry’. (See Simpson
1994, section 8 for more on these connections in the case of families of projective
varieties.)
The two standard descriptions of the abelian Gauss–Manin connection gen-
eralize to descriptions of this non-linear connection. In the Betti picture we
may identify two nearby fibres of MBetti simply by keeping the monodromy
representations (points of the fibres) constant: moving around in B amounts to
deforming the configuration of four points in P1 and it is easy to see how to
identify the fundamental groups of the four-punctured spheres as the punctures
are deformed – use the same generating loops. This ‘isomonodromic’ description,
 What is Painlevé VI? 217

preserving the monodromy representation, is the nonabelian analogue of keeping

the periods of one-forms constant.
On the De Rham side the non-linear connection can be described in terms of
extending a connection on a vector bundle over a fibre F, to a flat connection on
a vector bundle over a family of fibres and then restricting to another fibre, much
as the abelian case is described in terms of closed one-forms (linear connections
replacing one-forms and flatness replacing the notion of closedness).
Each of these descriptions has a use: the De Rham viewpoint lends itself to
giving an explicit form of the non-linear connection (essentially amounting to
the condition for the flatness of the connection over the family of fibres). The
Betti viewpoint is more global and allows us to study the monodromy of the
non-linear connection, as an explicit action on fibres of MBetti .

11.2.2 Explicit non-linear equations

The De Rham bundle MDe Rham is well approximated by the space of logarithmic
connections with four poles on the trivial rank 2 holomorphic bundle (with trivial
determinant) over P1 . Call the space of such connections M∗ and observe it
parametrizes connections of the form (A), and that these are determined by the
value of t ∈ B and the residues:

M∗ ∼= B × (A1 , . . . , A4 ) Ai ∈ g, Ai = 0 /G.

Here G = SL2 (C) does not act on B and acts by diagonal conjugation on the
residues Ai . In general this quotient will not be well behaved, but it has a natural
Poisson structure and the generic symplectic leaves will be smooth complex
symplectic surfaces. Clearly M∗ is trivial as a bundle over B (projecting onto the
configuration of poles), but the nonabelian Gauss–Manin connection is different
to this trivial connection and was computed about 100 years ago by Schlesinger
(essentially in the way stated above it seems). The non-linear connection is given
by Schlesinger’s equations, which in the case at hand are
dA1 [A2 , A1 ] dA3 [A2 , A3 ]
= , =
dt t dt t−1
together with a third equation for dA2 /dt easily deduced from the fact that A4
remains constant. If the residues of the connection satisfy these equations then
the corresponding monodromy representation remains constant as t varies. (They
are 
easily derived from the vanishing
 of the curvature of the ‘full’ connection
d − A1 z + A2 z−t + A3 z−1 .)
dz dz−dt dz

To get from here to PVI one chooses specific functions x, y on M∗ which

restrict to coordinates on each generic symplectic leaf and writes down the
connection in these (carefully chosen) coordinates (see Boalch 2005, pp.199–200
for a discussion of the formulae, which are from Jimbo and Miwa 1981). This
leads to two coupled non-linear first-order equations, and eliminating x leads
to the second-order Painlevé VI equation for y(t). It was first written down in
218 Towards a non-linear Schwarz’s list

full generality by R. Fuchs (1905) (whose father L. Fuchs was also the father of
‘Fuchsian equations’).

11.2.3 Monodromy of Painlevé VI

Since the Betti and De Rham realizations are analytically isomorphic, we see
the monodromy of solutions to PVI thus corresponds to the monodromy of the
connection on MBetti . This amounts to an action of the fundamental group of
the base B on a fibre, and this action can be described explicitly.
Let Mt = Hom(π1 (P1 \ {0, t, 1, ∞}), G)/G be the fibre of MBetti at some fixed
point t ∈ B. The key point is that π1 (B) ∼= F2 (the free nonabelian group on
two generators) may be identified with the pure mapping class group of the four
punctured sphere P1 \ {0, t, 1, ∞}. As such it has a natural action on Mt (by
pushing forward loops generating the fundamental group), and this action is the
desired monodromy action.
Explicitly, upon choosing appropriate generating loops of π1 (P1 \ {0, t, 1, ∞})
we see Mt may be described directly in terms of monodromy matrices:
 
Mt ∼
= (M1 , . . . , M4 ) Mi ∈ G, M4 · · · M1 = 1 /G

which in turn is simply the quotient G3 /G of three copies of G by diagonal

conjugation by G = SL2 (C). In fact this quotient has been studied classically:
the ring of G invariant functions on G3 has seven generators and one relation,
embedding the affine quotient variety as a hypersurface in C7 . The particular
equation for this hypersurface appears on p. 366 of Fricke and Klein (1897).
The Painlevé VI parameters essentially specify the conjugacy classes of the four
monodromies Mi , and serve here to fibre the six-dimensional hypersurface G3 /G
into a four-parameter family of surfaces. Looking at the explicit equation shows
they are affine cubic surfaces. In turn Iwasaki (2002) has recently pointed out
that this family of cubics may be quite simply related to the explicit family of
Cayley (1849) and so contains the generic cubic surface.
The desired action of the free group F2 on the Betti spaces is given by the
squares of the following ‘Hurwitz’ action:
 
ω1 (M1 , M2 , M3 ) = M2 , M2 M1 M2−1 , M3
 
ω2 (M1 , M2 , M3 ) = M1 , M3 , M3 M2 M3−1 .

More explicitly if we consider simple positive loops l1 , l2 in B based at 12 encircling

0, 1, respectively, then the monodromy of the connection   MBetti around
on  li is
given by ωi2 (with respect to certain generators of π1 P1 \ 0, 12 , 1, ∞ . In turn
it is possible to write this action directly as an action on the ring of invariant
function on G3 .
 Algebraic solutions from finite subgroups of SL2 (C) 219

11.3 Algebraic solutions from finite subgroups of SL2 (C)

11.3.1 What exactly is an algebraic solution?
The obvious definition is simply an algebraic function y(t) which solves PVI for
some value of the four parameters. Thus it will be specified by some polynomial
equation
F (y, t) = 0
and a four-tuple θ of parameters. In practice however such polynomials F can
be quite unwieldy and are difficult to transform under the affine Weyl symmetry
group, making it difficult to see if in fact two solutions are equivalent. This leads
to our preferred definition.
Definition . An algebraic solution of PVI is a compact, possibly singular, alge-
braic curve Π together with two rational functions y, t : Π → P1 :
y
Π −−−−→ P1
⏐
⏐
t+

P1
such that
r t is a Belyi map (i.e. its branch locus is a subset of {0, 1, ∞}).
r y, when viewed as a function of t away from the ramification points of t,
solves PVI for some value of the four parameters.
In principle it is straightforward to go between the two definitions, but in practice
it is useful to look for a good model of Π (and the model given by the closure of
the zero locus of the polynomial F is usually a bad choice).

11.3.2 (A) → (C)

Suppose we have a linear connection (A) with finite monodromy. Its monodromy
representation will be specified by a triple (M1 , M2 , M3 ) ∈ G3 generating a finite
subgroup Γ ⊂ G (where G = SL2 (C) as above). This linear connection specifies
the initial value (and first derivative) of a solution to PVI . This PVI solution will
have finite monodromy, since we know the branching of PVI solutions corresponds
to the F2 action on conjugacy classes of triples in G3 , and the orbit through
(M1 , M2 , M3 ) will be finite, since the action is within triples of generators of Γ.
Thus we see that finite F2 orbits (in G3 /G) correspond to PVI solutions with
a finite number of branches, and the points of such F2 orbits correspond to the
individual branches of the PVI solution. In particular the size of the orbit, the
number of branches, is the degree of the map t : Π → P1 . (Indeed the F2 action
on such a finite orbit itself gives the full permutation representation of the Belyi
map t : Π → P1 , and in particular, by the Riemann–Hurwitz formula, determines
the genus of the ‘Painlevé curve’ Π.)
220 Towards a non-linear Schwarz’s list

Said differently it is useful to define a topological algebraic PVI solution (or

henceforth for brevity a topological solution) to be a finite F2 orbit in G3 /G. (The
classification of such orbits is still open and is the main step in classifying all
finite branching PVI solutions.) In these terms the first paragraph above points
out that one obtains ‘obvious’ topological solutions upon taking any triple of
generators of any finite subgroup of G.
For example (omitting discussion of how they were actually constructed), here
are some solutions corresponding to certain triples of generators of the binary
tetrahedral and octahedral subgroups, due to Dubrovin (1995) and Hitchin
(2003) (in different but equivalent forms):
Tetrahedral solution of degree 3

(s − 1)(s + 2) (s − 1)2 (s + 2)
y= , t= , and θ = (2, 1, 1, 2)/3,
s(s + 1) (s + 1)2 (s − 2)

Octahedral solution of degree 4

(s − 1)2 (s + 1)(s − 1)3

y= , t= , and θ = (1, 1, 1, 1)/4.
s(s − 2) s3 (s − 2)
In both cases Π is a rational curve (with parameter s). Although written in
this compact form, one should bear in mind these formulae represent a whole
(isomonodromic) family of connections (A) as t varies. An explicit elliptic
solution appears in Hitchin (1995a) and may be written as
Elliptic dihedral solution
  
(3 s − 1) s2 − 4 s − 1 s2 + u (s (s + 2) − u)
y= ,
(3 s3 + 7 s2 + s + 1) (s2 − u) (s (s − 2) + u)
 2
s2 + u (s (s + 2) − u) (s (s − 2) − u)
t= 2 ,
(s2 − u) (s (s + 2) + u) (s (s − 2) + u)
 
where the pair (s, u) lives on the elliptic curve u2 = s s2 + s − 1 and θ =
(1, 1, 1, 1)/2. This solution has degree 12 and corresponds to a triple of generators
of the binary dihedral group of order 20.
It turns out (see Boalch 2006a, remark 16) that the icosahedral solutions of
Dubrovin and Mazzocco (2000) fit into this framework as well and correspond to
(certain) triples of generators of the binary icosahedral group, although in the
first instance they arose from the icosahedral reflection group as described earlier.
Note that remark 0.1 of Dubrovin and Mazzocco (2000) describes a relation
between their solutions of PVI and a certain folding of Schwarz’s list; this is
different to the relation just mentioned – in particular problem (A) demands an
extension of Schwarz’s list.
 Beyond Platonic Painlevé VI solutions 221

11.4 Beyond Platonic Painlevé VI solutions

My starting point in this project was simply the observation that there should
be more algebraic solutions to PVI than those coming from finite subgroups
of SL2 (C). Dubrovin (1995) had shown how to relate three-dimensional real
orthogonal reflection groups to a certain one-parameter family of the full four-
dimensional family of PVI equations (namely, the family having parameters
θ = (0, 0, 0, ∗)) and this was used in Dubrovin and Mazzocco 2000 to classify alge-
braic solutions having parameters in this one-parameter family. (Some aspects
of op. cit. were subsequently extended by Mazzocco 2001 to classify rational
solutions – i.e. those with only one branch, cf. also Yuan and Li 2002.) The
further observation was that if one is able to get away from the orthogonality
condition here then one will relate any PVI equation to a three-dimensional
complex reflection group.

Theorem 11.1 (Boalch 2003, 2005) The isomonodromic deformations of

type (B) connections (on rank 3 vector bundles) are also controlled by the
Painlevé VI equation, and all PVI equations arise in this way.

Thus a solution to PVI can also be viewed as specifying an isomonodromic

family of rank 3 Fuchsian connections. It turns out that the formulae to go from
a PVI solution y(t) to such an isomonodromic family are more symmetrical than
in the previous case (type (A)) so we will recall them here. (For the analogous
formulae for (A) see Jimbo and Miwa 1981 and in Harnad’s dual picture –
the formula for which should be compared to that below – see Harnad 1994
and also Mazzocco 2002, which was kindly pointed out by the referee.) First
the parameters: let λi = Tr(Bi ) for i = 1, 2, 3 and let µi be the eigenvalues, in
some
  of B1 + B2 + B3 (which is minus the residue at infinity), so that
order,
λ i = µi .

Theorem 11.2 (Boalch 2006b) If y(t) solves Painlevé VI with parameters θ

where

θ1 = λ1 − µ1 , θ2 = λ2 − µ1 , θ3 = λ3 − µ1 , and θ4 = µ3 − µ2 ,

and we define x(t) via


1 (t − 1)y  − θ1 y  − 1 − θ2 t y  + θ3
x= + −
2 y y−t y−1

then the family of logarithmic connections (B) will be isomonodromic as t varies,

where
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
λ1 b12 b13 0 0 0 0 0 0
B1 = ⎝ 0 0 0 ⎠ , B2 = ⎝ b21 λ2 b23 ⎠ , and B3 = ⎝ 0 0 0 ⎠
0 0 0 0 0 0 b31 b32 λ3
222 Towards a non-linear Schwarz’s list

b12 = λ1 − µ3 y + (µ1 − xy)(y − 1), b32 = (µ2 − λ2 − b12 )/t,

b13 = λ1 t − µ3 y + (µ1 − xy)(y − t), b23 = (µ2 − λ3 )t − b13 ,
µ3 (y − t) − µ1 (y − 1) + x(y − t)(y − 1)
b21 = λ2 + , b31 = (µ2 − λ1 − b21 )/t.
t−1

The implication of this for algebraic solutions should now be clear: the mon-
odromy of a PVI solution is also described by an action of the free group F2 on
(conjugacy classes of) triples of three-dimensional complex reflections (r1 , r2 , r3 )
(with the same formula as before, just replace Mi by ri ). Thus in this context the
‘obvious’ topological solutions (i.e. finite F2 orbits) come from taking a triple of
generating reflections of a finite complex reflection group in GL3 (C). Such finite
complex reflection groups were classified by Shephard and Todd (1954) and apart
from the familiar real orthogonal reflection groups there is an infinite family plus
four exceptional complex groups, the Klein reflection group (of order 336, a two-
fold cover of Klein’s simple group isomorphic to PSL2 (F7 ) → PGL3 (C)), two
Hessian groups, and the Valentiner group (of order 2,160, a six-fold cover of
A6 → PGL3 (C)).
The infinite family of groups and the two Hessian groups do not seem to
lead to interesting new solutions, but by computing the F2 orbits (determining
the topology of Π) it is easy to see that the Klein group yields a genus 0
degree 7 solution and the Valentiner group has three inequivalent triples of
generating reflections, each leading to genus 1 solutions with degrees 15,15,
and 24, respectively. These are new solutions, previously undetected. (The 24
appearing here led to a certain amount of trepidation, given that the 10 page
elliptic solution of Dubrovin and Mazzocco 2000 had degree 18.)

11.4.1 Construction
Of course finding the topological solution is not the same as finding an explicit
isomonodromic family of connections; one needs to solve a family of Riemann–
Hilbert problems inverting the transcendental Riemann–Hilbert map for each
value of t. (Indeed my original plan was to just prove the existence of new
interesting solutions, in Boalch (2003), but a certain stubbornness, and some
inspiration from reading about Klein’s work finding explicit 3 × 3 matrices
generating his simple group, convinced us to go further.)
The two main steps in the method we finally got to work are as follows. (This
is a generalization of the method used by Dubrovin and Mazzocco 2000.)
1. Jimbo’s asymptotic formulae. Jimbo (1982) found an exact for-
mula for the leading asymptotics at t = 0 of the branch of the PVI solution
y(t) corresponding to any sufficiently generic linear monodromy representation
(M1 , M2 , M3 ). (This formula was obtained by considering the degeneration of
the isomonodromic family of connections (A) as t → 0; in the limit the four-
punctured sphere degenerates into a stable curve with two components, each
 Beyond Platonic Painlevé VI solutions 223

B
0 1 ∞

Figure 11.1 Degeneration to two hypergeometric systems.

with three marked points. The connections (A) degenerate into hypergeometric
systems on each component, with known monodromy (Figure 11.1). Since these
are rigid it is easy to solve their Riemann–Hilbert problems explicitly and this
gives the leading asymptotics of the isomonodromic family and thus of the PVI
solution.)
This is useful for us because, as Jimbo mentions, one may substitute the leading
asymptotics back into the PVI equation to get arbitrarily many terms of the
precise asymptotic expansion of the solution at 0. If the solution is algebraic,
then this is its Puiseux expansion, a sufficient number of terms of which will
determine the entire solution.
It turns out there was a typo in Jimbo (1982), which meant the entire method
did not work (indeed the fact it did not work led to the questioning of Jimbo’s
formula and hence the correction in Boalch 2005). (Note the special parameters
of Dubrovin and Mazzocco 2000 are not covered by Jimbo’s result; rather they
adapted the argument of Jimbo 1982 to their case.)
2. Relating (A) and (B). Since Jimbo’s formula requires a monodromy
representation of a connection of type (A), and we are starting with a triple of
3 × 3 complex reflections (the monodromy representation of a connection of type
(B)), the second step is that we need to see how to go between these two pictures
(on both the De Rham and Betti sides of the Riemann–Hilbert correspondence).
This will be described in the following subsection.

11.4.2 Relating connections (A) and (B)

We wish to sketch how to convert a connection (B) on a rank 3 vector bundle
into a connection of the form (A) on a rank 2 bundle. On the other side of the
Riemann–Hilbert correspondence this amounts to an F2 -equivariant map from
triples of complex reflections to triples of elements of G = SL2 (C) (as in Boalch
2005, section 2).
224 Towards a non-linear Schwarz’s list

Of course the monodromy groups change in a highly non-trivial way under

this procedure. For example, the Klein reflection group becomes the triangle
group ∆237 ⊂ G, which is an infinite group, and the Valentiner group becomes
the binary icosahedral group (leading to an unexpected relation between A6
and A5 ).
After this procedure was put on the arxiv (Boalch 2005) we learnt (Dettweiler
and Reiter 2007) that it is essentially a case of the middle convolution functor
used by Katz (1996), although our construction using the complex analytic
Fourier–Laplace transform is different from that of Katz (using l-adic methods)
and from the work of Dettweiler and Reiter (2000).
The basic picture which emerges is as follows (see the figure below), and ought
to be better known. It was obtained essentially by a careful reading of Balser,
Jurkat and Lutz (1981), although the basic idea of relating irregular and Fuchsian
systems by the Laplace transform dates back to Birkhoff and Poincaré. (Dubrovin
1995, 1999 used an orthogonal analogue in relation to Frobenius manifolds, also
using Balser, Jurkat, and Lutz 1981. Moreover the top triangle is essentially a
case of ‘Harnad duality’ (Harnad 1994) so for n = 3 we knew we would obtain
all PVI equations.)

The idea is to describe a transcendental map from gln (C) to GLn (C) in two
different ways (the two paths down the left and the right from the top to the
bottom of the figure).
Choose n distinct complex numbers a1 , . . . , an and define A0 = diag
(a1 , . . . , an ). Roughly speaking (on a dense open patch) the left-hand column
arises by defining Ai = Ei A (setting to zero all but the ith row of A) and
 Beyond Platonic Painlevé VI solutions 225

 Ai
constructing the logarithmic connection d − z−ai
dz having rank 1 residues
at each ai . Then taking the monodromy of this yields n complex reflections ri
(and if bases of solutions are chosen carefully one can naturally define vectors ei
and one-forms αi such that ri = 1 + ei ⊗ αi and that the ei form a basis). Then
the map to GLn (C) is given by taking the product of rn · · · r1 of these reflections,
written in the ei basis.
Now the key algebraic fact, which dates back at least to Killing (1889) (see
Coleman 1989), is that any such product of complex reflection lies in the big cell
of GLn (C) and so may be factored as the product of a lower triangular and an
upper triangular matrix. We write this product as u−1 − hu+ with u± ∈ U± the
unipotent triangular subgroups, and h ∈ H diagonal:

rn · · · r2 r1 = u−1
− hu+ . (11.2)

Further, although this relation between the reflections and u± looks to be highly
non-linear, one can relate them in an almost linear fashion: the matrix hu+ − u−
is the matrix with entries αi (ej ).
On the other hand it turns out that the same  map can
 be defined by taking
A0
the Stokes data of the irregular connection d − z2 + z dz. Indeed the map on
A

the right-hand side generalizes (Boalch 2002) to any complex reductive group
G in place of GLn (C), but only for GLn (C) is the alternative ‘logarithmic’
viewpoint available. Thus u± are also the two Stokes matrices of this irregular
connection (the natural analogue of monodromy data for such connections); the
exact definition is not important here. (The element h is the so-called formal
monodromy, explicitly it is simply exp(2πiΛ) where Λ is the diagonal part of
A.) The two connections are related (see Balser, Jurkat, and Lutz 1981) by
the Fourier–Laplace transform: this is more than just formal, and by relating
bases of solutions on both sides the stated relation between the Stokes and
monodromy data is obtained. (In both cases the resulting element of GLn (C)
is the monodromy around z = ∞ in a suitable basis.) In summary we see that
the ‘Betti’ incarnation of the Fourier–Laplace transform is the relation of Killing–
Coxeter.
Now to apply this in the current context we consider the effect of adding a
scalar λ to A ∈ gln (C). On the right-hand side this corresponds to tensoring
the irregular connection by the logarithmic connection d − λdz/z on the trivial
line bundle, and Balser, Jurkat, and Lutz (1981) showed that the Stokes data is
changed only by scaling h by s := exp(2πiλ), fixing u± . On the logarithmic side
this corresponds to a non-trivial convolution operation, changing the monodromy
representation in a non-trivial way. Of course using the Killing–Coxeter identity
we now see precisely how the complex reflections vary. (It is perhaps worth
noting that this scalar shift is essentially the inverse of the spectral parameter
introduced by Killing 1889, p. 20, appearing in the characteristic polynomial of
the Killing–Coxeter matrix (11.2): det(u−1 − shu+ − 1) = det(shu+ − u− ).)
If we set n = 3 then the logarithmic connections appearing are of the form
(B), upon taking a1 , a2 , a3 = 0, t, 1. Then we may choose the scalar shift such
226 Towards a non-linear Schwarz’s list

that the resulting element of GL3 (C) has 1 as an eigenvalue. This implies that
the connections are reducible and we can take the irreducible rank 2 sub- or
quotient connection. Projecting to sl2 gives the desired connection of type (A)
(see Boalch 2005). (Note that there is a choice involved here, of which eigenvalue
to shift to 1.)

11.4.3 New solutions

Thus in summary the procedure now is as follows: take a triple of generating
reflections of a finite complex reflection group in GL3 (C). Push it down to
the 2 × 2 framework using the scalar shift to obtain a triple (M1 , M2 , M3 ) of
elements of SL2 (C) in an isomorphic F2 orbit. Apply Jimbo’s formula to get the
leading asymptotics of the corresponding PVI solutions at t = 0 on each branch
(i.e. for each triple in the F2 orbit). (Converting the values which arise into
exact algebraic numbers.) Substitute these leading terms back into PVI to obtain
arbitrarily many terms of the Puiseux expansion at 0 of each solution branch.
Use these expansions to determine the polynomial F (y, t) defining the solution
(assuming it is algebraic). Find a parametrization of the resulting algebraic curve
(e.g. using M. van Hoeij’s wonderful Maple algebraic curves package).
For example, for the Klein complex reflection group of order 336 this works
perfectly (Boalch 2005) and the resulting solution is
Klein solution
 2  
5 s − 8 s + 5 7 s2 − 7 s + 4
y=− ,
s (s − 2) (s + 1) (2 s − 1) (4 s2 − 7 s + 7)
 2 2
7s − 7s + 4
t = 2, and θ = (2, 2, 2, 4)/7
s3 (4 s2 − 7 s + 7)
which has 7 branches. One may of course now substitute this back into the
formula of Theorem 11.2 (with λ = (1, 1, 1)/2 and µ = (3, 5, 13)/14) to obtain
an explicit family of logarithmic connections having monodromy equal to the
Klein reflection group generated by reflections (see Boalch 2006b, section 3).
When converted to connections of type (A) these ‘Klein connections’ have
infinite (projective) monodromy group equal to the triangle group ∆237 (cf.
Boalch 2006c, appendix B). On the other hand it turns out (Boalch 2006a) that
for the Valentiner connections, even though they are much trickier to construct
directly, we can still compute immediately that they become connections of type
(A) with binary icosahedral monodromy. They are also inequivalent to those
appearing in the work of Dubrovin and Mazzocco related to the real orthogonal
icosahedral reflection group (which lead to unipotently generated monodromy
with one choice of the scalar shift, but finite binary icosahedral monodromy
with a different choice, cf. Boalch 2006a, remark 16).
Thus it seemed like a good idea to examine precisely what PVI solutions
arise upon taking arbitrary triples of generators (M1 , M2 , M3 ) of the binary
 Beyond Platonic Painlevé VI solutions 227

Table 11.2. Icosahedral solutions 11–52

Degree Genus Walls Type Degree Genus Walls Type

11 2 0 2 b2 c2 32 10 0 3 d4
12 2 0 2 b2 d2 33 12 0 0 abcd
13 2 0 2 c2 d2 34 12 1 1 a b c2
14 3 0 1 b c2 d 35 12 1 1 a b d2
15 3 0 1 b c d2 36 12 1 1 b2 c d
16 4 0 2 a c3 37 15 1 2 b3 c
17 4 0 2 a d3 38 15 1 2 b3 d
18 4 0 2 c3 d 39 15 1 2 b2 c2
19 4 0 2 c d3 40 15 1 2 b2 d2
20 5 0 1 b2 c d 41 18 1 3 b4
21 5 0 2 c2 d2 42 20 1 1 a b2 c
22 6 0 1 b c2 d 43 20 1 1 a b2 d
23 6 0 1 b c d2 44 20 1 3 a2 c2
24 8 0 1 a c2 d 45 20 1 3 a2 d 2
25 8 0 1 a c d2 46 24 1 2 a b3
26 9 1 2 b c3 47 30 2 2 a2 b c
27 9 1 2 b d3 48 30 2 2 a2 b d
28 10 0 2 a2 c d 49 36 3 3 a2 b2
29 10 0 2 b3 c 50 40 3 3 a3 c
30 10 0 2 b3 d 51 40 3 3 a3 d
31 10 0 3 c4 52 72 7 3 a3 b

icosahedral group. Thus we looked at all triples of generators and quotiented

by the relation coming from the affine F4 symmetries of PVI . The resulting
table has 52 rows (which is quite small considering there are 26,688 conjugacy
classes of generating triples). The first 10 rows correspond to the 10 icosahedral
rows of Schwarz’s list and thus the projective monodromy around one of the
four punctures is the identity (these correspond to the PVI solution y = t). The
remaining rows are as in Table 11.2 (this is abridged from Boalch 2006a). (Note
that the right notion of equivalence in the linear non-rigid problem (A) seems to
be the ‘geometric equivalence’ of Boalch 2006a, section 4 – however this coincides
with equivalence under the affine F4 Weyl group, in this case.)
Thus there are lots of other icosahedral solutions the largest having genus 7
and 72 branches. (The column ‘Type’ indicates the set of conjugacy classes of
local monodromy of the corresponding connections of type (A), as we marked on
Schwarz’s list. The column ‘Walls’ indicates the number of reflection hyperplanes
for the affine F4 Weyl group that the solution’s parameters θ lie on.) A few of
these solutions had appeared before: those with degree <5 are simple deforma-
tions of previous solutions, solutions 21 and 26 are in Kitaev (2005) and the
Dubrovin–Mazzocco icosahedral solutions are equivalent to those on rows 31,32,
and 41. On the other hand the Valentiner solutions are quite far down the list
on rows 37, 38, and 46.
The above method of constructing solutions using Jimbo’s asymptotic formula
applies only to sufficiently generic monodromy representations but it turns out
228 Towards a non-linear Schwarz’s list

that most of the rows of this table have some representative (in their affine F4
orbit) to which Jimbo’s formula maybe applied (on every branch). Thus we could
start working down the list constructing new solutions. An initial goal was to get
to solution 33: this solution purports to be on none of the reflection hyperplanes
and the folklore was that all explicit solutions to Painlevé equations must lie on
some reflection hyperplane. The folklore was wrong:
‘Generic’ solution/Icosahedral solution 33
9s(s2 + 1)(3s − 4)(15s4 − 5s3 + 3s2 − 3s + 2)
y=− ,
(2s − 1)2 (9s2 + 4)(9s2 + 3s + 10)

27s5 (s2 + 1)2 (3s − 4)3

t= , and
4(2s − 1)3 (9s2 + 4)2
θ = (2/5, 1/2, 1/3, 4/5).
So far this looks to be the only example of a ‘classical’ solution of any of
the Painlevé equations that does not lie on a reflection hyperplane (of the
full symmetry group). Apart from being in the interior of a Weyl alcove this
solution is generic in another sense: a randomly chosen triple of generators of
the binary icosahedral group is most likely to lead to it (more of the 26,688
triples of generators correspond to this row than to any other). Notice also that
this solution has type abcd; there is one local monodromy in each of the four
non-trivial conjugacy classes of A5 .
At this stage we were approaching solution 41 which we knew took 10 pages
to write down. So we stopped and looked around to see if there were other
interesting (even just topological) solutions. (The tetrahedral and octahedral
cases could all now be fully dealt with Boalch 2006c.)

11.5 Pullbacks
In his 1884 book on the icosahedron (see Klein 1956), Klein showed that all
second-order Fuchsian differential equations with finite monodromy are (essen-
tially) pullbacks of a hypergeometric equation along a rational map f :

In particular (k = 3) all the icosahedral entries on Schwarz’s list may be

obtained by pulling back the ‘235’ hypergeometric equation (on row VI of
Schwarz’s list).
 Pullbacks 229

In our context, an isomonodromic family of connections of type (A) amounts

to a family of Fuchsian equations with five singularities (at 0, t, 1, ∞, plus an
apparent singularity at another point y). 2 Klein’s theorem says each element of
this family arises as the pullback of the 235 hypergeometric equation along a
rational map, so the family corresponds to a family of rational maps.
Thus finding a PVI solution corresponding to a family of connections (A)
with finite monodromy amounts to giving a certain family of rational maps
f : P1 → P1 . To construct such PVI solutions one may try to find such families
of rational maps, such that each map pulls back a hypergeometric equation to
an equation with the right number of singular points – or to one that can be
put in this form after using elementary transformations to remove extraneous
apparent singularities. (This is not straightforward; e.g. given a finite monodromy
representation of a connection (A) it is not immediate even what degree such a
map f will have.)
An important further observation (due to C. Doran 2001 and A. Kitaev 2002)
is that any such family of rational maps will lead to algebraic solutions of Painlevé
VI regardless of whether or not the hypergeometric equation being pulled back
has finite monodromy (provided the equation upstairs has the right number of
poles); the algebraicity follows from that of the family of rational maps.
Andreev and Kitaev (2002); Kitaev (2002, 2005) have used this to construct
some PVI solutions, essentially by starting to enumerate all such rational maps
(this leads to a few new solutions, but most in fact turn out to be equivalent to
each other or to ones previously constructed – see Section 11.7).
On the other hand, Doran had the idea that interesting PVI solutions should
come from hypergeometric equations with interesting monodromy groups. Thus
(amongst other things) Doran (2001) studied the possible hypergeometric equa-
tions with monodromy a hyperbolic arithmetic triangle group which may be
pulled back to yield PVI solutions. Indeed in Doran (2001, corollary 4.6), he
lists such possible triangle groups and the degrees and ramification indices of
the corresponding rational maps f , although no new solutions were actually
constructed. We picked up on this thread in Boalch (2006c, section 5): it was
found that all but one entry on Doran’s list corresponded to a known explicit
solution (although were perhaps unknown when Doran 2001 was published). The
remaining entry was for a family of degree 10 rational maps f pulling back the
237 triangle group with ramification indices (partitions of 10):
[2, 2, 2, 2, 2], [3, 3, 3, 1], [7, 1, 1, 1]
over 0, 1, ∞ (where the hypergeometric system has projective monodromy of
orders 2, 3, and 7, respectively), as well as minimal ramification [18 , 2] over
another variable point. As explained in Boalch (2006c) one can get from here to

2 This is the same y appearing in P

VI – that is, the function y on the space of connections
(A) is the position of the apparent singularity that appears when the connection is converted
into a Fuchsian equation (Fuchs 1905).
230 Towards a non-linear Schwarz’s list

P1

2 2 3 7
1 2 3 1
f ·· 2
· 2 3 1
1 2 1 1

2 3 7
P1

Figure 11.2 237 degree 10 rational map f .

a topological PVI solution by drawing a picture: we wish to find such a rational

map f topologically – that is, describe the topology of a branched cover f :
P1 → P1 with this ramification data. This may be done by playing ‘join the
dots’ (completely in the spirit of Grothendieck’s Dessins d’Enfants) and yields a
covering figure as required. One figure so obtained is shown in Figure 11.2. (Note
that, in the context of Painlevé equations, the idea of drawing pictures such as
Figure 11.2 first appeared in Kitaev (2005).)
The upper copy of P1 is thus divided into 10 connected components and f
maps each component isomorphically onto the complement of the interval drawn
on the lower P1 (the lines and the vertices upstairs are the preimages of the lines
and vertices downstairs). In particular the figure shows how loops upstairs map to
words in the generators of the fundamental group π1 (P1 \ {0, 1, ∞}) downstairs.
In this way we can compute by hand the monodromy of the equation upstairs
obtained by pulling back a hypergeometric equation with monodromy ∆237 . This
yields the triple:

M1 = caca−1 c−1 , M2 = c, and M3 = c−1 a−1 cac

(where a, b, and c are lifts to SL2 (C) of standard generators of ∆237 with cba = 1),
which we know a priori lives in a finite F2 orbit. One finds immediately that the
 Final steps 231

orbit through the conjugacy class of this triple has size 18 and constitutes a
genus 1, degree 18 topological PVI solution.
Now it turns out that Jimbo’s formula may be applied to every branch of
this solution, and proceeding as before we obtain (Boalch 2006c) the solution
explicitly:
Elliptic 237 solution
 
1 3 s8 −2 s7 −4 s6 − 204 s5 −536 s4 −1738 s3 −5064 s2 −4808 s−3199 u
y= − ,
2 4 (s6 + 196 s3 + 189 s2 + 756 s + 154) (s2 + s + 7) (s + 1)
 9 
1 s − 84 s6 − 378 s5 − 1512 s4 − 5208 s3 − 7236 s2 − 8127 s − 784 u
t= − 2 2 ,
2 432 s (s + 1) (s2 + s + 7)
where u2 = s (s2 + s + 7) and θ = (2/7, 2/7, 2/7, 1/3). (This solution, or rather
an inequivalent ‘Galois conjugate’ of it, has also been obtained independently
by Kitaev 2006, p. 219 by directly computing such a family of rational maps –
apparently also influenced by Doran’s list.)

11.6 Final steps

11.6.1 Up to degree 24
We now have an example of a degree 18 elliptic solution to Painlevé VI with
a quite simple form. This leads immediately to the suspicion that the 10-page
Dubrovin–Mazzocco solution is just written at a bad value of the parameters.
Indeed using the method we have been ‘tweaking’ while working down the
icosahedral table enables us to guess good a priori choices of the parameters
θ within the corresponding affine F4 equivalence class (row 41 in Table 11.2)
i.e. so that the expression for the polynomial F will be ‘small’. Choosing such
parameters and constructing the solution from scratch at those parameters yields

Theorem 11.3 (Boalch 2006a) The Dubrovin–Mazzocco icosahedral solution

is equivalent to the solution
1 8 s7 − 28 s6 + 75 s5 + 31 s4 − 269 s3 + 318 s2 − 166 s + 56
y= − and
2 18 u (s − 1) (3 s3 − 4 s2 + 4 s + 2)
 
1 (s +1) 32 (s8 +1) − 320 (s7 +s) +1112 (s6 +s2 ) − 2420 (s5 +s3 ) + 3167 s4
t= +
2 54 u3 s (s − 1)
on the elliptic curve
u2 = s (8 s2 − 11 s + 8)
with θ = (1, 1, 1, 1)/3. In particular this elliptic curve is birational to that defined
by the 10-page polynomial.
232 Towards a non-linear Schwarz’s list

Substituting this into the formula of Theorem 11.2 with λ = (1, 1, 1)/2 and µ =
(1, 3, 5)/6 now gives explicitly the third (and trickiest) family of connections of
type (B) with monodromy the icosahedral reflection group.
This can be pushed further with more tweaking to get up to degree 24 (row
46 in Table 11.2), that is, to obtain the largest Valentiner solution (Boalch
2006a) (the main further tricks used are described in (Boalch 2006c, appen-
dix C). In particular this finishes the construction of all elliptic icosahedral
solutions. Intriguingly, one finds that the resulting elliptic icosahedral Painlevé
curves Π become singular only on reduction modulo the primes 2, 3, and 5
(except for rows 44 and 45 – we will see another reason in the following
subsection that these are abnormal). Similarly the elliptic Painlevé curve related
to the 237 triangle group becomes singular only on reduction modulo 2, 3,
and 7.

11.6.2 Quadratic/Landen/folding transformations

Now the happy fact is that the remaining icosahedral solutions may be obtained
from earlier solutions by a trick, first introduced in the context of PVI by Kitaev
(1991) and a simpler equivalent form was found by Ramani, Grammaticos, and
Tamizhmani (2000). Manin (1998) refers to some equivalent transformations as
Landen transformations. (Landen has clear precedence since the original Landen
transformations were rediscovered by Gauss!) Tsuda, Okamoto, and Sakai (2005)
call them folding transformations.
In any case the basic idea is simple: if one has a connection (A) with two
local projective monodromies of order 2 (say at 0, ∞) then one can pull it back
along the map z → z 2 and obtain a connection with only apparent singularities
at 0, ∞ (which can be removed) and four genuine singularities. This can be
normalized into the form (A), and the key point is that this works in families and
maps isomonodromic deformations of the original connections to isomonodromic
deformations of the resulting connections – that is, it transforms certain solutions
of PVI into different, generally inequivalent, solutions. Of course this is not a
genuine symmetry of PVI since special parameters are required, but it is precisely
what is needed to construct the remaining solutions.
Indeed observe that each of the rows of the icosahedral table with degree >24
have type a2 ξη for some ξ, η ∈ {a, b, c, d} – that is, they have two projective
monodromies of order 2. Pulling back along the squaring map will transform the
corresponding connections into connections of type ξ 2 η 2 . It turns out (in this
icosahedral case) the corresponding PVI solutions have half the degree, and we
obtain an algebraic relation between the solutions. This program is carried out
in Boalch (2007) and the remaining icosahedral solutions are obtained (see also
Kitaev and Vidūnas 2007). (Notice also that the elliptic solutions on rows 44 and
45 are related in this way to earlier, genus zero solutions.) For example, in Boalch
(2007) we found an explicit equation for the genus 7 algebraic curve naturally
attached to the icosahedron, on which the largest (degree 72) icosahedral solution
 Conclusion 233

is defined: it may be modelled as the plane octic with affine equation

Genus 7 icosahedral Painlevé curve
9 (p6 q 2 + p2 q 6 ) + 18 p4 q 4 +
4 (p6 + q 6 ) + 26 (p4 q 2 + p2 q 4 ) + 8 (p4 + q 4 ) + 57 p2 q 2 +
20 (p2 + q 2 ) + 16 = 0.

11.7 Conclusion
Thus in conclusion we have filled in a number of rows of what could be called
the non-linear Schwarz’s list. Whether or not there will be other rows remains
to be seen. So far this list of known algebraic solutions to PVI takes the following
shape (we will use the letters d and g to denote the degree and genus of solutions,
and consider solutions up to equivalence under Okamoto’s affine F4 symmetry
group. Some non-trivial work has been done to establish which of the published
solutions are equivalent to each other and which were genuinely new). See also
Boalch (2006d).
First there are the rational solutions (d = 1), studied by Mazzocco (2001)
and Yuan and Li (2002), which fit into the set of Riccati solutions classified by
Watanabe (1998). (Beware that ‘rational’ here means the solution is a rational
function of t, which implies, but is by no means equivalent to, having a rational
parameterization.)
Then there are three continuous
√ families of solutions g = 0, d = 2, 3, 4.
The degree 2 family is y = t which, as one may readily verify, solves PVI
for a family of possible parameter values. Similarly the degree 3 tetrahedral
solution, and the degree 4 octahedral and dihedral solutions (of Dubrovin 1995
and Hitchin 1995a, 2003) fit into such families, as discussed in Ben Hamed and
Gavrilov (2005); Boalch (2006a) and Cantat and Loray (2007). In general in
such a family y(t) may depend on the parameters of the family. Ben Hamed
and Gavrilov (2005) showed that any family with y(t) not depending on the
parameters is equivalent to one of the above cases and recently Cantat and Loray
(2007) showed that any solution with two, three, or four branches is in such
family.
Next there is one discrete family (d, g unbounded, θ = (0, 0, 0, 1) ∼
(1, 1, 1, 1)/2). Indeed this PVI equation was solved completely by Picard (1889,
p. 299), Fuchs (1905), and in a different way by Hitchin (1995b). Algebraic
(determinantal) formulae for the algebraic solutions amongst these appear in
Hitchin (1995a), using links with the Poncelet problem – in this framework they
are dihedral solutions (controlling connections of type (A) with binary dihedral
monodromy).
Finally there are 45 exceptional solutions, which collapse down to 30 if we
identify solutions related by quadratic transformations. The possible genera are
0, 1, 2, 3, 7, and the highest degree is 72. Of these 30 solutions 7 have previously
234 Towards a non-linear Schwarz’s list

appeared: 1 is due to Dubrovin (1995), 2 to Dubrovin and Mazzocco (2000), and

4 to Kitaev (3 in Kitaev 2005, plus – in Kitaev 2006 – a Galois conjugate of
the elliptic 237 solution already mentioned). Two of these exceptional solutions
are octahedral, 1 is the Klein solution, 3 are the elliptic 237 solution (and its 2
Galois conjugates), and the remaining 24 are icosahedral.

References
Andreev, F. V. and Kitaev, A. V. (2002). Transformations RS24 (3) of the ranks ≤ 4 and
algebraic solutions of the sixth Painlevé equation. Commun. Math. Phys. 228(1),
151–76. nlin.SI/0107074.
Balser, W., Jurkat, W. B. and Lutz, D. A. (1981). On the reduction of connection
problems for differential equations with an irregular singularity to ones with only
regular singularities, I., SIAM J. Math. Anal. 12(5), 691–721.
Ben Hamed, B. and Gavrilov, L. (2005). Families of Painlevé VI equations having a
common solution. Int. Math. Res. Not. No. 60, 3727–52. math/0507002.
Beukers, F. and Heckman, G. (1989). Monodromy for the hypergeometric function
n Fn−1 . Invent. Math. 95(2), 325–54.
Beukers, F. and van der Waall, A. (2004). Lamé equations with algebraic solutions. J.
Differential Equations 197(1), 1–25.
Boalch, P. P. (2002). G-bundles, isomonodromy and quantum Weyl groups. Int. Math.
Res. Not. No. 22, 1129–66, math.DG/0108152.
Boalch, P. P. (2003). Painlevé equations and complex reflections. Ann. Inst. Fourier
53(4), 1009–22.
Boalch, P. P. (2005). From Klein to Painlevé via Fourier, Laplace and Jimbo. Proc.
London Math. Soc. 90(3), no. 3, 167–208, math.AG/0308221.
Boalch, P. P. (2006a). The fifty-two icosahedral solutions to Painlevé VI. J. Reine
Angew. Math. 596, 183–214, math.AG/0406281.
Boalch, P. P. (2006b). Six results on Painlevé VI. Sémin. Congr., Vol. 14, Society of
Mathematics France, Paris, math.AG/0503043, pp. 1–20.
Boalch, P. P. (2006c). Some explicit solutions to the Riemann–Hilbert problem. IRMA
Lect. Math. Theor. Phys. 9, 85–112, math.DG/0501464.
Boalch, P. P. (2006d). Survey of the known algebraic solutions of Painlevé VI.
Two talks given at the Newton Institute 71pp of transparencies on their webpage
www.tinyurl.com/boalchINI1, 2.
Boalch, P. P. (2007). Higher genus icosahedral Painlevé curves. Funk. Ekvac. (Kobe)
50(1), 19–32, math.AG/0506407.
Cantat, S. and Loray, F. (2007). Holomorphic dynamics, Painlevé VI equation and
character varieties, arXiv:0711.1579. Preprint Nov.
Cayley, A. (1849). On the triple tangent planes of surfaces of the third order. Dublin
Math. J. No. 4, 118–38.
Coleman, A. J. (1989). Killing and Coxeter transformations of Kac-Moody algebras.
Invent. Math. 95, 447–77.
Dettweiler M. and Reiter, S. (2000). An algorithm of Katz and its application to the
inverse Galois problem. J. Symbolic Comput. 30(6), 761–98.
Dettweiler M. and Reiter, S. (2007). Painlevé equations and the middle convolution.
Adv. Geom. 7(3), 317–30. (See also preprint dated October 1, 2004).
Doran, C. F. (2001). Algebraic and geometric isomonodromic deformations. J. Differ-
ential Geom. 59(1), 33–85.
Dubrovin, B. (1995). Geometry of 2D topological field theories. Integrable Systems and
Quantum Groups, M. Francaviglia and S. Greco (eds.), Vol. 1620, Springer Lecture
Notes in Mathematics, pp. 120–348. hep-th/9407018.
 References 235

Dubrovin, B. (1999). Painlevé transcendents in two-dimensional topological field theory.

The Painlevé Property, Springer, New York, pp. 287–412, math/9803107.
Dubrovin B. and Mazzocco, M. (2000). Monodromy of certain Painlevé-VI transcen-
dents and reflection groups. Invent. Math. 141(1), 55–147, math.AG/9806056, SISSA
preprint 1997.
Fricke R. and Klein, F. (1897). Vorlesungen über die Theorie der automorphen Funk-
tionen. I, Druck und Verlag von B. G. Teubner, Leipzig, Germany, p. 366.
Fuchs, R. (1905). Sur quelques équations différentielles linéaires du seconde ordre. C.
R. Acad. Sci. Paris 141, 555–8.
Harnad, J. (1994). Dual isomonodromic deformations and moment maps to loop
algebras. Commun. Math. Phys. 166, 337–65. rep-th/9301076.
Hitchin, N. J. (1995a). Poncelet polygons and the Painlevé equations. Geometry and
analysis (Bombay, 1992), Tata Institute of Fundamental Research, Bombay, pp. 151–
85.
Hitchin, N. J. (1995b). Twistor spaces, Einstein metrics and isomonodromic deforma-
tions. J. Differential Geom. 42(1), 30–112.
Hitchin, N. J. (2003). A lecture on the octahedron. Bull. London Math. Soc. 35,
577–600.
Iwasaki, K. (2008). Finite branch solutions to Painlevé VI around a fixed singular point.
Adv. Math 217(5), 1889–1934. arXiv:0704.0679.
Iwasaki, K. (2002). A modular group action on cubic surfaces and the monodromy of
the Painlevé VI equation. Proc. Japan Acad., Ser. A 78, 131–5.
Jimbo, M. (1982). Monodromy problem and the boundary condition for some Painlevé
equations. Publ. Res. Inst. Math. Sci. 18(3), 1137–61, MR 85c:58050.
Jimbo, M. and Miwa, T. (1981). Monodromy preserving deformations of linear differ-
ential equations with rational coefficients II. Physica 2D, 407–48.
Katz, N. M. (1996). Rigid local systems. Annals of Mathematics Studies, Vol. 139,
Princeton University Press, Princeton, NJ.
Killing, W. (1889). Die Zusammensetzung der stetigen endlichen Transformationsgrup-
pen II. Math. Ann. 33, 1–48. Available via http://dz1.gdz-cms.de.
Kitaev, A. V. (1991). Quadratic transformations for the sixth Painlevé equation. Lett.
Math. Phys. 21(2), 105–11.
Kitaev, A. V. (2002). Special functions of isomonodromy type, rational transformations
of the spectral parameter, and algebraic solutions of the sixth Painlevé equation.
Algebra i Analiz 14(3), 121–39. nlin.SI/0102026.
Kitaev, A. V. (2005). Grothendieck’s dessins d’enfants, their deformations, and alge-
braic solutions of the sixth Painlevé and Gauss hypergeometric equations. Algebra i
Analiz 17(1), 224–75, nlin.SI/0309078.
Kitaev, A. V. (2006). Remarks towards a classification of RS24 (3)-transformations and
algebraic solutions of the sixth Painlevé equation. Sémin. Congr., Vol. 14, Society of
Mathematics France, Paris, pp. 199–227. math. CA/0503082.
Kitaev A. V. and Vidūnas, R. (2007). Quadratic transformations of the sixth Painlevé
equation with application to algebraic solutions, math.CA/0511149. Math. Nachr.
280(16), 1834–1855.
Klein, F. (1956). Lectures on the Icosahedron and the Solution of Equations of the Fifth
Degree, Dover Publications Inc., New York.
Manin, Yu. I. (1998). Sixth Painlevé equation, universal elliptic curve, and mirror of P2 ,
Geometry of differential equations. AMS Transl. Ser. 2, Vol. 186, AMS, Providence,
RI, pp. 131–51. alg-geom/9605010.
Mason L. J. and Woodhouse, N. M. J. (1996). Integrability, Self-Duality and Twistor
Theory, O.U.P., Oxford.
Mazzocco, M. (2001). Rational solutions of the Painlevé VI equation. J. Phys. A 34(11),
2281–94. See also nlin.SI/0007036 v2, March 2004.
236 Towards a non-linear Schwarz’s list

Mazzocco, M. (2002). Painlevé sixth equation as isomonodromic deformations equa-

tion of an irregular system. The Kowalevski Property (Leeds, 2000), CRM
Proc. Lecture Notes, Vol. 32, American Mathematical Society, Providence, RI,
pp. 219–38.
Nitsure, N. (1993). Moduli of semistable logarithmic connections. J. Am. Math. Soc.
6(3), 597–609.
Okamoto, K. (1987). Studies on the Painlevé equations. I. Sixth Painlevé equation PVI .
Ann. Mat. Pura Appl. (4) 146, 337–81.
Picard, E. (1889). Mémoire sur la théorie des fonctions algébriques de deux variables.
J. de Math. Pures et Appliquées (dit ‘Journal de Liouville’) 4(5), 135–320. Available
via Gallica-Math.
Ramani, A. Grammaticos, B. and Tamizhmani, T. (2000). Quadratic relations in
continuous and discrete Painlevé equations. J. Phys. A 33(15), 3033–44.
Saito, K. (1993). On a linear structure of the quotient variety by a finite reflexion group.
Publ. Res. Inst. Math. Sci. 29(4), 535–79.
Schwarz, H. A. (1873). Über diejenigen Fälle, in welchen die Gaussische hyperge-
ometrische reihe eine algebraische Funktion ihres vierten Elements darstellt. J. Reine
Angew. Math. 75, 292–335.
Shephard G. C. and Todd, J. A. (1954). Finite unitary reflection groups. Can. J. Math.
6, 274–304.
Simpson, C. T. (1994). Moduli of representations of the fundamental group of a smooth
projective variety, I, II, Publ. Math. I.H.E.S. 79 and 80, 47–129, and 5–79.
Tsuda, T., Okamoto, K. and Sakai, H. (2005). Folding transformations of the Painlevé
equations. Math. Ann. 331, 713–38.
Watanabe, H. (1998). Birational canonical transformations and classical solutions of
the sixth Painlevé equation. Ann. Scuola Norm. Sup. Pisa (4) 27(3–4), 379–425.
Yuan, W. and Li, Y. (2002). Rational solutions of Painlevé equations. Can. J. Math.
54(3), 648–70.
 XII
AN INTRODUCTION TO BUNDLE GERBES

Michael K. Murray
Dedicated to Nigel Hitchin on the occasion of his 60th birthday

Over the years I have had many interesting mathematical conversations with
Nigel and regularly came away with a solution to a problem or a new idea.
While preparing this chapter I was trying to recall when he first told me
about gerbes. I thought for awhile that age was going to get the better of
my memory as many conversations seemed to have blurred together. But then
I discovered that the annual departmental research reports really do have
their uses. In July 1992 I attended the ‘Symposium on gauge theories and
topology’ at Warwick and reported in the 1992 Departmental Research Report
that:

I . . . had discussions with Nigel Hitchin about ‘gerbes’. These are a generalisation
of line bundles . . .

Further searching of my electronic files revealed an order for Brylinski’s book

Loop Spaces, Characteristic Classes and Geometric Quantization on 29 April
1993. I recall that the book took some months to makes its away across the sea to
Australia during which time I pondered the advertising material I had which said
that gerbes were fibrations of groupoids. Trying to interpret this led to a paper on
bundle gerbes which I submitted to Nigel in his role as a London Mathematical
Society Editor on 25 July 1994. The Departmental Research Report of the same
year reports that:

This year I began some work on a geometric construction called a bundle gerbe.
These provide a geometric realisation of the three dimensional cohomology of a
manifold.

My sincere thanks to Nigel for introducing me to gerbes and for the many
other fascinating insights into mathematics that he has given me over the
years.
238 An introduction to bundle gerbes

12.1 Introduction
The theory of gerbes began with Giraud (1971) and was popularized in the book
by Brylinski (1993). A short introduction by Nigel Hitchin (2003) in the ‘What
is a gerbe?’ series can be found in the Notices of the AMS. Gerbes provide
a geometric realization of the three-dimensional cohomology of a manifold in
a manner analogous to the way a line bundle is a geometric realization of
two-dimensional cohomology. Part of the reason for their recent popularity is
applications to string theory in particular the notion of the B-field. Strings on
a manifold are elements in the loop space of the manifold and we would expect
their quantization to involve a Hermitian line bundle on the loop space arising
from a two class on the loop space. That two class can arise as the transgression
of some three class on the underlying manifold. Gerbes provide a geometrization
of this process. String theory however is not the only application of gerbes and
we refer the interested reader to the related work of Hitchin (2001, 2006) which
applies gerbes to generalized geometry and to reviews such as (Carey et al.
2000) and (Mickelsson 2006) which give applications of gerbes to other problems
in quantum field theory.
As with everything else in the theory of gerbes, the relationship of bundle
gerbes to gerbes is best understood by comparison with the case of Hermitian
line bundles or equivalently U (1) (principal) bundles. There are basically three
ways of thinking about U (1) bundles over a manifold M :

1. A certain kind of locally free sheaf on M

2. A co-cycle gαβ : Uα ∩ Uβ → U (1) for some open cover U = {Uα | α ∈ I}
of M
3. A principal U (1) bundle P → M

In the case of gerbes over M we can think of these as

1. A certain kind of sheaf of groupoids on M (Giraud and Brylinski)

2. A co-cycle gαβγ : Uα ∩ Uβ ∩ Uγ → U (1) for some open cover U = {Uα | α ∈
I} of M or alternatively a choice of U (1) bundle Pαβ → Uα ∩ Uβ for each
double overlap (Hitchin and Chatterjee)
3. A bundle gerbe (Murray)

Note that we are slightly abusing the definition of gerbe here as what we are
considering are gerbes with band the sheaf of smooth functions from M into U (1).
There are more general kinds of gerbes on M just as there are more general kinds
of sheaves on M beyond those arising as the sheave of sections of a Hermitian
line bundle.
Recall some of the basic facts about U (1) bundles on a manifold M :

1. If P → M is a U (1) bundle there is a dual bundle P ∗ → M and if Q → M

is another U (1) bundle there is a product P ⊗ Q → M .
 Introduction 239

2. If f : N → M is a smooth map there is a pullback bundle f ∗ (P ) → N and

this behaves well with respect to dual and product. That is, f ∗ (P ∗ ) and
(f ∗ (P ))∗ are isomorphic as also are f ∗ (P ⊗ Q) and f ∗ (P ) ⊗ f ∗ (Q).
3. Associated to a U (1) bundle P → M is a characteristic class, the chern
class, c(P ) ∈ H 2 (M, Z), which is natural with respect to pullback, that is,
f ∗ (c(P )) = c(f ∗ (P )) and additive with respect to product and dual, that
is, c(P ⊗ Q) = c(P ) + c(Q) and c(P ∗ ) = −c(P ).
4. P → M is called trivial if it is isomorphic to M × U (1) or equivalently
admits a global section. P is trivial if and only if c(P ) = 0.
5. There is a notion of a connection on P → M . Associated to a connection A
on P is a closed two-form FA called the curvature of A with the property
that FA /2πi is a de Rham representative for the image of c(P ) in real
cohomology.
6. If γ : S 1 → M is a loop in M and P → M a U (1) bundle with connection
A then parallel transport around γ defines the holonomy, hol(A, γ) of A
around γ which is an element of U (1). If γ is the boundary of a disc D ⊂ M
then we have

 
hol(A, ∂D) = exp FA .
D

A gerbe is an attempt to generalize all the above facts about U (1) bundles to
some new kind of mathematical object in such a way that the characteristic class
is in three-dimensional cohomology. Obviously for consistency other dimensions
then have to change. In particular the curvature should be a three-form and
holonomy should be over two-dimensional submanifolds. It turns out to be useful
to consider the general case of any dimension of cohomology which we call a
p-gerbe. For historical reasons a p-gerbe has a characteristic class in H p+2 (M, Z)
so the interesting values of p are −2, −1, 0, 1, . . . with U (1) bundles corresponding
to p = 0.
A p-gerbe then is some mathematical object which represents (p + 2)-
dimensional cohomology. To make completely precise what representing (p + 2)-
dimensional cohomology means would take us to far afield from the present
topic, but we give a sketch here to motivate the behaviour we are looking for
in p-gerbes. To this end we will assume our p-gerbes P live in some category G
and there is a (forgetful) functor Π : G → Man the category of manifolds. The
functor Π and the category G have to satisfy

1. If P is a p-gerbe there is a dual p-gerbe P ∗ and if Q is another p-gerbe

there is a product p-gerbe P ⊗ Q. In other words G is monoidal and has a
dual operation.
2. If f : N → M is a smooth map and Π(P ) = M there is a pullback p-gerbe
f ∗ (P ) and a morphism fˆ: f ∗ (P ) → P such that Π(f ∗ (P )) = N and π(fˆ) =
f . Pullback should behave well with respect to dual and product. That is,
240 An introduction to bundle gerbes

f ∗ (P ∗ ) and (f ∗ (P )∗ should be isomorphic as also should be f ∗ (P ⊗ Q) and

f ∗ (P ) ⊗ f ∗ (Q).
3. Associated to a p-gerbe P is a characteristic class c(P ) ∈ H p+2 (Π(P ), Z),
which is natural with respect to pullback, that is, f ∗ (c(P )) = c(f ∗ (P )) and
additive with respect to product and dual, that is, c(P ⊗ Q) = c(P ) + c(Q)
and c(P ∗ ) = −c(P ).
4. As well as the notion of P and Q being isomorphic there is a possibly
weaker notion of equivalence where P and Q are equivalent if and only if
c(P ) = c(Q). We say P is trivial if c(P ) = 0.
5. There is a notion of a connective structure A on P . Associated to a
connective structure A on P is a closed (p + 2)-form ω on Π(P ) called
the (p + 2)-curvature of A with the property that ω/2πi is a de Rham
representative for the image of c(P ) in real cohomology.
6. If X ⊂ Π(P ) is an oriented (p + 1)-dimensional submanifold of Π(P )
we should be able to define the holonomy of the connective structure
hol(A, X) ∈ U(1) over X. Moreover if Y ⊂ Π(M ) is an oriented (p + 2)-
dimensional submanifold with boundary then we want to have that

 
hol(A, ∂Y) = exp Fω .
Y

Clearly by construction the category of U (1) bundles, with the forgetful functor
which assigns to a U (1) bundle its base manifold, is an example of a 0-gerbe.
Before we consider other examples we need some facts about bundles with
structure group an abelian Lie group H. If P → M is an H bundle on M
then by choosing local sections of P for an open cover U = {Uα | α ∈ I} we
can construct transition functions hαβ : Uα ∩ Uβ → H and in the usual way this
defines a class c(P ) ∈ H 1 (M, H) where here we abuse notation and write H for
what is really the sheaf of smooth functions with values in H. It is a standard
fact that isomorphism classes of H bundles are in bijective correspondence with
H 1 (M, H) in this manner. If P → M is an H bundle we can define its dual as
follows. Let P ∗ be isomorphic to P as a manifold with projection to M and for
convenience let p∗ ∈ P ∗ denote p ∈ P thought of as an element of P ∗ . Then define
a new H action on P ∗ by p∗ h = (ph−1 )∗ . It is ovious that if hαβ are transition
functions for P then h∗αβ = h−1 ∗
αβ are transition functions for P . In particular we
have that c(P ∗ ) = −c(P ) if we write the group structure on H 1 (M, H) additively.
If Q is another H bundle we can form the fibre product P ×M Q and let H act
on it by (p, q)h = (ph, qh−1 ). Denote the orbit of (p, q) under this action by [p, q]
and define an H action by [p, q]h = [p, qh] = [ph, q]. The resulting H bundle
is denoted by P ⊗ Q → M . If hαβ are transition functions for P and kαβ are
transition functions for Q then hαβ kαβ are transition functions for P ⊗ Q and
thus c(P ⊗ Q) = c(P ) + c(Q). Notice that these constructions will not generally
work for non-abelian groups because in such a case the action of H on P ∗ is not
a right action and the action on P ⊗ Q is not even well-defined.
 Introduction 241

Example 12.1 The simplest example is that of functions f : M → Z where

we define the functor Π by Π(f ) = M . The degree of f is the class induced in
H 0 (M, Z) so functions from M to Z are −2-gerbes over M . Product and dual
are pointwise addition and negation. There is no sensible notion of connective
structure.

Example 12.2 Consider next principal Z bundles P → M . Clearly we want

the functor Π to be Π(P ) = M and pullbacks are well known to exist. As Z is
abelian the constructions above apply and there are duals and products. The
isomorphism class of a bundle is determined by a class in H 1 (M, Z) so Z bundles
are p = −1 gerbes. A Z bundle is trivial as a −1-gerbe if and only if it is trivial
as a Z bundle.
It is not immediately obvious what a connective structure on a Z bundle is but
it turns out that the correct notion is that of a Z equivariant map φ̂ : P → iR
where the action of n ∈ Z on iR is addition of 2πin so that φ̂(pn) = φ̂(p) + 2πin.
The map φ̂ then descends to a map φ : M → S 1 and the class of the bundle is
the degree of this map. The pullback of the standard one-form on R, that is, dφ̂
is a one-form on P which descends to a one-form φ−1 dφ on M . The de Rham
class (φ−1 dφ)/2πi is the image of the class of the bundle in real cohomology.
We expect holonomy to be over a −1 + 1 = 0-dimensional submanifold. If
m1 , . . . , mr is a collection
 of points in M with each mi oriented by some i ∈ {±1}
let us denote by i mi their union as an oriented zero-dimensional submanifold
of M . Then we define
   Fr
hol φ̂, m
i i = φ(mi )i .
i=1

In the case of an oriented one-dimensional submanifold X ⊂ M with ends −X0

and +X1 the fundamental theorem of calculus tells us that
 
 
hol φ̂, X1 − X0 = exp dφ .
X

Notice that if we express a Z bundle locally in terms of transitions functions

these are maps of the form fαβ : Uα ∩ Uβ → Z. That is, over each double overlap
we have a −2-gerbe.

Example 12.3 It is clear from the above example that maps φ : M → U (1)
are also −1-gerbes with a connective structure. The dual and product are
just pointwise inverse and pointwise product. The class is the degree and the
connective structure is included automatically as part of φ.
We can also forget that there is a natural connective structure and just regard
maps φ : M → U (1) as −1-gerbes. In that case the natural notion of isomorphism
between two maps φ, χ : M → U (1) would be equality. However two such maps
have the same degree if and only if they are homotopic. So the notion of
242 An introduction to bundle gerbes

equivalence of maps φ : M → U (1), thought of as −1-gerbes (without connective

structure), should be homotopy and is different to the notion of isomorphism.
Example 12.4 As we have remarked U (1) bundles are, of course, p = 0 gerbes.
Notice that locally a U (1) bundle is given by transition functions gαβ : Uα ∩
Uβ → U (1), that is, on each double overlap we have a −1-gerbe (with connective
structure).
We will see below that this pattern of a p-gerbe being defined as a (p − 1)-gerbe
on double overlaps of some open cover is exploited by Hitchin and Chatterjee
to give a definition of a 1-gerbe. But first we need some additional background
material.

12.2 Background
We will be interested in surjective submersions π : Y → M which we regard as
generalizations of open covers. In particular if U = {Uα | α ∈ I} is an open cover
we have the disjoint union
YU = {(x, α) | x ∈ Uα } ⊂ M × I
with projection map π(x, α) = x. The surjective morphism π : YU → M is called
the nerve of the open cover U.
A morphism of surjective submersions π : Y → M and p : X → M is a map
ρ : Y → X covering the identity, that is, p ◦ ρ = π. Any surjective submersion
π : Y → M admits local sections so there is an open cover U of M and local
sections sα : Uα → Y of π. These local sections define a morphism s : YU → Y
by s(x, α) = sα (x). Indeed any morphism YU → Y will be of this form. If V =
{Vα | α ∈ J} is a refinement of U, that is, there is a map ρ : J → I such that for
every α ∈ J we have Vα ⊂ Uρ (α), we have a morphism of surjective submersions
YV → YU defined by (α, x) → (ρ(α), x).
Given a surjective morphism π : Y → M we can form the p-fold fibre product
Y [p] = {(y1 , . . . , yp ) | π(y1 ) = · · · = π(yp )} ⊂ Y p .
The submersion property of π implies that Y [p] is a submanifold of Y p . There
are smooth maps πi : Y [p] → Y [p−1] , for i = 1, . . . , p, defined by omitting the ith
element. We will be interested in two particular examples.
[p]
Example 12.5 If U is an open cover of M then the pth fibre product YU is the
disjoint union of all the ordered p-fold intersections. For example, if U = {U1 , U2 }
[2]
is an open cover of M then YU is the disjoint union of U1 ∩ U2 , U2 ∩ U1 , U1 ∩ U1 ,
and U2 ∩ U2 .
Example 12.6 If P → M is a principal G bundle then P → M is a surjective
submersion. It is easy to show that P [p] = P × Gp−1 . In particular P [2] = P × G
and we shall need later the related fact that there is a map g : P [2] → G defined
by p1 g(p1 , p2 ) = p2 .
 Bundle gerbes 243

Let Ωq (Y [p] ) be the space of differential q-forms on Y [p] . Define


p
δ : Ωq (Y [p−1] ) → Ωq (Y [p] ) by δ = (−1)p−1 πi∗ .
i=1

These maps form the fundamental complex

π∗ δ δ δ
0 → Ωq (M ) → Ωq (Y ) → Ωq (Y [2] ) → Ωq (Y [3] ) → . . .
and from Murray (1996) we have
Proposition 12.1 The fundamental complex is exact for all q ≥ 0.
Note that if Y = YU then this proposition is a well-known result about the
Čech de Rham double complex (see e.g. Bott and Tu’s book 1982).
Finally we need some notation. Let H be an abelian group. If g : Y [p−1] → H
we define δ(g) : Y [p] → H by
δ(g) = (g ◦ π1 ) − (g ◦ π2 ) + (g ◦ π3 ) · · · .
If P → Y [p−1] is an H bundle we define an H bundle δ(P ) → Y [p] by
δ(P ) = π1∗ (P ) ⊗ (π2∗ (P ))∗ ⊗ π3∗ (P ) ⊗ · · · .
It is easy to check that δ(δ(g)) = 1 and that δ(δ(P )) is canonically trivial.

12.3 Bundle gerbes

Definition 12.1 A bundle gerbe (Murray 1996) over M is a pair (P, Y ) where
Y → M is a surjective submersion and P → Y [2] is a U (1) bundle satisfying
1. There is a bundle gerbe multiplication which is a smooth isomorphism
m : π3∗ (P ) ⊗ π1∗ (P ) → π2∗ (P )
of U (1) bundles over Y [3] .
2. This multiplication is associative, that is, if we let P(y1 ,y2 ) denote the fibre of
P over (y1 , y2 ) then the following diagram commutes for all (y1 , y2 , y3 , y4 ) ∈
Y [4] :
P(y1 ,y2 ) ⊗ P(y2 ,y3 ) ⊗ P(y3 ,y4 ) → P(y1 ,y3 ) ⊗ P(y3 ,y4 )
↓ ↓
P(y1 ,y2 ) ⊗ P(y2 ,y4 ) → P(y1 ,y4 )
We remark that for any (y1 , y2 , y3 ) ∈ Y [3] the bundle gerbe multiplication
defines an isomorphism:
m : P(y1 ,y2 ) ⊗ P(y2 ,y3 ) → P(y1 ,y3 )
of U (1) spaces.
We can show using the bundle gerbe multiplication that there are natural
isomorphisms P(y1 ,y2 ) ∼ ∗
= P(y2 ,y1 )
and P(y,y)  U (1).
244 An introduction to bundle gerbes

We can rephrase the existence and associativity of the bundle gerbe multi-
plication to an equivalent pair of conditions in the following way. The bundle
gerbe multiplication gives rise to a section s of δ(P ) → Y [3] . Moreover δ(s)
is a section of δ(δ(P )) → Y [4] . But δ(δ(P )) is canonically trivial so it makes
sense to ask that δ(s) = 1. This is the condition of associativity. The family of
spaces {Y [p] | p = 1, 2, . . . } is an example of a simplicial space (Dupont 1978)
and by comparing to Brylinski and McLaughlin (1994) we see that a bundle
gerbe is the same thing as a simplicial line bundle over this particular simplicial
space.
Example 12.7 If we replace Y in the definition by YU for some open cover U of
M we obtain the definition of gerbe given by Hitchin (2001) and by his student
Chatterjee (1998). This consists of choosing an open cover U of M and a family
of U (1) bundles P : Uα ∩ Uβ such that over triple overlaps we have sections
∗
sαβγ ∈ Γ(Uα ∩ Uβ ∩ Uγ | Pβγ ⊗ Pαγ ⊗ Pαβ )
and we require that δ(s) = 1 in the appropriate way.
Example 12.8 The simplest example of a line bundle is given by the clutching
construction on the two sphere S 2 . If U0 and U1 are the open neighbourhoods of
the north and south hemispheres we take the transition function g : U0 ∩ U1 →
U (1) to have winding number 1. As there are only two open sets there is no
condition on triple overlaps and we obtain the U (1) bundle over S 2 of chern
class 1. In a similar fashion we can consider U0 and U1 to be open neighbourhoods
of the north and south hemispheres of the three-sphere S 3 . Their intersection is
retractable to the two-sphere so we can choose over this the U (1) bundle P of
chern class 1. Again there are no additional conditions and we obtain the gerbe
of degree 1 over S 3 .
Example 12.9 Hitchin and Chatterjee also consider a gerbe as in Example
12.7 but with the added requirement that each Pαβ is trivial in the form Pαβ =
[2]
Uα ∩ Uβ × U (1). Writing elements of the disjoint union YU as (α, β, x) where
x ∈ Uα ∩ Uβ we see that the bundle gerbe multiplication must take the form
((α, β, x), z) ⊗ ((β, γ, x), w) → ((α, γ, x), zwgαβγ (x))
for some gαβγ : Uα ∩ Uβ ∩ Uγ → U (1) and will be associative precisely when gαβγ
is a co-cycle.
We will refer to gerbes of the forms in Examples 12.7 and 12.9 as Hitchin–
Chatterjee gerbes. The connection with bundle gerbes is simple. For clarity we
define
Definition 12.2 A bundle gerbe (P, Y ) over M is called local if Y = YU for
some open cover U of M .
We then obviously have
 Bundle gerbes 245

Proposition 12.2 A Hitchin–Chatterjee gerbe is the same thing as a local

bundle gerbe.
If (P, Y ) is a bundle gerbe over M then associated to every point m of M we
have a groupoid constructed as follows. The objects are the elements of the fibre
Ym and the morphisms between y1 and y2 in Ym are P(y1 ,y2 ) . Composition comes
from the bundle gerbe multiplication. If we call a groupoid a U (1) groupoid if
it is transitive and the group of morphisms of a point is isomorphic to U (1),
then the algebraic conditions on the bundle gerbe (i.e. the multiplication and its
associativity) are captured precisely by saying that a bundle gerbe is a bundle
of U (1) groupoids over M .
We now consider the properties given in Section 12.1 which we would like a
1-gerbe to satisfy and show how they are satisfied by bundle gerbes.

12.3.1 Pullback
If f : N → M then we can pullback Y → M to f ∗ (Y ) → N with a map
fˆ: f ∗ (Y ) → Y covering f . There is an induced map fˆ[2] : f ∗ (Y ) → Y [2] . Let
[2]

f ∗ (P, Y ) = (fˆ[2]∗ (P ), f ∗ (Y )).

To see this is a bundle gerbe notice that all this is doing is pulling back the
U (1) groupoid at f (n) ∈ M and placing it at n ∈ N so we have a bundle of U (1)
groupoids over N and thus a bundle gerbe.

12.3.2 Dual and product

If (P, Y ) is a bundle gerbe then (P, Y )∗ = (P ∗ , Y ) is also a bundle gerbe called
the dual of (P, Y ).
If (P, Y ) and (Q, X) are bundle gerbes we can form the fibre product Y ×M
X → M , a new surjective submersion and then define a U (1) bundle
P ⊗ Q → (Y ×M X)[2] = Y [2] ×M X [2]
by
(P ⊗ Q)((y1 ,x1 ),(y2 ,x2 )) = P(y1 ,y2 ) ⊗ Q(x1 ,x2 ) .
We define (P, Y ) ⊗ (Q, X) = (P ⊗ Q, Y ×M X).

12.3.3 Characteristic class

The characteristic class of a bundle gerbe is called the Dixmier–Douady class.
We construct it as follows. Choose a good cover U of M (Bott and Tu 1982) with
sections sα : Uα → Y . Then
(sα , sβ ) : Uα ∩ Uβ → Y [2]
is a section. Choose a section σαβ of Pαβ = (sα , sβ )∗ (P ). That is, some
σαβ : Uα ∩ Uβ → P
246 An introduction to bundle gerbes

such that σαβ (x) ∈ P(sα (x),sβ (x)) . Over triple overlaps we have
m(σαβ (x), σβγ (x)) = gαβγ (x)σαγ (x) ∈ P(sα (x),sγ (x))
for gαβγ : Uα ∩ Uβ ∩ Uγ → U (1). This defines a co-cycle which is the Dixmier–
Douady class:
DD((P, Y )) = [gαβγ ] ∈ H 2 (M, U (1)) = H 3 (M, Z).
Example 12.10 If U is an open cover and gαβγ a U (1) co-cycle then we can
build a Hitchin–Chatterjee gerbe or local bundle gerbe of the type considered in
Example 12.9. It is easy to see that this has Dixmier–Douady class given by the
Čech class [gαβγ ].
Notice that this example shows that every class in H 3 (M, Z) arises as the
Dixmier–Douady class of some Hitchin–Chatterjee gerbe or of some (local)
bundle gerbe.
It is straightforward to check that if f : N → M and (P, Y ) is a bundle gerbe
over M then f ∗ (DD(P, Y )) = DD(f ∗ (P, Y )). Moreover we have
1. DD((P, Y )∗ ) = −DD(P, Y )
2. DD((P, Y ) ⊗ (Q, X)) = DD(P, Y ) + DD(Q, X)
We will defer the question of triviality of a bundle gerbe until the next section
and consider next the notion of a connective structure on a bundle gerbe.

12.3.4 Connective structure

As P → Y [2] is a U (1) bundle we can pick a connection A. Call it a bundle gerbe
connection if it respects the bundle gerbe multiplication. That is, if the section
s of δ(P ) → Y [3] satisfies s∗ (δ(A)) = 0, that is, if it is flat for δ(A). We would
like bundle gerbe connections to exist. This is a straightforward consequence of
the fact that the fundemental complex is exact. Indeed if A is any connection
consider s∗ (δ(A)); we have δ(s∗ (δ(A))) = δ(s)∗ (δδ(A)) = 0 because δδ(A) is the
flat connection on the canonically trivial bundle δδ(P ). Hence there is a one-form
a on Y [2] such that δ(a) = s∗ (δ(A)) and thus A − a is a bundle gerbe connection.
If A is a bundle gerbe connection then the curvature FA ∈ Ω2 (Y [2] ) satisfies
δ(FA ) = 0. From the exactness of the fundamental complex there must be an
f ∈ Ω2 (Y ) such that FA = δ(f ). As δ commutes with d we have δ(df ) = dδ(f ) =
dFA = 0. Hence df = π ∗ (ω) for some ω ∈ Ω3 (M ). So π ∗ (dω) = dπ ∗ (ω) = ddf = 0
and ω is closed. In fact it is a consequence of standard Čech de Rham theory
that:
1
ω = r(DD(P, Y )) ∈ H 3 (M, R).
2πi
We call f a curving for A, the pair (A, f ) a connective structure for (P, Y ),
and ω is called the three-curvature of the connective structure (A, f ). In string
theory the two-form f is called the B-field.
 Triviality 247

We can give a local description of the connective structure as follows. Assume

we have an open cover U of M with local sections sα : Uα → Y and sections over
double-overlaps σαβ of (sα , sβ )∗ (P ) → Uα ∩ Uβ . We define
Aαβ = (sα , sβ )∗ (A) ∈ Ω1 (Uα ∩ Uβ )
and
fα = s∗α (f ) ∈ Ω2 (Uα ).
These satisfy
−1
Aβγ − Aαγ + Aαβ = gαβγ dgαβγ
fβ − fα = dfαβ

and the three-curvature ω restricted to Uα is dfα .

Example 12.11 We can use this local description of the connective structure
to calculate the Dixmier–Douady class of the Hitchin–Chatterjee gerbe on the
three-sphere defined in Example 12.8. Stereographic projection from either pole
identifies S 3 − {(1, 0, 0)} and S 3 − {(−1, 0, 0)} with R3 and maps the equator
to the unit sphere S 2 ⊂ R3 . Let U0 and U1 be the pre-images of the interior of
a ball of radius 2 in R3 under both stereographic projections. We can identify
U0 ∩ U1 with S 2 × (−1, 1). Pull back the line bundle of chern class k on S 2 , with
connection A and curvature F , to U0 ∩ U1 . Because there are no triple overlaps
this is a bundle gerbe connection. If we choose a partition of unity ψ0 and ψ1
for U0 and U1 then f0 = −ψ1 F and f1 = ψ0 F define two-forms on U0 and U1 ,
respectively, satisfying F = f1 − f0 on U0 ∩ U1 . These two-forms define a curving
for the bundle gerbe connection. The curvature is the globally defined three-form
ω whose restriction to U0 and U1 is −dψ1 ∧ F and dψ0 ∧ F , respectively. The
integral of ω over the three-sphere reduces, by Stokes theorem, to the integral
of F over the two-sphere. Hence this bundle gerbe has Dixmier–Douady class
k ∈ H 3 (M, Z) = Z.
Holonomy will need to wait until we have a considered the notion of triviality
which we turn to now.

12.4 Triviality
Recall that a U (1) bundle P → M is trivial if it is isomorphic to the bundle
M × U (1) or, equivalently, has a global section. This occurs if and only if P → M
has zero Chern class. If sa : Uα → P are local sections then P is determined by
a transition function g : Uα ∩ Uβ → U (1) given by sα = sβ gαβ and P → M is
trivial if and only if there exist hα : Uα → U (1) such that
gαβ = hβ h−1
α .
248 An introduction to bundle gerbes

In an analogous way Hitchin (2001) and Chatterjee (1998) define a gerbe Pαβ →
Uα ∩ Uβ to be trivial if there are U (1) bundles Rα → Uα and isomorphisms
φαβ : Rα ⊗ Rβ∗ → Pαβ on double-overlaps in such a way that the multiplication
becomes the obvious contraction
Rα ⊗ Rβ∗ ⊗ Rβ ⊗ Rγ∗ → Rα ⊗ Rγ∗ .
In the bundle gerbe formalism this idea takes the following form (Murray and
Stevenson 2000). Let R → Y be a U (1) bundle and let δ(R) → Y [2] be defined
as above. Note that δ(R) has a natural associative bundle gerbe multiplication
given by
δ(R)(y1 ,y2 ) ⊗ δ(R)(y2 ,y3 ) = Ry1 ⊗ Ry∗2 ⊗ Ry2 ⊗ Ry∗3  Ry1 ⊗ Ry∗3 = δ(R)(y1 ,y3 ) .
Definition 12.3 A bundle gerbe (P, Y ) over M is called trivial if there is a
U (1) bundle R → Y such that (P, Y ) is isomorphic to (δ(R), Y ). In such a case
we call a choice of R and the isomorphism δ(R)  P a trivialization of (P, Y ).
Example 12.12 Let (P, Y ) be a bundle gerbe and assume that Y → M admits
a global section s : M → Y . Define R → Y by Ry = P(s(π(y)),y) . Then we have
an isomorphism
∗
δ(R)(y1 ,y2 ) = P(s(π(y2 )),y2 ) ⊗ P(s(π(y1 )),y1 )

= P(s(π(y2 )),y2 ) ⊗ P(y1 ,s(π(y1 )))

 P(y1 ,y2 )
using the bundle gerbe multiplication and the fact that s(π(y1 )) = s(π(y2 )). It
is easy to check that this isomorphism preserves the respective bundle gerbe
multiplications and we have shown that if Y admits a global section then any
bundle gerbe (P, Y ) is trivial. Notice that the converse is not true. Just take
an open cover with more than one element and gαβγ = 1 to obtain a Hitchin–
Chatterjee gerbe which has zero Dixmier–Douady class but for which YU → M
has no global section.
Consider the Dixmier–Douady class of δ(R). If sα : Uα → Y are local sections
for a good cover choose local sections ηα of s∗α (R). Then we can take as local
sections of (sα , sβ )∗ (δ(R)) the sections σαβ = ηα ⊗ ηβ∗ and it follows that the
corresponding gαβγ = 1 and δ(R) has Dixmier–Douady class equal to zero. The
converse is also true. Consider a bundle gerbe with Dixmier–Douady class zero.
So we have an open cover and σαβ such that
gαβγ = hβγ h−1
αγ hαβ .

By replacing σαβ by σαβ /hαβ we can assume that gαβγ = 1. Let Yα = π −1 (Uα )
and define Rα → Yα by letting the fibre of Rα over y ∈ Yα be P(y,sα (π(y))) .
Construct an isomorphism χαβ (y) from the fibre of Rα over y to the fibre of
 Triviality 249

Rβ over y by noting that

P(y,sα (π(y))) = P(y,sβ (π(y))) ⊗ P(sα (π(y))sβ (π(y)))

and using σαβ (π(y)) ∈ P(sα (π(y))sβ (π(y))) . Because σβγ σαγ σαβ = 1 we can show
that χαβ (y) ◦ χβγ (y) = χαγ (y) and hence the Rα clutch together to form a global
U (1) bundle R → Y . It is straightforward to check that δ(R) = P .
Consider now a U (1) bundle R → Y and assume that δ(R) → Y [2] has a section
s with δ(s) = 1 with respect to the canonical trivialization of δ(δ(R)) → Y [3] . The
section s is called descent data for R and is equivalent to R being the pullback
of a U (1) bundle on M . Indeed s constitutes a family of isomorphisms

s(y1 , y2 ) : Ry1 → Ry2

and δ(s) = 1 is equivalent to s(y2 , y3 ) ◦ s(y1 , y2 ) = s(y1 , y3 ) from which it is easy

to define a bundle on M whose pullback is R.
Assume that a bundle gerbe (P, Y ) over M is trivial and that R1 and R2
are two trivializations. Then we have δ(R1 )  P  δ(R2 ) and hence a section of
δ (R1∗ ⊗ R2 ) which is descent data for R1∗ ⊗ R2 . Thus R1 = R2 ⊗ π ∗ (Q) for some
U (1) bundle Q → M . It is easy to show the converse that if R is a trivialization
and Q → M a U (1) bundle then R ⊗ π ∗ (Q) is another trivialization. We have
now proved

Proposition 12.3 Let (P, Y ) be a bundle gerbe over M . Then

1. (P, Y ) is trivial if and only if DD(P, Y ) = 0.
2. If DD(P, Y ) = 0 then any two trivializations of (P, Y ) differ by a U (1)
bundle on M .

This should be compared to the case of U (1) bundles (0-gerbes) where two
trivializations or sections of the bundle differ by a map into U (1) which is a −1-
gerbe. The general pattern is that we expect two trivializations of a p-gerbe to
differ by a (p − 1)-gerbe. Notice also that whereas any two trivial U (1) bundles
are isomorphic there are many trivial bundle gerbes which are not isomorphic.
This leads us to the notion of stable isomorphism.

Definition 12.4 If (P, Y ) and (Q, X) are bundle gerbes over M we say they are
stably isomorphic (Murray and Stevenson 2000) if (P, Y )∗ ⊗ (Q, X) is trivial. A
choice of a trivialization is called a stable isomorphism from (P, Y ) to (Q, X).

We have

Proposition 12.4 Bundle gerbes (P, Y ) and (Q, X) over M are stably isomor-
phic if and only if DD(P, Y ) = DD(Q, X). The Dixmier–Douady class defines
a bijection between stable isomorphism classes of bundle gerbes on M and
H 3 (M, Z).
250 An introduction to bundle gerbes

Proof. Bundle gerbes (P, Y ) and (Q, X) over M are stably isomorphic if and
only if (P, Y )∗ ⊗ (Q, X) is trivial which occurs if and only if −DD(P, Y ) +
D(Q, X) = 0. We have already seen that every three class arises as the Dixmier–
Douady class of some bundle gerbe on M .

It follows that the correct notion of equivalence for bundle gerbes is stable
isomorphism. It is actually possible to compose stable isomorphisms and the
details are given in work of Stevenson (2000) where the structure of the two
category of all bundles gerbes on M is discussed (see also Waldorf 2007).
Note that we also have
Proposition 12.5 Every bundle gerbe is stably isomorphic to a Hitchin–
Chatterjee gerbe.

12.4.1 Holonomy
Consider now a bundle gerbe (P, Y ) with connective structure (A, f ) over a
surface Σ. Because H 3 (Σ, Z) = 0 we know that (P, Y ) is trivial. So there is a
U (1) bundle R → Y with δ(R) = P . Choose a connection a for R and note that
δ(a) is a connection for P using the isomorphism δ(R) = P . But δ(δ(a)) is flat
so δ(a) is a bundle gerbe connection. Hence A = δ(a) + α for a one-form α on
Y [2] with δ(α) = 0. Using the exactness of the fundamental complex we can solve
α = δ(α ) and hence show that δ(a + α ) = A. So without loss of generality we
can choose a connection a on R with δ(a) = A. Consider the two-form f − Fa .
This satisfies δ(f − Fa ) = FA − Fδ(a) = 0 so f − FA = π ∗ (µa ) for some two-form
µa on Σ. Define the holonomy of (A, f ) over Σ by

 
hol((A, f ), Σ) = exp µa
Σ

and note that this is independent of the choice of trivialization R and connection
a. Indeed any two trivializations with connection will differ by a U (1) bundle
on M with connection and the corresponding µa will differ by the curvature
of the connection on that U (1) bundle. But the integral of the curvature of a
U (1) bundle over a closed surface is in 2πiZ so the two definitions of holonomy
agree.
If (P, Y ) is a bundle gerbe with connective structure (A, f ) on a general
manifold M and Σ ⊂ M is a submanifold we can pull (P, Y ) and (A, f ) back
to Σ and define hol((A, f ), Σ) as above. In this more general setting if X ⊂ M is
a three-dimensional submanifold with boundary ∂X also a submanifold of M we
can trivialize (P, Y ) over all of X and repeat the construction above. We then
have dµa = ω, the three-curvature of (A, f ), and thus

 
hol((A, f ), ∂X) = exp ω
X

the final property in Section 12.1 which we wanted a 1-gerbe to satisfy.

 Examples of bundle gerbes 251

Assume that we have a local description for (P, Y ) and (A, f ) as in Section
12.3.4 in terms of gαβγ , Aαβ , and fα . Then there is a remarkable formula
(Gawedzki and Reis 2002) for the holonomy, first proposed in Alvarez (1985)
‘
and also Gawedzki (1988) and subsequently derived by a number of authors,
‘
which can be described as follows. Choose a triangulation $ of Σ and a map
χ : $ → I such that for any simplex σ ∈ $ we have σ ⊂ Uχ(σ) . Write σ2 , σ 1 ,
and σ 0 for two-, one-, and zero-dimensional simplices, that is, faces, edges, and
vertices, respectively. Then

hol((A, f ), Σ) =
   
   F
exp fχ(σ 2 ) exp Aχ(σ2 )χ(σ 1 ) gχ(σ2 )χ(σ 1 )χ(σ 0 ) (σ 0 ).
σ2 σ2 σ 1 ⊂σ 2 σ1 σ 0 ⊂σ1 ⊂σ 2

12.4.2 Obstructions to certain kinds of Y → M

When we consider the examples in the next section it will become apparent that
they tend to cluster into two kinds: either Y → M has infinite-dimensional fibres
or Y → M has discrete fibres as in the case of Hitchin–Chatterjee gerbes. There
is a reason for this which is the following result:
Proposition 12.6 Let (P, Y ) be a bundle gerbe over M with Y → M a finite-
dimensional fibration with both M and the fibres of Y → M one connected. Then
the Dixmier–Douady class of (P, Y ) is torsion.
Proof. The proof uses the result from Gotay et al. (1983) which shows that if
Y → M satisfies the hypothesis we have stated in the proposition and moreover
there is a smooth two-form µ defined on the vertical tangent bundle of Y → M
which is closed on each fibre then we can extend µ to a global, closed two-
form on Y . To apply this choose a connection and curving f for (P, Y ). If we
restrict f to any fibre of Y → M it agrees with F and hence is closed in the
fibre directions. So there exists a global two-form µ on Y → M which agrees
with f in the vertical directions. Consider ρ = f − µ. Both ρ and dρ are zero in
the vertical directions so that ρ = π ∗ (χ) for some two-form χ on M . But then
π ∗ (dρ) = d(f − µ) = π ∗ (ω) where ω is the three-curvature. Hence the image of
the Dixmier–Douady class in real cohomology is zero so the Dixmier–Douady
class is torsion.

12.5 Examples of bundle gerbes

We expect to find bundle gerbes on manifolds M which have three-dimensional
cohomology. There are two, related, cases we will discuss here. The first comes
from lifting problems for principal bundles, the so-called lifting bundle gerbe, and
the second is to take M a simple, compact Lie group.
252 An introduction to bundle gerbes

12.5.1 Lifting bundle gerbe

Consider a central extension of Lie groups
ρ
0 → U (1) → Ĝ → G → 0

so that U (1) is the kernel of p and is in the centre of G. If Y → M is a G principal

bundle then we can ask if it lifts to a Ĝ bundle Ŷ → M . That is, can we find a Ĝ
bundle Ŷ → M and a bundle morphism f : Ŷ → Y such that f (pg) = f (p)ρ(g)
for all p ∈ Ŷ and g ∈ Ĝ? There is a well-known topological obstruction to the
existence of Ŷ which we can calculate as follows. Choose a good open cover U of
M and local sections of Y that give rise to a transition function

gαβ : Uα ∩ Uβ → G

in the usual way. Because these double overlaps are all contractible we can choose
lifts of each gαβ to ĝαβ : Uα ∩ Uβ → Ĝ. Note that
−1
αβγ = ĝβγ ĝαγ ĝαβ

takes values in U (1). It is not difficult to show that αβγ is a U (1) valued co-cycle
and that the class

[ αβγ ] ∈ H 2 (M, U (1)  H 3 (M, Z)

vanishes if and only if P lifts to a Ĝ bundle.

Given a G bundle Y → M there is a map g : Y [2] → G defined by y1 g(y1 , y2 ) =
y2 . Notice that Ĝ → G is a U (1) bundle so we can pull it back to define a U (1)
bundle Q → Y [2] whose fibres are cosets of U (1) in Ĝ defined by

Q(y1 ,y2 ) = U (1)g(y1 , y2 ).

Because g(y1 , y2 )g(y2 , y3 ) = g(y1 , y3 ) the product of an element in the coset

containing g(y1 , y2 ) with an element in the coset containing g(y2 , y3 ) will be
an element in the coset containing g(y1 , y3 ) which defines a bundle gerbe multi-
plication:

Q(y1 ,y2 ) ⊗ Q(y2 ,y3 ) → Q(y1 ,y3 ) .

The bundle gerbe (Q, Y ) is called the lifting bundle gerbe of Y (Murray 1996).
It is easy to check that the lifting bundle gerbe has Dixmier–Douady class
precisely the obstruction to lifting the bundle Y → M . Indeed if we follow
through the construction in Section 12.3.3 we find that σ̂αβ = ĝαβ . It follows
from the discussion above that the lifting bundle gerbe is trivial if and only if
the bundle Y → M lifts to Ĝ. In fact this follows directly because a lift Ŷ → Y
will be a U (1) bundle over Y and actually a trivialization of the lifting bundle
gerbe defined above.
 Examples of bundle gerbes 253

12.5.2 Projective bundles

Let H be a Hilbert space, possibly finite-dimensional. A Hilbert bundle with
fibre H can be regarded locally as a collection of transition functions

gαβ : Uα ∩ Uβ → U (H)

satisfying the co-cycle condition

−1
gβγ gαγ gαβ = 1.

In some situations we have slightly less than this, namely, a collection of

transition functions

gαβ : Uα ∩ Uβ → U (H)

satisfying
−1
gβγ gαγ gαβ = αβγ 1U (H) .

If we denote by ρ : U (H) → P U (H) the projection onto the projective unitary

group we see that have
 −1 
ρ(gβγ )ρ gαγ ρ(gαβ ) = ρ( αβγ 1) = 1P U (H)

so there is a well-defined bundle of projective spaces or a principal P U (H)

bundle. Given a projective bundle a natural question is to ask when it is the
projectivization of a global Hilbert bundle. This is equivalent to lifting the
P U (H) bundle to U (H). The obstruction to this lifting in the class is

[ αβγ ] ∈ H 2 (M, U (1))

arising from the sequence

0 → U (1) → U (H) → P U (H) → 0.

In the finite-dimensional class we can also consider the obstruction to lifting

from P U (n) to SU (n) via the exact sequence

0 → Zn → SU (n) → P U (n) → 0.

In this case the lifting bundle gerbe is really a Zn ⊂ U (1) bundle gerbe and the
Dixmier–Douady class is a torsion class in the image of the Bockstein map

H 2 (M, Zn ) → H 3 (M, Z).

For further details on Zn bundle gerbes see Carey et al. (2000). Note finally
that if Y → M is a P U (n) bundle then the lifting bundle gerbe has finite-
dimensional fibres so the fact that its Dixmier–Douady class is torsion is implied
by Proposition 12.6.
254 An introduction to bundle gerbes

12.5.3 Bundle gerbes on Lie groups

If G is a compact, simple, Lie group then H 3 (G, Z)  Z so we expect to find
bundle gerbes on G. There are, in fact, a number of constructions of bundle
gerbes on G and we have seen one already for SU (2)  S 3 in Section 12.3.2. For
convenience let us fix an orientation of G and call the bundle gerbe over G of
Dixmier–Douady class 1 the basic bundle gerbe on G.
Example 12.13 Let P G be the based path space of all smooth maps g of
the interval [0, 1] into G with g(0) = 1. Let ev : PG → G be the evaluation map
ev(g) = g(1). The kernel of ev is the space of based loops Ω(G), that is, all smooth
maps g : [0, 1] → G for which g(0) = 1 = g(1). In the smooth Fréchet topology
P G → G is a principal Ω(G) bundle. Moreover there is a well-known central
extension (Pressley and Segal 1986)
G
0 → U (1) → Ω(G) → Ω(G) → 0
where Ω̂(G) is the Kac–Moody group. Hence there is a corresponding lifting
bundle gerbe over G which is the basic bundle gerbe. It is possible to give an
explicit construction of this gerbe over G (see Murray 1996 and also Stevenson
2000).
If we wish to avoid the infinite-dimensional spaces there is a construction of
the basic bundle gerbe over G in the simply connected case due to Meinrenken
(2003) and see also Gawedzki and Reis (2004) for the non-simply connected
‘
case. This construction has disconnected fibres for Y → G and uses the standard
structure theory of compact, simple, simply connected Lie groups.
Example 12.14 For G = SU (n) there is a simple construction of the basic
bundle gerbe due to Meinrenken (2003) (see also Mickelsson 2003) using an open
cover of G which can be presented without the cover as follows. Let
Y = {(g, z) | det(g − z1) = 0} ⊂ G × U (1).
For convenience let us write an element ((g, z), (g, w)) ∈ Y [2] as (g, w, z). If u ∈
U (1) and u = w and u = z let us say that u is between w and z if an anticlockwise
rotation of z into w passes through u. Then let W(g,w,z) be the sum of all the
eigenspaces of g for eigenvalues between z and w and define P(g,w,z) to be the
U (1) frame bundle of det W(g,w,z) . To define the bundle gerbe product notice
that if u is between w and z then
W(g,w,u) ⊕ W(g,u,z) ⊕ W(g,z,w) = Cn
so that
det W(g,w,u) ⊗ det W(g,u,z) ⊗ det W(g,z,w) = C
and thus
∗
det W(g,w,u) ⊗ det W(g,u,z) = det W(g,z,w) .
 Applications of bundle gerbes 255

Similarly
W(g,w,z) ⊕ W(g,z,w) = Cn
so that
det W(g,w,z) ⊗ det W(g,z,w) = C
and
∗
det W(g,w,z) = det W(g,z,w) .
There are a number of other cases that can be dealt with in a similar fashion and
putting all these facts together gives a bundle gerbe multiplication on P → Y [2] .
A construction of the curving on (P, Y ) appears in Murray and Stevenson (2008).

12.6 Applications of bundle gerbes

12.6.1 Wess–Zumino–Witten term
The Wess–Zumino–Witten term associates to a smooth map g of a surface Σ
into a compact, simple, Lie group G an invariant Γ(g) ∈ U (1). As noted by a
number of authors (Carey et al. 2000; Hitchin 2002; Gawedzki and Reis 2002)
‘
this can be understood as the holonomy of a connection and curving on the
basic gerbe on the group. The original definition of Witten (1984) is that we
choose a three-manifold X with ∂X = Σ and extend g to ĝ : X → G. We then
consider

ĝ∗ (ω)
X

where ω is a three-form on G representing a generator of H 3 (G, 2πiZ). If we

choose a different extension g̃ : X → G then the pair can be combined to define
a map from the manifold X ∪Σ X, formed by joining two copies of X (with
opposite orientations) along Σ, into G. Call this map ĝ ∪ g̃. Then
  
∗ ∗
ĝ (ω) − g̃ (ω) = (ĝ ∪ g̃)∗ (ω) ∈ 2πiZ.
X X X∪Σ X

It follows that

 
Γ(g) = exp ĝ ∗ (ω) ∈ U (1)
X

depends only on g.
If we choose a suitable connection and curving (A, f ) on the basic gerbe on
G, so that it has curvature ω then we see that
Γ(g) = hol(Σ, g∗ (A, f )).
The gerbe approach to the Wess–Zumino–Witten term has two advantages.
Firstly, it removes the topological restriction on M of two-connectedness
256 An introduction to bundle gerbes

necessary in Witten’s definition so that the map g can be extended to the three-
manifold X. Secondly, we can use the local formula for the holonomy given in
Section 12.4.1. For details see Gawedzki and Reis (2002).
‘
12.6.2 Faddeev–Mickelsson anomaly
We follow Carey and Murray (1996) and Segal (1985). Let X be a compact,
Riemannian, spin, three-manifold and denote by A the space of connections on
a complex vector bundle over M and by G the space of gauge transformations.
For any A ∈ A the chiral Dirac operator DA coupled to A has discrete spectrum.
Let
Y = {(A, t) | t ∈
/ spec(DA )}
be considered as a submersion over A. Note that G acts on Y and Y /G → A/G
is another submersion. Following Segal (1985) for (A, s) ∈ Y we can decompose
the Hilbert space H of coupled spinors into a direct sum of eigenspaces of DA for
eigenvalues greater than s and a direct sum of eigenspaces of DA for eigenvalues
− +
less than s. Denote these by H(A,s) and H(A,s) , respectively. We can then form
the Fock space
H H ∗
−
F(A,s) = +
H(A,s) ⊗ H(A,s)

which is a bundle F → Y . Notice that G acts on F and gives rise to a bundle

F/G → Y /G.
If we choose another t < s we have
−
H = H(A,t) ⊕ V(A,t,s) ⊕ H(A,s)
+

where V(A,t,s) is the sum of all the eigenspaces of DA for eigenvalues between t
and s. If we use the canonical isomorphism
H H
∗
V(A,t,s) ⊗ det V(A,t,s) = V(A,t,s)

then we can show that

F(A,s) = F(A,t) ⊗ det V(A,t,s) .
It follows that the projective spaces of F(A,s) and F(A,t) are canonically isomor-
phic and descend to a projective bundle P → A and as G acts there is also a
projective bundle P/G → A/G. The question of interest (Segal 1985) is whether
there is a Hilbert bundle H → A/G whose projectivization is P/G. Notice that
as A is contractible the answer to the equivalent question on A is clearly
positive.
As noted in Section 12.5.2 we could give a bundle gerbe interpretation of this
question via the lifting bundle gerbe for the central extension
0 → U (1) → P U (H̃) → U (H̃) → 0
 Other matters 257

for a suitable Hilbert space H̃. However there is a more direct approach as follows.
Let P(A,s,t) be the unitary frame bundle of det V(A,t,s) . Notice that if r < s < t
then
V(A,r,t) ⊕ V(A,t,s) = V(A,r,s)
so that
det V(A,r,t) ⊗ det V(A,t,s) = det V(A,r,s)
gives rise to a bundle gerbe multiplication similar to the SU (n) case in Example
12.14. Again G acts so this descends to a bundle gerbe on A/G. If this bundle
gerbe is trivial then there is a line bundle L → Y such that
det V(A,t,s) = L(A,s) ⊗ L∗(A,t)
and moreover as it is the bundle gerbe on A/G which is trivial these are G
equivariant isomorphisms. It follows that
F(A,s) = F(A,t) ⊗ L(A,s) ⊗ L∗(A,t)
or
F(A,s) ⊗ L∗(A,s) = F(A,t) ⊗ L∗(A,t) ,
and these are G-equivariant isomorphisms. Hence F(A,s) ⊗ L∗(A,s) descends to a
G-equivariant Hilbert bundle on A whose projectivization is P. It follows that
there is a Hilbert bundle on A/G whose projectivization is P/G.

12.6.3 String structures

In Killingback (1987), he introduced the notion of a string structure. Given a G
bundle Q → X he considers the corresponding loop bundle LQ → LX which is
an LG bundle. As noted above we have the Kac–Moody central extension
I → LG → 0.
0 → U (1) → LG
I calls a choice of
Killingback says that Q → X is string if LQ → LX lifts to LG,
such a lift a string structure, and defines the obstruction class to the lift to be
the string class. In Murray and Stevenson (2003) we consider the more general
situation of an LG bundle on a manifold M and the question of whether it
I There is an equivalence between LG bundles on M and G bundles
lifts to LG.
on S 1 × M which is exploited to define a connection and curving on the lifting
bundle gerbe associated to the LG bundle. This enables the derivation of a de
Rham representative for the image of the string class in real cohomology. The
same approach could be applied to the other level Kac–Moody central extensions.

12.7 Other matters

In the interests of brevity nothing has been said about the original approach
to gerbes as sheaves of groupoids. Details are given in Brylinski (1993) and the
258 An introduction to bundle gerbes

relationship with bundle gerbes is discussed in Murray (1996) and in more detail
in Stevenson (2000). In this same context I would like to thank Larry Breen
for pointing out that in the work of Ulbrich (1990, 1991) the notion of cocycle
bitorsors can be interpreted as a form of bundle gerbe.
In the definitions and theory above we could replace U (1) by any abelian
group H and there would only be obvious minor modifications such as the
Dixmier–Douady class being in H 2 (M, H). If we want to replace H by a non-
abelian group things become more difficult as we mentioned in the introduction
because we cannot form the product of two H bundles if H is non-abelian.
To get around this difficulty we need to replace principal bundles by principal
bibundles which have a left and right group actions. The resulting theory becomes
more complicated although closer to Giraud’s original aim of understanding non-
abelian cohomology. For details see Aschieri et al. (2005) and Breen and Messing
(2005).
We have motivated gerbes by the idea of replacing the transition function gαβ
by a U (1) bundle Pαβ on double-overlaps. It is natural to consider what happens
if we take the next step and replace Pαβ by a bundle gerbe on each double-
overlap. This gives rise to the notion of bundle 2-gerbes whose characteristic
class is a four class. In particular there is associated to any principal G bundle
P → M a bundle 2-gerbe whose characteristic class is the pontrjagin class of P .
For details of the theory see Stevenson (2000, 2004) and for an application to
Chern–Simons theory see Carey et al. (2005).
It is clear that we can continue on in this fashion and consider bundle 2-gerbes
on double-overlaps and more generally inductively use p-gerbes on double-
overlaps to define (p + 1)-gerbes. However the theory becomes increasingly com-
plex for a reason that we have not paid much attention to in the discussion above.
In the case of U (1) bundles the transition function has to satisfy one condition,
the co-cycle identity. In the case of gerbes the U (1) bundles over double-overlaps
satisfy two conditions, the existence of a bundle gerbe multiplication and its
associativity. In the case of bundle 2-gerbes there are three conditions and the
complexity continues to grow in this fashion.

Acknowledgements
The author acknowledges the support of the Australian Research Council and
would like to thank the organizers of the conference for the invitation to
participate. The referee is also thanked for useful remarks on the manuscript.

References
Alvarez, O. (1985). Topological quantization and cohomology. Commun. Math. Phys.
100, 279–309.
Aschieri, P., Cantini, L., and Jurčo, B. (2005). Nonabelian bundle gerbes, their differ-
ential geometry and gauge theory. Commun. Math. Phys. 254(2), 367–400.
 References 259

Bott, R and Tu, L. W. (1982). Differential Forms in Algebraic Topology. Vol. 82 of

Graduate Texts in Mathematics, Springer-Verlag, New York/Berlin.
Breen, L and Messing, W. (2005). Differential geometry of gerbes. Adv. Math. 198(2),
732–846.
Brylinski, J.-L. (1993). Loop Spaces, Characteristic Classes and Geometric Quantiza-
tion, Vol. 107 of Progress in Mathematics, Birkhäuser Boston Inc., Boston, MA.
Brylinski, J.-L. and McLaughlin, D. A. (1994). The geometry of degree-four character-
istic classes and of line bundles on loop spaces. I. Duke Math. J. 75(3), 603–38.
Carey, A. L. and Murray, M. K. (1996). Faddeev’s anomaly and bundle gerbes. Lett.
Math. Phys. 37(1), 29–36.
Carey, A. L., Johnson, S., Murray, M. K., Stevenson, D., and Wang, B.-L. (2005).
Bundle gerbes for Chern-Simons and Wess-Zumino-Witten theories. Commun. Math.
Phys. 259(3), 577–613.
Carey, A. L., Mickelsson, J., and Murray, M. K. (2000). Bundle gerbes applied to
quantum field theory. Rev. Math. Phys. 12(1), 65–90.
Chatterjee, D.S. (1998). On gerbs, PhD thesis, Trinity College, Cambridge, UK.
Available at http://www.maths.ox.ac.uk/hitchin/hitchinstudents/chatterjee.pdf
Dupont, J. L. (1978). Curvature and Characteristic Classes. Vol. 640 of Lecture Notes
in Mathematics, Springer–Verlag, Berlin, Germany.
Gawedzki, K. (1988). Topological actions in two-dimensional quantum field theories.
‘
In G. t’Hooft, A. Jaffe, G. Mack, P. K. Mitter, and R. Stora (eds.), Nonperturbative
Quantum Field Theories Vol. 185 of NATO Series, pp. 101–41. Plenum Press
New York/London.
Gawedzki, K. and Reis, N. (2002). WZW branes and gerbes. Rev. Math. Phys. 14(12),
‘
1281–334.
Gawedzki, K. and Reis, N. (2004). Basic gerbe over non-simply connected compact
‘
groups. J. Geom. Phys. 50(1–4), 28–55.
Gotay, M. J., Lashof, R., Śniatycki, J., and Weinstein, A. (1983) Closed forms on
symplectic fibre bundles. Comment. Math. Helv. 58(4), 617–21.
Giraud, J. (1971). Cohomologie non-abélienne. Grundl. 179, Springer Verlag,
Berlin/Heidelberg/New York.
Hitchin, N. (2001). Lectures on special Lagrangian submanifolds. In Winter School on
Mirror Symmetry, Vector Bundles and Lagrangian Submanifolds (Cambridge, MA,
1999) Vol. 23 of AMS/IP Studies on Advanced Mathematics, pp. 151–82, American
Mathematical Society, Providence, RI.
Hitchin, N. (2002). The Wess-Zumino term for a harmonic map. J. Reine Angew. Math.
543, 83–101.
Hitchin, N. (2003). What is a gerbe?, Notices Am. Math. Soc. 50(2), 218–19.
Hitchin (2006). Instantons, Poisson structures and generalized Kähler geometry. Com-
mun. Math. Phys. 265(1), 131–64.
Killingback, T. (1987). World-sheet anomalies and loop geometry. Nucl. Phys. B 288,
578–88.
Meinrenken., E. (2003). The basic gerbe over a compact simple Lie group. Enseign.
Math. (2) 49(3–4), 307–33. math.DG/0209194.
Mickelsson, J. (2003). Gerbes on quantum groups. math.DG/0308235.
Mickelsson, J. (2006). Gerbes and quantum field theory. math-ph/0603031. In J.-P.
Francoise, G. L. Naber and T. S. Tsun (eds), Encyclopedia of Mathematical Physics,
Elsevier Academic Press.
260 An introduction to bundle gerbes

Murray, M. K. (1996). Bundle gerbes. J. London Math. Soc. (2) 54(2), 403–16.
Murray, M. K. and Stevenson, D. (2000). Bundle gerbes: stable isomorphism and local
theory. J. London Math. Soc. (2) 62(3), 925–37.
Murray, M. K. and Stevenson, D. (2003). Higgs fields, bundle gerbes and string
structures. Commun. Math. Phys. 243(3), 541–55.
Murray, M. K. and Stevenson, D. (2008). The basic bundle gerbe on unitary groups.
J. Geometry Phys. 58(11), 1571–90.
Pressley, A. and Segal, G. (1986). Loop Groups. Oxford Mathematical Monographs,
The Clarendon Press, Oxford.
Segal, G. (1985). Faddeev’s anomaly in Gauss’s Law. Preprint, Oxford University,
Oxford.
Stevenson, D. (2000). The geometry of bundle gerbes, PhD thesis, University of
Adelaide, Adelaide, Australia 2000. math.DG/0004117.
Stevenson, D. (2004). Bundle 2-gerbes. Proc. London Math. Soc. (3) 88(2), 405–35.
Ulbrich, K.-H. (1990). On the correspondence between gerbes and bouquets. Math.
Proc. Camb. Philos. Soc. 108(1), 1–5.
Ulbrich, K.-H. (1991). On cocycle bitorsors and gerbes over a Grothendieck topos.
Math. Proc. Camb. Philos. Soc. 110(1), 49–55.
Waldorf, K. (2007). More morphisms between bundle gerbes. Theory and Applications
of Categories 18(9), 240–273.
Witten, E. (1984). Non-abelian bosonization in two dimensions. Commun. Math. Phys.
92(4), 455–72.
 XIII
PROJECTIVE LINKING AND BOUNDARIES OF
POSITIVE HOLOMORPHIC CHAINS IN
PROJECTIVE MANIFOLDS, PART I

F. Reese Harvey and H. Blaine Lawson, Jr.

Dedicated to Nigel Hitchin in celebration of his 60th birthday

13.1 Introduction
In 2000 H. Alexander and J. Wermer published the following result:

Theorem 13.1 (Alexander–Wermer) Let Γ be a compact-oriented smooth

submanifold of dimension 2p − 1 in Cn . Then Γ bounds a positive holomorphic
p-chain in Cn if and only if the linking number

Link(Γ, Z) ≥ 0

for all canonically oriented algebraic subvarieties Z of codimension p in Cn − Γ.

The linking number is an integer-valued topological invariant defined by the

intersection Link(Γ, Z) ≡ N • Z with any 2p-chain N having ∂N = Γ in Cn
(see Section 13.3). A positive holomorphic p-chain is a finite sum of canonically
oriented complex subvarieties of dimension p and finite volume in Cn − Γ (see
Definition 13.7). Here the notion of boundary is taken in the sense of currents,
that is, Stokes’ theorem is satisfied. However, for smooth Γ there is boundary
regularity almost everywhere, and if Γ is real analytic, one has complete boundary
regularity in the sense of immersions (see Harvey and Lawson 1975).
The main point of this chapter is to formulate and prove an analogue of the
Alexander–Wermer theorem for oriented (not necessarily connected) curves in
a projective manifold. In the sequel we shall study the corresponding result for
submanifolds of any odd dimension.
Before stating the main result we remark that a key ingredient in the proof
of the Alexander–Wermer theorem is the following classical theorem and its
generalizations (Wermer 1958):

Theorem 13.2 (Wermer) Let Γ ⊂ Cn be a compact real analytic curve and

denote by
G poly = {z ∈ Cn : |p(z)| ≤ sup |p|
Γ for all polynomials p}
Γ
262 Projective linking and boundaries of positive holomorphic chains

G poly − Γ is a one-dimensional complex analytic

its polynomial hull. Then Γ
subvariety of C − Γ.
n

For compact subsets K of complex projective n-space Pn , the authors recently

introduced the notion of the projective hull
G proj = {z ∈ Pn : ∃C s.t. σz  ≤ C d sup σ ∀σ ∈ H 0 (Pn , O(d)), d > 0}
K
K

and defined K to be stable if the constant C can be chosen independently of

G proj were established in
G proj . A number of basic properties of K
the point z ∈ K
Harvey and Lawson (2006a), and the following analogue of Wermer’s theorem
was proved in Harvey, Lawson, and Wermer 2008.
Theorem 13.3 (Harvey–Lawson–Wermer) Let Γ ⊂ Pn be a stable real
G proj − Γ is a one-dimensional complex analytic subvariety
analytic curve. Then Γ
of P − Γ.
n

It is interesting to note that while Wermer’s theorem holds for curves with only
weak differentiability properties (see Alexander and Wermer 1998 or Dinh and
Lawrence 2003 for an account), its projective analogue fails even for C ∞ -curves.
On the other hand there is much evidence for the following:
Conjecture 13.1 Every real analytic curve in Pn is stable.
This brings us to the notion of projective linking numbers. Suppose that Γ ⊂
Pn is a compact-oriented smooth curve, and let Z ⊂ Pn − Γ be an algebraic
subvariety of codimension 1. The projective linking number of Γ with Z is defined
to be

LinkP (Γ, V ) ≡ N • Z − deg(Z) ω
N

where ω is the standard Kähler form on Pn and N is any integral 2-chain with
∂N = Γ in Pn . Here Z is given the canonical orientation, and • : H2 (Pn , Γ) ×
H2n−2 (Pn − Γ) → Z is the topologically defined intersection pairing. This defin-
ition is independent of the choice of N (see Section 13.3). The associated reduced
linking number is defined to be

J P (Γ, Z) ≡ 1
Link LinkP (Γ, Z).
deg(Z)
The basic result proved here is the following:
Theorem 13.4 Let Γ be a oriented, stable, real analytic curve in Pn with a pos-
itive integer multiplicity on each component. Then the following are equivalent:
1. Γ is the boundary of a positive holomorphic 1-chain of mass ≤ Λ in Pn .
J P (Γ, Z) ≥ −Λ for all algebraic hypersurfaces Z in Pn − M .
2. Link
 Introduction 263

If Γ bounds any positive holomorphic 1-chain, then there is a unique such

chain T0 of least mass. (All others are obtained by adding algebraic 1-cycles
to T0 .) Note that Λ0 ≡ M (T0 ) is the smallest positive number such that
(2) holds.

Corollary 13.1 Let Γ be as in Theorem 13.4 and suppose T is a positive

holomorphic 1-chain with dT = Γ. Then T is the unique holomorphic chain of
least mass with dT = Γ if and only if
< =
T •Z
inf =0
Z degZ
where the infimum is taken over all algebraic hypersurfaces in the complement
Pn − Γ.

Condition (2) in Theorem 13.4 has several equivalent formulations. The first
is in terms of projective winding numbers. Given a holomorphic section σ ∈
H 0 (Pn , O(d)), the projective winding number of σ on Γ is defined as the integral

WindP (Γ, σ) ≡ dC logσ,
Γ

and we set
 P (Γ, σ) ≡ 1 WindP (Γ, σ).
Wind
d
Another formulation involves the cone P SH ω (Pn ) of quasi-
plurisubharmonic functions. These are the upper semi-continuous
functions f : Pn → [−∞, ∞) for which ddC f + ω is a positive (1,1)-current
on Pn .

Proposition 13.1 Let Γ be an oriented smooth curve in Pn with a positive

integer multiplicity on each component. Then for any Λ > 0 the following are
equivalent:

1. J P (Γ, Z) ≥ −Λ for all algebraic hypersurfaces Z ⊂ Pn − Γ.

Link

2.  C P (Γ, σ) ≥ −Λ for all holomorphic
Wind sections σ of O(d), and all d > 0.
3. Γ
d u ≥ −Λ for all u ∈ P SH ω (Pn
).

Any smooth curve Γ in Pn lies in some affine chart Cn ⊂ Pn , and it is natural

to ask for a reformulation of condition (1) in terms of the conventional linking
numbers of Γ with algebraic hypersurfaces in that chart. This is done explicitly
in Theorem 13.4.
The results above extend to any projective manifold X. Given a very
ample Hermitian line bundle λ on X there are intrinsically defined λ-linking
numbers Linkλ (Γ, Z), λ-winding numbers Windλ (Γ, σ) for σ ∈ H 0 (X, λd ), and
ω = c1 (λ)-quasi-plurisubharmonic functions P SH ω (X). With these notions,
264 Projective linking and boundaries of positive holomorphic chains

Proposition 13.1 and Theorem 13.4 carry over to X. This is done in

Section 13.7.
Theorem 13.4 leads to the following interesting result:
Theorem 13.5 Let γ ⊂ Pn be a finite disjoint union of real analytic curves
and assume γ is stable. Then a class τ ∈ H2 (Pn , γ; Z) is represented by a positive
holomorphic chain with boundary on γ if and only if
τ •u ≥ 0
for all u ∈ H2n−2 (Pn − γ; Z) represented by positive algebraic hypersurfaces in
Pn − γ.
This result expands to a duality between the cones of relative and absolute
classes which are representable by positive holomorphic chains (Harvey and
Lawson 2006d).
The arguments given in Section 13.6 show that there is even more evidence
for the following:
Conjecture 13.2 Every oriented, real analytic curve in Pn which satisfies the
equivalent conditions of Proposition 13.1 is stable.

In Harvey and Lawson 2006c we prove that if Conjecture 13.2 holds for curves
in P2 , then all the results above continue to hold for real analytic Γ of any odd
dimension 2p − 1. (No stability hypothesis is needed.)
Remark 13.1 There are several quite different characterizations of the bound-
aries of general (i.e. not necessarily positive) holomorphic chains in projective
and certain quasi-projective manifolds (see, e.g., Harvey and Lawson 1977, 2005,
Dolbeault 1983, and Dolbeault and Henkin 1994, 1997).
Note. To keep formulas simple throughout the chapter we adopt the convention
that
i
dC = (∂ − ∂).
2π

13.2 Projective hulls

In this section we recall the definition and basic properties of the projective hull
introduced in Harvey and Lawson (2006a). This material is not really necessary
for reading the rest of the chapter.
Let O(1) → Pn denote the holomorphic line bundle of Chern class 1 endowed
with the standard unitary-invariant metric, and let O(d) be its dth tensor power
with the induced tensor product metric.
Definition 13.1 Let K ⊂ Pn be a compact subset of complex projective n-space.
A point x ∈ Pn belongs to the projective hull of K if there exists a constant
 Projective hulls 265

C = C(x) such that

σ(x) ≤ C d sup σ (13.1)
K

for all global sections σ ∈ H 0 (Pn , O(d)) and all d > 0. This set of points is
G
denoted K.
The set KG is independent of the choice of metric on O(1).
The projective hull possesses interesting properties. It and its generalizations
function in projective and Kähler manifolds much as the polynomial hull and
its generalizations function in affine and Stein manifolds. The following were
established in Harvey and Lawson (2006a):
1. If Y ⊂ Pn is an algebraic subvariety and K ⊂ Y , then K G ⊂ Y . That is, KG is
contained in the Zariski closure of K. Furthermore, if Y ⊂ Pn is a projective
manifold and λ = O(1) Y , the λ-projective hull of K ⊂ Y , defined as in
(13.1) with O(1) replaced by λ, agrees with K. G The same is true of λk for
any k.
2. If K is contained in an affine open subset Ω ⊂ Pn , then (K) G poly,Ω ⊆ K.
G
3. If K = ∂C, where C is a holomorphic curve with boundary in P , then n

C ⊆ K.G
G −
4. {K} − K satisfies the maximum modulus principle for holomorphic func-
tions on open subsets of Pn − K. ({K} G − denotes the closure of K.)
G
G
5. {K} −
− K is 1-pseudoconcave in the sense of Dinh and Lawrence (2003).
In particular, for any open subset U ⊂ Pn − K, if the Hausdorff 2-measure
G − ∩ U ) < ∞, then K
H 2 (K G − ∩ U is a complex analytic subvariety of dimen-
sion 1 in U .
6. If K is a real analytic curve, then the Hausdorff dimension of K G is 2.
7. KG is pluripolar if and only if K is pluripolar. 1 Thus there exist smooth
closed curves Γ ⊂ P2 with Γ G = P2 . However, real analytic curves are always
pluripolar.
The projective hull has simple characterizations in both affine and homo-
geneous coordinates. For example, if π : Cn+1 − {0} −→ Pn is the standard
projection, and we set S(K) ≡ π−1 (K) ∩ S 2n+1 , then
< =
G 
K = π S(K)poly − {0}


poly is the polynomial hull of S(K) in C
n+1
where S(K) .
There is also a best constant function C : KG → R+ defined at x to be the
G the set
least C = C(x) for which the defining property (13.1) holds. For x ∈ K,
−1 
π (x) ∩ S(K)poly is a disk of radius ρ(x) = 1/C(x). One deduces that K G is
compact if C is bounded.
1 A set K is pluripolar if it is locally contained in the −∞ set of a plurisubharmonic function.
266 Projective linking and boundaries of positive holomorphic chains

Definition 13.2 G → R+ is
A compact subset K ⊂ Pn is called stable if C : K
bounded.

Combining a classical argument of E. Bishop with (5) and (6) above gives the
following:

Theorem 13.6 (Harvey, Lawson, and Wermer) Let Γ ⊂ Pn be a compact

G − Γ is a one-
stable real analytic curve (not necessarily connected). Then Γ
dimensional complex analytic subvariety of P − Γ.
n

As mentioned in the introduction, there is much evidence for the conjecture

that any compact real analytic curve in Pn is stable, and therefore the stability
hypothesis could be removed from Theorem 13.6. Interestingly, the conclusion of
Theorem 13.6 fails to hold in general for smooth curves (see Harvey and Lawson
2006a, Section 4), but may hold for smooth curves which are pluripolar or, say,
quasi-analytic.

Remark 13.2 The close parallel between polynomial and projective hulls is
signaled by the existence of a projective Gelfand transformation, whose relation
to the classical Gelfand transform is analogous to the relation between Proj of
a graded ring and Spec of a ring in modern algebraic geometry. To any Banach
graded algebra A∗ = d≥0 Ad (a normed graded algebra which is a direct sum
of Banach spaces) one can associate a topological space XA∗ and a Hermitian
line bundle λA∗ → XA∗ with the property that A∗ embeds as a closed subalgebra
9  
A∗ ⊆ Γcont XA∗ , λdA∗
d≥0

of the algebra of continuous sections of powers of λ with the sup-norm. When

K ⊂ Pn is a compact subset and Ad = H 0 (Pn , O(d)) K with the sup-norm on
K, there is a natual homeomorphism
G ∼
K = X A∗ and O(1) ∼
= λA ∗ .

This parallels the affine case where, for a compact subset K ⊂ Cn , the Gelfand
spectrum of the closure of the polynomials on K in the sup-norm corresponds
to the polynomial hull of K.

Furthermore, when A∗ is finitely generated, the space XA∗ can be realized,

essentially uniquely, by a subset XA∗ ⊂ Pn with λA∗ ∼
= O(1) and
GA = XA .
X ∗ ∗

This parallels the classical correspondence between finitely generated Banach

algebras and polynomially convex subsets of Cn . Details are given in Harvey and
Lawson (2006a).
 Projective linking and projective winding numbers 267

13.3 Projective linking and projective winding numbers

In this section we introduce the notion of projective linking numbers and
projective winding numbers for oriented curves in Pn .
Let M = M 2p−1 ⊂ Pn be a compact oriented submanifold of dimension 2p − 1,
and recall the intersection pairing
• : H2p (Pn , M ; Z) × H2(n−p) (Pn − M ; Z) −→ Z (13.2)
which under Alexander duality corresponds to the Kronecker pairing:
κ : H 2(n−p) (Pn − M ; Z) × H2(n−p) (Pn − M ; Z) −→ Z.
When homology classes are represented by cycles which intersect transversally
in regular points, the map (13.2) is given by the usual algebraic intersection
number.
Definition 13.3 Fix M as above and let Z ⊂ Pn − M be an algebraic subva-
riety of dimension n − p. Then the projective linking number of M and Z is
defined as follows. Choose a 2p-chain N in Pn with dN = M and set

LinkP (M, Z) ≡ N • Z − deg(Z) ωp (13.3)
N

form on P . This definition extends by linearity

where ω is the standard Kähler  n

to algebraic (n − p)-cycles Z = j nj Zj supported in Pn − M .

Lemma 13.1 The linking number LinkP (M, Z) is independent of the choice of
the cobounding chain N .
Proof. Let N  be another choice and set W = N − N  . Then dW = 0 and

W • Z − deg(Z) ω p = deg(W ) · deg(Z) − deg(Z) · deg(W ) = 0.
W

For now we shall be concerned with the case p = 1. Here there is a naturally
related notion of projective winding number defined as follows:
Definition 13.4 Suppose Γ ⊂ Pn is a smooth, closed, oriented curve, and let
σ be a holomorphic section of O() over Pn which does not vanish on Γ. Then
the projective winding number of σ on Γ is defined to be

WindP (Γ, σ) ≡ dC logσ (13.4)
Γ

Proposition 13.2 Let Γ and σ be as in Definition 13.4, and let Z be the divisor
of σ. Then

WindP (Γ, σ) = LinkP (Γ, Z).

268 Projective linking and boundaries of positive holomorphic chains

Proof. We recall the fundamental formula

ddC logσ = Z − ω (13.5)
which follows by writing σ = ρ|σ| in a local holomorphic trivialization of O()
and applying the Chern and Poincaré–Lelong formulas: ddC logρ = −ω and
ddC log|σ| = Div(σ) = Z. We now write Γ = ∂N for a rectifiable 2-current N
in Pn and note that
  
dC logσ = ddC logσ = N • Z − deg(Z) ω
Γ N N

by (13.5).

It will be useful in what follows to normalize these linking numbers:
Definition 13.5 For Γ and Z as above, we define the reduced projective linking
number to be
J P (Γ, Z) ≡ 1
Link LinkP (Γ, Z)
deg(Z)
 P (Γ, Z) ≡ WindP (Γ, Z)/
and the reduced projective winding number to be Wind
deg(Z).
Note that if Z = Div(σ) for σ ∈ H 0 (Pn , O()), then by Proposition 13.2 we
have

J P (Γ, Z) =
Link dC logσ .
1
(13.6)
Γ

Remark 13.3 (Relation to affine linking numbers) Note that a smooth

curve Γ ⊂ Pn does not meet the generic hyperplane Pn−1 and is therefore
contained in the affine chart Cn = Pn − Pn−1 . Let N be an integral 2-chain with
support in Cn and dN = Γ. Then
 
LinkP (Γ, Pn−1 ) = N • Pn−1 − ω=− ω, (13.7)
N N

and for any divisor Z of degree  which does not meet Γ we have
LinkP (Γ, Z) − LinkP (Γ, Pn−1 ) = N • Z = LinkCn (Γ, Z),
the classical linking number of Γ and Z.
Remark 13.4 (Relation to sparks and differential characters) The
de Rham–Federer approach to differential characters (cf. Gillet and Soulé 1989,
Harris 1989, and Harvey, Lawson, and Zweck 2003) is built on objects called
sparks. These are generalized differential forms (or currents) α which satisfy the
equation
dα = R − φ
 Quasi-plurisubharmonic functions 269

where R is rectifiable and φ is smooth. By (13.5) the generalized 1-form dC logσ

is such a creature. Its associated Cheeger–Simons character [α] : Z1 (Pn ) −→ R/Z
on smooth 1-cycles is defined by the formula

[α](Γ) ≡ ω (mod Z)
N

where N is a 2-chain with dN = Γ. Thus we see that the “correction term” in

the projective linking number is the value of a secondary invariant related to
differential characters.
1
Remark 13.5 Note that by (13.5) the function u ≡ logσ which appears in
(13.6) has the property that
1
ddC u + ω = Z ≥ 0 (13.8)

by the fundamental equation (13.5). This leads us naturally to the next
section.

13.4 Quasi-plurisubharmonic functions

The following concept, due to Demailly (1982) and developed systematically by
Guedj and Zeriahi (2003), is central to the study of projective hulls, and it is
intimately related to projective linking numbers.

Definition 13.6 An upper semi-continuous function u : X −→ [−∞, ∞) on a

Kähler manifold (X, ω) is quasi-plurisubharmonic (or ω-quasi-plurisubharmonic)
if u ≡ −∞ and

ddC u + ω ≥ 0 on X. (13.9)

The convex set of all such functions on X will be denoted by P SH ω (X).

These functions enjoy many of the properties of classical plurisubharmonic

functions and play an important role in understanding various capacities in
projective space (cf. Guedj and Zeriahi 2003). One of the appealing geometric
properties of this class is the following. Suppose ω is the curvature form of
a holomorphic line bundle λ → X with Hermitian metric g. Then a smooth
function u : X → R is quasi-plurisubharmonic if and only if the Hermitian metric
eu g has curvature ≥ 0 on X.
The quasi-plurisubharmonic functions are directly relevant to projective hulls,
as the next result shows (cf. Guedj and Zeriahi 2003, proof of Theorem 4.2 and
Theorem 13.8 below).

Theorem 13.7 Let ω denote the standard Kähler form on Pn . Then the
G of a compact subset K ⊂ Pn , is exactly the subset of points
projective hull K
270 Projective linking and boundaries of positive holomorphic chains

x ∈ Pn for which there exists a constant Λ = Λ(x) with

u(x) ≤ sup u + Λ f or all u ∈ P SH ω (X) (13.10)
K
This enables one to generalize the notion of projective hull from projective
algebraic manifolds to general Kähler manifolds.
Note that the least constant Λ(x) for which (13.10) holds is exactly Λ(x) =
logC(x) where C(x) is the best constant function discussed in Section 13.2.
1
By considering functions of the form u = log{σ d } for sections σ ∈
H (P , O(d)), one immediatly sees the necessity of the condition in Theorem
0 n

13.7. Sufficiency follows from the fact that such functions are the extreme points
of the cone P SH ω (Pn ).
One importance of quasi-plurisubharmonic functions is that they enable us to
establish Poisson–Jensen measures for points in K G (Harvey and Lawson 2006a,
Theorem 11.1).

13.5 Boundaries of positive holomorphic chains

Let Γ be a smooth, oriented, closed curve in Pn . We recall that (even if Γ is only
class C 1 ) any irreducible complex analytic subvariety V ⊂ Pn − Γ of dimension 1
has finite Hausdorff 2-measure and defines a current [V ] of dimension 2 in Pn by
integration on the canonically oriented manifold of regular points. This current
satisfies d[V ] = 0 in Pn − Γ (see e.g. Harvey 1977).
Definition 13.7  By a positive holomorphic 1-chain with boundary Γ we mean
a finite sum T = k nk [Vk ] where each nk ∈ Z+ and each Vk ⊂ Pn − Γ is an
irreducible subvariety of dimension 1, so that
dT = Γ (as currents on Pn ).
We shall be interested in conditions on Γ which are necessary and sufficient
for it to be such a boundary.
Proposition 13.3 Let Γ ⊂ Pn be a smooth, oriented, closed curve (not neces-
sarily connected) with a positive integer multiplicity on each component. Then
the following are equivalent:
1. J
Link(Γ, Z) ≥ −Λ for all algebraic hypersurfaces Z ⊂ Pn − Γ.
2.  σ) ≥ −Λ for all σ ∈ H 0 (Pn , O()),  > 0, with no zeros on Γ.
Wind(Γ,
 C
3. Γ
d u ≥ −Λ for all u ∈ P SH ω (Pn ).
Proof. The equivalence of (1) and (2) follows immediately from Proposition
13.2. That (3) ⇒ (1) follows from (13.6) and (13.8). To see that (1) ⇒ (3) we
use the following non-trivial fact due essentially to Demailly (1982) and Guedj
(1999).

1
Proposition 13.4 The functions of the form log(σ ) for σ ∈ H 0 (Pn , O()),
 > 0, are weakly dense in P SH ω (Pn ) modulo the constant functions.
 Projective Alexander–Wermer theorem for curves 271

Proof. Fix u ∈ P SH ω (Pn ) and consider the positive (1,1)-current T ≡ ddC u +

ω. By Demailly (1982) and Guedj (1999, Theorem 0.1), there exist sequences
σj ∈ H 0 (Pn , O(j )) and Nj ∈ R+ , j = 1, 2, 3, ... such that
< =
1 j  1 
T = lim Div(σj ) = lim dd log σj 
C j +ω . (13.11)
j→∞ Nj j→∞ Nj
 
Since T (ωn−1 ) = Pn (ddC u ∧ ω n−1 + ω n ) = Pn ω n = 1 = limj→∞ j /Nj by
(13.11), we 1 may assume that Nj = j for all j. Therefore, setting uj =
ddC logσj  j , we have by (13.11) that limj ddC uj = ddC u and, in particular,
limj ∆uj = ∆u. Renormalizing
  each uj by an additive constant, we may also
assume that limj uj = u.
Recall the formula Id = H + G ◦ ∆ on C ∞ (X) where H is harmonic projection
and G is the Green’s function. This decomposition
 carries over to distributions

D (X) by adjoint. It follows that uj − uj = G(∆uj ) → G(∆u) = u − u and
so uj → u as claimed.

Theorem 13.8 Let Γ be as above. Suppose Γ = dT where T is a positive holo-
morphic chain with mass M(T ) ≤ Λ. Then Γ satisfies the equivalent conditions
(1), (2), and (3) of Proposition 13.3.
Proof. Suppose u ∈ P SH ω (Pn ). Since ddC u + ω ≥ 0, we have
 
dC u = dC u = T (ddC u) = T (ddC u + ω) − T (ω) ≥ −T (ω)
Γ dT

= −M(T ) ≥ −Λ
as asserted.

13.6 Projective Alexander–Wermer theorem for curves

We now examine the question of whether the equivalent necessary conditions,
given in Theorem 13.8, are in fact sufficient.
Theorem 13.9 Let Γ ⊂ Pn be an embedded, oriented, real analytic closed
curve, not necessarily connected and with a positive integer multiplicity on each
component. Suppose the underlying curve supp (Γ) is stable. If Γ satisfies the
equivalent conditions (1), (2), and (3) of Proposition 13.3, then there exists a
positive holomorphic 1-chain T in Pn with M(T ) ≤ Λ such that
dT = Γ.
Note. When it is clear from context, we also use Γ to denote the underlying
curve supp Γ.
Proof. Fix a compact set K ⊂ Pn and consider the sets
CK ≡ {ddC T : T ∈ P1,1 , M(T ) ≤ 1 and supp T ⊂ K}
272 Projective linking and boundaries of positive holomorphic chains

SK ≡ {u ∈ C ∞ (Pn ) : ddC u + ω ≥ 0 on K}

where P1,1 ⊂ E2 (Pn ) denotes the convex cone of positive currents of bidimen-
sion (1,1) on Pn . Note that CK is a weakly closed convex subset which con-
tains 0 (since the set of T ∈ P1,1 with M(T ) ≤ 1 and supp T ⊂ K is weakly
compact).
Suppose that K is a weakly closed convex subset containing 0 in a topological
vector space V . Then the polar of K is defined to be the subset of the dual space
given by

K0 ≡ {L ∈ V  : (L, v) ≥ −1 for v ∈ K}.

Similarly given a subset L ⊂ V  we define

L0 ≡ {v ∈ V : (L, v) ≥ −1 for L ∈ L}.

The Bipolar Theorem (Schaefer 1999) states that

(K0 )0 = K.

Proposition 13.5

SK = (CK )0

Proof. Suppose that u ∈ C ∞ (Pn ) satisfies

u(ddC T ) = T (ddC u) ≥ −1 (13.12)

for all T ∈ P1,1 with M(T ) ≤ 1 and supp T ⊂ K. Consider T = δx ξ where x ∈ K

and ξ is a positive (1,1)-vector at x with mass-norm ξ = 1. Then

T (ddC u) = (ddC u)(ξ) = (ddC u + ω)(ξ) − 1 ≥ −1

by (13.12), and so u ∈ SK . This proves that (CK )0 ⊂ SK .

For the converse, let u ∈ SK and fix T ∈ P1,1 with M(T ) ≤ 1 and supp T ⊂ K.
Then

(ddC T )(u) = T (ddC u) = T (ddC u + ω) − T (ω) ≥ T (ω) = −M(T ) ≥ −1,

and so u ∈ (CK )0 .

As an immediate consequence we have the following. If we set

Λ · SK ≡ {ddC T : T ∈ P1,1 , M(T ) ≤ Λ and supp T ⊂ K}, then

 
(Λ · SK ) = u ∈ C ∞ (Pn ) : ddC u + Λ1 ω ≥ 0 on K
0

(13.13)
 Projective Alexander–Wermer theorem for curves 273

Recall from Section 13.4 that for each Λ ∈ R we have the compact set
G
Γ(Λ) = {x ∈ Pn : u(x) ≤ sup u + Λ ∀u ∈ P SH ω (Pn )}
Γ

G = K Γ(Λ).
and that Γ G The following lemma is established in Harvey and Lawson
Λ
(2006a, 18.7):
Lemma 13.2 Let u be a C ∞ function which is defined and quasi-
G − . Fix Λ > 0. Then there is a C ∞
plurisubharmonic on a neighborhood of {Γ}
, which is defined and quasi-plurisubharmonic on all of Pn and agrees
function u
G
with u on a neighborhood of Γ(Λ).
G − denotes the closure of Γ.
Here {Γ} G By our stability assumption, Γ G = Γ(Λ)
G
G G −
for some Λ, and therefore Γ = {Γ} . However, we shall keep the notation {Γ}G −,
when appropriate, in order to prove the more general result mentioned in Remark
13.6 below.
The lemma above leads to the following:
Proposition 13.6 If Γ satisfies condition (3) in Proposition 13.3, that is, if

(−dC Γ)(u) = dC u ≥ −Λ
Γ

for all u ∈ P SH ω (P ), then there exists T ∈ P1,1 with M(T ) ≤ Λ and supp T ⊂
n

G − , such that
{Γ}
ddC T = −dC Γ (13.14)
Proof. If this is not true, then it must fail on some compact neighborhood K of
G − . (Otherwise, there exists a sequence of positive currents {Tj }j , satisfying
{Γ}
(13.14), with M (Tj ) ≤ Λ and supp Tj ⊂ Kj where Kj are compact neighborhoods
squeezing down to {Γ} G − . By the standard compactness theorem for positive
currents, there would then be a convergent subsequence whose limit T would
satisfy the conclusion of Proposition 13.6.)
By (13.13) and the Bipolar Theorem, we then conclude that there is a smooth
function u which is quasi-plurisubharmonic on K with −dC Γ(u) < −Λ. Applying
Lemma 13.2 contradicts the hypothesis.

Now let T be the current given by Proposition 13.6. Let V denote the projective
hull of Γ and recall from Harvey, Lawson, and Wermer (Theorem 4.1) that V has
the following strong regularity. There exists a Riemann surface S with finitely
many components, a compact region W ⊂ S with real analytic boundary, and a
holomorphic map ρ : S → Pn which is generically injective and satisfies

1. ρ(W ) = V .
2. ρ is an embedding on a tubular neighborhood of ∂W in S.
3. ρ(∂W ) is a union of components of the support of Γ.
274 Projective linking and boundaries of positive holomorphic chains

Let Σ denote an -tubular neighborhood of ∂W on S (with respect to some

analytic metric), with chosen so that
4. ρ is injective on Σ.
5. ρ(∂ + Σ) ∩ V = ∅ where ∂ + Σ denotes the “outer” boundary of Σ, that is,
the union of components of ∂Σ not contained in W .
Write Σ = Σ+ ∪ ∂W ∪ Σ− where Σ+ denotes the union of components of Σ −
∂W which are not contained in W . Then we have d[Σ+ ] = [∂ + Σ] − [∂W ] (with
standard orientations coming from the domains). Hence we have
ddC [Σ+ ] = −dC [∂ + Σ] + dC [∂W ].
Let ν : Σ → Z denote the locally constant, integer-valued function with the
property that
ρ∗ (ν[∂W ]) = Γ.
Then
    def
ddC ρ∗ ν[Σ+ ] = −dC ρ∗ ν[∂ + Σ] + dC Γ = −dC Γ+ + dC Γ. (13.15)
We now define
 
T + = T + ρ∗ ν[Σ+ ]
and note that by (13.14) we have
ddC T + = −dC Γ+ . (13.16)
Now in the open set Pn − Γ+ , the current T + is a positive (1,1)-current which
is supported in the analytic subvariety V + ≡ ρ(W + ) ≡ ρ(Σ+ ∪ W ) and satisfies
ddC T + = 0. We note that ρ : W + → V + is just the normalization of V + . It
follows from Harvey and Lawson (1983, Lemma 32) that there is a harmonic
function h : W → R+ so that
T + = ρ∗ (h[W + ]) in Pn − Γ+ .
This function is evidently constant (= ν) in Σ+ , and hence h is constant on every
component of W + . Thus T = ρ∗ (h[W ]) is a positive holomorphic chain.
Now since supp dT ⊂ supp Γ and supp Γ is a regular curve, it follows from
the Federer flat support
 theorem (Federer 1969, Theorem 4.1.15) that dT is
of the form dT = k ck [Γk ] where the ck ’s are constants and the Γk ’s are the
oriented,
 connected components of supp Γ. The current Γ itself has the form Γ=
k νk [Γk ] for integers νk . Since dd T = −d Γ, we have d (dT − Γ) = (ck −
C C C

νk )dC [Γk ] = 0. It follows that ck = νk for all k, that is, dT = Γ as claimed.

Remark 13.6 Theorem 13.9 remains true if one replaces the second statement
G − of Γ is locally contained
with the assumption that the closed projective hull {Γ}
G − − Γ.
in a 1-dimensional complex subvariety at each point of {Γ}
 Theorems for general projective manifolds 275

Theorem 13.9 can be restated in terms of the classical (affine) linking

numbers:
Theorem 13.10 Let Γ ⊂ Cn ⊂ Pn be an oriented, closed curve as in Theorem
13.9. Then there exists a constant Λ0 so that the classical linking number
LinkCn (Γ, Z) ≥ −Λ0 degZ
for all algebraic hypersurfaces in Cn − Γ if and only if Γ is the boundary of a
positive holomorphic 1-chain T in Pn = Cn ∪ Pn−1 with (projective) mass

1
M(T ) ≤ Λ0 + dC log(1 + z2 ).
2 Γ
Proof. This follows from Remark 13.3 and the fact that the projective Kähler
form ω = 12 ddC log(1 + z2 ) in Cn .

Note. The case Λ0 = 0 corresponds to the theorem of Alexander and Wermer
(2000). A proof of the full Alexander–Wermer theorem (for C 1 -curves with no
stability assumption) in the spirit of the arguments above is given in Harvey and
Lawson (2006b).
Note that the current of least mass among those provided by Theorem 13.9 is
uniquely determined by Γ. All others differ from this one by adding a positive
algebraic cycle.
Corollary 13.2 Let Γ be as in Theorem 13.9 and suppose T is a positive
holomorphic 1-chain with dT = Γ. Then T is the unique holomorphic chain of
least mass with dT = Γ if and only if
< =
T •Z
inf =0
Z degZ
where the infimum is taken over all algebraic hypersurfaces in the complement
Pn − Γ.
Proof. Suppose inf Z {T • Z/degZ} = c > 0. Then Link J P (Γ, Z) = T • Z/
deg(Z) − T (ω) ≥ c − T (ω) for all positive divisors in P − Γ. Hence, by Theorem
n

13.9 there exists a positive holomorphic chain T  with dT  = Γ and M (T  ) ≤

M (T ) − c.
On the other hand suppose that T is not the unique positive holomorphic chain
T0 of least mass. Then T = T0 + W where W is a positive algebraic cycle, and
one has T • Z/degZ = T0 • Z/degZ + W • Z/degZ ≥ W • Z/degZ = degW .

13.7 Theorems for general projective manifolds

The results established above generalize from Pn to any projective manifold. Let
X be a compact complex manifold with a positive holomorphic line bundle λ.
Fix a Hermitian metric on λ with curvature form ω > 0, and give X the Kähler
276 Projective linking and boundaries of positive holomorphic chains

metric associated to ω. Let Γ be a closed curve with integral weights as in

Theorem 13.9, and assume [Γ] = 0 in H1 (X; Z).

Definition 13.8 Let Z = Div(σ) be the divisor of a holomorphic section σ ∈

H 0 (X, O(λ
)) for some  ≥ 1. If Z does not meet Γ, we can define the linking
number and the reduced linking number by

Linkλ (Γ, Z) ≡ N • Z −  ω and J λ (Γ, Z) ≡ 1 Link(Γ, Z),
Link
N 
respectively, where N is any 2-chain in Z with dN = Γ and where the intersection
pairing • is defined as in (13.2) with Pn replaced by X.

To see that this is well defined

 suppose that N  is another 2-chain with dN  =
Γ. Then (N − N  ) • Z −  N −N  ω = (N − N  ) • (Z − [ω]) = 0 because Z − ω
is cohomologous to zero in X.

Definition 13.9 The winding number of a section σ ∈ H 0 (X, O(λ
)) with

σ > 0 on Γ is defined to be

Windλ (Γ, σ) ≡ dC logσ.
Γ

The reduced winding number is


 λ (Γ, σ) ≡ 1 Windλ (Γ, σ) =
Wind dC logσ
1

 Γ

From the Poincaré–Lelong equation

ddC logσ = Div(σ) − ω (13.17)
we see that
Windλ (Γ, σ) = Linkλ (Γ, Div(σ)). (13.18)
1
From (13.17) we also see that logσ belongs to the class P SH ω (X) of quasi-
plurisubharmonic functions on X defined in (13.9).

Proposition 13.7 The following are equivalent:

J λ (Γ, Z) ≥ −Λ for all divisors Z of holomorphic sections of λ
, and
1. Link
all  > 0.
 λ (Γ, σ) ≥ −Λ for all holomorphic sections σ of λ
, and all  > 0.
2. Wind
 C
3. Γ
d u ≥ −Λ for all u ∈ P SH ω (X).

Proof. That (1) ⇔ (2) ⇒ (3) is clear. To see that (3) ⇒ (2) we use results
of Demailly. Fix u ∈ P SH ω (X) and consider the positive closed (1,1)-current
T ≡ ddC u + ω. Note that [T ] = [ω] = c1 (λ) ∈ H 2 (X; Z). It follows from Demailly
 Relative holomorphic cycles 277

(1982) that T is the weak limit

1
T = lim Div(σj )
j→∞ Nj

where σj ∈ H 0 (X, O(λ
j )) and Nj > 0. We can normalize this sequence by

scalars so that M( N1j Div(σj )) = M(T ) for all j. Set Ω = (n−1)!
1
ω n−1 . Then



M(T ) = M( N1j Div(σj )) = ( N1j Div(σj ), Ω) = N1j (j ω, Ω) = Njj (ω, Ω) = Njj M(T ).
Therefore, j = Nj for all j. Since Div(σj ) = ddC logσj  + j ω, we conclude
that
< =
1 1
ddC u = T − ω = lim ddC logσj  j + ω − ω = lim ddC logσj  j .
j→∞ j→∞

The remainder of the argument replicates the one given for Proposition 13.4.


For a compact subset K ⊂ X the authors introduced the notion of the λ-hull
KG λ of K and showed that K Gλ = KG for any embedding X → PN by sections of
some power of λ. We shall say that K is λ-stable if K G λ is compact.
M
Theorem 13.11 Let Γ = a=1 mα Γα be an embedded, oriented, real analytic
closed curve with integer multiplicities in X, and assume Γ is λ-stable. Then
there exists a positive holomorphic 1-chain T in X with dT = Γ and M(T ) ≤ Λ
if and only if any of the equivalent conditions of Proposition 13.6 is satisfied.
Proof. If dT = Γ and M(T ) ≤ Λ, then (1) follows as in the proof of Theorem
13.8 above.
For the converse we recall from Harvey and Lawson (2006a, (4.4)) that the
G λ of a compact subset K ⊂ X consists exactly of the points x where the
hull K
extremal function
ΛK (x) ≡ sup{u(x) : u ∈ P SH ω (X) and u ≤ 0 on X}
is finite. This enables one to directly carry through the arguments for Theorem
13.9 in this case.

13.8 Relative holomorphic cycles

Our main Theorem 13.9 has a nice reinterpretation in terms of the Alexander–
Lefschetz duality pairing discussed in Section 13.3.
Theorem 13.12 Let γ ⊂ Pn be a finite disjoint union of real analytic curves
with γ-stable. Then a class τ ∈ H2 (Pn , γ; Z) is represented by a positive holo-
morphic chain with boundary on γ if and only if
τ •u ≥ 0 (13.19)
for all u ∈ H2n−2 (P − γ; Z) represented by positive algebraic hypersurfaces in
n

Pn − γ.
278 Projective linking and boundaries of positive holomorphic chains

Proof. The implication ⇒ is clear from the positivity of complex intersections.

For the converse, consider the short exact sequence
δ
0 −→ H2 (Pn ; Z) −→ H2 (Pn , γ; Z) −→ H1 (γ; Z) −→ 0.
Note that



δτ = mk −
→
γk ≡ Γ
k=1

where →−
γ 1 , ..., −
→
γ
are the connected components of γ with a chosen orientation
and the mk ’s are integers which we can assume to be positive. We are assuming
that (13.19) holds for any positive algebraic class u. If Γ = 0, the desired
conclusion is immediate, so we assume that Γ = 0. Now let Z ⊂ Pn − γ be any
positive algebraic hypersurface, and note that
τ •Z τ •Z
0 ≤ = − τ (ω) + τ (ω) = LinkP (Γ, Z) + τ (ω).
degZ degZ
Therefore, by Theorem 13.9, there exists a positive holomorphic 1-chain T with
dT = Γ and M (T ) ≤ τ (ω). Note that δ([T ] − τ ) = 0 and ([T ] − τ )(ω) ≥ 0. Hence,
[T ] − τ is a positive class in H2 (Pn ; Z) and is represented by a positive algebraic
1-cycle, say W . Therefore, τ = [T + W ] is represented by a positive holomorphic
chain as claimed.

This result leads to a nice duality between the cone in H2 (Pn , γ; Z) of those
classes which are represented by positive holomorphic chains, and the cone in
H2n−2 (Pn − γ; Z) of classes represented by positive algebraic hypersurfaces. This
duality, in fact, extends to more general projective manifolds. All this is discussed
in detail in Harvey and Lawson (2006d).

Acknowledgment
This chapter was partially supported by the NSF.

References
Alexander, H. and Wermer, J. (1998). Several Complex Variables and Banach Algebras,
Springer-Verlag, New York.
Alexander, H. and Wermer, J. (2000). Linking numbers and boundaries of varieties.
Ann. Math. 151, 125–150.
Demailly, J.-P. (1982). Estimations L2 pour l’opérateur ∂ d’un fibré vectoriel holomor-
phe semi-positif au-dessus d’une variété kählérienne complète. Ann. Sci. E. N. S.
15(4), 457–511.
Demailly, J.-P. (1982). Courants positifs extrêmaux et conjecture de Hodge. Invent.
Math. 69(4), 347–74.
Dinh ,T.-C. and Lawrence, M. (2003). Polynomial hulls and positive currents. Ann.
Fac. Sci. de Toulouse 12, 317–34.
 References 279

Dolbeault, P. (1983). On holomorphic chains with given boundary in CPn . Springer

Lect. Notes Math. 1089, 1135–40.
Dolbeault, P. and Henkin, G. (1994). Surfaces de Riemann de bord donné dans CPn ,
pp. 163–87 in Contributions to Complex Analysis and Analytic Geometry, H. Skoda
and J.-M. Trépreau (eds.), Aspects of Mathematics, Vol. 26, Vieweg, Wiesbaden,
Germany.
Dolbeault, P. and Henkin, G. (1997). Chaı̂nes holomorphes de bord donné dans CPn .
Bull. Soc. Math. France 125, 383–445.
Duval, J. and Sibony, N. (1995). Polynomial convexity, rational convexity and currents.
Duke Math. J. 79, 487–513.
Federer, H. (1969). Geometric Measure Theory, Springer–Verlag, New York.
Gillet H. and Soulé, C. (1989). Arithmetic chow groups and differential characters,
pp. 30–68, in Algebraic K-theory; Connections with Geometry and Topology,
J. F. Jardine and V. P. Snaith (eds.), Kluwer Academic Publishers Amsterdam.
Guedj, V. (1999). Approximation of currents on complex manifolds. Math. Ann. 313,
437–474.
Guedj V. and Zeriahi, A. (2003). Intrinsic capacities on compact Kähler manifolds,
Preprint, Université de Toulouse.
Harris, B. (1989). Differential characters and the Abel-Jacobi map, pp. 69–86, in
Algebraic K-theory; Connections with Geometry and Topology, J. F. Jardine and
V. P. Snaith (eds.), Kluwer Academic Publishers Amsterdam.
Harvey, F. R. (1977). Holomorphic chains and their boundaries, pp. 309–32, in Several
Complex Variables, Proceedings of Symposia in Pure Mathematics XXX Part 1,
A.M.S., Providence, RI.
Harvey, F. R. and Lawson, H. B. Jr., (1975). On boundaries of complex analytic
varieties, I. Ann. Math. 102, 223–90.
Harvey, F. R. and Lawson, H. B. Jr., (1977). On boundaries of complex analytic
varieties, II. Ann. Math. 106, 213–38.
Harvey, F. R. and Lawson, H. B. Jr., (1983). An intrinsic characterization of Kähler
manifolds. Invent. Math. 74, 169–98.
Harvey, F. R. and Lawson, H. B. Jr., (2005). Boundaries of varieties in projective
manifolds. J. Geom. Analysis 14(4), 673–95. ArXiv:math.CV/0512490.
Harvey, F. R. and Lawson, H. B. Jr., (2006a). Projective hulls and the projec-
tive Gelfand transformation. Asian J. Math. 10(2), 279–318. ArXiv:math.CV/
0510286.
Harvey, F. R. and Lawson, H. B. Jr., (2006b). Remarks on the Alexander-Wermer the-
orem for curves. Matemática Contemporânea 31, 53–64. ArXiv: math.CV/0610611.
Harvey, F. R. and Lawson, H. B. Jr., (2006c). Projective linking and boundaries of
holomorphic chains in projective manifolds, Part II, pp. 365–80, in Inspired by S. S.
Chern, P. Griffiths (ed.), Nankai Tracts in Mathematics, Vol. 11, World Scientific
Publishing, Singapore.
Harvey, F. R. and Lawson, H. B. Jr., (2006d). Relative holomorphic cycles and duality,
Preprint, Stony Brook.
Harvey, F. R. Lawson, H. B. Jr., and Zweck, J. (2003). A de Rham-Federer theory of
differential characters and character duality. Am. J. Math. 125, 791–847.
Harvey, F. R. Lawson, H. B. Jr., and Wermer, J. (2008). The projective hull of certain
curves in C2 . Astérisque 322, 241–54. ArXiv:math.CV/0611482.
280 Projective linking and boundaries of positive holomorphic chains

Harvey, F. R. and Polking, J. (1975). Extending analytic objects. Commun. Pure Appl.
Math. 28, 701–27.
Schaefer, H. H. (1999). Topological Vector Spaces, Springer Verlag, New York.
Wermer, J. (1958). The hull of a curve in Cn . Ann. Math. 68, 550–61.
Wermer, J. (2001). The argument principle and boundaries of analytic varieties.
Operator Theory: Adv. Appl. 127, 639–59.
 XIV
SKYRMIONS AND NUCLEI

Nicholas S. Manton
Dedicated to Nigel Hitchin on the occasion of his 60th birthday

14.1 Skyrmions
T. H. R. Skyrme (1961, 1962) proposed that the interior of a nucleus is dominated
by a non-linear semiclassical medium formed from the three pion fields, and
he introduced the Skyrme model, a version of the Lorentz-invariant, non-linear
sigma model, in which the pion fields π = (π1 , π2 , π3 ) are combined into an
SU (2)-valued scalar field
U = (1 − π · π)1/2 1 + iπ · τ , (14.1)
where τ are the Pauli matrices. There is an associated current, with spatial
components Ri = (∂i U )U † . For static fields, the energy in the Skyrme model is
given by
 < =
1 1 1
E= − Tr(R R
i i ) − Tr([R i , R j ][R i , R j ]) + m2
Tr(1 − U ) d3 x ,
12π 2 2 16
(14.2)
and the vacuum is U = 1. E is invariant under translations and rotations in R
3

and also under SO(3) isospin rotations given by the conjugation

U → OU O † , O ∈ SU (2) . (14.3)
This rotates the pion fields among themselves. Stationary points of E satisfy the
Skyrme field equation, and we shall mostly consider minima of E. (For a review,
see Manton and Sutcliffe 2004.)
The expression (14.2) is in ‘Skyrme units’ and m is a dimensionless pion
mass parameter. We will discuss below the calibration of the energy and length
units by comparison with physical data. Traditionally m has been given a value
of approximately 0.5 (Adkins and Nappi 1984), but recent work suggests a
higher value, m = 1 (Battye et al. 2005; Battye and Sutcliffe 2006) or m = 1.125
(Manton and Wood 2006).
The model has a conserved, integer-valued topological charge B, the baryon
number. This is the degree of the map U : R → SU (2), which is well-defined
3
282 Skyrmions and nuclei

1 : O(3) 2 : Dµh 3 : Td 4 : Oh

5 : D2d 6 : D4d 7: Yh 8 : D6d

Figure 14.1 Skyrmions for 1 ≤ B ≤ 8, with m = 0. A surface of constant

baryon density is shown, together with the baryon number and symmetry.

because U → 1 at spatial infinity. B is the integral of the baryon density

1
B=− ijk Tr(Ri Rj Rk ), (14.4)
24π 2
which is proportional to the Jacobian of the map U .
In the above units there is the Faddeev–Bogomolny energy bound, E ≥ |B|.
The minimal energy solutions for each B are called Skyrmions. (More loosely,
local minima and saddle points of E with nearby energies are also sometimes
called Skyrmions.) Skyrme argued that these topological soliton solutions can
be identified with nucleons and nuclei.
The B = 1 Skyrmion has the spherically symmetric, hedgehog form:
U (x) = exp {if (r)G
x · τ } = cos f (r)1 + i sin f (r)G
x·τ . (14.5)
f is a radial profile function obeying an ordinary differential equation (ODE)
with the boundary conditions f (0) = π and f (∞) = 0. Skyrmions with baryon
numbers greater than 1 all have interesting shapes (see Figure 14.1); they are not
spherical like the basic B = 1 Skyrmion. The B = 2 Skyrmion is toroidal, and the
B = 3 Skyrmion tetrahedral. The B = 4 Skyrmion is cubic and can be obtained
by bringing together two B = 2 toroids along their common axis. The B = 6
solution has D4d symmetry and can be formed from three toroids stacked one
above the other, and the B = 7 Skyrmion has icosahedral symmetry. The toroidal
structure of the B = 2 Skyrmion has some phenomenological support from
nuclear physics (Forest et al. 1996). Table 14.1 presents, for m = 0 and baryon
numbers 1–8, the symmetries and energies of the Skyrmions, computed from
numerical solutions of the field equation (Battye and Sutcliffe 1997, 2001, 2002).
 Rational map ansatz 283

Table 14.1. Symmetry K and energy per baryon

E/B for numerically computed Skyrmions

B K E/B

1 O(3) 1.2322
2 D∞h 1.1791
3 Td 1.1462
4 Oh 1.1201
5 D2d 1.1172
6 D4d 1.1079
7 Yh 1.0947
8 D6d 1.0960

SU (2) Yang–Mills–Higgs monopoles, satisfying the Bogomolny equation, have

been constructed with very similar symmetries (and baryon number replaced
by monopole charge) (Hitchin et al. 1995), and there was, historically, an inter-
esting interplay between the discovery of symmetric monopoles and symmetric
Skyrmions.
When m = 0, the Skyrmions for baryon numbers from B = 8 up to B = 22
(Battye and Sutcliffe 2001, 2002), and various larger values of B (Battye et al.
2003) are hollow polyhedra, rather like carbon fullerenes. The baryon density is
concentrated in a shell of roughly constant thickness, surrounding a region whose
volume increases like B 3/2 where the energy and baryon density are very small.
Such a hollow structure clearly disagrees with the rather uniform baryon density
observed in the interior of real nuclei.
However, for baryon numbers B ≥ 8, the hollow polyhedral Skyrmions do not
remain stable when the pion mass parameter m is of order 1 (Battye and Sutcliffe
2005, 2006). This is because, in the interior of the hollow polyhedra, the Skyrme
field is very close to U = −1 (i.e. the antipode to the vacuum value), and here the
pion mass term gives the field a maximal potential energy density. Consequently,
this interior region tends to pinch off and separate into smaller sub-regions.
The new, stable Skyrmions have a denser structure. Some of them are clusters
of B = 4 Skyrmions; these are described in Section 14.3. Others, for 10 ≤ B ≤ 16,
have a planar character. It should be possible to interpret these solutions as
fragments of an infinite crystalline sheet with hexagonal (or perhaps in some
cases, square) symmetry. This would be a two-layer version of the solution
presented in Battye and Sutcliffe (1998) (which by itself has the wrong boundary
conditions).

14.2 Rational map ansatz

The observed similarities between Skyrmions and monopoles suggests there could
be an approximate construction of Skyrmions from monopoles. As yet, there is
no known direct transformation between the fields of a monopole and those of
284 Skyrmions and nuclei

a Skyrmion, but there is an indirect transformation via rational maps between

Riemann spheres. It is known that there is a precise one-to-one correspondence
between charge N monopoles and degree N rational maps (we have in mind here
the Jarvis 2000 maps); thus a Skyrme field constructed from a rational map is
indirectly constructed from a monopole.
The rational map ansatz of Houghton et al. (1998) separates the angular from
the radial dependence of the Skyrme field U . One introduces a complex (Riemann
sphere) coordinate z = tan θ2 eiφ , where θ and φ are the usual spherical polar
coordinates, and constructs the Skyrme field from a rational function of z:

p(z)
R(z) = (14.6)
q(z)

where p and q are polynomials with no common root, and a radial profile function
f (r) satisfying f (0) = π and f (∞) = 0. One should think of R as a smooth map
from a two-sphere in space (at a given radius) to a two-sphere in the target SU (2)
(at a given distance from the identity). By standard stereographic projection, the
point z corresponds to the Cartesian unit vector
1
Gz =
n (z + z̄, i(z̄ − z), 1 − |z|2 ) . (14.7)
1 + |z|2

Similarly, an image point R can be expressed as a unit vector

1
GR =
n (R + R̄, i(R̄ − R), 1 − |R|2 ) . (14.8)
1 + |R|2

The ansatz for the Skyrme field is then

nR(z) · τ ) ,
U (r, z) = exp(if (r)G (14.9)

generalizing the hedgehog formula (14.5). The ansatz is a suspension of the map
R : S 2 → S 2 to produce a map U : R → S 3 , the suspension points being the
3

origin and infinity in R , which are mapped to U = −1 and U = 1, respectively.

3

The baryon number, B, of this Skyrme field equals the topological degree of
the rational map R : S 2 → S 2 , which is the higher of the algebraic degrees of
p and q.
An SU (2) Möbius transformation on the domain S 2 of the rational map
corresponds to a spatial rotation, whereas an SU (2) Möbius transformation
on the target S 2 corresponds to a rotation of n G R , and hence to an isospin
rotation of the Skyrme field. Thus if a rational map R has some symmetry
(i.e. a rotation of the domain can be compensated by a rotation of the target),
then the resulting Skyrme field has that symmetry (i.e. a spatial rotation can be
compensated by an isospin rotation).
An important feature of the rational map ansatz is that when one substitutes
it into the Skyrme energy function (14.2), the angular and radial parts decouple.
 Rational map ansatz 285

The energy simplifies to

 ∞

1 sin4 f
E= r2 f 2 + 2B(f 2 + 1) sin2 f + I + 2m2 2
r (1 − cos f ) dr,
3π 0 r2
(14.10)
where I denotes the angular integral

4
1 1 + |z|2 dR 2i dzdz̄
I= , (14.11)
4π 1 + |R|2 dz (1 + |z|2 )2
which only depends on the rational map R. I is an interesting function on the
space of rational maps. To minimize the energy (for given B), it is sufficient to
first minimize I with respect to the coefficients occurring in the rational map,
and then to solve an ODE for f (r) whose coefficients depend on the rational map
only through B and the minimized I. Optimal rational maps, and the associated
profile functions, have been found for many values of B, and often have a high
degree of symmetry. The optimized fields within the rational map ansatz are good
approximations to Skyrmions, and they are also used as a starting point for a
numerical relaxation to true Skyrmion solutions (which almost always have the
same symmetry, but no exact separation of the angular and radial dependence
of U ).
The simplest degree 1 rational map is R(z) = z, which is spherically symmet-
ric. The ansatz (14.9) then reduces to the hedgehog field (14.5). For B = 2, 3, 4, 7,
the symmetries of the numerically computed Skyrmions are D∞h , Td , Oh , Yh ,
respectively. In each of these cases there is a unique rational map with this
symmetry, up to rotations and isorotations, namely,
√ √
z 3 − 3iz z 4 + 2 3iz 2 + 1
R(z) = z , R(z) = √
2
, R(z) = √ ,
3iz 2 − 1 z 4 − 2 3iz 2 + 1
z 7 − 7z 5 − 7z 2 − 1
R(z) = , (14.12)
z 7 + 7z 5 − 7z 2 + 1
and these also minimize I. For B = 5, 6, and 8, rational maps with dihedral
symmetries are required, and these involve one or two coefficients that need
to be determined numerically. Table 14.2 lists the energies of the approximate
solutions obtained using the rational map ansatz, together with the values of I,
again for m = 0 (Battye and Sutcliffe 2001, 2002).
The Wronskian of a rational map R(z) = p(z)/q(z) of degree B is the polyno-
mial
W (z) = p (z)q(z) − q  (z)p(z) (14.13)
of degree 2B − 2. Where W is zero, the derivative dR/dz is zero, so only the
radial derivative of U is non-vanishing. The baryon density therefore vanishes
along the entire radial half-line in the direction of a zero of W , and the energy
density is also low. This explains why the Skyrmion baryon density contours
286 Skyrmions and nuclei

Table 14.2. Symmetry K, the value of the angular

integral I, and the energy per baryon E/B of the
approximate Skyrmions obtained using the rational
map ansatz

B K I E/B

1 O(3) 1.0 1.232

2 D∞h 5.8 1.208
3 Td 13.6 1.184
4 Oh 20.7 1.137
5 D2d 35.8 1.147
6 D4d 50.8 1.137
7 Yh 60.9 1.107
8 D6d 85.6 1.118

look like polyhedra with holes in the directions given by the zeros of W , and
why there are 2B − 2 such holes, precisely the structures seen in Figure 14.1.
As an example, the icosahedrally symmetric degree 7 map in (14.12) has
Wronskian
W (z) = 28z(z 10 + 11z 5 − 1), (14.14)
which is proportional to one of the icosahedral Klein polynomials, and vanishes
at the twelve face centres of a dodecahedron (including z = ∞).
The solutions we have described so far are for m = 0, but it is found that
qualitatively similar solutions with 10–20% higher energy exist for m up to 1
and a little beyond, provided B ≤ 7. There is, however, a qualitative change for
B > 7, as we discuss next.

14.3 Skyrmions and α-particles

The 4 He nucleus, or α-particle, is particularly stable and can be regarded as a
building block of larger nuclei. The α-particle model (Blatt and Weisskopf 1952;
Brink et al. 1970; Wuosmaa et al. 1995; Von Oertzen et al. 2006) has considerable
success describing the nuclei 8 Be, 12 C, 16 O, etc. with baryon numbers a multiple
of four and having equal numbers of protons and neutrons, as ‘molecules’ of
pointlike α-particles. For m = 1, new Skyrmion solutions with baryon number
a multiple of four have recently been found (Battye et al. 2007), which make
contact with the α-particle model. These new solutions are clusters of cubic
B = 4 Skyrmions, and for B ≥ 12 they are energetically more stable than the
hollow polyhedral Skyrmions, the effect being marginal for B = 8.

14.3.1 B = 8
When m = 0, the B = 8 Skyrmion is a hollow polyhedron with D6h symmetry,
with no obvious relation to a pair of cubic B = 4 Skyrmions, and this solution
persists at m = 1. However, motivated by the α-particle model, one might expect
 Skyrmions and α-particles 287

Figure 14.2 Baryon density contours for (a) two B = 4 cubes with one of the
cubes rotated by 90◦ around the line joining them, (b) the B = 8 truncated
octahedron, and (c) the relaxed B = 8 Skyrmion with m = 1.

that at m = 1, the lowest energy solution is a ‘molecule’ of two cubic, B = 4

Skyrmions. Two such Skyrmions, placed initially in the same orientation and next
to each other, have a weak quadrupole–quadrupole attraction (Manton 1994;
Manton et al. 2004). Because of a significant short-range octupole interaction
in the single pion field component that has no quadrupole moment, it is best
to also twist one cube by 90◦ relative to the other around the axis joining
them (see Figure 14.2a). An alternative starting configuration has the shape
of a truncated octahedron and is obtained using the rational map ansatz with a
cubically symmetric degree 8 map (see Figure 14.2b).
Numerical relaxation from either starting point produces the stable solution
displayed in Figure 14.2c, which has D4h symmetry. For m = 1, the energy per
baryon of this new Skyrmion and also of the old D6h -symmetric Skyrmion is
1.294. The change of structure therefore has a marginal effect in this case, but
one expects that for m > 1 and for larger B, clusters of B = 4 cubes will be the
more stable solutions. Note that there is a clear attraction between B = 4 cubes,
because the energy per baryon of the B = 4 cube is 1.307 for m = 1. Numerical
errors are estimated as 0.5% or less.

14.3.2 B = 12
In the α-particle model, three α-particles form an equilateral triangle. This moti-
vates the search for a triangular B = 12 solution in the Skyrme model, composed
of three B = 4 cubes. A configuration with approximate D3h symmetry can be
obtained with each cube related to its neighbour by a spatial rotation through
120◦ followed by an isorotation by 120◦ . The isorotation cyclically permutes the
values of the pion fields on the faces of the cube, so that these values match
on touching faces, and the cubes attract. It is fairly easy to see that around the
centre of the triangle the field has a winding equivalent to a single Skyrmion, and
is unstable to this Skyrmion moving down and merging with the bottom face
of the triangle; in fact, it fills a hole in the baryon density. This C3 -symmetric
Skyrmion with E/B = 1.288 is presented in Figure 14.3.
288 Skyrmions and nuclei

Figure 14.3 Top and bottom views of the B = 12 Skyrmion with triangular
symmetry.

The top view looks similar to the minimal energy B = 11 Skyrmion with
m = 0, and suggests that the initial arrangement of three cubes can also be
viewed as the m = 1 version of this B = 11 Skyrmion with a single Skyrmion
placed inside at the origin. Such a field configuration can be constructed with
exact D3h symmetry using the double rational map ansatz (Manton and Piette
2001) (see Appendix). This involves a D3h -symmetric outer map of degree 11,
Rout , and a spherically symmetric degree 1 inner map, Rin , together with an
overall radial profile function. The maps are

z 2 (1 + az 3 + bz 6 + cz 9 )
Rout (z) = (14.15)
c + bz 3 + az 6 + z 9
1
Rin (z) = − , (14.16)
z
where a = −2.47, b = −0.84, and c = −0.13. Note that the orientation of Rin has
to be chosen compatibly with the D3h symmetry of Rout . Numerically relaxing
the field yields a solution with exact D3h symmetry, but as we discussed above,
this is only a saddle point and not a local energy minimum.
In Battye and Sutcliffe (2006) another B = 12 solution with C3 symmetry
was found, with E/B = 1.289. It is a general observation that rearrangements
of clusters have only a tiny effect on the energy of a Skyrmion, so as B increases
one expects an increasingly large number of local minima with extremely close
energies.
Rearranged solutions are analogous to the rearrangements of the α-particles
which model excited states of nuclei. An example is the Skyrme model analogue
of the three α-particles in a chain configuration modelling the 7.65 MeV excited
state of 12 C (Morinaga 1956; Friedrich et al. 1971). This is obtained from three
B = 4 cubes placed next to each other in a line, with the middle cube twisted
relative to the other two by 90◦ around the axis of the chain. The relaxed solution
is displayed in Figure 14.4 and has E/B = 1.285. This is apparently the lowest
energy of the B = 12 solutions, but note that the energy difference we are trying
 Skyrmions and α-particles 289

Figure 14.4 B = 12 Skyrmion formed from three cubes in a line, with the
middle cube being rotated by 90◦ around the line of the cubes.

12
to understand, 7.65 MeV, is less than 0.1% of the total energy of a C nucleus,
smaller than the numerical errors in the Skyrmion energies.

14.3.3 B = 16
There is a tetrahedrally symmetric B = 16 solution which is an arrangement
of four B = 4 cubes. It is created using the double rational map ansatz as a
starting point, with an outer map of degree 12, Rout , and an inner map of degree
4, Rin , with compatible symmetries, and an overall radial profile function. There
is a Td -symmetric map Rout , and this can be combined with the Oh -symmetric
map familiar from the B = 4 Skyrmion, giving Td symmetry overall. The maps
are

ap3+ + bp3−
Rout = (14.17)
p2+ p−
p+
Rin = , (14.18)
p−
√
where p± (z) = z 4 ± 2 3 iz 2 + 1, a = −0.53, and b = 0.78. Letting the field U
relax, preserving the Td symmetry, results in the solution displayed in Figure
14.5a, in which U = −1 at 16 points clustered into groups of four close to the
centre of each cube. The solution has E/B = 1.288.
This tetrahedral solution is only a saddle point. It is energetically more
favourable for the two cubes on a pair of opposite edges of the tetrahedron
to open out, leading to the D2d symmetric solution in Figure 14.5b, which has
the slightly lower energy, E/B = 1.284. An α-particle molecule of similar shape
has also been found, where it is termed a bent rhomb (Bauhoff et al. 1984).
A stable tetrahedral solution would be preferable, since the closed shell
structure of 16 O is known to be compatible with clustering into a tetrahedral
arrangement of four α-particles. Moreover, the 16 O ground state and the excited
states at 6.1 MeV and 10.4 MeV, with spin/parity 0+ , 3− , and 4+ , and some
higher states, look convincingly like a rotational band for a tetrahedral intrinsic
structure (Dennison 1954; Robson 1979).
290 Skyrmions and nuclei

Figure 14.5 B = 16 Skyrmions composed of four cubes: (a) Tetrahedral

arrangement, (b) bent square, (c) flat square, and (d) another tetrahedral
arrangement.

There is a further metastable solution of low energy, in which four B = 4 cubes

all have the same orientation, and are connected together to form a flat square
(see Figure 14.5c). The solution has E/B = 1.293.
Yet another tetrahedral saddle-point solution of four cubes exists (see Fig-
ure 14.5d). Its E/B is 1.295. A perturbation that breaks the tetrahedral sym-
metry results in the bent square solution of Figure 14.5b.

14.3.4 B = 24
Here one may begin with six B = 4 cubes on the vertices of an octahedron, all
with the same spatial orientation (see Figure 14.6a). The cubes above and below
the four in a square are given an isospin rotation by 180◦ , as this aids their
attraction. The relaxed solution, shown in Figure 14.6b, resembles the B = 16
bent square Skyrmion of Figure 14.5b, joined with the B = 8 Skyrmion of Figure
14.2c. The fact that these B = 16 and B = 8 Skyrmions appear as substructures
is further evidence that they are the minimal energy arrangements of B = 4
cubes. The energy per baryon of this B = 24 Skyrmion is 1.282.

Figure 14.6 (a) Initial configuration of six cubes in a square bipyramid arrange-
ment and (b) the final relaxed B = 24 Skyrmion.
 Skyrmions and α-particles 291

Figure 14.7 (a) Initial configuration of four cubes in a square with three extra
cubes placed above the square but not aligned with the cubes below and (b) the
relaxed B = 28 Skyrmion.

14.3.5 B = 28
A starting configuration is shown in Figure 14.7a, with four cubes in a square
and three more placed above, all with the same space and isospace orientations.
Figure 14.7b displays the solution after relaxation. This B = 28 Skyrmion clearly
resembles the cubic B = 32 Skyrmion (see below) with one B = 4 cube removed.
The energy per baryon is 1.279, which is slightly lower than might have been
expected given the energies for B = 16 and B = 24.

14.3.6 B = 32
Even for relatively small values of the pion mass parameter m, the B = 32
Skyrmion is cubic, and has lower energy than the minimal energy, hollow
polyhedral structure (Battye and Sutcliffe 2006). The solution may be thought
of as eight B = 4 cubic Skyrmions placed on the vertices of a cube, each with the
same spatial and isospin orientations, and it may also be created by cutting out a
cubic B = 32 chunk from the infinite, triply periodic Skyrme crystal (Baskerville
1996).
Alternatively, it may be obtained beginning with the double rational map
ansatz. One places a B = 4 cube inside a B = 28 configuration with cubic
symmetry using the maps
p+ (ap6+ + bp3+ p3− − p6− )
Rout = (14.19)
p− (p6+ − bp3+ p3− − ap6− )
p+
Rin = , (14.20)
p−
where a = 0.33 and b = 1.64, and p± (z) are as before. This is displayed in Figure
14.8a. Numerical relaxation yields the solution in Figure 14.8b, which is the
B = 32 Skyrmion, with E/B = 1.274.
292 Skyrmions and nuclei

Figure 14.8 (a) Initial condition of the cubic B = 4 Skyrmion inside a cubic
B = 28 configuration and (b) the final relaxed B = 32 Skyrmion, which is a
chunk of the Skyrme crystal.

Note that slicing the B = 32 Skyrmion in half produces the square B = 16

solution of Figure 14.5c, whereas if one removes every alternate B = 4 cube then
the result is the unstable tetrahedral solution of Figure 14.5d.
Recall that a Skyrme field created using the rational map ansatz with a degree
B̃ map has a polyhedral structure with 2B̃ − 2 holes in the baryon density.
Figure 14.8b shows that the B = 32 crystal chunk contains 54 exterior holes and
this corresponds to a degree 28 map. This motivates the above double rational
map construction. The next crystal chunk with cubic symmetry contains 27
B = 4 cubes and therefore has B = 108. To create this from a triple rational
map ansatz, wrapping a third shell around the B̃ = 4 and B̃ = 28 shells requires
a map with degree 76. A degree 76 map has 150 holes, the correct number of
exterior holes for the B = 108 crystal chunk. However, there is a six-parameter
family of Oh -symmetric degree 76 maps and it has not yet been possible to
obtain an appropriate member of this family with a suitable distribution of the
150 holes, either analytically or numerically.

14.4 Quantization and calibration

Quantization is a vital issue for Skyrmions, because Skyrmions are supposed
to model physical nucleons (protons and neutrons) and nuclei, and a nucleon
is a spin 12 fermion. In the approach of Finkelstein and Rubinstein (1968), one
quantizes a soliton as a fermion by lifting the classical field configuration space
to its simply connected covering space. In the SU (2) Skyrme model, this is a
double cover for any value of B. States should be multiplied by a factor of −1
when acted upon by any operation corresponding to a circuit around a non-
contractible loop in the configuration space. Equivalently, the wavefunction has
opposite signs on the two points of the covering space that cover one point in
the configuration space. Finkelstein and Rubinstein showed that the exchange of
 Quantization and calibration 293

two B = 1 Skyrmions is a loop which is homotopic to a 2π rotation of one of the

B = 1 Skyrmions, in agreement with the spin–statistics result, and it was verified
by Williams (1970) that a 2π rotation of a B = 1 Skyrmion is a non-contractible
loop, thus requiring the Skyrmion to be quantized as a fermion. This result was
generalized by Giulini (1993), who showed that a 2π rotation of a Skyrmion of
baryon number B is a non-contractible loop if B is odd and contractible if B is
even.
A practical, approximate quantum theory of Skyrmions is achieved by a rigid
body quantization of the spin and isospin rotations. Vibrational modes, whose
excited states often have considerably higher energy, are ignored.
For the B = 1 Skyrmion, this quantization was carried out by Adkins, Nappi,
and Witten (1983), who showed that the lowest energy states (compatible with
the Finkelstein–Rubinstein [FR] constraints) have spin 12 , and may be identified
with the proton/neutron isospin doublet. The next lowest states are identified
with spin 32 delta resonances. The masses of the nucleons and deltas were used
to calibrate the Skyrme model. However, the delta is a broad resonance about
300 MeV above the nucleon ground state, and it strongly radiates pions. Related
to this, it has been recently observed that the spin of the delta strongly deforms
the B = 1 Skyrmion (Battye et al. 2005; Houghton and Magee 2006), and if this
is taken into account, it has a rather complicated effect on the calibration of the
model.
The Skyrmions with baryon numbers B = 2, 3, 4, and 6 have the right proper-
ties to model the deuteron 2 H, the isospin doublet 3 H/3 He, the α-particle 4 He,
and the nucleus 6 Li (Braaten and Carson 1988; Carson 1991; Walhout 1992; Irwin
2000). In each case, the collective coordinate quantization is constrained by the
symmetries of the classical solution, and the need to impose FR constraints. The
resulting lowest energy states for B = 2, 3, 4, and 6 have spin/parity, respectively,
+
J P = 1+ , 12 , 0+ , and 1+ (with isospin T = 0 for B = 2, 4, 6 and T = 12 for
B = 3), agreeing with those of real nuclei. However, in the traditional calibration
of the model, the binding energies are too large and the sizes too small.
Collective coordinate quantization of the dodecahedral B = 7 Skyrmion leads
to a lowest spin of J = 72 when T = 12 (Irwin 2000; Krusch 2003), disagreeing
with the experimental value J = 32 for the ground state of the isospin doublet
7
Li/7 Be. This suggests that the Skyrmion is too symmetric to model the ground
state and it would be preferable if a less symmetric solution existed, which could
have a larger classical energy, but be quantized with a lower spin (Manko and
Manton 2007). Quantum states of the B = 5 Skyrmion also differ from those of
5
He/5 Li.
Finding the allowed spin and isospin states of Skyrmions has never been easy.
There is a tricky interplay of the representations of the symmetry group of the
Skyrmion and the FR constraints. Krusch (2003, 2006) has been able to simplify
part of the calculation by exploiting the known topology of the space of rational
maps (Segal 1979). Suppose a Skyrmion (approximately described by the rational
map ansatz) is invariant under a combined rotation and isorotation. Krusch has
294 Skyrmions and nuclei

found a simple formula for determining whether the FR factor associated with
this symmetry operation is +1 or −1, in terms of the angles of rotation and the
baryon number.
Recently, a recalibration of the Skyrme model around the properties of the
6
Li nucleus has been performed (Manton and Wood 2006). This is better suited
to nuclear physics applications. The motivation for the choice of 6 Li is that it
is a small nucleus of isospin zero, with a pair of levels that can reasonably be
interpreted as a rotational band. Because the isospin is zero, the electric charge
density is half the baryon density. The ground state of spin 1 and the first
excited state of spin 3 are separated by just a few MeV, whereas the mass of
the 6 Li nucleus is approximately 5600 MeV. The rotational motion is therefore
quite non-relativistic, not leading to strong pion radiation, nor to significant
Skyrmion deformation, which could affect the calibration. The B = 6 Skyrmion
is also well-known, and can be approximated by the rational map ansatz, which is
useful when estimating the energy and size. In the new calibration, the conversion
from Skyrme units to physical energy and length units is fixed by fitting the 6 Li
mass and charge radius. m is determined as usual from the physical pion mass,
138 MeV, but because of the change of units, m is roughly doubled and is now
m = 1.125, which is in the range where the solutions described in Section 14.3,
constructed from B = 4 cubes, are favoured.
In the new calibration, the masses of Skyrmions and nuclei are rather similar
to those in the traditional calibration of Adkins and Nappi (1984), but the length
scale is roughly doubled. This means that Skyrmion moments of inertia are four
times larger, and hence the energy splitting of rotational levels is four times
smaller. These moments of inertia (both rotational and isorotational) have been
estimated for a range of baryon numbers, using the rational map ansatz where
appropriate, and used to determine the energies of spin and isospin excited
states for several nuclei (Manko et al. 2007). The results appear much closer
to experimental nuclear energy levels than those found previously. For example,
the spectrum of excited states of the double cube B = 8 Skyrmion matches well
the experimental spectrum of 8 Be and its isobars 8 Li, 8 B, etc. There are also
some detailed suggestions for a Skyrme model understanding of the physically
observed spin states for B = 5 and B = 7, which, as mentioned above, have so
far been problematic.

14.5 Conclusions
For small baryon numbers, the rational map ansatz gives good approximations to
Skyrmions, some of which have the symmetries of Platonic solids. The ansatz has
recently been shown to be helpful both in classifying the allowed spin and isospin
states of quantized Skyrmions, and also for estimating the energies, radii, and
moments of inertia of Skyrmions. The Skyrme model with pion mass parameter
of order 1 has qualitatively new solutions for baryon numbers B ≥ 8. These
are not hollow polyhedra, the solutions found for m = 0 or m = 0.528 (the
 Conclusions 295

traditional value), but more dense structures often with clear clustering into
B = 4 cubes, the Skyrme model analogue of α-particle molecules. This gives an
improved fit of the model to nuclei like 8 Be and 12 C, provided the energy scale
and especially the length scale of the model are recalibrated. The double rational
map ansatz gives some insight into the shapes of Skyrmions for baryon numbers
in the range 8 ≤ B ≤ 32, and an improved version of this ansatz is proposed
(see Appendix).
The interpretation of the Skyrmions discussed in this chapter has been that of
nucleons and nuclei. However, there has been considerable interest in recent years
in studying variants of non-linear scalar field theories in the context of condensed
matter physics, the classic example being the continuum limit of the Heisenberg
model of ferromagnets. The Skyrmions that occur in such theories are often
localized in two dimensions, which means they can become extended vortex-like
structures in a three-dimensional material. For a recent analysis of one model and
further references, see Rössler et al. (2006). If the topology of the model is as in
the original Skyrme model, then Skyrmions with three-dimensional localization
should occur in the interior of the material, and it would be interesting if evidence
could be found for the polyhedral structures of higher charge Skyrmions shown
in Figure 14.1.

Appendix: Double rational map ansatz

Here, we briefly review the double rational map ansatz for Skyrmions (Manton
and Piette 2001), which generalizes the original rational map ansatz (Houghton
et al. 1998). This uses two rational maps Rin (z) and Rout (z), with a profile
function f (r) satisfying f (0) = 2π and f (∞) = 0. It is assumed that f decreases
monotonically as r increases, passing through π at a radius r0 . The ansatz
for the Skyrme field is again (14.9), with the understanding that for r ≤ r0 ,
R(z) = Rin (z), and for r > r0 , R(z) = Rout (z). Notice that now U = 1 both at
the origin and at spatial infinity. It can be shown that the total baryon number
is the sum of the degrees of the maps Rin and Rout . The ansatz is optimized
by adjusting the coefficients of both maps, allowing variations of r0 , and solving
for f (r). All this is quite hard, but easier if Rin and Rout share a substantial
symmetry.
It has recently been realized that the double rational map ansatz is a special
case of Skyrme’s product ansatz (Skyrme 1962), in which Skyrme fields (often
Skyrmion solutions, but not here) U1 (x) with baryon number B1 and U2 (x) with
baryon number B2 are simply multiplied, giving the field U (x) = U1 (x)U2 (x)
with baryon number B1 + B2 . To express the double rational map ansatz in this
way, define U1 by the original rational map ansatz (14.9), with rational map Rin
and profile function f1 (r) = f (r) − π for r ≤ r0 , f1 (r) = 0 for r ≥ r0 , and define
U2 similarly, with rational map Rout and profile function f2 (r) = π for r ≤ r0 ,
f2 (r) = f (r) for r ≥ r0 . U2 takes the value −1 in the region r ≤ r0 , and this
explains why f1 differs by π from f .
296 Skyrmions and nuclei

A problem with the double rational map ansatz is that U takes the value −1
on the entire sphere at radius r0 , and true Skyrmions do not have this behaviour.
But the formulation above suggests how this problem can be avoided. One simply
needs to relax the non-overlapping character of the product of U1 and U2 by
allowing both f1 and f2 to decrease freely from π at r = 0 to 0 at r = ∞, without
further constraint at an intermediate radius r0 . One would expect f1 to decrease
more rapidly than f2 if the Skyrmion has a genuine inner and outer structure.
This version of the product ansatz preserves the joint rotational symmetries of
U1 and U2 , but not any inversion or reflection symmetries. As usual with the
product ansatz, the field U is sensitive to the order in which U1 and U2 are
multiplied, and the loss of reflection symmetry is related to this.
It is easy to verify that a small perturbation of the profiles f1 and f2 away
from their initial, non-overlapping form will tend to reduce the energy. This is
because the trajectory of the field U along a generic radial line will no longer
pass through U = −1, but will rather take a short cut, which reduces the radial
derivative of U without a significant increase in the angular derivatives, and
also reduces the potential energy. The non-generic lines are those for which
Rin (z) = Rout (z), and there are B of these, counted with multiplicity. Therefore,
U = −1 at B points, the number required topologically if they all have positive
multiplicity. The points lie on these special lines and are all at the same distance
from the origin. Further investigations are needed. With S. Krusch, the author
is attempting to understand this improved double rational map ansatz more
systematically, in order to find its optimal form. Alternatively, one could try ad
hoc ansätze for f1 and f2 , perhaps with one scale parameter each, and seek to
minimize the energy of U numerically.

Acknowledgements
This chapter is largely based on Battye et al. (2007) and part of Manton and
Sutcliffe (2004). I would like to acknowledge the contributions of Richard Battye
and Paul Sutcliffe to these joint works, and especially thank them for all the
figures that are reproduced here.
This is also an occasion to express my debt to Nigel Hitchin’s mathemati-
cal insight. His work on monopoles, in particular, has inspired several of the
developments discussed here.

References
Adkins, G. S. and Nappi, C. R. (1984). The Skyrme model with pion masses. Nucl.
Phys. B 233, 109.
Adkins, G. S., Nappi, C. R., and Witten, E. (1983). Static properties of nucleons in
the Skyrme model. Nucl. Phys. B 228, 552.
Baskerville, W. K. (1996). Making nuclei out of the Skyrme crystal. Nucl. Phys. A 596,
611.
 References 297

Battye, R. A. and Sutcliffe, P. M. (1997). Symmetric Skyrmions. Phys. Rev. Lett. 79,
363.
Battye, R. A. and Sutcliffe, P. M. (1998). A Skyrme lattice with hexagonal symmetry.
Phys. Lett. B 416, 385.
Battye, R. A. and Sutcliffe, P. M. (2001). Solitonic fullerene structures in light atomic
nuclei. Phys. Rev. Lett. 86, 3989.
Battye, R. A. and Sutcliffe, P. M. (2002). Skyrmions, fullerenes and rational maps. Rev.
Math. Phys. 14, 29.
Battye, R. A. and Sutcliffe, P. M. (2005). Skyrmions and the pion mass. Nucl. Phys. B
705, 384.
Battye, R. A. and Sutcliffe, P. M. (2006). Skyrmions with massive pions. Phys. Rev. C
73, 055205.
Battye, R. A., Houghton, C. J., and Sutcliffe, P. M. (2003). Icosahedral Skyrmions.
J. Math. Phys. 44, 3543.
Battye, R. A., Krusch, S., and Sutcliffe, P. M. (2005). Spinning Skyrmions and the
Skyrme parameters. Phys. Lett. B 626, 120.
Battye, R. A., Manton, N. S., and Sutcliffe, P. M. (2007). Skyrmions and the α-particle
model of nuclei. Proc. R. Soc. A 463, 261.
Bauhoff, W., Schultheis, H., and Schultheis, R. (1984). Alpha cluster model and the
spectrum of 16 O. Phys. Rev. C 29, 1046.
Blatt, J. M. and Weisskopf, V. F. (1952). Theoretical Nuclear Physics. Wiley, New York,
p. 292.
Braaten, E. and Carson, L. (1998). Deuteron as a toroidal Skyrmion. Phys. Rev. D 38,
3525.
Brink, D. M., Friedrich, H., Weiguny, A., and Wong, C. W. (1970). Investigation of the
alpha-particle model for light nuclei. Phys. Lett. B 33, 143.
Carson, L. (1991). B = 3 nuclei as quantized multi-Skyrmions. Phys. Rev. Lett. 66,
1406.
Dennison, D. M. (1954). Energy levels of the 16 O nucleus. Phys. Rev. 96, 378.
Finkelstein, D. and Rubinstein, J. (1968). Connection between spin, statistics and kinks.
J. Math. Phys. 9, 1762.
Forest, J. L., Pandharipande, V. R., Pieper, S. C., Wiringa, R. B., Schiavilla, R.,
and Arriaga, A. (1996). Femtometer toroidal structures in nuclei. Phys. Rev. C 54,
646.
Friedrich, H., Satpathy, L., and Weiguny, A. (1971). Why is there no rotational band
based on the 7.65 MeV 0+ state in 12 C?. Phys. Lett. B 36, 189.
Giulini, D. (1993). On the possibility of spinorial quantization in the Skyrme model.
Mod. Phys. Lett. A 8, 1917.
Hitchin, N. J., Manton, N. S., and Murray, M. K. (1995). Symmetric monopoles.
Nonlinearity 8, 661.
Houghton, C. J. and Magee, S. (2006). A zero-mode quantization of the Skyrmion.
Phys. Lett. B 632, 593.
Houghton, C. J., Manton, N. S., and Sutcliffe, P. M. (1998). Rational maps, monopoles
and Skyrmions. Nucl. Phys. B 510, 507.
Irwin, P. (2000). Zero mode quantization of multi-Skyrmions. Phys. Rev. D 61,
114024.
Jarvis, S. (2000). A rational map for Euclidean monopoles via radial scattering. J. Reine
Angew. Math. 524, 17.
298 Skyrmions and nuclei

Krusch, S. (2003). Homotopy of rational maps and the quantization of Skyrmions. Ann.
Phys. 304, 103.
Krusch, S. (2006). Finkelstein–Rubinstein constraints for the Skyrme model with pion
masses. Proc. R. Soc. A 462, 2001.
Manko, O. V. and Manton, N. S. (2007). On the spin of the B = 7 Skyrmion. J. Phys.
A 40, 3683.
Manko, O. V., Manton N. S., and Wood, S. W. (2007). Light nuclei as quantized
Skyrmions. Phys. Rev. C 76, 055203.
Manton, N. S. (1994). Skyrmions and their pion multipole moments. Acta Phys. Pol.
B 25, 1757.
Manton, N. S. and Piette, B. M. A. G. (2001). Understanding Skyrmions using rational
maps. Prog. Math. 201, 469.
Manton, N. and Sutcliffe, P. (2004). Topological Solitons. Cambridge University Press,
Cambridge, UK, Chapter 9.
Manton, N. S. and Wood, S. W. (2006). Reparametrizing the Skyrme model using the
lithium-6 nucleus. Phys. Rev. D 74, 125017.
Manton, N. S., Schroers, B. J., and Singer, M. A. (2004). The interaction energy of
well-separated Skyrme solitons. Commun. Math. Phys. 245, 123.
Morinaga, H. (1956). Interpretation of some of the excited states of 4n self-conjugate
nuclei. Phys. Rev. 101, 254.
Robson, D. (1979). Evidence for the tetrahedral nature of 16 O. Phys. Rev. Lett. 42,
876.
Rössler, U. K., Bogdanov, A. N., and Pfleiderer, C. (2006). Spontaneous Skyrmion
ground states in magnetic metals. Nature 442, 797.
Segal, G. (1979). The topology of the space of rational maps. Acta Math. 143, 39.
Skyrme, T. H. R. (1961). A nonlinear field theory. Proc. R. Soc. A 260, 127.
Skyrme, T. H. R. (1962). A unified field theory of mesons and baryons. Nucl. Phys. 31,
556.
Von Oertzen, W., Freer, M., and Kanada-En’yo, Y. (2006). Nuclear clusters and nuclear
molecules. Phys. Rep. 432, 43.
Walhout, T. S. (1992). Quantizing the four-baryon Skyrmion. Nucl. Phys. A 547, 423.
Williams, J. G. (1970). Topological analysis of a nonlinear field theory. J. Math. Phys.
11, 2611.
Wuosmaa, A. H., Betts, R. R., Freer M., and Fulton, B. R. (1995). Recent advances in
the study of nuclear clusters. Ann. Rev. Nucl. Part. Sci. 45, 89.
 XV
MIRROR SYMMETRY OF FOURIER–MUKAI
TRANSFORMATION FOR ELLIPTIC CALABI–YAU
MANIFOLDS

Naichung Conan Leung and Shing-Tung Yau

Dedicated to Nigel Hitchin on the occasion of his 60th birthday

15.1 Introduction
Mirror symmetry conjecture says that for any Calabi–Yau (CY) manifold M
near the large complex/symplectic structure limit, there is another CY manifold
X, called the mirror manifold, such that the B-model superstring theory on
M is equivalent to the A-model superstring theory on X, and vice versa.
Mathematically speaking, it roughly says that the complex geometry of M is
equivalent to the symplectic geometry of X, and vice versa. It is conjectured
(Strominger et al. 1996) that this duality can be realized as a Fourier-type
transformation along fibers of special Lagrangian fibrations on M and X, called
the SYZ mirror transformation F SY Z .
Suppose M has an elliptic fibration structure,

p : M → S,

then there is another manifold W with a dual elliptic fibration over S. Since
any elliptic curve is isomorphic to its dual, we have actually M ∼ = W provided
that p has irreducible fibers. There is also a Fourier–Mukai transformation (FM
transform) FcxFM
between the complex geometries of M and W . On the level of
derived category of coherent sheaves, Fcx FM
is an equivalence of categories. On
the level of cycles, this can be described as a spectral cover construction and it
is a very powerful tool in the studies of holomorphic vector bundles over M .
In this chapter we address the following two questions: (i) What is the SYZ
transform of the elliptic fibration structure on M ? (ii) What is the SYZ transform
of the FM transform Fcx FM
?
The answer to the first question is a twin Lagrangian fibration structure on
the mirror manifold X, coupled with a superpotential. To simplify the matter, we
will ignore the superpotential in our present discussions. Similarly there is a twin
Lagrangian fibration structure on the mirror manifold Y to W . We will explain
several important properties of twin Lagrangian fibrations. In particular, we show
that the twin Lagrangian fibration on Y is dual to the twin Lagrangian fibration
300 Mirror symmetry of Fourier–Mukai transformation

on X. There is also an identification between X and Y , which is analogous to

the identification between total spaces of dual elliptic fibrations M and W .
For the second question, the SYZ transform of Fcx FM
should be a symplectic
FM transform Fsym from X to Y . We will argue that this is actually the identity
FM

transformation! A naive explanation of this is because Fsym FM

= F SY Z ◦ Fcx
FM
◦
F SY Z
and each of the two F SY Z
transforms undo half of the complex FM
transform.
The plan of the chapter is as follow: In Section 15.2 we review the SYZ mirror
transformation and show that the mirror manifold to an elliptically fibered CY
manifold has a twin Lagrangian fibration structure. In Section 15.3 we review
the FM transform in complex geometry in general and also for elliptic manifolds.
In Section 15.4 we first define the symplectic FM transform between Lagrangian
cycles on X and Y . Then we define twin Lagrangian fibrations, give several
examples of them, and study their basic properties. In Section 15.5 we show
that the SYZ transformation of the complex FM transform between M and W
is the symplectic FM transform between X and Y , which is actually the identity
transformation.

15.2 Mirror symmetry and SYZ transformation

15.2.1 Geometry of Calabi–Yau manifolds
A real 2n-dimensional Riemannian manifold M is a CY manifold if the holonomy
group of its Levi-Civita connection is a subgroup of SU (n). Equivalently, a
CY manifold is a Kähler manifold with a parallel holomorphic volume form. A
theorem (Yau 1978) of the second author says that any compact Kähler manifold
with trivial canonical line bundle admits such a structure.
The complex geometry of M includes the study of (i) the moduli space of
complex structures on M , (ii) complex submanifolds, holomorphic vector bun-
dles, and Hermitian Yang–Mills metrics, and (iii) the derived category Db (M ) of
coherent sheaves on M . The symplectic geometry of M includes the study of (i)
the moduli space of (complexified) symplectic structures on M , (ii) Lagrangian
submanifolds and their intersection theory, and (iii) the Fukaya–Floer category
F uk (M ) of Lagrangian submanifolds in M . The complex geometry is more
non-linear in nature, whereas the symplectic geometry requires the inclusion
of quantum corrections, in which contributions from holomorphic curves in M
needed to be included.

15.2.2 Mirror symmetry conjectures

Roughly speaking, the mirror symmetry conjecture says that for mirror CY
manifolds M and X, the complex geometry of M is equivalent to the symplectic
geometry of X and vice versa:

Complex geometry (M ) Mirror symmetry Symplectic geometry (X) .

←−−−−−−−−−−−−→
 Mirror symmetry and SYZ transformation 301

This conjecture has far-reaching consequences in many different parts of math-

ematics and physics. For instance, (i) Candelas et al. (1991) studied the iden-
tification between the moduli of complex structures on M with the moduli of
complexified symplectic structures on X and derived an amazing formula which
enumerative the number of rational curves of each degree in the quintic CY
threefold. This mirror formula has been proved mathematically by Liu, Lian,
and the second author (1997) and Givental (1996) independently. (ii) A theorem
of Donaldson (1985) and Uhlenbeck and the second author (1986) related the
existence of Hermitian Yang–Mills metrics to the stability of holomorphic vector
bundles. Thomas and the second author (2002) conjectured a mirror phenomenon
to this for special Lagrangian submanifolds. (iii) Kontsevich’s homological mirror
conjecture (1995) identifies Db (M ) with F uk (X).
In this chapter, we use the SYZ transform to study the mirror of an elliptic
fibration structure on M and the FM transform associated to it.

15.2.3 SYZ transform

In order to explain the origin of this duality, Strominger, Yau, and Zaslow
(1996) used physical reasonings to argue that (i) both M and X admit spe-
cial Lagrangian tori fibrations with sections, which are fiberwise dual to each
other. These are called SYZ fibrations. (ii) The equivalence between these two
types of geometries is given by a geometric version of the Fourier–Mukai type
transformation between M and X. We call this the SYZ transformation (see e.g.
Leung 1998, 2000, 2005).
To see this, note that the manifold M itself is the moduli space of certain
complex cycles in M , namely, points in M . Therefore M should also be the
moduli space of certain special Lagrangian submanifolds with flat U (1) bundles
in X. These special Lagrangian submanifolds form a Lagrangian (tori) fibration
on X since points form a fibration on M in a trivial way. By considering all those
points in M which correspond to the same special Lagrangian torus in X but
with different flat U (1) connections, we know that M also admit a tori fibration
which is naturally dual to the one on X. This is because the moduli space of flat
U (1) connections on a torus is naturally its dual torus.
T T∗
↓ ↓
M X
↓ ↓
B B
Dual Lagrangian fibrations on mirror manifolds

Similarly the mirror of the complex cycle M itself is a special Lagrangian section
for X → B. This heuristic reasoning is only expected to hold true asymptotically
near the large complex structure limit (LCSL) (Strominger et al. 1996). We can
apply this SYZ transform to other coherent sheaves on M . For example, suppose
302 Mirror symmetry of Fourier–Mukai transformation

L is a holomorphic line bundle on M , which restricts to a flat bundle on each fiber

Tb ⊂ M of the special Lagrangian fibration. Such a flat connection determines
a point in the dual torus Tb∗ ⊂ X. By varying b ∈ B, we obtain a Lagrangian
section to X → B, which is the mirror to L. This fiberwise Fourier transform
forms the backbone of the SYZ mirror transform.

15.2.4 Mirror of elliptic fibrations

In this section we continue our reasonings to explain why the mirror manifold
X to a CY manifold M with an elliptic fibration p : M → S admits another
special Lagrangian fibration which is compatible with the original SYZ fibration
on X. If the elliptic fibration on M has a section, then the corresponding special
Lagrangian fibration on X also admits an appropriate section. Recall that the
way we obtain the SYZ fibration π : X → B on X is by viewing the identity map
M → M as a fibration on M by complex cycles and with a holomorphic section,
both in trivial manners.
We consider the moduli space of coherent sheaves on M whose generic member
has the same Hilbert polynomial as ι∗ OF where F is an elliptic fiber of M →
S and ι : F → M is the inclusion morphism. Geometrically speaking, this is a
moduli space of elliptic curves in M together with holomorphic line bundles over
them which are trivial topologically. This moduli space is nothing but the total
space of the dual elliptic fibration p : W → S.
By the principle of mirror symmetry, we should view W also as the moduli
space of certain Lagrangian cycles in the mirror manifold X. As dimC W = n
and W consists of geometric cycles which foliate M , we can argue as before that
X should have another Lagrangian fibration p : X → C such that W is also the
moduli space of its Lagrangian fibers together with flat U (1) connections over
them. Similarly the section σ of the elliptic fibration p : M → S determines a
Lagrangian section of p : X → C:
p
X → C
π↓
B

Given any elliptic fiber F in M , there is an one complex parameter family

of points in M which intersect F , namely, those points in F . Homologically
speaking, suppose S is any coherent sheaf on M of the form S = ι∗ L with ι :
F → M the natural inclusion and L ∈ P ic0 (F )  F ∗ , the dual elliptic curve,
then Ext∗OM (S, Om ) = 0 exactly when m ∈ F . Translating this to the mirror
side, given any Lagrangian fiber π −1 (b), it should intersect an one real parameter
family of p−1 (c)’s where c ∈ C.
On the other hand, given any point m ∈ M , there is a unique elliptic curve
F that passes through m. But the coherent sheaves S = ι∗ L with L ∈ F ∗ also
intersect m homologically and parametrized by an one complex parameter family,
 Mirror symmetry and SYZ transformation 303

namely, by F ∗ . On the mirror side, it says that given any Lagrangian fiber p−1 (c)
to p : X → C, it should intersect an one real parameter family of π −1 (b)’s where
b ∈ B.
Furthermore if m ∈ M and S1 , S2 ∈ W satisfying Ext∗OM (S1 , Om ) = 0 =
Ext∗OM (S2 , Om ), then m ∈ F1 = F2 with Fi = SuppSi . On the mirror side, this
means that if π −1 (b) ∩ p−1 (ci ) = 0 for i = 1, 2 then π(p−1 (c1 )) = π(p−1 (c2 )).
Similarly if π−1 (bi ) ∩ p−1 (c) = 0 for i = 1, 2 then p(π −1 (b1 )) = p(π −1 (b2 )).
Namely, π(p−1 (c))’s in B form a fibration over some space D, which is also
the base space of a fibration on C given by p(π −1 (b))’s. That is,
X→C
↓ ↓
B →D
We might also expect that these fibrations on B and C over D are both affine
fibrations. Such a structure on X will be called a twin Lagrangian fibration on
X. On the first sight, it seems that these two Lagrangian fibrations on X are
on equal footing. But these arguments are only valid outside singular fibers
of M → S. Recall that these two Lagrangian fibrations on X are mirror to
holomorphic fibrations id : M → M and p : M → S. The two base manifolds are
quite different in nature: M is CY but S is not.
The Lagrangian fibers to X → C are mirror to the smooth elliptic curve fibers
of M → S. The situation near a singular elliptic curve fiber could be quite
different. Their locus in S, called the discriminant locus D, causes S fails to be
CY because of the formula KS−1 = 12 1
D. In particular KS−1 is an effective divisor
on S, which is indeed ample in many cases, namely S, is a Fano manifold.
There is a version of the mirror symmetry conjecture for Fano manifolds, and
their mirror involve Lagrangian fibrations together with a holomorphic function,
called the superpotential. It is reasonable to expect that the Lagrangian fibration
X → C should also interact with this superpotential corresponding to S. We
hope to come back to further discuss this issue in the future.
15.2.4.1 Large complex structure limits
Mirror symmetry for M is expected to work well near the LCSL. In terms of
Hodge theory, it means that M is a member of an one-parameter family of CY
n-folds Mt with 0 < |t| < 1 such that the monodromy operator T : H n (M ) →
H n (M ) is of maximally unipotent, that is, N = log(I − T ) satisfies N n = 0
but N n+1 = 0. On the mirror side (Deligne 1997), this corresponds to the hard
Lefschetz action L = ∧ωX : ⊕H p,p (X) → ⊕H p,p (X) which satisfies Ln = 0 but
Ln+1 = 0.
We now assume that M has an elliptic fibration structure. In our above
discussions, we need to require each member Mt in the family also have an
elliptic fibration. When n ≥ 3 the existence of an elliptic fibration is invariant
under deformations of complex structures. Wilson (1997) showed that if M is
304 Mirror symmetry of Fourier–Mukai transformation

a CY threefold then the existence of an elliptic fibration structure on M can

be determined by cohomological conditions. When n = 2 existence of an elliptic
fibration is not a deformation invariant property because of H 2,0 (M ) = 0. The
existence of an elliptic fibration on a K3 surface M is equivalent to finding
f ∈ H 2 (M, Z) ∩ H 1,1 (M ) satisfying f 2 = 0. On the mirror side, this corresponds
to the fact that the ωX -Lagrangian fibration X → C on X is indeed Lagrangian
with respect to tωX for every t > 0.
15.2.4.2 Calabi–Yau fibrations
Some of the discussions in this section work for the mirror X of a CY n-fold M
with a fibration by q-dimensional complex subvarieties. Every smooth fiber of
such a fibration on M is automatically CY. In this case there will again be two
Lagrangian fibrations on X with the property that each nonempty intersection
of fibers of X → B and X → C should have codimension q in one of the fiber,
and hence both.
We can also consider a CY manifold M with more than one fibration struc-
tures. For example, when M is a CY threefold with an elliptic fibration M → S
over a Hirzebruch surface S. Then the P1 -bundle structure on S gives a K3
fibration M → P1 on M . In this circumstance, the mirror manifold X should
admit three Lagrangian fibrations which are compatible to each other in certain
large structure limit.

15.3 Fourier–Mukai transform and elliptic CY manifolds

15.3.1 General Fourier–Mukai transform
Suppose M and W are smooth projective varieties. Given any coherent sheaf
P on M × W we can define a FM transform F between derived categories of
coherent sheaves on M and W as follow (Orlov 1996):
Fcx
FM
: Db (M ) → Db (W )
Fcx
FM
(−) = R• p∗ (P ⊗ p∗ (−)) ,
where p : M × W → M and p : M × W → W are projection maps. It was orig-
inally introduced by Mukai in the situation when M and W are dual Abelian
varieties and P is the Poincaré bundle. In this case Fcx
FM
is an equivalence of
derived categories. We will need to use Fcx for families of Abelian varieties
FM

situations. In general the FM transform is a useful tool to verify equivalences

of derived categories, for example, under flops or the McKay correspondence.
On the other hand, a theorem of Orlov (1996) says that any triangle-preserving
equivalence Φ : Db (M ) → Db (W ) is given by a FM transform.

15.3.2 Elliptic fibrations and their duals

In this subsection we recall basic facts about elliptic fibrations (see Donagi 1997,
1998 and Friedman et al. 1997 for details). Suppose M is a CY manifold with
 Fourier–Mukai transform and elliptic CY manifolds 305

an elliptic fibration p : M → S with section σM and with connected fibers. We

denote its dual elliptic fibration as p : W → S, which is again an elliptic CY with
a section σW . For example, a CY hypersurface in a Fano toric variety which is a
P2 -bundle always admits an elliptic fibration structure.
Recall that the Weierstrass model (e.g. Andreas et al. 2001) of a (reduced
irreducible) elliptic curve Tτ in P2 is of the form y 2 z = 4x3 − g2 xz 2 − g3 z 3 with
g2 , g3 constants. The point [0, 1, 0] ∈ P2 is the origin of the elliptic curve Tτ . It is
smooth if and only if g23 − 27g32 is nonzero. In our situation, we have a family of
elliptic curves p : M → S. We assume that if all fibers are isomorphic to reduced
irreducible cubic plane curves and S is smooth, then we can express M in a
similar fashion with g2 and g3 varying over S. To describe it, we let L = R1 p∗ OM
be a line bundle over S, which can also be identified  ⊗4
 as OM (−σ) |σ where
σ = σM (S) is the given section. Then g2 ∈  H 0
S, L and
 g 3 ∈ H 0
S, L⊗6 ,
and the Weierstrass model defines M ⊂ P OS ⊕ L⊗2 ⊕ L⊗3 . Furthermore the
discriminant locus in S is the zero locus of the section g23 − 27g32 of the bundle
L⊗12 .
The dual elliptic fibration W is the compactified relative Jacobian of M ,
namely, W parametrizes rank 1 torsion-free sheaves of degree zero on fibers
of p : M → S. There is a natural identification between M and W because every
elliptic curve Tτ is canonically isomorphic to its Jacobian Jac (Tτ ) given by
Tτ → Jac (Tτ ), p −→ O (p − p0 ) where p0 is the origin of Tτ .

15.3.3 FM transform and spectral cover construction

Given any elliptic curve Tτ , or more generally a principally polarized Abelian
variety, the dual elliptic curve Tτ∗ is its Jacobian, which parametrizes topologi-
cally trivial holomorphic line bundles on Tτ . Mukai (1987) used an analog of the
Fourier transformation to define an equivalence of derived categories of coherent
sheaves on Tτ and Tτ∗ , Fcx
FM
: Db (Tτ ) → D b (Tτ∗ ), called the FM transform. This
can be generalized to the family version as follow: Consider the relative Poincaré
line bundle, or more precisely a divisorial sheaf, P on M ×S W which is given
by

P = O (∆ − σM × W − M × σW ),

where ∆ is the relative diagonal in M ×S W and σM (respectively, σW ) is the

section of p : M → S (respectively, p : W → S). We define the following Fourier-
Mukai functor Fcx FM
: Db (M ) → D b (W ) as Fcx FM
(−) = R• p∗ (P ⊗ p∗ (−)),
where the one in Section 15.3.1 is a generalization of this. Then this can be
proven to give an equivalence of derived categories (Bridgeland and Maciocia
2002; Hu et al. 2002; Donagi and Pantev 2008). Indeed
 FM 
Fcx,P
FM
∗ ⊗K
S
Fcx,P (S) = S [−1]
 FM 
Fcx,P
FM
Fcx,P ∗ ⊗KS (S) = S [−1] .
306 Mirror symmetry of Fourier–Mukai transformation

Besides working on the level of derived categories, we can also study the FM
transform of a stable bundle, the so-called spectral cover construction (Donagi
1997, 1998; Friedman et al. 1997, 1999). The basic idea is any stable bundle over
an elliptic curve is essentially a direct sum of line bundles. In the family situation,
a stable bundle on M gives a multi-section for p : W → S, together with a line
bundle over it. Such construction is important in describing the moduli space of
stable bundles and it can also be generalized to construct principal G-bundles
on M , which play an important role in the duality between F-theory and String
theory (Friedman et al. 1997).

15.4 Symplectic FM transform and twin Lagrangian fibrations

15.4.1 Symplectic Fourier–Mukai transform

Definition 15.1 A Lagrangian cycle in a symplectic manifold (X, ωX ) is a pair

(L, L) with L a Lagrangian submanifold in X and L a unitary flat line bundle
over L. We denote C (X) the collection of Lagrangian cycles in X.

Lagrangian cycles are the objects that form the sophisticated Fukaya–Floer
category F uk (X), where morphisms are Floer homology groups which counts
holomorphic disks bounding cycles of Lagrangian submanifolds. We could also
generalize the notion of Lagrangian cycles to allow L to be a stratified Lagrangian
submanifold and to allow L to be a higher rank flat bundle.
Suppose (X, ωX ) and (Y, ωY ) are symplectic manifolds of dimensions 2m and
2n, respectively. Then (X × Y, ωX − ωY ) is again a symplectic manifold. Given
any Lagrangian cycle (P, P) ∈ C (X × Y ) we can construct a Fourier-Mukai type
transformation, or simply symplectic FM transform, defined as follow (Weinstein
1979):
Fsym
FM
: C (X) → C (Y )
 
Fsym
FM
(L, L) = L̂, L̂ .

Claim 15.1 Suppose that L × Y intersects P transversely. Then the projection

of (L × Y ) ∩ P ⊂ X × Y to Y is a Lagrangian (immersed) submanifold in Y ,
which we denote it as L̂.

Remark 15.1 If L × Y and P do not intersect transversely, then the image

might have bigger dimension. But the vanishing of ω should still imply that it is
a Lagrangian subspace, possibly singular.

Since L × Y intersects P transversely,

 mfor any point in L, by Darboux theorem,
there exists a local coordinate xi , y i i=1 centered at the point such that ωX =
m
i=1 dx ∧ dy and L is given by y = 0 for all i. Suppose this point is also the
i i i

X component of a transversal intersection

 npoint of L × Y and P . If we use an
n
appropriate Darboux coordinates uk , v k k=1 with ωY = k=1 duk ∧ dv k and
 Symplectic FM transform and twin Lagrangian fibrations 307

P is locally determined by
P : xi = xi (u, y) and vk = v k (u, y) for all i, k.
Then (L × Y ) ∩ P ⊂ X × Y is locally given by
⎧ i
⎪
⎪ x = xi (u, 0) , y i = 0,
⎨
(L × Y ) ∩ P :
⎪
⎪ uk = uk , v k = vk (u, 0) .
⎩
(i.e. no restriction)
Therefore its projection to Y becomes
L̂ : uj = uj v k = vk (u, 0) .
The Lagrangian condition for P ⊂ X × Y is
∂xi ∂xj ∂v k ∂v l
+ =0 and + = 0,
∂y j ∂y i ∂ul ∂uk
for any (u, y) and for all i, j = 1, ..., m and k, l = 1, ..., n. By setting y = 0 for the
second equation, this implies that L̂ is a Lagrangian submanifold in Y .
To obtain the flat bundle L̂ over L̂, in the generic case, we can identify L̂ ⊂ Y
with (L × Y ) ∩ P ⊂ X × Y . The flat bundle L on L induced one on L̂ by pullback
to L × Y and then restrict to L̂. By restriction, P also determine a flat connection
on L̂. The tensor product of these two flat bundles is defined as L̂ over L̂. More
work will be needed to handle the pushforward in the non-generic case though.
Examples of Lagrangian cycles in products of symplectic manifolds include
(i) the graph of any symplectic map f : X → Y , that is, f ∗ ωy = ωX ; (ii) if M
(respectively N ) is a Lagrangian submanifold in X (respectively Y ), then P =
M × N is obviously a Lagrangian submanifold in X × Y ; and (iii) let C be any
coisotropic submanifold in (X, ωX ). It induces a canonical isotropic foliation on
C and such that its leaf space C/∼ has a natural symplectic structure, provided
that it is smooth and Hausdorff. This is called the symplectic reduction. Then the
natural inclusion C ⊂ X × (C/∼) is a Lagrangian submanifold in the product
symplectic manifold. Thus we can use this to obtain a symplectic FM transform
Fsym
FM
: C (X) → C (C/ ∼).

15.4.2 Twin Lagrangian fibrations

15.4.2.1 Review of Lagrangian fibrations
Suppose (X, ω) is a symplectic manifold with a Lagrangian fibration
π:X→B
and with a Lagrangian section. Away from the singularities of π, we have an
short exact sequence
0 → Tvert X → T X → π ∗ T B → 0,
308 Mirror symmetry of Fourier–Mukai transformation

where Tvert X is the vertical tangent bundle. Using this and the Lagrangian
condition, we have a canonical identification between Tvert X and the pullback
cotangent bundle:

Tvert X ∼
= π ∗ T ∗ B.

Therefore every cotangent vector at b ∈ B determines a vector field on the fiber

Xb = π −1 (b). By integrating these constant vector fields on Xb , we obtain a
natural affine structure on Xb . This implies that Xb must be an affine torus,
that is, Xb = Tp Xb /Λb ∼= T b for some lattice Λb in Tp Xb = Tb∗ B, provided that
π is proper. This lattice structure on T ∗ B in turn defines an affine structure on
the base manifold B. Of course, this affine structure on B is only defined outside
the discriminant locus, that is, those b ∈ B with Xb singular.
Before we discuss twin Lagrangian fibrations in details, let us first explain the
linear aspects of them.

15.4.2.2 Linear algebra for twin Lagrangian fibrations

 
Suppose V  R2n , ω is a symplectic vector space and Tb and Tc are two
Lagrangian subspaces in V . They give two Lagrangian fibrations V → B and
V → C with B = V /Tb and C = V /Tc . If B and C intersect transversely, then
there is a natural isomorphism

C∼
= B∗

given by the following composition

C → V → V ∗  B ∗ .
ω

Suppose Tbc := Tb ∩ Tc has dimension n − q and we write D = V / (Tb + Tc )

then we have the following diagram of affine morphisms

V →C
↓ ↓
B→D

The fibers of the columns (respectively, rows) are Tb and Tb /Tbc (respectively,
Tc and Tc /Tbc ). The Lagrangian conditions implies the existence of a natural
isomorphism Tc /Tbc ∼
∗
= (Tb /Tbc ) , namely, the two affine bundles B → D and
C → D are fiberwise dual to each other.
The usual SYZ transform which switches the fibers of a Lagrangian fibration
V → B to their duals will interchange complex geometry and symplectic geome-
try. In order to stay within the symplectic geometry, we should take the fiberwise
dual to both Lagrangian fibrations V → B and V → C.
Taking dual to both Tb and Tc has the same effect as taking dual
to (Tb + Tc ) /Tbc while keeping Tbc fixed. This gives us the following new
 Symplectic FM transform and twin Lagrangian fibrations 309

commutative diagram:
U → C
↓ ↓
B → D
Here B  (respectively C  ) is the total space of taking fiberwise dual to the
fibration B → D (respectively C → D). The fiber Tb of U → B  is obtained by
taking dual along the base of the fibration Tb → Tb /Tbc . This is the same as
taking fiberwise dual, up to conjugation with the duality of total spaces. The
∗
fiber of B  → D is (Tc /Tbc ) and likewise for C  → D.
A more intrinsic way to describe this double dual process is as follow: The
fiber of V → D is a coisotropic subspace in V , say Vd . The symplectic reduction
Vd /∼ (see e.g. Weinstein 1979) is another symplectic vector space. Then U is
obtained by replacing Vd / ∼ by its dual symplectic space from
 V. 
In terms of an explicit coordinate system on V given by xi , xα , yi , yα with
1 ≤ i ≤ n − q and n − q + 1 ≤ α ≤ n, then we have
y i , yα y
V →C → αi
xα , xi x
↓ ↓ with coordinates
↓ ↓
B→D
xα , xi → xi
and
yi , yα∗ yα∗
U → C α∗ i →
x ,x xi
↓ ↓ with coordinates
↓ ↓
B → D
xα∗ , xi → xi
The Lagrangian conditions actually imply that B  ∼
= C, C  ∼
= B, and U ∼
=V.
We now return back to the general symplectic manifolds situation.
15.4.2.3 Twin Lagrangian fibrations
First we recall that every fiber of a Lagrangian fibration, say π : X → B, has a
natural affine structure.
Definition 15.2 Let (X, ω) be a symplectic manifold of dimension 2n and
π : X → B and p : X → C are two Lagrangian fibrations on X with Lagrangian
sections. We call this a twin Lagrangian fibration of index q if for general b ∈ B
and c ∈ C with p−1 (c) ∩ π −1 (b) nonempty, then it is an affine subspace of π −1 (b)
π p
of codimension q. We denote such a structure as B ← X → C.
Since π −1 (b) is the union of affine subspaces p−1 (c) ∩ π −1 (b), they form an
affine foliation of π −1 (b) of codimension q. In the above definition we assume that
when p−1 (c) ∩ π −1 (b) is nonempty, then it is a codimension q affine submanifold
of π −1 (b). We did not assume that this affine structure is compatible with the
one on p−1 (c) which comes from the other Lagrangian fibration p. Nevertheless,
310 Mirror symmetry of Fourier–Mukai transformation

under suitable assumptions, we will show that this is indeed the case and the
above definition is symmetric with respect to B and C. Furthermore we have a
commutative diagram of affine morphisms:
X→C
↓ ↓.
B →D
π p
Claim 15.2 Suppose that B ← X → C is a twin Lagrangian fibration. Then
B admits a natural (singular) foliation whose leaves are q-dimensional subspaces
π(p−1 (c))’s with c ∈ C.
Proof. Since p−1 (c) ∩ π −1 (b) is always of codimension k in p−1 (c) if nonempty,
π(p−1 (c)) is a q dimensional subspace in B. Suppose b is a smooth point in
π(p−1 (c)), we claim that its tangent space Tb (π(p−1 (c))) ⊂ Tb B is independent of
the choice of c. Assuming this, we obtain a q-dimensional (singular) distribution
on B. Furthermore π(p−1 (c))’s with c ∈ C are leaves of this distribution, thus
we have the required foliation on B.
To prove the claim, we recall that π −1 (b) has a natural affine structure and
π (b) admits an affine foliation whose leaves are p−1 (c) ∩ π −1 (b) with c ∈ C, by
−1

the assumption of a twin Lagrangian fibration. The key observation is these imply
that for any x ∈ p−1 (c) ∩ π −1 (b), under the natural identification Tx∗ (π −1 (b)) 
Tb B, the conormal bundle of p−1 (c) ∩ π −1 (b) in π −1 (b) at x is the same linear
subspace in Tb B of dimension q. Furthermore this coincides with π(Tx (p−1 (c))).
Hence the result.

We will assume that this foliation on B is indeed a fibration which we denote
as
p : B → D.
As a corollary of the above claim, we have the following immediate result:
π p
Corollary 15.1 Suppose that B ← X → C is a Lagrangian twin fibration.
Then we have a commutative diagram
X→C
↓ ↓
B →D
 
where p : B → D is the above fibration fibered by π p−1 (c) ’s on B.
π p
Claim 15.3 Suppose that B ← X → C is a twin Lagrangian fibration with
sections and π is a proper map. Then D has a natural affine structure and
p : B → D is an affine morphism.
Proof. Since π : X → B is a Lagrangian fibration, its general fiber π −1 (b) has
a natural affine structure, thus as an affine manifold π −1 (b)  Rn /Λ for some
 Symplectic FM transform and twin Lagrangian fibrations 311

lattice Λ  Zr in Tb∗ B. We have r = n because π is proper, that is, π −1 (b) is

compact. This full rank lattice Λ ⊂ Tb∗ B determines an integral affine structure
on B, away from the locus of singular fibers. Thus the proposition is equivalent
to the fibration of π(p−1 (c))’s being an affine fibration on B.
From the proof of the previous proposition, we know that p−1 (c) ∩ π −1 (b)
is an affine subtorus of π −1 (b)  T n . This implies that Tb (p−1 (c)) ⊂ Tb B is
an integral affine subspace. Because of the integral structure, these subspaces
Tb (p−1 (c))’s are affinely equivalently to each other locally. Hence the foliation
they determine is an affine foliation in B. In particular the leave space D
inherits an affine structure such that p : B → D is an affine morphism. Hence the
result.

This implies that, outside the singular locus of p : B → D, there exists a rank q
ε
vector bundle Rq → E → D with affine gluing functions, together with a multi-
section EZ such that p is affine isomorphic to the projection E/EZ → D. In
particular B = E/EZ .
Given a general x ∈ X with b = π (x), c = p (x), and d = p (b) = π (b), we write
Xd = p−1 (c) ∩ π −1 (b). From π and p both being Lagrangian fibrations on X,
we have
∗ ∗
0 → NX d /π
−1 (b),x → Tb B → Tx Xd → 0

and
∗ ∗
0 → NX d /p
−1 (c),x → Tc C → Tx Xd → 0.

As we have shown earlier NXd /p−1 (c),x can be identified with the fiber of E →
∗
D over d ∈ D, denoted as Ed . Moreover the exact sequence 0 → NX d /π
−1 (b),x →
∗ ∗ ∗
Tb B → Tx Xd → 0 is equivalent to 0 → ε E → T E → ε T D → 0 for the vector
bundle ε : E → D. This implies that Tc C is naturally isomorphic to the tangent
space of E ∗ at d.
Notice that the affine
 structure
 of the fibers of the Lagrangian fibration p :
X → C is given by T p−1 (c)  p∗ (Tc∗ C). When p is proper, these data also
determine an affine structure on C. Thus having such a natural identification
between Tc C and Td (E ∗ ), all the affine structures involved are compatible with
each other. Such a claim can be checked by a direct diagram chasing method.
Thus we have obtained the following result:
π p
Claim 15.4 Suppose B ← X → C is a twin Lagrangian fibration with π and
p π
p proper. Then C ← X → B is also a twin Lagrangian fibration and
X→C
↓ ↓
B →D
is a commutative diagram of affine morphisms.
312 Mirror symmetry of Fourier–Mukai transformation

Similar to B = E/EZ for a vector bundle E → D, we have C = F/FZ for

another vector bundle F → D of rank q and multi-section FZ of it. Furthermore
the two bundles E and F over D are fiberwise dual to each other.
Similar discussions can be applied to Lagrangian sections to π : X → B and
p : X → C and we can obtain affine sections to affine fibrations B → D and C →
D and we can also prove that the following diagram of sections is commutative:
X←C
↑ ↑
B ←D
Remark 15.2 The composition map X → D is a coisotropic fibration with
fiber dimension equals n + q. If we apply the symplectic reduction on each
coisotropic fiber, then we obtain a symplectic fibration over D which can be
naturally identified with B ×D C = (E ⊕ F ) / (EZ ⊕ FZ ).
Conversely, suppose that we have two proper Lagrangian fibrations π : X → B
and p : X → C and a commutative diagram of maps
X→ C
↓ ↓
B → D0
for some space D0 satisfying for any b ∈ B and c ∈ C with the same image
d ∈ D0 , then the preimage of d in X for the composition map is equal to π −1 (b) ∩
p−1 (c). We leave it as an exercise for readers to show that this gives a twin
Lagrangian fibration structure on X.

15.4.3 Dual twin Lagrangian fibration and its FM transform

Suppose (X, ω) is a symplectic manifold with a twin Lagrangian fibration with
sections:
X→C
↓ ↓
B →D
The dual of such a structure is obtained by taking the fiberwise dual to both
fibrations X → B and X → C and we obtain the following commutative diagram:
Y → C
↓ ↓
B → D
Recall that a general fiber of X → D is a coisotropic submanifold in X which
is a torus of dimension n + q, denoted Td . Its symplectic reduction (Td /∼) is
a symplectic torus of dimension 2q. We can view Y as being obtained from
X by replacing those directions along the symplectic torus (Td /∼) by the dual
∗
symplectic torus (Td /∼) . It is an important problem to describe Y near singular
 Symplectic FM transform and twin Lagrangian fibrations 313

fibers. Such a Y is called the dual twin Lagrangian fibration to X. It is clear that
the dual twin Lagrangian fibration to Y is X again.
As we have explained in the linear situation, the fibration B  → D (respectively

C → D) is the dual fibration to B → D (respectively C → D). Furthermore the
Lagrangian conditions imply that there are natural identifications B  ∼ = C and
C ∼= B and also
X∼
= Y.
It is interesting to know whether this identification will continue to hold true if
superpotentials are also included in our discussions.
 FM FM
Since X ∼ = Y, we choose the Lagrangian cycle Psym , Psym on X × Y given
by the diagonal Lagrangian submanifold together with the trivial flat bundle
over it. We call this the Lagrangian Poincaré cycle on X × Y . The symplectic
FM transform Fsym FM
for the twin Lagrangian fibration on X and its dual twin
Lagrangian fibration on Y is defined using this Lagrangian cycle:
Fsym
FM
: C (X) → C (Y )
 
Fsym
FM
(L, L) = L̂, L̂ .

Notice that this is actually an identity transformation.

15.4.4 Examples of twin Lagrangian fibrations

We are going to describe some examples of symplectic manifolds X with twin
Lagrangian fibrations. First we notice that this property is stable under taking
the product with another symplectic manifold with a Lagrangian fibration.
The trivial example is the product of a flat torus with its dual, T × T ∗ . The
complex projective space CPn also has a twin Lagrangian fibration: its toric
fibration
µ : CPn → Rn
 
|z1 |2 |z2 |2 |zn |2
µ [z0 , ..., zn ] = n 2 , n 2 , ..., n 2
j=0 |zj | j=0 |zj | j=0 |zj |

is a Lagrangian fibration. If we consider an automorphism of CPn defined by

f ([z0 , ..., zn ]) = [z0 + z1 , z0 − z1 , z2 , ..., zn ], then
µ ◦ f : CPn → Rn
gives another Lagrangian fibration on CPn . It is easy to check that the non-trivial
intersection of any two generic fibers of µ and µ ◦ f is an (n − 1)-dimensional
torus T n−1 . This example can be generalized to the Gelfand–Zeltin system on
Grassmannians and partial flag varieties.
Taub-NUT example: Besides trivial examples of products of Lagrangian fibra-
tions, we can write down explicit twin Lagrangian fibrations on four-dimensional
314 Mirror symmetry of Fourier–Mukai transformation

hyperkähler manifolds X with S 1 -symmetry. These spaces are classified by an

positive integer n, denoted An , with explicit Taub-NUT metrics. Topologically
An is the canonical resolution of C2 /Zn+1 for a finite subgroup Zn+1 ⊂ SU (2).
The hyperkähler moment map

µ = (µ1 , µ2 , µ3 ) : X → Im H = R3

is a singular S 1 -fibration on X. For any unit vector t ∈ S 2 ⊂ R3 , it determines a

complex structure Jt on X. Furthermore the composition of µ with the projection
to the orthogonal plane to t gives a ΩJt -complex Lagrangian fibration on X.
Corresponding to i, j, k ∈ Im H we have complex structures I, J, K on X. Thus
(µ2 , µ3 ) : X → R2 = C is a ΩI -complex Lagrangian fibration with coordinate
(y, z) ∈ C. Similarly (µ1 , µ3 ) : X → R2 = B is a ΩJ -complex Lagrangian fibra-
tion with coordinate (x, z) ∈ B. Then we have a ωK twin Lagrangian fibration
structure on X with D being the z-axis.
This construction can be easily generalized to any 4n-dimensional hyperkähler
manifold X with a tri-Hamiltonian T n action with a hyperkähler moment map

µ : X → Rn ⊗ R3 .

In the above example, the generic fiber of X → B, or X → C, is topologically

a cylinder, thus noncompact. Therefore the base of the Lagrangian fibration
only has a partial affine structure. For Lagrangian fibrations with compact
fibers, there are usually singularities for the affine structures for the base spaces
corresponding to singular fibers. When dim B = 2, the affine structure near a
generic singularity has been studied in Gross and Siebert (2003) and Kontsevich
and Soibelman (2006). The monodromy across the “slit,” namely, the positive x-
axis, is given by (x, y) → (x, x + y). Near the singular point 0 ∈ B = R2 , the only
affine fibration is given by horizontal lines. Other attempts to obtain fibrations
on B will get overlapping fibers.
K3 surface with an elliptic fibration M → P1 can be constructed as an anti-
canonical divisor of a Fano toric threefold P∆ which admits a toric P2 -bundle
structure, P2 → P∆ → P1 . The mirror to M is another K3 surface X inside a
Fano toric threefold P∇ associated to a polytope which is a two-sided cone over
a triangle.
Generically X = {f = 0} admits a Lagrangian fibration over B = ∂∇, homeo-
morphic to the two-sphere S 2 , with 24 singular fibers. B admits a natural affine
structure with singularity along the base points of these singular fibers (see Gross
and Siebert 2003 for its construction), which all lie on the edges of ∇, and their
“slits” are also along these edges. We consider the restriction of the projection
of R3 to the x-axis to B = ∂∇, the image is the interval D in R corresponding
to the polytope of the base P1 :

p : B → D.
 SYZ transformation of FM transform 315

Since the slits for all the singular points in the middle triangle in ∂∇ are
vertical with respect to p, we obtain an affine fibration around there. But p will
cease to be an affine fibration around other singular points of B. However if
we choose the defining function f for M appropriately, then we can arrange all
other singular points to be far away from this middle triangle. Thus we obtain an
affine fibration on a large portion of B. As a matter of fact, if we allow M to be
singular, then we can make p to be an affine fibration on the whole B\Sing (B).
This picture can be generalized to certain higher dimensional manifolds, for
instance CY hypersurfaces in toric varieties.

15.5 SYZ transformation of FM transform

Given an elliptically fibered CY manifold M, we argued in Section 15.2.4 that its
mirror manifold X should admit a twin Lagrangian fibration (with superpotential
which we neglect in our present discussions). In this section, we first argue that
the mirror manifold to the dual elliptic fibration W to M is the dual twin
Lagrangian fibration Y to X. Second, we will explain the mirror to the universal
Poincaré sheaf that defines the FM transform between complex geometries of M
and W is the universal Lagrangian Poincaré cycle that defines the FM transform
between symplectic geometries of X and Y . Third, we show that these FM
transforms commute with the SYZ transforms. Namely, the SYZ transform of
the complex FM transform is the symplectic FM transform, which is actually an
identity transformation.

15.5.1 SYZ transform of dual elliptic fibrations

As we mentioned above, we are going to argue that the following diagram
commutes:
M ←
Mirror
−−−−−−manifolds
−−−−−−→ X

Dual elliptic L L Dual twin Lagrangian

+ +
fibrations fibrations

W ←
Mirror
−−−−−−manifolds
−−−−−−→ Y
Indeed our arguments only work in the large structure limit but we expect that
a modified version of such will work in more general situations. To be precise,
we are only dealing with the flat situation: We assume that M is CY n-fold
with a q-dimensional Abelian varieties fibration p : M → S which is compatible
with a Lagrangian fibration π : M → B in the sense that the restriction of π to
any smooth Abelian variety fiber of p determines a Lagrangian fibration on the
Abelian variety. We are going to impose a very restrictive assumption, which we
expect to be the first-order approximation of what happens at the large structure
limit. Namely, M is the product of an Abelian variety with S.
316 Mirror symmetry of Fourier–Mukai transformation

1 n
Let√ z , ..., z be a local complex coordinate system on M such that z =
x + −1y with y 1 , ..., yn (respectively x1 , ..., xn ) be an affine coordinate system
on fibers (respectively base) of the special Lagrangian fibration on M → B.
According to the SYZ proposal, the mirror manifold X to M is the total space
of the dual Lagrangian torus fibration. For the affine coordinate system y 1 , ..., yn
on the torus fiber, we denote the dual coordinate system on its dual torus as
y1 , ..., yn . Thus X has a Darboux coordinate system given by y1 , ..., yn , x1 , ..., xn .
We assume the q-dimensional Abelian variety fibration M → S is compatible
with the Lagrangian fibration structure in the sense that z n−q+1 , ..., z n gives
a complex affine coordinate system on fibers to M → S. In particular, the
dual coordinate system on the dual Abelian variety is given by zn−q+1 , ..., zn .
Therefore such a coordinate system on M induces a similar coordinate system on
the dual Abelian variety fibration W → S, labeled (z 1 , ..., z n−q , zn−q+1 , ..., zn ).
Note that on a fiber of the Abelian variety fibration M → S, yn−q+1 , ..., yn
(respectively, xn−q+1 , ..., xn ) are those coordinates belonging to Lagrangian
fibers (respectively, base) for the special Lagrangian fibration on M . Since fibers
are much smaller in scale when comparing to the base near the LCSL (Strominger
et al. 1996), when we take the dual Abelian variety, yn−q+1 , ..., yn will become
much larger in scale when comparing to xn−q+1 , ..., xn . As a consequences,
the special Lagrangian fiber (respectively, base) coordinates for the CY man-
ifold Y are y 1 , ..., y n−q ,xn−q+1 , ..., xn (respectively, x1 , ..., xn−q , yn−q+1 , ..., yn ).
Therefore the special Lagrangian fibration on the mirror manifold Y to W has
fiber (respectively, base) coordinates as y1 , ..., yn−q , xn−q+1 , ..., xn (respectively,
x1 , ..., xn−q , yn−q+1 , ..., yn ).
Next we need to check that Y is indeed the total space of the dual twin
Lagrangian fibration to X. For the Abelian variety fiber in M , with coordi-
nates xn−q+1 , ...xn , y n−q+1 , ...y n , its mirror Lagrangian cycle in X has coor-
dinates xn−q+1 , ...xn , y1 , ...yn−q and it is the fiber of the other Lagrangian
fibration on X. Similarly, for the Abelian variety fibers in W , with coordinates
xn−q+1 , ...xn , yn−q+1 , ...yn , its mirror Lagrangian cycle in Y has coordinates
y1 , ...yn−q , yn−q+1 , ...yn and it is the fiber of the other Lagrangian fibration on
Y . It is now clear that X and Y are dual twin Lagrangian fibrations. All these
fiber/base coordinates systems are summarized in the following diagram:

yi yα y yi
M: → X : αi
←
xi xα x xα
) )
yi xα xα yi
W : i ←
→ Y : i
x yα x yα

Our conventions are i = 1, ..., n − q and α = n − q + 1, ..., n. For each space in the
above diagram, the coordinates in the top (respectively, bottom) row are for the
SYZ special Lagrangian fibers (respectively, base). Also the coordinates in the left
 SYZ transformation of FM transform 317

(respectively, right) column are for the fibers (respectively, base) of the Abelian
variety fibrations for M or W and the other Lagrangian fibrations for X or Y .

15.5.2 SYZ transform of the universal Poincaré sheaf

Recall that the FM transform on derived categories of manifolds with elliptic
fibrations is given by
F M : D b (M ) → D b (W )
F M (−) = R• p∗ (Pcx ⊗ p∗ (−)) ,
where Pcx is a coherent sheaf on M × W with support
Supp (Pcx ) = M ×S W ⊂ M × W,
and
Pcx = O (∆ − σM × W − M × σW ) ,
with ∆ the relative diagonal in M ×S W .
We note that M × W is again a CY manifold with its mirror manifold being
X × Y . We are going to describe the SYZ transform of Pcx from Db (M × W )
to F uk (X × Y ). We will apply the transformation on the level of object, as
proposed in (Strominger et al. 1996).

Claim 15.5 Corresponding to the mirror symmetry between CY (2n)-folds

M × W and X × Y
Complex Symplectic
(M × W ) SYZ −−−−−−→ geometry (X × Y )
←−−−transform
geometry
the mirror of the coherent sheaf Pcx on M × W is the diagonal Lagrangian cycle
(Psym , Psym ) ∈ C (X × Y ) as described in Section 15.4.3.

Proof. We will continue to use the same coordinate systems as before. The
Poincaré sheaf Pcx for the dual Abelian variety fibrations
 M → S and W →  S
has support M ×S W ⊂ M × W which has coordinates xi , y i , xα , y α , xα , yα i,α
(where xi and y i are diagonal coordinates in the product space). Over M ×S
W , Pcx is a complex line bundle, indeed only a divisorial sheaf, with an U (1)
connection:
√ 
FM
Dcx = d + −1 α (xα dxα + xα dxα − yα dy α − y α dyα ) .
The reason that the signs for terms involving x and y are different is the following:
For Abelian varieties, the dual complex coordinates are z α and zα , which induces
the dual coordinates for xα , y α as xα , −yα because Re (z α zα ) = xα xα − y α yα .
To apply the SYZ mirror transform to Pcx from the special Lagrangian fibra-
tion M × W → B × B ∗ to the one X × Y → B × B ∗ , we first need to describe
318 Mirror symmetry of Fourier–Mukai transformation

the Poincaré bundle P SY Z over

(M × W ) × (X × Y ) ⊂ (M × W ) × (X × Y ) .
B×B ∗

To describe the coordinates on this space, we need to rename the (x, y) coordi-
nates on W and Y to (u, v) coordinates. That is,

W Y
v i uα uα vi
SYZ
←−−−mirror
−−−→
ui vα ui v α

Now the universal U (1) connection on the line bundle

C →P SY Z → (M × W ) × (X × Y )
B×B ∗

is given by
√  i 
DSY Z = d + −1 y dyi + yi dy i + y α dyα + yα dy α
√  i 
− −1 v dvi + vi dv i + uα duα + uα duα .

Note that we have used different signs for the SYZ transform between M and X
and SYZ transform between W and Y .
In this newly named coordinates on W , the universal connection on the
Poincaré bundle

C → Pcx
FM
→ M ×S W

is given by
√ 
Dcx
FM
=d− −1 α (uα dxα + xα duα − vα dy α − y α dvα ) .

To apply the SYZ transform, we need to first pullback the bundle Pcx FM
(with
its connection Dcx ) from M ×S W ⊂ M × W to (M × W ) ×B×B (X × Y ) ⊂
FM ∗

(M × W ) × (X × Y ) and tensors it with P SY Z (with its connection DSY Z ).

Then we pushforward along the projective map (M × W ) ×B×B ∗ (X × Y ) →
X ×Y.
We can separate our discussions into two parts: (i) perform the SYZ transform
along those directions with indexes i = 1, ..., n − q and (ii) perform the SYZ
transform along those directions with indexes α = n − q + 1, ..., n.

Part (i) with i = 1, ..., n − q: This part is easy because the Poincaré bundle for
the
 i FM itransform
  does
 not involve here. In fact it is the trivial line bundle over
u = x ∩ v i = y i ⊆ M × W . In these coordinates, the SYZ Poincaré bundle
P SY Z has support (M × W ) ×B×B ∗ (X × Y ) ⊂ (M × W ) × (X × Y ), which in
our coordinate systems means the xi ’s coordinates for M and X are the same
 SYZ transformation of FM transform 319

and similarly the ui ’s coordinates for W and Y are the same:

√ 
DSY Z = d + −1 i y i dyi + yi dy i − v i dvi − vi dv i
√ 
= d + −1 i y i (dyi − dvi ) + (yi − vi ) dy i .

To pushforward to X × Y , we integrate along yi ’s directions.

√  Along these direc-
tions, the restriction of the above connection is d + −1 i (yi − vi ) dy i , which
has a (unique up to scaling) flat section precisely when yi − vi = 0 for all i.
When this happens, we also have y i (dyi− dvi ) = 0. Hence the SYZ transform
 of
Pcx
FM
is given by the trivial bundle over ui = xi and yi = vi for all i ⊂ X × Y .
It is a Lagrangian submanifold in X × Y with the symplectic form ωX×Y =
Σdxi ∧ dyi + Σdui ∧ dvi .

Part (ii) with α = n − q + 1, ..., n: In these coordinates, the support for the SYZ
Poincaré bundle P SY Z imposes a constraint which says that the xα ’s coordinates
for M and X are the same and the vα ’s coordinates for W and Y are the same.
Now we need to integrate the directions y α ’s and uα ’s. First we rearrange the
terms
√ 
Dcx
FM
⊗ DSY Z = d − −1 α (uα dxα + xα duα − vα dy α − y α dvα )
√ 
+ −1 α (y α dyα + yα dy α − uα duα − uα duα )
√ 
= d + −1 α ((−vα + yα ) dy α + (xα − uα ) duα )
√ 
+ −1 α (uα dxα − y α dvα + y α dyα − uα duα ) .

When we integrate along the y α ’s and uα ’s directions, the terms uα dxα , y α dvα ,
y dyα , and uα duα has non-trivial Fourier modes in these variables and therefore
α

they contribute zero to Psym , the mirror cycle of Pcx .

α
When we pushforward √ along
 y ’s direction, the restriction of the above con-
nection becomes d + −1 α (−vα + yα ) dy α along any fiber and it admits a
(unique up to scaling) parallel section precisely when yα = vα for all α. Similarly
we obtain uα = xα for all α when we pushforward along uα ’s directions. Namely,
the support of the mirror object to Pcx FM
is given by

{yα = vα and xα = uα for all α} = Psym ⊂ X × Y.

Namely, this is precisely given by the diagonal Lagrangian submanifold as in the

example in Section 15.4.3.
There is no remaining component of Dcx FM
⊗ DSY Z and therefore the unitary
flat connection on Psym is trivial.
By putting parts (i) and (ii) together,
 we have shown that the mirror transfor-
mation of the complex Poincaré cycle
 F M FcxP FM
 = M ×S W, Pcx FM
∈ C (M × W ) is
the Lagrangian Poincaré cycle Psym , Psym M
∈ C (X × Y ) in the flat limit. Hence
the claim.

320 Mirror symmetry of Fourier–Mukai transformation

15.5.3 SYZ transforms commute with FM transforms

Recall from the last section that when Pcx
FM
is the Poincaré sheaf for the FM
transform between the elliptic CY manifold M and its dual elliptic CY manifold
W , then its mirror Psym
FM
is the diagonal Lagrangian submanifold in X × Y .
Thus it defines a symplectic FM transform FsymFM
from X to Y . We claim that
the complex/symplectic FM transforms commute with the SYZ transforms:

Fsym
FM
◦ F SY Z = F SY Z ◦ Fcx
FM

as depicted in the following diagram:

Complex Symplectic
(M ) F SY Z
geometry ←−−−→ geometry (X)
L L
⏐ ⏐
⏐
FM +
⏐ FM
Fcx + Fsym

Complex Z Symplectic
F−SY
(W ) ← −−→ geometry (Y )
geometry

Since Fsym
FM
is essentially an identity transformation, we can regard the FM
transform for elliptic CY as a square of the SYZ transforms! Note that even
though one can identify M with W and X with Y , nevertheless, F(M,X) SY Z
and
F(W,Y ) do not correspond to each other under these identifications of spaces. For
SY Z

instances, the whole manifold M (respectively, W ) transforms to the zero section

of the SYZ fibration on X (respectively, Y ). However, they are sections of two
different fibrations on X  Y as it admits a twin Lagrangian fibration structure.
In fact our claim follows from a more general statement: If M and W are two
CY manifolds, possibly of different dimensions and X and Y are their mirror
manifolds, respectively. Let Pcx be any complex cycle in M × W and Psym be its
mirror Lagrangian cycle in X × Y . They define general complex/symplectic FM
transforms Fcx FM
and Fsym
FM
, respectively. Then these FM transforms commute
with the SYZ transforms.
The key point is the SYZ transform is an involution, that is, F SY Z ◦ F SY Z =
id. Suppose S is a complex cycle in M , we want to show that
   FM 
Fsym
FM
F(M,X)
SY Z
(S) = F(W,X)
SY Z
Fcx (S) .

Equivalently,

(πX )∗ (πM )∗ SM ⊗ P(M,X)

SY Z
⊗ Psym
FM
= (πW )∗ (πM )∗ SM ⊗ Pcx
FM
⊗ P(W,Y
SY Z
).

Therefore, it suffices to prove that

(πX )∗ P(M,X)
SY Z
⊗ Psym
FM
= (πW )∗ Pcx
FM
⊗ P(W,Y
SY Z
),
 Conclusions and discussions 321

over M × Y . However,

Psym
FM
= (πM )∗ (πW )∗ Pcx
FM
⊗ P(M
SY Z
×W,X×Y )

= (πM )∗ (πW )∗ Pcx

FM
⊗ P(M,X)
SY Z
⊗ P(W,Y
SY Z
).

Hence

(πX )∗ P(M,X)
SY Z
⊗ Psym
FM

 
= (πX )∗ P(M,X)
SY Z
⊗ (πM )∗ (πW )∗ Pcx
FM
⊗ P(M,X)
SY Z
⊗ P(W,Y
SY Z
)
 
= (πW )∗ Pcx
FM
⊗ P(W,Y
SY Z
) ⊗ (πM )∗ (πX )∗ P(M,X) ⊗ P(M,X)
SY Z SY Z

= (πW )∗ Pcx
FM
⊗ P(W,Y
SY Z
).

The last equality holds because SYZ transforms are involutive. Hence the
result.

15.6 Conclusions and discussions

In this chapter we have introduced the notion of a twin Lagrangian fibration
and explained several of its properties. We argued via the SYZ proposal that
the mirror manifold of an elliptic CY manifold should admit such a structure,
possibly coupled with a non-trivial superpotential which we have not fully
understood yet.
An important tool in the study of the complex geometry of elliptic manifolds is
the FM transform. We argued that under the SYZ transform, this FM transform
will become the identity transformation for the symplectic geometry between
dual twin Lagrangian fibrations. Even though we are mostly interested in the
elliptic fibration situation, these arguments can be applied to Abelian varieties
fibrations, and possibly to general CY fibrations with suitable adjustments.
We could also study mirror of the symplectic geometry of the elliptic CY
manifolds. For example, let ωM (respectively, ωS ) be any Kähler form on M
(respectively, S). The pullback of ωS under the elliptic fibration M → S, denoted
as ωS again, is only nef but not ample. Obviously it satisfies ωSn−1 = 0 and
ωSn = 0. For any t > 0, ωM,t = tωM + (1 − t) ωS is always a Kähler form on M .
n n+1
In particular it satisfies ωM,t = 0 and ωM,t = 0.
On the mirror side, these correspond to a two-parameter family of complex
structures on X where (i) the generic monodromy is of maximally unipotent
and the corresponding vanishing cycles are the Lagrangian fibers of X → B;
and (ii) a special monodromy T0 for this family has the property N0n−1 = 0 but
N0n = 0 where N0 = log (I − T0 ) and the corresponding vanishing cycles are the
(n + 1)-dimensional coisotropic fibers of the composition map X → B → D.
322 Mirror symmetry of Fourier–Mukai transformation

Acknowledgments
The second author has known Nigel for almost forty years, having met him at
the Institute of Advanced Study. In those days, we had many dinners together,
and I (S.-T. Yau) always enjoyed talking to him. His many original ideas in
mathematics influenced me tremendously. This is especially the case for K3
surfaces, in which his deep insight affected my thinking of the Calabi–Yau
conjecture.
The research of N.C. Leung is partially supported by RGC Earmarked grants
of Hong Kong. The work of S.-T. Yau is supported in part by NSF grants DMS-
0354737, DMS-0306600, and DMS-0628341.

References
Andreas, B., Curio, G., Ruiperez, D. H. and Yau, S.-T. (2001). Fibrewise T-duality for
D-branes on elliptic Calabi-Yau. J. High Energy Phys. No. 3, Paper 20, 13 pp.
Bridgeland, T. and Maciocia, A. (2002). Fourier-Mukai transforms for K3 and elliptic
fibrations. J. Algebraic Geom. 11(4), 629–57.
Candelas, P., de la Ossa, X. C., Green, P. S. and Parkes, L. (1991). A pair of Calabi-Yau
manifolds as an exactly soluble superconformal theory. Nucl. Phys. B 359, 21–74.
Deligne, P. (1997). Local behavior of Hodge structures at infinity. In Mirror Symmetry,
II, pp. 683–699, AMS/IP Studies in Advanced Mathematics, Vol. 1, American
Mathematical Society, Providence, RI.
Donaldson, S. K. (1985). Anti self-dual Yang-Mills connections over complex algebraic
surfaces and stable vector bundles. Proc. London Math. Soc. (3) 50(1), 1–26.
Donagi, R. (1998). Taniguchi lectures on principal bundles on elliptic fibrations. In
Integrable Systems and Algebraic Geometry (Kobe/Kyoto, 1997), pp. 33–46, World
Scientific Publishing, River Edge, NJ.
Donagi, R. (1997). Principal bundles on elliptic fibrations. Asian J. Math. 1(2), 214–23.
Donagi R. and Pantev, T. (2008). Torus fibrations, gerbes, and duality. With an
appendix by Dmitry Arinkin. Mem. Amer. Math. Soc. 193(901), vi–90.
Friedman, R., Morgan, J. and Witten, E. (1997). Vector bundles and F theory.
Commun. Math. Phys. 187(3), 679–743.
Friedman, R., Morgan, J. and Witten, E. (1999). Vector bundles over elliptic fibrations.
J. Algebraic Geom. 8(2), 279–401.
Givental, A. (1996). Equivariant Gromov-Witten invariants. Int. Math. Res. Notices,
No. 13, 613–63.
Gross M. and Siebert, B. (2003). Affine manifolds, log structures, and mirror symmetry.
Turkish J. Math. 27(1), 33–60.
Hu, Y., Liu, C. H. and Yau, S.-T. (2002). Toric morphisms and fibrations of toric
Calabi-Yau hypersurfaces. Adv. Theor. Math. Phys. 6(3), 457–506.
Kontsevich, M. (1995). Homological algebra of mirror symmetry. In Proceedings of the
1994 International Congress of Mathematicians, Vol. I, Birkäuser, Zürich, p. 120.
Kontsevich, M. and Soibelman, Y. (2006). Affine structures and non-Archimedean
analytic spaces. In The Unity of Mathematics, pp. 321–385, Progress of Mathematics,
Vol. 244, Birkhauser Boston, Boston, MA.
 References 323

Leung, N. C. (2005). Mirror symmetry without correction. Commun. Anal. Geom.

13(2), 287–331.
Leung N. C. and Vafa, C. (1998). Branes and toric geometry. Adv. Theor. Math. Phys.
2, 91–118.
Leung, N. C., Yau, S.-T. and Zaslow, E. (2000). From special Lagrangian to Hermitian-
Yang-Mills via Fourier-Mukai transform. Adv. Theor. Math. Phys. 4(6), 1319–41.
Lian, B. H., Liu, K. and Yau, S. T. (1997). Mirror principle I. Asian J. Math. 1(4),
729–63.
Mukai, S. (1987). Fourier functor and its application to the moduli of bundles on an
abelian variety. In Algebraic Geometry, Sendai, 1985, pp. 515–50, Advanced Studies
in Pure Mathematics, Vol. 10, North-Holland, Amsterdam, the Netherlands.
Orlov, D. O. (1996). Equivalences of derived categories and K3 surfaces, Preprint.
alg.geom/9606006.
Strominger, A., Yau, S.-T. and Zaslow, E. (1996). Mirror symmetry is T-duality. Nucl.
Phys. B 479, 243–59.
Thomas, R. and Yau, S. T. (2002). Special Lagrangians, stable bundles and mean
curvature flow. Commun. Anal. Geom. 10(5), 1075–113.
Uhlenbeck, K. and Yau, S. T. (1986). On the existence of Hermitian Yang-Mills
connections in stable vector bundles. Commun. Pure Appl. Math. 39, 257–93.
Weinstein, A. (1979). Lectures on symplectic manifolds. Corrected reprint. CBMS
Regional Conference Series in Mathematics, Vol. 29, American Mathematical Society,
Providence, RI.
Wilson, P. M. H. (1997). The role of c 2 in Calabi-Yau classification: a preliminary
survey. In Mirror Symmetry II, pp. 381–92, AMS Studies in Advanced Mathematics,
Vol. 1, American Mathematical Society, Providence, RI.
Yau, S.-T. (1978). On the Ricci curvature of a compact Kahler manifold and the
complex Monge-Ampere equation. I. Commun. Pure Appl. Math. 31(3), 339–411.
 XVI
S-DUALITY IN HYPERKÄHLER HODGE THEORY

Tamás Hausel
To Nigel Hitchin for his 60th birthday

16.1 Introduction
In this chapter we survey the motivations, related results, and progress made
towards the following problem, raised by Hitchin in 1995:

Problem 16.1 What is the space of L2 harmonic forms on the moduli space
of Higgs bundles on a Riemann surface?

The moduli space MdDol (SLn ) of stable rank n Higgs bundles with fixed
determinant of degree d on a Riemann surface was introduced and studied
in Hitchin (1987), Nitsure (1991), and Simpson (1991). The Betti numbers of
this space for n = 2 were determined in Hitchin (1987b) while for n = 3 in
Gothen (1994). The above problem raised two new directions to study. First
is the Riemannian geometry of MdDol (SLn ), or more precisely the asymptotics
of the natural hyperkähler metric, and its connection with Hodge theory. The
second one, which can be considered the topological side of Problem 16.1,
is to determine the intersection form on the middle-dimensional compactly
supported cohomology of MdDol (SLn ). While the first question seems still
out of reach, although we will report on some modest progress below, the
second is more approachable and we offer a conjecture at the end of this
survey.
Problem 16.1 was motivated by S-duality conjectures emerging from the string
theory literature about Hodge theory on certain hyperkähler moduli spaces,
which are close relatives of MdDol (SLn ).
In the physics literature S-duality stands for a strong–weak duality between two
quantum field theories. The interest from the physics point of view is that it gives
a tool to study physical theories with a large coupling constant via a conjectured
equivalence with a theory with a small coupling constant where perturbative
methods give a good understanding. The S-duality conjecture relevant for us
is based on the Montonen–Olive electromagnetic duality proposal from 1977 in
four-dimensional Yang–Mills theory (Montonen and Olive 1977). It was noted in
(Witten and Olive 1978) that this duality proposal is more likely to hold in a
supersymmetric version of the theory, and in Osborn (1979) it was argued that
N = 4 supersymmetry is a good candidate. Hyperkähler Hodge theory is relevant
 Introduction 325

in N = 4 supersymmetry as the space of differential forms on a hyperkähler

manifold possesses an action of the N = 4 supersymmetry algebra via the various
operators in hyperkähler Hodge theory.
In this chapter our interest lies in the mathematical predictions of such
S-duality conjectures in physics. Sen (1994), using S-duality arguments in
N = 4 supersymmetric
 Yang–Mills theory, predicted the dimension of the spaces
d 0 0 of the hyperkähler
H Mk of L harmonic d-forms on the universal cover M
2
k
moduli space Mk0 of certain SU (2) magnetic monopoles on R3 . In the interpre-
tation of Sen (1994) the L2 harmonic forms on M 0 can be thought of as bound
k
states of the
 theory,
 and the conjectured S-duality implies an action of SL(2, Z)
 ∗ 0
on kH Mk . By further physical arguments Sen managed to predict this
representation of SL(2, Z) completely, implying the following:

Conjecture 16.1 0 is
The dimension of the space of L2 harmonic forms on Mk
   <
0 0 d = mid
dim Hd M k =
φ(k) d = mid,

where φ(k) = ki=1 δ1(i,k) is the Euler φ function, and mid = 2k − 2 is half of
0 .
the dimension of Mk

Similar S-duality arguments led Vafa and Witten (1994) to get a conjecture
on the space of L2 harmonic forms on a certain smooth completion Mφk,c1 ,
constructed in Kronheimer (1990) and Nakajima (1998), of the moduli space
of U (n) Yang–Mills instantons of first Chern class c1 , energy k, and framing
φ on one of Kronheimer’s ALE spaces, which are four-dimensional complete
hyperkähler manifolds, with an asymptotically locally Euclidean metric.
Conjecture 16.2 The dimension of the space of L2 harmonic forms on Mφk,c1
is

   0       d = mid
dim Hd Mφk,c1 = k,c1 k,c1
mid
dim im Hcpt Mφ →H mid
Mφ d = mid,

where mid now denotes half of the dimension of Mφk,c1 .

Vafa and Witten (1994) further argue that Conjecture 16.2 implies, via the work
of Nakajima (1998) and Kac (1990), that
   
Zφ (q) = q k−c/24 dim Hmid Mφk,c1 (16.1)
c1 ,k

is a modular form, which, as was speculated in Vafa and Witten (1994), might
be a consequence of S-duality.
This chapter will introduce the reader to various mathematical aspects of these
three problems and offer mathematical techniques and results relating to them.
326 S-duality in hyperkähler Hodge theory

16.2 Hyperkähler quotients

A Riemannian manifold (M, g) is hyperkähler if it is Kähler with respect
to three integrable complex structures I, J, K ∈ Γ(End(T M )), which satisfy
I 2 = J 2 = K 2 = IJK = −1, with Kähler forms ωI , ωJ , and ωK . Known compact
examples are scarce (see e.g. Joyce 2000, section 7). Non-compact complete
examples however are much more abundant. This is mostly because there is a
widely applicable 1 hyperkähler quotient construction, due to Hitchin et al. (1987).
The construction itself is an elegant quaternionization of the Marsden–Weinstein
symplectic (or more precisely Kähler) quotient construction (see Mumford et al.
1994, chapter 8 for an introduction for the latter).
Let M be a hyperkähler manifold, G a Lie group, with Lie algebra g, and assume
G acts on M preserving the hyperkähler structure (i.e. it acts by triholomorphic
isometries). Let us further assume that we have moment maps µI : M → g∗ ,
µJ : M → g∗ , and µK : M → g∗ with respect to the symplectic forms ωI , ωJ ,
and ωK , respectively. We combine them into a single hyperkähler moment map:
µH = (µI , µJ , µK ) : M → R3 ⊗ g∗ .
One takes ξ ∈ R3 ⊗ (g∗ )G and constructs the hyperkähler quotient at level ξ by
M////ξ G := µ−1
H (ξ)/G.

The main result of Hitchin et al. (1987) is that the natural Riemannian metric
on the smooth points of this quotient is hyperkähler.
Now we list three important examples of this construction, where the original
hyperkähler manifold M and Lie group G are both infinite dimensional.

16.2.1 Moduli of Yang–Mills instantons on R4

Here we follow Hitchin (1987a, I example 3.6), compare also with Atiyah (1978).
Let G be a compact connected Lie group, which will be U (n) or SU (n) in this
chapter. Let P → R4 be a G-principal bundle over R4 . Let M be the space of
G-connections A on P of class C ∞ , such that the energy

Tr(FA ∧ ∗FA ) < ∞
R4

is finite. Write
A = A1 dx1 + A2 dx2 + A3 dx3 + A4 dx4
in a fixed gauge, where Ai ∈ Ω0 (R4 , ad(P )). Let G = Ω(R4 , Ad(P )) be the gauge
group of P . An element g ∈ G acts on A ∈ M by the formula g(A) = g −1 Ag +
g−1 dg, preserving the hyperkähler structure. One finds that the hyperkähler

1 Some colleagues even suggest, due to the success of this construction, that HyperKähLeR

is in fact just a pronouncable version of the acronym HKLR.

 Hyperkähler quotients 327

moment map equation

µH (A) = 0 ⇔ FA = ∗FA

is just the self-dual Yang–Mills equation. Define the hyperkähler quotient

M(R4 , P ) = µ−1
H (0)/G, the moduli space of finite-energy self-dual Yang–Mills
instantons on P . By its construction it has a natural hyperkähler metric.
A similar construction (Kronheimer and Nakajima 1990) for G = U (n) yields
a hyperkähler metric on moduli spaces of U (n) Yang–Mills instantons on
certain four-dimensional complete hyperkähler manifolds, the ALE spaces of
Kronheimer (1989). These moduli spaces will have natural completions and
various components of them will be the spaces Mφk,c1 which were mentioned
in the introduction. They will resurface later as examples for Nakajima quiver
varieties.

16.2.2 Moduli space of magnetic monopoles on R3

The following construction can be considered as a dimensional reduction of the
previous example. Here we follow Hitchin (1987a, I example 3.5) and Atiyah and
Hitchin (1988).
Assume that G = SU (2) and the matrices Ai are independent of x4 . Then we
have

A = A1 dx1 +A2 dx2 +A3 dx3

a connection on R3 and A4 = φ ∈ Ω0 (R3 , ad(P )) becomes the Higgs field. The

gauge group now is G = Ω(R3 , Ad(P )) and M is the space of configurations (A, φ)
satisfying certain boundary conditions. (The boundary condition is chosen to
ensure finite energy.) The gauge group G acts on M by gauge transformations,
preserving the natural hyperkähler metric on M. The corresponding hyperkähler
moment map equation

µH (A, φ) = 0 ⇔ FA = ∗dA φ

is equivalent to the Bogomolny equation.

Now by construction M = µ−1 H (0)/G, the moduli space of magnetic monopoles
on R3 has a natural hyperkähler metric. It has infinitely many components M =
∪∞k=1 Mk labeled by the magnetic charge k of the monopole.
Mk is acted upon by R3 by translations and by U (1) by rotating the phase of
the monopole. The quotient Mk0 is still a smooth complete hyperkähler manifold
of dimension 4k − 4, with fundamental group Zk . We will denote by M J0 its
k
universal cover. In Atiyah and Hitchin (1985) they find the hyperkähler metric
explicitly on the four-manifold M20 and subsequently describe the scattering of
two monopoles.
328 S-duality in hyperkähler Hodge theory

16.2.3 Hitchin moduli space

This example can be considered as a two-dimensional reduction of Section 16.2.1.
We follow Hitchin (1987b, section 1; 1987a I example 3.3).
Now we assume that G = U (n) and the matrices Ai in Section 16.2.1 are inde-
pendent of x3 , x4 . We have now the connection A = A1 dx1 +A2 dx2 on the U (n)
principal bundle P on R2 . We introduce Φ = (A3 − A4 i)dz ∈ Ω1,0 (R2 , ad(P )⊗ C)
the complex Higgs field. The gauge group now is G = Ω(R2 , Ad(P )), which acts
by gauge transformations on the space M of configurations (A, Φ) preserving the
natural hyperkähler metric on M. The moment map equations

F (A) = −[Φ, Φ∗ ],
µH (A, Φ) = 0 ⇔
dA Φ = 0

are then equivalent with Hitchin’s self-duality equations. There are no solutions
of finite energy on R2 , but as the equations are conformally invariant, we can
replace R2 with a genus g compact Riemann surface C in the above definitions,
and define M(C, P ) = µ−1H (0)/G, the Hitchin moduli space, which has a natural
hyperkähler metric by construction. There are different ways to think about
this space with the different complex structures, which will be explained in
Section 16.5.2.

16.3 Hodge theory

16.3.1 L2 harmonic forms on complete manifolds
Let M be a complete Riemannian manifold of dimension n. We say that a smooth
differential k-form α ∈ Ωk (M ) is harmonic if and only if dα = d∗α = 0, where
∗ : Ωk (M ) → Ωn−k (M ) is the Hodge star operator. It is L2 if and only if

α ∧ ∗α < ∞.
M

We denote by H∗ (M ) the space of L2 harmonic forms.

A fundamental theorem of Hodge theory is the Hodge (orthogonal) decom-
position theorem of de Rham (1984, section 32 theorem 24, Section 35 theorem
26):
   
Ω∗L2 = d Ω∗cpt ⊕ H∗ ⊕ δ Ω∗cpt , (16.2)

where δ is the adjoint of d. When M is compact this implies the celebrated

Hodge theorem, which says that H∗ (M ) ∼ = H ∗ (M ), that is, that there is a
unique harmonic representative in every de Rham cohomology class. When M
is non-compact we only have a topological lower bound. Namely, the Hodge
decomposition theorem implies that the composite map
∗
Hcpt (M ) → H∗ (M ) → H ∗ (M )
 Hodge theory 329

is just the forgetful map. (In the compact case these maps are isomorphisms,
which gives the Hodge theorem mentioned above.) Thus
∗
im(Hcpt (M ) → H ∗ (M )) (16.3)

is a “topological lower bound” for H∗ (M ). By Poincaré duality the map

∗
Hcpt (M ) → H ∗ (M ) is equivalent with the intersection pairing on Hcpt
∗
(M ).
In the cases most relevant for us M will be a hyperkähler manifold (some-
times orbifold) so dim(M ) = 4k and we will additionally have H i (M ) = 0 for
∗
i > 2k. Therefore the possible non-trivial image in im(Hcpt (M ) → H ∗ (M )) will
be concentrated in the middle 2k dimension. (We will use the notation mid =
dim(M )/2 for the middle dimension of a manifold.) For such a hyperkähler
manifold we denote
  mid 
χL2 (M ) = dim im Hcpt (M ) → H mid (M )
  ∗ 
= dim im Hcpt (M ) → H ∗ (M ) (16.4)

the dimension of this image. χL2 (M ) can be thought of either as a “topological

lower bound” for dim(H∗ (M )) or the Euler characteristic of topological L2
cohomology.

16.3.2 Results on L2 harmonic forms

There were few general theorems on describing H∗ (M ) for a non-compact com-
plete manifold M (see however Hausel et al. 2004, introduction for an overview).
It was thus a surprising development when Sen (1994), using arguments from
S-duality, managed to predict the dimension of L2 harmonic forms on M k as
0
was explained in Conjecture 16.1 in the Introduction. In particular, according to
Sen’s Conjecture 16.1 the space H2 (M0 ) should be one dimensional. Using the
2
explicit description of Atiyah and Hitchin (1985) of the metric on M02 in Sen
(1994) he was able to find an explicit L2 harmonic two-form, called the Sen two-
form, on M20 . This was perhaps the strongest mathematical support exhibited
for Conjecture 16.1 in Sen (1994).
More general mathematical support for Conjecture 16.1 came in 1996. Segal
mid 0
and Selby (1996) showed that the intersection form on Hcpt (Mk ) is definite.
Moreover they obtained for the topological lower bound (16.3) for Hmid (M 0 )
k

0 ) = dim(H mid (M
χL2 (M 0 )) = φ(k).
k k

0 ) in Sen’s Conjecture 16.1.

This agrees with the predicted dimension of Hmid (M k
Motivated by Problem 16.1 and Segal–Selby’s topological lower bound for
Conjecture 16.1, the author calculated in Hausel (1998) that the intersection
mid
pairing on the g-dimensional space Hcpt (M1Dol (SL2 )) is trivial, in other words
 
χL2 M1Dol (SL2 ) = 0 (16.5)
330 S-duality in hyperkähler Hodge theory

for g > 1. This thus gave the surprising result that there are no L2 harmonic
forms on M1Dol (SL2 ) plainly by topological reasons. The technique used in the
proof of (16.5) was imitating Kirwan’s proof (1992) of Mumford’s conjecture on
the cohomology ring of the moduli space of stable rank 2 bundles of degree 1 on
the Riemann surface C. Therefore the extension of (16.5) to higher rank Higgs
bundle moduli spaces MdDol (SLn ) was not straightforward.
The next advance towards Sen’s Conjecture 16.1 came in 2000. Hitchin (2000)
showed that Hd (M ) = 0 unless d = dim(M )/2 for a complete hyperkähler mani-
fold M of linear growth. Examples include all our hyperkähler quotients discussed
in this chapter. The proofs in Hitchin (2000) use techniques inspired by Jost and
Zuo’s extension (2000) of ideas of Gromov (1991). It is interesting to note that
some of the proofs in Hitchin (2000) also exploit the operators in hyperkähler
Hodge theory, which are relevant in N = 4 supersymmetry. Using the symmetries
of the Atiyah–Hitchin metric (Hitchin 2000) proves Sen’s conjecture for k = 2,
that up to a scalar the only L2 harmonic form on M 0 is Sen’s two-form.
2
A more topological approach was introduced in Hausel et al. (2004). Hausel
et al. (2004) proves for fibered boundary manifolds M
 
Hmid (M ) ∼ mid
= im IHm mid
(M )→IHm̄ (M ) , (16.6)

where M is a certain compactification of M , dictated by the asymptotics of the

mid
fibered boundary metric on M . Moreover IHm (M ) denotes the intersection
cohomology in dimension mid = dim(M )/2 with lower middle perversity m and
mid
IHm̄ (M ) denotes the intersection cohomology in the middle dimension with
upper middle perversity m̄ of the possibly badly singular (i.e. not necessarily
a Witt space) compactification M . To illustrate (16.6) we take the compact-
ification of M 0 , which happens to be the smooth space CP2 (with the non-
2
standard orientation), where the above cohomologies in (16.6) all coincide, giving
H2 (M0 ) ∼ 2 2
2 = H (CP ). This provides a topological explanation for the existence
and uniqueness of the Sen two-form.
The assumption that the metric is fibered boundary in Hausel et al. (2004) is
fairly restrictive. Among hyperkähler quotients only a few examples satisfy this
property (see the discussion in Hausel et al. 2004, section 7). Examples include
all ALE gravitational instantons of Kronheimer (1989) and all known ALF (see
Cherkis and Kapustin 1999) and some ALG gravitational instantons (see Cherkis
and Kapustin 2002). In general our hyperkähler quotients have some kind of
stratified asymptotic behaviour at infinity. For example, the metric on Mk0 is
fibered boundary only when k = 2, for higher k it is known to behave differently
at different regions of infinity. The first result which could handle Hodge theory
on Riemannian manifolds with such a stratified behaviour at infinity appeared
recently in a work by Carron. It proves for a QALE space M that:
 mid 
Hmid (M ) ∼
= im Hcpt (M )→H mid (M ) .
 Mixed Hodge theory 331

A QALE space (Joyce 2000, section 9) by definition is a certain Calabi–Yau

metric on a crepant resolution of Ck /Γ, where Γ ⊂ SU (k) is a finite subgroup.
The asymptotics of the metric on such a QALE space is reminiscent to the
asymptotics of the natural hyperkähler metric on Mφk,c1 appearing in the Vafa–
Witten Conjecture 16.2. It is thus reasonable to hope that the Vafa–Witten
Conjecture 16.2 will be decided soon.
As there have been extensive studies starting with Gibbons and Manton
(1995) and more recently Bielawski (2008) on the asymptotics of the Riemannian
metric on Mk0 , it is conceivable that we will have a precise understanding of the
asymptotic behaviour of this metric, and in turn the Hodge theory of L2 harmonic
forms on M 0 , perhaps extending techniques from Carron. Thus one may be
k
optimistic that Sen’s Conjecture 16.1 will be decided in the foreseeable future.
Finally, one must admit that the description of the asymptotics of the metric
at infinity on MdDol (SLn ) is still lacking, thus calculation of H∗ MdDol (SLn )
is presently
 hopeless.
 The topological side of Problem 16.1, that is, to determine
χL2 MdDol (SLn ) , when (d, n) = 1, is more reasonable. After introducing a
new arithmetic technique to study Hodge structures on the cohomology of our
hyperkähler manifolds, we will be able to offer a general conjecture on the
intersection form on Higgs moduli spaces, in particular that (16.5) holds for
any n.

16.4 Mixed Hodge theory

As explained above there have been some limited successes of calculating H∗ (M )
for a hyperkähler quotient and understanding its relation to the cohomology
H ∗ (M ) or more generally the cohomology of an appropriate compactification
H ∗ (M̄ ). Another extension of Hodge theory yields some different and in some
ways more detailed insight into the cohomology of our hyperkähler quotients.
This technique is Deligne’s mixed Hodge structure on the cohomology of any
complex algebraic variety. Instead of the global analysis on the Riemannian
geometry of the complex algebraic variety it will relate to the arithmetic of
the variety over finite fields.

16.4.1 Mixed Hodge structure of Deligne

Motivated by the (then still unproven) Weil conjectures and Grothendieck’s
“yoga of weights”, which drew cohomological conclusions about complex varieties
from the truth of those conjectures, Deligne (1971, 1974) proved the existence
of mixed Hodge structures on the cohomology H ∗ (M, Q) of a complex algebraic
variety M . Here we give a quick introduction, for more details see Hausel and
Rodriguez-Villegas (section 2.2) and the references therein. Deligne’s mixed
Hodge structure entails two filtrations on the rational cohomology of M . The
increasing weight filtration

0 = W−1 ⊆ W0 ⊆ · · · ⊆ W2j = H j (X, Q)

332 S-duality in hyperkähler Hodge theory

and a decreasing Hodge filtration

H j (X, C) = F 0 ⊇ F 1 ⊇ · · · ⊇ F m ⊇ F m+1 = 0.

We can define mixed Hodge numbers obtained from these two filtrations by the
following formula:
 
hp,q;j (X) := dimC GrpF Grp+q
W
H j (X)C . (16.7)

From these numbers we form


H(M ; x, y, t) = hp,q;k (M )xp y q tk ,
p,q,k

the mixed Hodge polynomial. By virtue of its definition it has the property that
the specialization

P (M ; t) = H(M ; 1, 1, t)

gives the Poincaré polynomial of M . When M is smooth of dimension n we take

another specialization

E(M ; x, y) := xn y n H(1/x, 1/y, −1), (16.8)

the so-called E-polynomial of a smooth variety M .

Deligne’s construction of mixed Hodge structure is complex geometrical: for
a smooth variety M it is defined by the log geometry of a compactification M
with normal crossing divisors. In particular a global analytical description, like
the Hodge theory of harmonic forms on a smooth complex projective manifold,
of the mixed Hodge structure on a smooth variety is missing, which causes some
difficulty in finding the meaning of mixed Hodge numbers in physical contexts
(see the remark after Conjecture 16.3).

16.4.2 Arithmetic and topological content of the E-polynomial

The connection of the E-polynomial to the arithmetic of the variety is provided
by the following theorem of Katz (Hausel and Rodriguez-Villegas, Appendix).
Here we give an informal version of Katz’s result for precise formulation (see
Hausel and Rodriguez-Villegas 2008, theorem 6.1.2(3), theorem 2.1.8):

Theorem 16.1 Let M be a smooth quasi-projective variety defined over Z (i.e.

given by equations with integer coefficients). Assume that the number of points
of M over a finite field Fq , that is,

E(q) := #{M (Fq )}

is a polynomial in q. Then the E-polynomial can be obtained from the count

polynomial as follows:

E(M ; x, y) = E(xy).
 Applications of mixed Hodge theory 333

This theorem is especially useful when we further have hp,q;k (M ) = 0 unless

p + q = k. In this case we say that the mixed Hodge structure on H ∗ (M ) is pure.
In this case


−1 −1
H(M ; x, y, t) = (xyt2 )n E ,
xt yt
and so the Poincaré polynomial can be recovered from the E-polynomial as
follows:


−1 −1
P (M ; t) = H(M ; 1, 1, t) = t2n E , .
t t
Examples of varieties with pure MHS on their cohomology include smooth
projective varieties (in this case we get the traditional Hodge structure, which is
by definition pure), the moduli space of Higgs bundles MDol , the moduli space
of flat connections MDR on a Riemann surface, and Nakajima’s quiver varieties.
In general we can define the pure part of H(M ; x, y, t) as
 
P H(M ; x, y) = Coeff T 0 H(M ; xT, yT, tT −1 ) .

More generally we can define the pure part of the cohomology of M as

P H ∗ (M ) := Wn H n (M ) ⊂ H ∗ (M ),

which is a subring P H ∗ (M ) ⊂ H ∗ (M ) of the cohomology of M . For a smooth

M , the pure part of H ∗ (M ) is always the image of the cohomology of a smooth
compactification (see Deligne 1971, corollaire 3.2.17). It is in fact this result
which can be used to show that the spaces mentioned in the previous paragraph
have pure mixed Hodge structure. That is, one can prove that they admit
a smooth compactification which surjects on cohomology. Prototypes of such
compactifications were constructed in Simpson (1997) for MDR and in Hausel
(1998) for MDol .

16.5 Applications of mixed Hodge theory

Using the method sketched in the previous section the strongest results on
cohomology can be achieved when the variety has a pure MHS on its cohomology,
consequently the E-polynomial determines the mixed Hodge polynomial, and
additionally it has polynomial-count so that Theorem 16.1 gives an arithmetic
way to determine the E-polynomial. This is the case for Nakajima quiver
varieties, where our method gives complete results.

16.5.1 Nakajima quiver varieties

Nakajima quiver varieties are constructed (Nakajima 1998) by a finite-
dimensional hyperkähler quotient construction. Here we review the affine
algebraic-geometric version of this construction.
334 S-duality in hyperkähler Hodge theory

Let Γ be a quiver (oriented graph) with vertex set I = {1, . . . , n} and edges
E ⊂ I × I. Let
v = (v1 , . . . , vn ), w = (w1 , . . . , wn ) ∈ NI
be two-dimensional vectors and Vi and Wi corresponding complex vector spaces,
that is, dim(Vi ) = vi and dim(Wi ) = wi . We define the vector spaces
9 9
Vv,w = Hom(Vt(a) , Vh(a) ) ⊕ Hom(Vi , Wi )
a∈E i∈I

of framed representations of the quiver Γ , and the action

F
ρ : GL(v) := GL(Vi ) → GL(Vv ),
i∈I

with derivative
F
 : gl(v) := gl(Vi ) → gl(Vv ).
i∈I

The complex moment map

µ : Vv,w × V∗v,w → gl∗v
of ρ is given at X ∈ glv by
µ(v, w), X = (X)v, w. (16.9)
For ξ = 1v ∈ gl(v) GL(v)
we define the (always smooth) Nakajima quiver vari-
ety by
 
M(v, w) = µ−1 (ξ)//GL(v) = Spec C[µ−1 (ξ)]GL(v)

as an affine GIT quotient. Alternatively one can construct the manifold underly-
ing M(v, w) as a hyperkähler quotient of Vv,w × V∗v,w by the maximal compact
subgroup U (v) ⊂ GL(v). This shows that M(v, w) possesses a hyperkähler
metric. The holomorphic symplectic quotient we presented above is the one where
the arithmetic technique of Section 16.4 is applicable. Before we explain that,
let us recall the following fundamental theorem of Nakajima (1998) about the
cohomology of these Nakajima quiver varieties:
Theorem 16.2 Assume that the quiver Γ has no edge-loops. Then there is an
irreducible representation of the Kac–Moody algebra g(Γ) of highest weight w
on ⊕v H mid (M(v, w)). In particular the Weyl–Kac character formula gives the
middle Betti numbers of Nakajima quiver varieties. Furthermore the intersection
form on Hcmid (M(v, w)) is definite, thus χL2 (M(v, w)) equals the middle Betti
number of M(v, w).
Remark 16.1 When Γ is an affine Dynkin diagram M(v, w) could be identi-
fied with one of the spaces Mφk,c1 of certain Yang–Mills instantons on a ALE
 Applications of mixed Hodge theory 335

space XΓ . In Kac (1990) he explains that the Weyl–Kac character formula for
an affine Dynkin diagram has certain modular properties. This was the line of
argument in Vafa and Witten (1994) that (16.1) is a modular form provided
Conjecture 16.2 holds.

In Hausel (2006) a simple Fourier transform technique was found to enumerate

the rational points of M(v, w) over a finite field Fq . The corresponding count
function E(q) turned out to be polynomial, and as the mixed Hodge structure is
pure on H ∗ (M(v, w)) the technique of Section 16.4 applies in its full strength
to give a formula for the Betti numbers of the varieties M(v, w). The result is
the following formula from Hausel (2006):

Theorem 16.3 For any quiver Γ, the Betti numbers of the Nakajima quiver
varieties are given by the following generating function, with the notation as in
Hausel (2006, theorem 3):
M  M 
  (i,j)∈E t−2λ
i ,λj 
i∈I t−2λ
i ,(1wi )

Tv M


M M mk (λi )

 v∈NI λ∈P(v) i∈I t−2λ
i ,λi 
k j=1 (1−t2j )
−d(v,w) v
Pt (M(v, w))t T =   M j ,
t−2λi ,λ 
v∈NI Tv
(i,j)∈E

M i i M M mk (λi )
i∈I t−2λ ,λ  k j=1 (1−t2j )
v∈NI λ∈P(v)

(16.10)

Remark 16.2 When Γ has no edge-loops Nakajima’s Theorem 16.1 implies

that the right-hand side of (16.10) is a deformation of the Weyl–Kac character
formula. Simple reasoning gives the same result about the denominator of the
right-hand side of (16.10) and the Kac denominator. Moreover, Kac’s denomina-
tor formula and Hua’s formula (2000, theorem 4.9) expressing the denominator of
(16.10) as an infinite product imply a conjecture of Kac (cf. Hua 2000, corollary
4.10). Namely, if AΓ (v, q) denotes the number of absolutely indecomposable
representations of Γ of dimension vector v over the finite field Fq , then it turns
out to be a polynomial in q and Kac’s conjecture 1 (1983) says that the constant
coefficient

AΓ (v, 0) = mv (16.11)

equals with the multiplicity of the weight v in the Kac–Moody algebra g(Γ).
This can be proved, as sketched above and announced in Hausel (2006), to be a
consequence of (16.10) and the above-mentioned results of Nakajima and Hua.

Remark 16.3 When the quiver is affine ADE and the RHS becomes an infinite
product (indications that this can happen are the infinite product in Hausel
2006, section 3 and the infinite products in the recent Sasaki) we could get
an alternative proof of the modularity of (16.1) in the Vafa–Witten S-duality
conjecture.
336 S-duality in hyperkähler Hodge theory

In the remaining part of this survey we will motivate and study another
application of the technique in Section 16.4, which will be less powerful as the
mixed Hodge structure will fail to be pure, but will also open new interesting
directions by the study of this more complicated mixed Hodge structure.

16.5.2 Spaces diffeomorphic to the Hitchin moduli space M(C, PU (n) )

Among the spaces discussed in this chapter it is the Hitchin moduli space
M(C, PU (n) ) as defined in Section 16.2.2 which exhibits perhaps the most
plentiful structures many of which are rooted in its hyperkähler quotient origin.
In particular there are three distinct complex algebraic variety structures on
M(C, PU (n) ). These can be thought of (Simpson 1997) as the three types of
non-Abelian (first) cohomology: Dolbeault, De Rham, and Betti, of the Riemann
surface C. The survey paper Hausel (2005) gives a quick introduction to these
spaces and some of the cohomological implications to be discussed below.
In this chapter the ground field is always C unless otherwise indicated.
Following Hitchin (1987b) and Simpson (1997) we define a component of the
twisted GLn = GLn (C) Dolbeault cohomology of C as
< =
Moduli space of semistable rank n
MdDol (GLn ) :=
degree d Hitchin pairs on C
the GLn De Rham cohomology as
< =
Moduli space of flat GLn -connections
MdDR (GLn ) := 2πid
on C \ {p}, with holonomy e n Id around p
and the GLn Betti cohomology

MdB (GLn ) := A1 , B1 , . . . , Ag , Bg ∈ GLn |

A−1 −1 −1 −1 2πid
1 B1 A1 B1 · · · Ag Bg Ag Bg = e Id //GLn
n

as a twisted GLn character variety of C.

When d = 0 these three varieties are diffeomorphic to the Hitchin moduli
space M(C, PU (n) ). However we prefer to consider the twisted versions, when
(d, n) = 1, because then all the varieties are smooth. In this case these
three varieties are all diffeomorphic to a twisted version Md (C, PU (n) )
of Hitchin moduli space
 and so to each other.
 The mixed
 Hodge struc-
ture is pure on H ∗ MdDol (GLn ) and H ∗ MdDR (GLn ) , while it is not
pure on H ∗ MdB (GLn ) . As the mixed
 Hodge structures are different on
H ∗ MdDR (GLn ) and H ∗ MdB (GLn ) , the spaces MDR d
(GLn ) and MdB (GLn )
cannot be isomorphic as complex algebraic varieties. Nevertheless as complex
analytic manifolds the Riemann–Hilbert monodromy map
RH
MdDR (GLn ) → MdB (GLn ) (16.12)
sending a flat connection to its holonomy gives an isomorphism.
 Applications of mixed Hodge theory 337

We will also consider the varieties MdDol (SLn ), MdDR (SLn ), and M0B (SLn ),
which can be defined by replacing GLn with SLn in the above definitions.
Moreover M0Dol (GL1 ), M0DR (GL1 ), and M0B (GL1 ) turn out to be Abelian
groups. Then M0Dol (GL1 ), M0DR (GL1 ), and M0B (GL1 ) will act on MdDol (GLn ),
MdDR (GLn ), and MdB (GLn ), respectively, by an appropriate form of ten-
sorization. Finally we denote the corresponding (affine GIT) quotients by
MdDol (P GLn ), MdDR (P GLn ) and MdB (P GLn ). In our case, when (d, n) = 1,
they will turn out to be orbifolds. For more details on the construction of these
varieties see Hausel (2005).
In the next section we explain the original motivation to consider the
E-polynomials of these three complex algebraic varieties. The motivation is
mirror symmetry, and most probably the same S-duality we discussed in the
Introduction in connection with the Hodge cohomology of the moduli spaces of
Yang–Mills instantons in four dimension and magnetic monopoles in three. S-
duality ideas relating to mirror symmetry for Hitchin spaces have appeared in
the physics literature (Bershadsky et al. 1995; Kapustin and Witten 2007).

16.5.3 Topological mirror test

For our mathematical considerations the relationship to mirror symmetry stems
from the following observation of Hausel and Thaddeus (2003). It uses the famous
Hitchin map (Hitchin 1987c), which makes the moduli space of Higgs bundles
MDol into a completely integrable Hamiltonian system, so that the generic fibers
are Abelian varieties.

Theorem 16.4 In the following diagram

MdDol (P
⏐ GLn ) MDol⏐(SLn )
d
⏐χP GL ⏐χSL
+ n + n
HP GLn ∼
= HSLn

the generic fibers of the Hitchin maps χP GLn and χSLn are dual Abelian
varieties.

Remark 16.4 If we change complex structures and consider MdDR (P GLn )

and MdDR (SLn ), then the Hitchin map on them becomes special Lagrangian
fibrations, and consequently the pair of MdDR (P GLn ) and MdDR (SLn ) satisfies
the requirements of the SYZ construction (Strominger et al. 1996) for a pair of
mirror symmetric Calabi–Yau manifolds (see Hausel and Thaddeus 2001, 2003
for more details).

This motivates the calculation of Hodge numbers of MdDR (P GLn ) and

MdDR (SLn ) to see if there is any relationship between them, which one would
expect in mirror symmetry. Based on calculations in the n = 2, 3 cases Hausel
and Thaddeus (2003) proposed
338 S-duality in hyperkähler Hodge theory

Conjecture 16.3 For all d, e ∈ Z, satisfying (d, n) = (e, n) = 1,

Be
  B̂ d
Est x, y; MdDR (SLn ) = Est (x, y; MeDR (P GLn )) ,

where B e and B̂ d are certain gerbes on the corresponding Hitchin spaces and
the E-polynomials above are stringy E-polynomials for orbifolds twisted by the
relevant gerbe as defined in Hausel and Thaddeus (2003).
Morally, this conjecture should be related to the S-duality considerations of
Kapustin and Witten (2007) and in turn to the geometric Langlands programme
of Beilinson and Drinfeld (1995). However the lack of global analytical inter-
pretation of the mixed Hodge numbers (16.7) appearing in Conjecture 16.3
prevents a straightforward physical interpretation. Nevertheless the agreement
of certain Hodge numbers for Hitchin spaces for Langlands dual groups is an
interesting direction from a purely mathematical point of view. In particular,
if we change our focus from MDR and MDol to MB we will uncover some
surprising connections to the representation theory of finite groups of Lie type.

16.5.4 Mirror symmetry for finite groups of Lie type

As MDR and MB are complex analytically identical via the Riemann–Hilbert
map (16.12), the complex analytical structure of dual special Lagrangian fibra-
tions of Theorem 16.4 are present on the pair MdB (SLn ) and MeB (P GLn ). We
might as well try to think of this pair as mirror symmetric in the SYZ picture.
The mixed Hodge numbers of MB are however different from the mixed Hodge
numbers of MDR so the corresponding topological mirror test (Hausel 2005) will
also be different from Conjecture 16.3:
Conjecture 16.4 For all d, e ∈ Z, satisfying (d, n) = (e, n) = 1,
e   B̂ d
B
Est x, y, MdB (SLn ) = Est (x, y, MeB (P GLn )) .
For this conjecture however there is a powerful arithmetic method to calculate
these E-polynomials. Using this technique we have already managed to check this
conjecture (Hausel 2005) when n is a prime and n = 4. This arithmetic method
is based on the technique explained in Section 16.4 and the following character
formula from Hausel and Rodriguez-Villegas (2008):
Theorem 16.5 Let G = SLn or GLn , let G(Fq ) be the corresponding finite
group of Lie type
√ √     |G(F )|2g−2  
E q, q, MdB (G) = # MdB (G(Fq )) = χ∈Irr(G(Fq )) χ(1)q 2g−1 χ ξnd ,
where the sum is over all irreducible characters of the finite group of Lie type
G(Fq ).
This character formula combined with Conjecture 16.4 implies certain rela-
tionships between the character tables of P GLn (Fq ) and SLn (Fq ). An intriguing
 Applications of mixed Hodge theory 339

way to formulate it is to say that certain differences between the character tables
of P GLn (Fq ) and its Langlands dual SLn (Fq ) are governed by mirror symmetry.
This kind of consideration could be interesting because the character tables of
P GLn (Fq ) or more generally those of GLn (Fq ) have been known for a long time
starting with the work of Green (1955), while the character tables of SLn (Fq )
have just recently been completed (Bonnafé 2006; Shoji 2006). It is especially
enjoyable to follow the effect of the mirror symmetry proposal of Conjecture 16.4
by comparing the character tables of GL2 (Fq ) and SL2 (Fq ) first calculated a
hundred years ago by Jordan 1907 and Schur 1907.

16.5.5 Conjectural answer

Finally, we can put all our observations and conjectures together to state a
conjectural answer to the topological side of Problem 16.1.
As we already noted the mixed Hodge structure on H ∗ (MB ) is not pure.
Therefore we are losing information by considering only E(MB ; x, y). It turns out
that it is interesting to consider the full mixed Hodge polynomial H(MB ; x, y, t).
When n = 2 it can be calculated via the explicit description of H ∗ (MB ) in Hausel
and Thaddeus (2003). We get Hausel and Rodriguez-Villegas (Theorem 1.1.3):

H(MB (P GL2 ); x, y, t)
(q 2 t3 + 1)2g q 2g−2 t4g−4 (q2 t + 1)2g 1 q 2g−2 t4g−4 (qt + 1)2g
= + −
(q 2 t2 − 1)(q 2 t4 − 1) (q 2 − 1)(q 2 t2 − 1) 2 (qt2 − 1)(q − 1)
1 q 2g−2 t4g−4 (qt − 1)2g
− ,
2 (q + 1)(qt2 + 1)
where q = xy and the four terms correspond to the four types of irreducible
characters of GL(2, Fq ). When g = 3 this equals

t12 q 12 + t12 q 10 + 6 t11 q 10 + t12 q 8 + t10 q 10 + 6 t11 q 8 + 16 t10 q 8 + 6 t9 q 8 + t10 q 6

+ t8 q 8 + 26 t9 q 6 + 16 t8 q 6 + 6 t7 q 6 + t8 q 4 + t6 q 6 + 6 t7 q 4 + 16 t6 q 4
+ 6 t5 q 4 + t4 q 4 + t4 q 2 + 6 t3 q 2 + t2 q 2 + 1.

In particular we see that the pure part is 1 + q2 t4 + q 4 t8 . These terms correspond

to the cohomology classes 1, β, and β 2 , and the term q 6 t12 is not present because
by the Newstead relation β g = β 3 = 0 holds (Hausel and Thaddeus 2003). In
particular there is no pure part in the middle = 12-dimensional cohomology.
The same argument holds for all g, which shows that there is no pure part in the
middle-dimensional cohomology of M1B (P GL2 ). It is however easy to see that
the intersection form on middle cohomology can only be non-trivial on the pure
part and so this implies Hausel and Rodriguez-Villegas (2008, Corollary 5.4.1):
∗
 1 
Corollary 16.1 The intersection form on Hcpt MB (P GL2 ) is trivial.
340 S-duality in hyperkähler Hodge theory

This gives an alternative proof of (16.5) as the equation

   
χL2 M1B (SL2 ) = χL2 M1B (P GL2 )
is easy to prove. Moreover this approach is more promising to generalize for any n.
We will offer a conjecture about the pure part of the cohomology of MdB (P GLn )
below and in turn that will yield a conjecture for the intersection form on the
middle-dimensional compactly supported cohomology, answering the topological
side of Problem 16.1.
To state our conjecture in its full generality we introduce character varieties
on Riemann surfaces with k punctures and parabolic type µ = (µ1 , . . . , µk ) at
the punctures, where µi is a partition of n. In other words we fix semisimple
conjugacy classes C1 , . . . , Ck ⊂ GLn , which are generic and have type µ (in other
words µij is the multiplicity of the jth eigenvalue of a matrix in Ci ). One can
prove as in Hausel et al. (2008, lemma 2.1.2) that there exists generic semisimple
conjugacy classes for every type µ = (µ1 , . . . , µk ). For a generic {C1 , . . . , Ck } of
type µ we define
MµB := {A1 , B1 , . . . , Ag , Bg ∈ GLn , C1 ∈ C1 , . . . , Ck ∈ Ck |
[A1 , B1 ] · · · [Ag , Bg ]C1 · · · Ck = In }//GLn
as an affine GIT quotient by the diagonal adjoint action of GLn . The generic
choice of the semisimple conjugacy classes implies that MµB is smooth. The torus
GL2g µ
1 acts on MB by multiplying the matrices Ai and Bi by a scalar. We can
define the quotient
M̄µB := MµB //GL2g
1

as the corresponding P GLn character variety. The variety M̄µB is an orbifold.

By studying the Riemann–Hilbert map on the level of cohomologies we are
led (Hausel, in preparation) to consider the comet-shaped quiver Γ associated to
g and µ. Namely, we can put g loops on a central vertex, and k legs oflength
l
l(µj ). We also equip Γ with a dimension vector v, which has dimension i=1 µji
at the lth vertex on the ith leg. Consider now the number AΓ (q, v) of absolutely
indecomposable representations of Γ of dimension v over the finite field Fq . Kac
(1983, proposition 1.15) proved that AΓ (q, v) is a polynomial in q with integer
coefficients. We have the following conjecture from Hausel (in preparation):
Conjecture 16.5 The pure part of the cohomology of M̄µB is given by
 
P H M̄µB , x, y = (xy)dµ /2 AΓ (v, 1/(xy)),
where (Γ, v) is the star-shaped quiver and dimension vector given by the parabolic
type µ, and dµ is the dimension of MµB .
This conjecture gives a cohomological interpretation of AΓ (v, q) and in partic-
ular implies that it has non-negative coefficients confirming (Kac 1983, conjecture
2) in the case when Γ is comet-shaped. When µ is indivisible Conjecture 16.5 can
 Applications of mixed Hodge theory 341

be proved to follow from the master conjecture in Hausel et al. (in preparation),
which expresses the mixed Hodge polynomials of all the character varieties M̄µB
as a generating function generalizing the Cauchy formula for Macdonald poly-
2
nomials.
 µIt also has the following consequence on the topological L cohomology
χL2 M̄B of (16.4).
Conjecture 16.6 The topological L2 cohomology of the manifold M̄µB is
given by
χL2 (M̄µB ) = 0, when g > 1 (16.13)
χL2 (M̄µB ) = 1, when g = 1 (16.14)
χL2 (M̄µB ) = mv , when g = 0, (16.15)
where mv is the multiplicity of the weight v in the Kac–Moody algebra g(Γ),
which are encoded by the Kac denominator formula for the star-shaped quiver Γ.
When g > 1 and the parabolic type is µ = ((n)), that is, we have only one
puncture with central conjugacy class, then one can identify M̄µB = MdB (P GLn ),
with some d such that (d, n) = 1. In this case (16.13) says that
 
χL2 MdB (P GLn ) = 0,
which appeared as (Hausel and Rodriguez-Villegas Conjecture 4.5.1). It follows
from the mirror symmetry Conjecture 16.3 that
 d   d 
mid
Hcpt MB (SLn ) ∼ mid
= Hcpt MB (P GLn )
and then the intersection forms also agree. This and (16.15) then imply that
(16.5) holds for any n, that is, that the intersection form on the compactly
supported cohomology of MdB (SLn ) is trivial. This gives a conjectural answer
to the topological side of Problem 16.1.
When g = 1 the conjectured (16.14) follows from Conjecture 16.5 and the
observation that the coefficient of q in the A-polynomial AΓ (q) for a g = 1 comet-
shaped quiver Γ is always 1.
When g = 0 the varieties MµB = M̄µB coincide. Conjecture 16.5 then implies
that
χL2 (MµB ) = AΓ (v, 0).
Conjecture (16.15) is a combination of this and the equality AΓ (v, 0) = mv , that
is, Kac’s conjecture 1, in Kac (1983), which, as discussed in Remark 16.2, follows
from Theorem 16.3.
Finally one can define M̄µDol the moduli space of stable parabolic P GLn -
Higgs bundles with quasi-parabolic type µj ∈ P(n) and generic weights at the
jth puncture on the Riemann surface (Boden and Yokogawa 1996, Garcı́a-Prada
et al. 2007). Then one can prove that M̄µB is diffeomorphic to M̄µDol . Thus
Conjecture 16.6 also calculates the intersection form on the compactly supported
342 S-duality in hyperkähler Hodge theory

cohomology of the moduli space M̄µDol of stable parabolic P GLn -Higgs bundles
of any rank.
Example 16.1 Consider the genus 0 Riemann surface P1 with four punctures.
Consider the moduli space Mtoy of stable rank 2 parabolic Higgs bundles on
P1 , with generic parabolic weights on the full parabolic flag at the punctures
(see Boden and Yokogawa 1996). This is a complex surface and the intersection
form on Hc2 (Mtoy ) was discussed in Hausel (1998, example 2 for theorem 7.13).
Hc2 (Mtoy ) is five-dimensional but χL2 (Mtoy ) is only four. (The cohomology class
of the generic fiber of the Hitchin map is the one in the kernel.)
Mtoy is diffeomorphic to the character variety M̄µB where g = 0 and µ =
((1, 1),
 (1,1), (1, 1), (1, 1)). Thus by Conjecture 16.6 we should be able to calculate
χL2 M̄µB in terms of the representation theory of the corresponding quiver Γ.
The corresponding quiver Γ in this case will be the affine D̃4 Dynkin diagram,
with v = (2, 1, 1, 1, 1) the minimal positive imaginary root. Its multiplicity mv
in the affine Kac–Moody algebra associated to Γ is known to be 4. Alternatively
it is known (Kac 1983 example b to conjecture 2) that AΓ (v, q) = q + 4, which
by (16.11) gives mv = 4. Thus indeed χL2 (MµB ) = mv = 4 checking (16.15) in
this case via Hausel (1998, Example 2 for Theorem 7.13).

Acknowledgements
This chapter is a write-up of the author’s talk at the Geometry Conference in
Honour of Nigel Hitchin in Madrid in September 2006. Problem 16.1 was raised
by Nigel Hitchin in 1995, then the author’s PhD supervisor, as a project for the
author’s PhD thesis. This chapter would like to show the impact of this modest-
looking question on the author’s subsequent research. The author’s research has
been supported by a Royal Society University Research Fellowship, NSF grant
DMS-0604775 and an Alfred Sloan Fellowship. The visit to Madrid was supported
by a Royal Society International Joint Project between the United Kingdom and
Spain.

References
Atiyah, M. F. (1978). Geometry of Yang-Mills fields. In Mathematical Problems in
Theoretical Physics (Proceeding of the International Conference, University of Rome,
Rome, 1977), Vol. 80 of Lecture Notes in Physics. pp. 216–21. Springer, Berlin,
Germany.
Atiyah, M. F. and Hitchin, N. J. (1985). Low energy scattering of nonabelian monopoles.
Phys. Lett. A 107(1), 21–5.
Atiyah, M. F. and Hitchin, N. J. (1988). The Geometry and Dynamics of Magnetic
Monopoles. M. B. Porter Lectures. Princeton University Press, Princeton, NJ.
Beilinson, A. and Drinfeld, V. Quantization of Hitchin’s integrable system and Hecke
eigensheaves. (384 pages, unpublished preprint, ca. 1995).
Bershadsky, M., Johansen, A., Shadov, V. and Vafa, C. (1995). Topological reduction
of 4D SYM to 2D σ-models. Nuclear Phys. B 448(1–2), 166–86.
 References 343

Bielawski, R. (2008). Monopoles and clusters. arXiv:hep-th/0702190. Commun. Math.

Phys. 284, 675–712.
Boden, H. U. and Yokogawa, K. (1996). Moduli spaces of parabolic Higgs bundles and
parabolic K(D) pairs over smooth curves. I. Int. J. Math. 7(5), 573–98.
Bonnafé, C. (2006). Sur les caractères des groupes réductifs finis à centre non connexe:
applications aux groupes spéciaux linéaires et unitaires. Astérisque (306), vi, 165.
Carron, G. Cohomologie L2 des variétés QALE. arXiv: math.DG/0501290.
Cherkis, S. A. and Kapustin, A. (1999). Singular monopoles and gravitational instan-
tons. Commun. Math. Phys. 203(3), 713–28.
Cherkis S. A. and Kapustin, A. (2002). Hyper-Kähler metrics from periodic monopoles.
Phys. Rev. D (3) 65(8), 084015, 10.
de Rham, G. (1984). Differentiable Manifolds, Vol. 266 of Grundlehren der Mathema-
tischen Wissenschaften. Springer-Verlag, Berlin, Germany.
Deligne, P. (1971). Théorie de Hodge. II. Inst. Hautes Études Sci. Publ. Math. (40),
5–57.
Deligne, P. (1974). Théorie de Hodge. III. Inst. Hautes Études Sci. Publ. Math. (44),
5–77.
Garcı́a-Prada, O., Gothen, P. and Muñoz, V. (2007). Betti numbers of the moduli space
of rank 3 parabolic Higgs bundles. Memoirs Am. Math. Soc. 187(879), 80.
Gibbons, G. W. and Manton, N. S. (1995). The moduli space metric for well-separated
BPS monopoles. Phys. Lett. B 356(1), 32–8.
Gothen, P. B. (1994). The Betti numbers of the moduli space of stable rank 3 Higgs
bundles on a Riemann surface. Int. J. Math. 5(6), 861–75.
Green, J. A. (1955). The characters of the finite general linear groups. Trans. Am.
Math. Soc. 80, 402–47.
Gromov, M. (1991). Kähler hyperbolicity and L2 -Hodge theory. J. Differential Geom.
33(1), 263–92.
Hausel, T. Arithmetic harmonic analysis, Macdonald polynomials and the topology of
the Riemann-Hilbert monodromy map (in preparation).
Hausel, T. (1998). Compactification of moduli of Higgs bundles. J. Reine Angew. Math.
503, 169–92.
Hausel, T. (1998). Vanishing of intersection numbers on the moduli space of Higgs
bundles. Adv. Theor. Math. Phys. 2(5), 1011–40.
Hausel, T. (2005). Mirror symmetry and Langlands duality in the non-abelian Hodge
theory of a curve. In Geometric Methods in Algebra and Number Theory, Vol. 235 of
Progress in Mathematics. pp. 193–217. Birkhäuser Boston, Boston, MA.
Hausel, T. (2006). Betti numbers of holomorphic symplectic quotients via arithmetic
Fourier transform. Proc. Natl. Acad. Sci. USA 103(16), 6120–4.
Hausel, T., Hunsicker, E. and Mazzeo, R. (2004). Hodge cohomology of gravitational
instantons. Duke Math. J. 122(3), 485–548.
Hausel, T., Letellier, E. and Rodriguez-Villegas, F. Arithmetic harmonic analysis on
character and quiver varieties, preprint arXiv:0810.2076.
Hausel, T. and Rodriguez-Villegas, F. (2008). Mixed Hodge polynomials of character
varieties. Invent. Math. 174(3), 555–624. arXiv:math.AG/0612668.
Hausel, T. and Thaddeus, M. (2001). Examples of mirror partners arising from inte-
grable systems. C. R. Acad. Sci. Paris Sér. I Math. 333(4), 313–18.
Hausel, T. and Thaddeus, M. (2003). Mirror symmetry, Langlands duality, and the
Hitchin system. Invent. Math. 153(1), 197–229.
344 S-duality in hyperkähler Hodge theory

Hausel, T. and Thaddeus, M. (2003). Relations in the cohomology ring of the moduli
space of rank 2 Higgs bundles. J. Am. Math. Soc. 16(2), 303–27.
Hitchin, N. J. (1987a). Monopoles, Minimal Surfaces and Algebraic Curves, Vol. 105
of Séminaire de Mathématiques Supérieures. Presses de l’Université de Montréal,
Montreal, QC.
Hitchin, N. J. (1987b). The self-duality equations on a Riemann surface. Proc. London
Math. Soc. (3) 55(1), 59–126.
Hitchin, N. J. (1987c). Stable bundles and integrable systems. Duke Math. J. 54(1),
91–114.
Hitchin, N. J. (2000). L2 -cohomology of hyperkähler quotients. Commun. Math. Phys.
211(1), 153–65.
Hitchin, N. J., Karlhede, A., Lindström, U. and Roček, M. (1987). Hyper-Kähler metrics
and supersymmetry. Commun. Math. Phys. 108(4), 535–89.
Hua, J. (2000). Counting representations of quivers over finite fields. J. Algebra 226(2),
1011–33.
Jordan, H. E. (1907). Group-characters of various types of linear groups. Am. J. Math.
29(4), 387–405.
Jost, J. and Zuo, K. (2000). Vanishing theorems for L2 -cohomology on infinite coverings
of compact Kähler manifolds and applications in algebraic geometry. Commun. Anal.
Geom. 8(1), 1–30.
Joyce, D. D. (2000). Compact Manifolds with Special Holonomy. Oxford Mathematical
Monographs. Oxford University Press, Oxford.
Kac, V. G. (1983). Root systems, representations of quivers and invariant theory. In
Invariant Theory (Montecatini, 1982), Vol. 996 of Lecture Notes in Mathematics.
pp. 74–108. Springer, Berlin, Germany.
Kac, V. G. (1990). Infinite-Dimensional Lie Algebras. 3rd edition. Cambridge Univer-
sity Press, Cambridge.
Kapustin, A. and Witten, E. (2007). Electric-magnetic duality and the geometric Lang-
lands program. Commun. Number Theory Phys. 1(1), 1–236. arXiv:hep-th/0604151.
Kirwan, F. (1992). The cohomology rings of moduli spaces of bundles over Riemann
surfaces. J. Am. Math. Soc. 5(4), 853–906.
Kronheimer, P. B. (1989). The construction of ALE spaces as hyper-Kähler quotients.
J. Differential Geom. 29(3), 665–83.
Kronheimer, P. B. and Nakajima, H. (1990). Yang-Mills instantons on ALE gravita-
tional instantons. Math. Ann. 288(2), 263–307.
Montonen, C. and Olive, C. D. I. (1977). Magnetic monopoles as gauge particles? Phys.
Lett. B B72, 117–20.
Mumford, D., Fogarty, J. and Kirwan, F. (1994). Geometric Invariant Theory. Vol. 34 of
Ergebnisse der Mathematik und ihrer Grenzgebiete (2), 3rd edition. Springer-Verlag,
Berlin, Germany.
Nakajima, H. (1998). Quiver varieties and Kac-Moody algebras. Duke Math. J. 91(3),
515–60.
Nitsure, N. (1991). Moduli space of semistable pairs on a curve. Proc. London Math.
Soc. (3) 62(2), 275–300.
Osborn, H. (1979). Topological charges for N=4 supersymmetric gauge theories and
monopoles of spin 1. Phys. Lett. B 83, 321.
Sasaki, T. O(−2) blow-up formula via instanton calculus on C 2 /Z and Weil conjecture.
2
arXiv:hep-th/0603162.
 References 345

Schur, I. (1907). Untersuchungen über die darstellung der endlichen gruppen durch
gebrochene lineare substitutionen. J. Reine Angew. Math. 132, 85.
Segal, G. and Selby, A. (1996). The cohomology of the space of magnetic monopoles.
Commun. Math. Phys. 177(3), 775–87.
Sen, A. (1994). Dyon-monopole bound states, self-dual harmonic forms on the multi-
monopole moduli space, and SL(2, Z) invariance in string theory. Phys. Lett. B
329(2–3), 217–21.
Shoji, T. (2006). Lusztig’s conjecture for finite special linear groups. Represent. Theory
10, 164–222 (electronic).
Simpson, C. (1997). The Hodge filtration on nonabelian cohomology. In Algebraic
Geometry—Santa Cruz 1995, Vol. 62 of Proceedings of the Symposium on Pure
Mathematics. pp. 217–81. American Mathematical Society, Providence, RI.
Simpson, C. T. (1991). Nonabelian Hodge theory. In Proceedings of the International
Congress of Mathematicians, Vols. I, II (Kyoto, 1990), pp. 747–56. Mathematical
Society Japan, Tokyo, Japan.
Strominger, A., Yau, S.-T. and Zaslow, E. (1996). Mirror symmetry is T -duality.
Nuclear Phys. B 479(1–2), 243–59.
Vafa, C. and Witten, E. (1994). A strong coupling test of S-duality. Nuclear Phys. B
431(1–2), 3–77.
Witten, E. and Olive, D. (1978). Supersymmetry algebras that include topological
charges. Phys. Lett. B 78, 97.
 XVII
NON-EMBEDDING AND NON-EXTENSION RESULTS
IN SPECIAL HOLONOMY

Robert L. Bryant
Dedicated to Nigel Hitchin with great admiration on the occasion of his 60th birthday

17.1 Introduction
In the early analyses of metrics with special holonomy in dimensions 7
and 8, particularly in regards to their existence and generality, heavy use was
made of the Cartan–Kähler theorem, essentially because the analyses were
reduced to the study of overdetermined PDE systems whose natures were
complicated by their diffeomorphism invariance. The Cartan–Kähler theory is
well suited for the study of such systems and the local properties of their
solutions. However, the Cartan–Kähler theory is not particularly well suited
for studies of global problems for two reasons: first, it is an approach to
PDE that relies entirely on the local solvability of initial value problems
and, second, the Cartan–Kähler theory is only applicable in the real-analytic
category.
Nevertheless, when there are no other adequate methods for analyzing the
local generality of such systems, the Cartan–Kähler theory is a useful tool and it
has the effect of focusing attention on the initial value problem as an interesting
problem in its own right. The point of this chapter is to clarify some of the
existence issues involved in applying the initial value problem to the problem
of constructing metrics with special holonomy. In particular, the role of the
assumption of real-analyticity will be discussed and examples will be constructed
to show that one cannot generally avoid such assumptions in the initial value
formulations of these problems.
The general approach can be outlined as follows: As is well known (cf. Bryant
1987), the problem of understanding the local Riemannian metrics in dimen-
sion n whose holonomy is contained in a specified connected group H ⊂ SO(n)
is essentially equivalent to the problem of understanding the H-structures in
dimension n whose intrinsic torsion vanishes, or, equivalently, that are parallel
with respect to the Levi-Civita connection of the Riemannian metric associated
to the underlying SO(n)-structure. In this chapter, an n-manifold M endowed
with an H-structure B → M with vanishing intrinsic torsion will be said to be
 Introduction 347

an H-manifold. 1 In particular, an H-manifold is a manifold M endowed with an

H-structure B that is flat to first order.
It frequently happens (as it does for all of the cases to be considered here)
that H acts transitively on S n−1 , with stabilizer subgroup K ⊂ H, where
 
K = {1}×SO(n−1) ∩ H.

In this case, an oriented hypersurface N ⊂ M in an H-manifold M inherits, in a

natural way, a K-structure B  → N . Typically, this K-structure will not, itself,
be torsion-free (unless N is a totally geodesic hypersurface in M ), but will satisfy
some weaker condition on its intrinsic torsion, essentially that its intrinsic torsion
can be expressed in terms of the second fundamental form of N as a submanifold
of M . 2 The problem then becomes to determine whether these weaker conditions
on a given K-structure B  → N n−1 are sufficient to imply that (N, B  ) can be
induced by immersion into an H-manifold (M, B).
It turns out that the concept of real-analyticity has an important bearing on
this question. As was pointed out in DeTurck and Kazdan (1981), a metric g
on M can only be real-analytic with respect to at most one real-analytic
structure on M : one considers the (non-maximal) atlas H(g) that consists of
the g-harmonic coordinate charts. If the overlaps of H(g) are not real-analytic,
then g is not real-analytic with respect to any real-analytic structure on M ,
while, if the overlaps of H(g) are real-analytic, it lies in a unique maximal real-
analytic atlas A(g), that is, a real-analytic structure on M , and this is the only
real-analytic structure on M with respect to which g could be real-analytic.
Because H ⊂ O(n), an H-structure B → M induces a canonical “underlying”
metric gB on M . In particular, if the H-structure B → M is real-analytic with
respect to some real-analytic structure A, then H(gB ) must be a real-analytic
atlas and A must equal A(gB ). Thus, it makes sense to say that an H-structure
is real-analytic, meaning that H(gB ) is a real-analytic atlas and that B is real-
analytic with respect to the underlying real-analytic structure A(gB ).
In the three cases to be considered in this chapter, in which H is one
of SU (2) ⊂ SO(4), G2 ⊂ SO(7), or Spin(7) ⊂ SO(8), it will be shown that the
weaker intrinsic torsion conditions on a hypersurface structure are sufficient to
induce an embedding (in fact, essentially a unique one) provided that the given
K-structure is real-analytic. It will also be shown, in each case, that there are
K-structures B  → N n−1 that satisfy these weaker intrinsic torsion conditions
that, nevertheless, cannot be induced by an immersion into an H-manifold. Of
course, such structures are not real-analytic.

1 While, strictly speaking, an H-manifold is a pair (M, B), it is common to refer to M as

an H-manifold when the torsion-free H-structure B can be inferred from context.

2 This observation that the intrinsic torsion of the induced K-structure on N can be

expressed in terms of the second fundamental form can be found, for example, in Conti and
Salamon (2007).
348 Non-embedding and non-extension results in special holonomy

The existence of the desired embedding in the analytic case is a consequence

of the Cartan–Kähler theorem and, for the cases of G2 ⊂ SO(7) and Spin(7) ⊂
SO(8), is implicit in the original analyses of G2 and Spin(7) structures to be
found in Bryant (1987). The case of SU (2) ⊂ SO(4) appears not to have been
treated in this manner before and will be discussed in detail in this chapter. 3
The examples constructed below, showing that existence can fail when one
does not have real-analyticity, appear to be new.
There is another interpretation of the initial value problem that has been
considered by a number of authors, in particular, Hitchin (2001): The idea of a
“flow” of K-structures that gives rise to a torsion-free H-structure. Let M be a
connected n-manifold endowed with an H-structure and let g be the underlying
metric. Let N ⊂ M be an embedded, normally oriented hypersurface, and let
r : M → [0, ∞) be the distance (in the metric g) from the hypersurface N . As
is well known, there is an open neighborhood U ⊂ M of N on which there is
a smooth function t : U → R satisfying |t| = r and dt = 0 on U as well as the
condition that its gradient along N is the specified oriented normal. There is
then a well-defined smooth embedding (t, f ) : U → R × N , where, for p ∈ U , the
function f (p) is the closest point of N to p. In this way, U can be identified with
an open neighborhood of {0} × N in R × N and, in particular, each level set of t
in U can be identified with an open subset of N . (When N is compact, there
will be an > 0 such that the level sets t = c for |c| < will be diffeomorphic
to N .) Thus, at least locally, one can think of the H-structure on U as a one-
parameter family of K-structures on N . When one imposes the condition that the
H-structure on U ⊂ R × N be torsion-free, this can be expressed as a first-order
initial value problem with the given K-structure on {0} × N as the initial value.
This first-order initial value problem is sometimes described as a “flow,” but this
can be misleading, especially if it causes one to think in terms of parabolic or
hyperbolic PDE.
Indeed, the character of this PDE problem is more like that of trying to use the
Cauchy–Riemann equations ux = vy and vx = −uy to extend a complex-valued
function defined on the imaginary axis x = 0 to a holomorphic function on a
neighborhood of the imaginary axis. One knows that a necessary and sufficient
condition for being able to do this is that the given function on the imaginary
axis must be real-analytic. In the cases to be considered in this chapter, the
requirements are not this strong since there will be cases in which the initial
K-structure is not real-analytic and yet a solution of the initial value problem
will exist. However, as will be seen, real-analyticity is a sufficient condition.
For background on the use of exterior differential systems in this chapter, the
reader might consult Bryant et al . (1991).
I have included the case of SU (2)-structures on four-manifolds because of its
historical interest (it was the first case of special holonomy to be analyzed) and

3 These three cases do not exhaust the possibilities; for the example of SU (3) ⊂ SO(6),

see Conti and Salamon (2007).

 Beginnings 349

because the algebra is simpler. Also, because other approaches, based on the
existence of local holomorphic coordinates, have been employed in this case,
there is an instructive comparison to be made between those methods and the
Cartan–Kähler approach. For this reason, I go into the SU (2)-case in some detail.
I hope that the reader will find this as interesting as I have.

17.2 Beginnings
17.2.1 Holonomy
Let (M n , g) be a connected Riemannian n-manifold.
17.2.1.1 Parallel transport
To g, one associates its Levi-Civita connection ∇, which defines, for a piecewise-
C 1 curve γ : [0, 1] → M , a parallel transport
Pγ∇ : Tγ(0) M → Tγ(1) M, (17.1)
which is a linear g-isometry between the two tangent spaces.
17.2.1.2 Group structure
In Schouten (1918), he considered the set
 
Hx = Pγ∇ γ(0) = γ(1) = x ⊆ O(Tx M ) (17.2)
and called its dimension the number of degrees of freedom of g.
It is easy to establish the identities
 −1
Pγ̄∇ = Pγ∇ and Pγ∇2 ∗γ1 = Pγ∇2 ◦ Pγ∇1 (17.3)
where γ̄ is the reverse of γ and γ2 ∗γ1 is the concatenation of paths γ1 and γ2
satisfying γ1 (1) = γ2 (0).
Consequently, Hx ⊂ O(Tx M ) is a subgroup and
 −1
Hγ(1) = Pγ∇ Hγ(0) Pγ∇ . (17.4)
In particular, fixing a linear isometry u : Tx M → Rn , the conjugacy class
of Hu = uHx u−1 in O(n) is well-defined, independent of the choice of x ∈ M
or the isometry u : Tx M → Rn . By abuse of terminology, we say that H is the
holonomy of the metric g if H ⊂ O(n) is a group conjugate to some (and hence
any) of the groups Hu .
For later reference, if u : Tx M → Rn is fixed, we let
 
Bu = u ◦ Pγ∇ γ(1) = x . (17.5)
This Bu is an Hu -subbundle of the orthonormal coframe bundle of g, that is, it is
an Hu -structure on M . By its very construction, it is invariant under ∇-parallel
translation and, since ∇ is torsion-free, it follows that this Hu -structure admits
a torsion-free compatible connection.
350 Non-embedding and non-extension results in special holonomy

Conversely, let H ⊂ O(n) be a subgroup and let B → M be an H-structure

on M . If B → M admits a compatible, torsion-free connection ∇, then B is
said to be torsion-free. In this case, ∇ (necessarily unique) is the Levi-Civita
connection of the underlying metric g on M and B is invariant under ∇-parallel
translation, implying that Hu ⊂ H for all u ∈ B. Thus, finding torsion-free
H-structures on M provides a way to find metrics on M whose holonomy lies
in H.
In this chapter, I will use the term H-manifold to denote an n-manifold M n
endowed with a torsion-free H-structure B → M . (The intended embedding H ⊂
O(n) is to be understood from context.)
17.2.1.3 Cartan’s early results
In Cartan (1925), he made the following statements:
1. Hx is a Lie subgroup of O(Tx M ), connected if M is simply connected.
2. If Hx acts reducibly on Tx M , then g is locally a product metric.
Cartan’s first statement was eventually proved (to modern standards of rigor)
by Borel and Lichnerowicz (1952) and the second statement was globalized and
proved by G. de Rham.

17.2.2 A non-trivial case

In dimensions 2 and 3, the above facts suffice to determine the possible holonomy
groups of simply connected manifolds.
Cartan (1925) 4 studied the first non-trivial case, namely, n = 4 and Hx 
SU (2), and observed that such metrics g
1. Have vanishing Ricci tensor
2. Are what we now call “self-dual”
3. Locally (modulo diffeomorphism) depend on two functions of three variables
While the first two observations are matters of calculation and/or definition,
the third observation is non-trivial. However, Cartan gave no indication of his
proof and, to my knowledge, never returned to this example again.
While I cannot be sure, I believe that it is likely that Cartan had a proof in
mind along the following lines. 5
The associated SU (2)-structure B → M of such a metric g satisfies structure
equations of the form
dω = − θ ∧ ω,
 
dθ = − θ ∧ θ + R ω ∧ ω , (17.6)
dR = − θ.R + R (ω),
4 Especially note chapter VII, section II.
5 He had developed all of the tools necessary for this proof in his famous series of papers
on pseudo-groups and the equivalence problem and, using those results, it would have been a
simple observation for him.
 Beginnings 351

where
1. The tautological one-form ω takes values in R4
2. The connection one-form θ takes values in su(2) ⊂ so(4)
3. The curvature function R takes values in W4 , the five-dimensional
  (real)
irreducible representation of SU (2) that lies in Hom Λ2 (R4 ), su(2)
4. The derived curvature function R takes values in V5 , the six-dimensional
complex irreducible representation of SU (2) that lies in Hom(R4 , W4 )
Calculation shows that the subspace V5 is an involutive tableau in Hom(R4 , W4 ),
with character sequence (s1 , s2 , s3 , s4 ) = (5, 5, 2, 0). Since the last non-zero char-
acter of this tableau is s3 = 2, Cartan’s generalization of the third fundamental
theorem of Lie applies to the structure equation (17.6) to yield his third obser-
vation above.

17.2.3 Hyper-Kähler viewpoint

Riemannian manifolds (M 4 , g) with
Hx  SU (2) ⊂ SO(4)
are nowadays said to be hyper-Kähler and we understand them as special cases
of Kähler manifolds. In fact, using our understanding of complex and Kähler
geometry, we now arrive at Cartan’s result by a somewhat different route.
Because the subgroup SU (2) ⊂ SO(4) acts trivially on the space of self-dual
two-forms on R4 , when (M 4 , g) has Hx  SU (2), then there exist three g-parallel
(and, hence, closed) self-dual two-forms on M , say Υ1 , Υ2 , and Υ3 , such that
Υi ∧ Υj = 2δij dVg . (17.7)
Conversely, a triple Υ = (Υ1 , Υ2 , Υ3 ) of closed two-forms on M that satisfies
Υ1 2 = Υ2 2 = Υ3 2 = 0 while Υ2 ∧ Υ3 = Υ3 ∧ Υ1 = Υ1 ∧ Υ2 = 0 (17.8)
is easily shown to be g-parallel and self-dual with respect to a unique metric g
on M for which (17.7) holds.
Suppose given a triple Υ = (Υ1 , Υ2 , Υ3 ) of two-forms on M satisfying (17.8).
The complex-valued two-form Ω1 = Υ2 + i Υ3 satisfies Ω1 2 = 0 because of (17.8)
and hence is of type (2, 0) with respect to a unique almost-complex structure J1
on M . Again appealing to (17.8), one sees that Υ1 ∧Ω1 = 0, implying that  the two-
form Υ1 is of type (1, 1) with respect to J1 . Moreover, since Υ21 = 12 Ω1 ∧ Ω1 ,
the two-form Υ1 is either J1 -positive or J1 -negative. In the former case, one says
that Υ satisfying (17.8) is positive, in the latter case, one says that Υ is negative.
Clearly, by replacing Υ1 by −Υ1 , one can convert any positive Υ into a negative Υ
and vice versa.
Thus, a positive triple Υ defines a unique SU (2)-structure BΥ → M for which
the component forms Υi are a positive, orthonormal basis for the self-dual two-
forms preserved by the SU (2)-structure.
352 Non-embedding and non-extension results in special holonomy

Now, given a positive triple (Υ1 , Υ2 , Υ3 ) of closed two-forms on M satisfy-

ing (17.8), one can prove (using the Newlander–Nirenberg theorem) that each
point of M lies in a local coordinate chart z = (z 1 , z 2 ) : U → C2 for which there
exists a real-analytic function φ : z(U ) → R so that

Υ2 + i Υ3 = dz 1 ∧ dz 2 and ¯
Υ1 = 12 i ∂ ∂φ, (17.9)

where φ satisfies the elliptic Monge–Ampère equation


∂2φ
det =1 (17.10)
∂z i ∂ z̄ j

and the strict pseudoconvexity condition with respect to z given by


∂2φ
> 0. (17.11)
∂z i ∂ z̄ j

In fact, if one fixes an open set U ⊂ C2 and chooses a smooth, strictly

pseudoconvex function φ : U → R satisfying (17.10), then the formulae (17.9)
define a positive triple (Υ1 , Υ2 , Υ3 ) on U consisting of gφ -parallel, self-dual two-
forms where

∂2φ
gφ = dz i ◦ dz̄ j . (17.12)
∂z i ∂ z̄ j

Thus, the holonomy of gφ is (conjugate to) a subgroup of SU (2) and it is not

difficult to show that, for a generic strictly pseudoconvex φ satisfying (17.10),
the holonomy of gφ is (conjugate to) the full SU (2).
Note that, because (17.10) is an analytic elliptic PDE at a strictly pseudo-
convex solution φ, all of its strictly pseudoconvex solutions are real-analytic. In
particular, a Riemannian metric g on M 4 with holonomy SU (2) is real-analytic
in harmonic coordinates (as would have followed anyway from its Ricci-flatness
and a result of DeTurck and Kazdan 1981).
Now, one does not normally think of solving an elliptic PDE by an initial
value problem, but, of course, in the analytic category, there is nothing wrong
with such a procedure. Indeed, the Cauchy–Kovalewskaya theorem implies that,
in this particular case, one can specify φ and its normal derivative along a
hypersurface, say, Im(z 2 ) = 0, as essentially arbitrary real-analytic functions
(subject only to an open condition that guarantees the strict pseudoconvexity
of the resulting solution) and thereby determine a unique strictly pseudoconvex
solution of (17.10). Thus, in this sense, one sees, again, that the “general” metric
with holonomy in SU (2) modulo diffeomorphism depends on two functions of
three variables, in agreement with Cartan’s claim.
 Hyper-Kähler four-manifolds 353

17.3 Hyper-Kähler four-manifolds

17.3.1 An exterior differential systems proof
One can use the characterization of SU (2) as the stabilizer of three two-forms
in dimension 4 as the basis of another analysis of the existence problem via an
EDS (exterior differential system). 6
Let M 4 be an analytic manifold and let Υ be the tautological two-form
on Λ2 (T ∗ M ). Let
 3
X 17 ⊂ Λ2 (T ∗ M ) (17.13)
be the submanifold consisting of positive triples (β1 , β2 , β3 ) ∈ Λ2 (Tx∗ M ) such
that
β1 2 = β2 2 = β3 2 = 0 and β1 ∧ β2 = β3 ∧ β1 = β2 ∧ β3 = 0. (17.14)
∗
Let πi : X → Λ (T M ) for 1 ≤ i ≤ 3 denote the projections onto the three
2

factors. The pullbacks Υi = πi∗ (Υ) define an EDS on X

I = {dΥ1 , dΥ2 , dΥ3 }. (17.15)
The basepoint projection π : X → M makes X into a bundle over M
whose fibers are diffeomorphic to the thirteen-dimensional homogeneous
space GL(4, R)/SU (2) and whose sections σ : M → X correspond to the SU (2)-
structures on M .
An I-integral manifold Y 4 ⊂ X transverse to π : X → M then represents a
choice of three closed two-forms Υi on an open subset U ⊂ M that satisfy the
algebra conditions needed to define an SU (2)-structure on U .
Calculation shows that I is involutive, with character sequence
(s0 , s1 , s2 , s3 , s4 ) = (0, 0, 3, 6, 4). (17.16)
In particular, a three-dimensional real-analytic I-integral manifold P ⊂ X that
is transverse to the fibers of π can be “thickened” to a four-dimensional integral
manifold that is transverse to the fibers of π. This “thickening” will not be
unique, however, because of the invariance of the ideal I under the obvious
action induced by the diffeomorphisms of M .

17.3.2 A sharper result

Suppose that (M 4 , g) has holonomy SU (2) and let Υi be three g-parallel two-
forms on M satisfying
Υi ∧ Υj = 2δij dVg (17.17)
such that Υ = (Υ1 , Υ2 , Υ3 ) is positive.
6 The argument to be given in this subsection for SU (2) ⊂ SO(4) is the analog of the

arguments given in Bryant (1987) for the cases G2 ⊂ SO(7) and Spin(7) ⊂ SO(8). I omit the
calculations in this case, since they are straightforward.
354 Non-embedding and non-extension results in special holonomy

If N 3 ⊂ M is an oriented hypersurface, with oriented normal n, then there is

a coframing η of N defined by
⎛ ⎞ ⎛ ⎞
η1 n Υ1
η = ⎝η2 ⎠ = ⎝n Υ2 ⎠ (17.18)
η3 n Υ3
and (because of the positivity condition) it satisfies
⎛ ⎞ ⎛ ⎞
Υ1 η2 ∧η3
N ∗ ⎝Υ2 ⎠ = ⎝η3 ∧η1 ⎠ = ∗η η. (17.19)
Υ3 η1 ∧η2
where ∗η is the Hodge star associated to the metric gη = η1 2 + η2 2 + η3 2 and
orientation η1 ∧η2 ∧η3 > 0.
In particular, note that the coframing η is not arbitrary, but satisfies the
system of three first-order PDE:
d(∗η η) = N ∗ dΥ = 0. (17.20)
An alternative manifestation of the involutivity of the system I that is better
adapted to the initial value problem then becomes the following existence and
uniqueness result:
Theorem 17.1 Let η be a real-analytic coframing of N such that d(∗η η) = 0.
There exists an essentially unique embedding of N into a SU (2)-holonomy
manifold (M 4 , g) that induces the given coframing η in the above manner.
Remark 17.1 (Essential uniqueness) The meaning of this term is as follows:
If N can be embedded into two different SU (2)-manifolds (M, Υ) and (M̄ , Ῡ) in
such a way that both embeddings induce the same coframing η on N by the above
pullback formula, then there are open neighborhoods U ⊂ M and Ū ⊂ M̄ of the
respective images of N in the two manifolds and a diffeomorphism f : U → Ū
that is the identity on the image of N and that satisfies f ∗ Ῡ = Υ.
Proof. Write dη = −θ∧η where θ = −t θ. (This θ exists and is unique by the
fundamental lemma of Riemannian geometry.)
On N × GL(3, R) define 7
ω = g −1 η and γ = g −1 dg + g −1 θg, (17.21)
so that dω = −γ ∧ω. On X = R × N × GL(3, R) define the three two-forms
⎛ ⎞ ⎛ ⎞
Υ1 dt∧ω1 + ω2 ∧ω3
⎝Υ2 ⎠ = ⎝dt∧ω2 + ω3 ∧ω1 ⎠ = dt ∧ ω + ∗ω ω. (17.22)
Υ3 dt∧ω3 + ω1 ∧ω2
7 In the following formulae, I regard g : N × GL(3, R) → GL(3, R) as the projection onto

the second factor. Also, to save writing, I will write η for π1∗ η where π1 : N × GL(3, R) → N
is the projection onto the first factor.
 Hyper-Kähler four-manifolds 355

Let I be the ideal on X generated by {dΥ1 , dΥ2 , dΥ3 }. One calculates that
 
dΥ = t γ − (tr γ)I3 ∧ ∗ω ω + γ ∧ ω ∧ dt. (17.23)
Consequently, I is involutive, with characters (s0 , s1 , s2 , s3 , s4 ) = (0, 0, 3, 6, 0).
Since d(∗η η) = 0, the locus L = {0} × N × {I3 } ⊂ X is a regular, real-analytic
integral manifold of the real-analytic ideal I. (Note that L is just a copy of N .)
By the Cartan–Kähler theorem, L lies in a unique four-dimensional I-integral
manifold M ⊂ X. The Υi thus pull back to M to be closed and to define the
desired SU (2)-structure forms Υi on M inducing η on L = N .

It is natural to ask whether it is necessary to assume that η be real-analytic

for the conclusion of Theorem 17.1. The following result shows that one cannot
weaken this assumption to “smooth” and still get the same conclusion:

Theorem 17.2 If η is a coframing on N 3 that is not real-analytic in any local

coordinate system and
d(∗η η) = 0 and ∗η (t η ∧ dη) = 2C (17.24)
for some constant C, then η cannot be induced by immersing N into an SU (2)-
manifold.
Smooth-but-not-real-analytic coframings η satisfying (17.24) do exist locally.

Proof. Suppose that Υ = (Υ1 , Υ2 , Υ3 ) is a positive triple of parallel, self-dual

two-forms on an (M 4 , g) with holonomy SU (2) and let N 3 ⊂ M be an oriented
hypersurface.
Calculation yields that the induced co-closed coframing η on N satisfies
∗η (t η ∧ dη) = 6H (17.25)
where H is the mean curvature of N in M .
Now, by DeTurck and Kazdan (1981), since g is Ricci-flat, it is real-analytic
in g-harmonic coordinates. In particular, such coordinate systems can be used
to define a real-analytic structure on M , which is the one that will be meant
henceforth. In particular, since the forms Υi are g-parallel, they, too, are real-
analytic with respect to this structure. If H is constant, then elliptic regularity
implies that N must be a real-analytic hypersurface in M and hence η must also
be real-analytic.
Thus, if η is a non-real-analytic coframing on N 3 that satisfies (17.24) for
some constant C, then η cannot be induced on N by an embedding into an
SU (2)-manifold.
To finish the proof, I will now show how to construct a coframing η on an
open subset of R3 that is not real-analytic in any coordinate system and yet
satisfies (17.24).
To begin, note that, if a coframing η on N 3 is real-analytic in any coordinate
system at all, it will be real-analytic in η-harmonic coordinates, that is, local
356 Non-embedding and non-extension results in special holonomy

coordinates x : U → R3 satisfying
 
d ∗η dx = 0. (17.26)
Now, fix a constant C and consider a coframing η = h(x)−1 dx on U ⊂ R3
where h : U → GL(3, R) is a mapping satisfying the first-order, quasi-linear
system:
 
d(∗η η) = 0, ∗η (t η ∧ dη) = 2C, and d ∗η dx = 0. (17.27)
The system (17.27) consists of seven equations for the nine unknown entries of h.
Calculation shows this first-order system to be underdetermined elliptic (i.e.
its symbol is surjective at every real covector ξ). By standard theory, it has
smooth local solutions that are not real-analytic.
Taking a non-real-analytic solution h, the resulting η will not be real-analytic
in the x-coordinates, which, by construction, are η-harmonic. Thus, such an η is
not real-analytic in any local coordinate system.

Remark 17.2 (“Flow” interpretation) The condition d(dt∧ω + ∗ω) = 0
has sometimes been described as an “SU (2)-flow” on coframings of N . In fact,
this closure condition can be written in the “evolutionary” form
d
ω = ∗ω (dω) − 1
2 ∗ω (t ω ∧ dω) ω. (17.28)
dt
By Theorem 17.1, if η on N 3 is real-analytic and satisfies d(∗η η) = 0, then (17.28)
has a solution in a neighborhood of t = 0 in R × N that satisfies the initial
condition
ω t=0 = η. (17.29)
One does not normally think of evolution equations as having to have real-
analytic initial data. However, Theorem 17.2 shows that some such regularity
assumption must be made. 8
Remark 17.3 (Co-closed coframings with specified metric) Given a
Riemannian three-manifold (N, g), it is an interesting question as to when there
exist (either locally or globally) a g-orthonormal coframing η = (η1 , η2 , η3 ) that
is co-closed. 9
One can formulate this as an EDS for the section η : N → B of the g-
orthonormal coframe bundle B → N . The natural EDS for this is not involutive,
but a slight extension is and is worth describing here.
8 On the other hand, the reader should be aware that real-analyticity, while sufficient, is

certainly not necessary for the initial-value problem to have a (local) solution. For example, it
suffices to let N be a smooth hypersurface in an SU (2)-manifold M 4 that is not real-analytic
with respect to the real-analytic structure provided by the holomorphic coordinates of the
underlying integrable complex structure. The induced co-closed coframing on N will then be
non-real-analytic initial data for which the initial-value problem will have a (local) solution.
9 Of course, closed g-orthonormal coframing exist locally if and only if g is flat.
 Hyper-Kähler four-manifolds 357

Let ωi and ωij = −ωji (where 1 ≤ i, j ≤ 3) be the tautological and Levi-Civita

connection forms on B. In particular, g pulls back to B to be ω1 2 +ω2 2 +ω3 2 and
these forms satisfy the structure equations:
dωi = −ωij ∧ ωj and dωij = −ωik ∧ ωkj + 12 Rijkl ωk ∧ ωl . (17.30)
Because of their tautological reproducing property, one has η ∗ (ωi ) = ηi . Con-
sequently, the g-orthonormal coframing η is co-closed if and only if, when
regarded as a section η : N → B, it is an integral manifold of the EDS I0
generated by the three closed three-forms Υij = d(ωi ∧ωj ) with 1 ≤ i < j ≤ 3.
Computation shows that this ideal has (s1 , s2 , s3 ) = (0, 2, 1), while an integral
manifold on which Ω = ω1 ∧ω2 ∧ω3 is non-vanishing (an obvious requirement for
a section η(M ) ⊂ B) must satisfy
η ∗ ωij = ijk Skl η ∗ ωl (17.31)
where ijk is fully antisymmetric in its indices, with 123 = 1, and Skl = Slk . In
particular, the space of admissible integral elements of (I0 , Ω) at each point of B
has dimension 6 < s1 +2s2 +3s3 . Hence, the system (I0 , Ω) is not involutive.
However, (17.31) show that any integral of (I0 , Ω) is also an integral of the
two-form
1
Υ = ω1 ∧ ω23 + ω2 ∧ ω31 + ω3 ∧ ω12 = 2 ijk ωi ∧ ωjk . (17.32)
Moreover, since
d(ωi ∧ ωj ) = ijk ωk ∧ Υ, (17.33)
the differential ideal I generated by Υ contains I0 . Calculation using (17.30)
yields
2 dΥ = ijk ωi ∧ ωlj ∧ ωlk − R Ω (17.34)
where R is the scalar curvature of the metric g. Thus, the ideal I is generated
algebraically by Υ and dΥ. An integral element of (I, Ω) is now cut out by
equations of the form
ωij − ijk Skl ωl = 0 (17.35)
where S = (Skl ) is a 3 × 3 symmetric matrix satisfying σ2 (S) = 12 R (where σ2 (S)
is the second elementary function of the eigenvalues of S). The characteristic
variety of such an integral element consists of the null (co-)vectors of the
quadratic form
QS = tr(S) g − Sij ωi ωj . (17.36)
In particular, except for the case S = 0 (which can only occur when R = 0),
this integral element is Cartan-regular, with Cartan character sequence
(s0 , s1 , s2 , s3 ) = (0, 1, 2, 0). In particular, the system (I, Ω) is involutive, and,
358 Non-embedding and non-extension results in special holonomy

in the real-analytic case, the general integral depends on two functions of two
variables.
Note, by the way, that when QS is positive (or negative) definite, the
linearization around such a solution is elliptic and hence such co-closed
coframings are as regular as the metric g. In particular, this always happens
when R > 0, that is, when the scalar curvature is positive. 10 Thus, for a real-
analytic metric with positive scalar curvature, all of its co-closed coframings are
real-analytic.

17.4 G2 -Manifolds
For background on the group G2 ⊂ SO(7) and G2 -manifolds, the reader can
consult Bryant (1987), Salamon (1989), Joyce (2000), and Bryant (2006). I will
generally follow the notation in Bryant (2006).
The crucial point is that the group G2 can be defined as the stabilizer
in GL(7, R) of the three-form

φ = e123 + e145 + e167 + e246 − e257 − e347 − e356 . (17.37)

where eijk = ei ∧ej ∧ek and the ei are a basis of linear forms on R7 . The Lie
group G2 is connected, has dimension 14, preserves the metric and orientation
for which the ei are an oriented orthonormal basis, and acts transitively on the
unit sphere S 6 ⊂ R7 . The G2 -stabilizer of e1 is the subgroup SU (3) ⊂ SO(6)
that preserves the two-form e23 + e45 + e67 and the three-form
 
e246 − e257 − e347 − e356 = Re (e2 + i e3 ) ∧ (e4 + i e5 ) ∧ (e6 + i e7 ) . (17.38)

The GL(7, R)-orbit of φ in Λ3 (R7 ) is an open (but not convex) cone Λ3+ (R7 ) ⊂
Λ (R7 ) that consists precisely of the three-forms on R7 whose stabilizers are
3

isomorphic to G2 . Consequently, for any smooth seven-manifold M 7 , there is

a well-defined open subset Λ3+ (T ∗ M ) ⊂ Λ3 (T ∗ M ) consisting of the three-forms
whose stabilizers are isomorphic to G2 .
A G2 -structure on M is thus specified by a three-form σ on M with the
property that σx lies in Λ3+ (Tx∗ M ) for all x ∈ M . Such a three-form σ will be said
to be definite. Explicitly, the G2 -structure B = Bσ consists of the linear isomor-
phisms u : Tx → R7 that satisfy u∗ φ = σx . Conversely, given a G2 -structure B →
M , there is a unique definite three-form σ on M such that B = Bσ . Put another
way, the G2 -structure bundle F(M )/G2 is naturally identified with Λ3+ (M ) by
identifying [u] = u · G2 with u∗ φ ∈ Λ3+ (Tπ(u)
∗
M ) for all u ∈ F(M ).
A σ ∈ Ω+ (M ) determines a unique metric gσ and orientation ∗σ by requiring
3

that the corresponding coframings u ∈ Bσ be oriented isometries.

10 It is interesting to note that, when R > 0, the ideal I on the six-manifold B is algebraically

equivalent at each point to the special Lagrangian ideal on C3 .

 G2 -Manifolds 359

It is a fact (Bryant 1987) that Bσ is torsion-free if and only if σ is gσ -parallel,

which, in turn, holds if and only if

dσ = 0 and d(∗σ σ) = 0. (17.39)

Thus, a G2 -manifold can be regarded as a pair (M 7 , σ) where σ ∈ Ω3+ (M ) satisfies

the non-linear system of PDE (17.39).
By a theorem of Bonan (see Besse 1987, chapter X), for any G2 -
manifold (M, σ), the associated metric gσ has vanishing Ricci tensor. In par-
ticular, by a result of DeTurck and Kazdan (1981), the metric gσ is real-analytic
in gσ -harmonic coordinates. Since σ is gσ -parallel, it, too, must be real-analytic
in gσ -harmonic coordinates.
There is a natural differential ideal on Λ3+ (T ∗ M ) = F (M )/G2 defined as
follows. For [u] ∈ F(M )/G2 , define
 
σ[u] = π ∗ (u∗ φ) and τ[u] = π ∗ u∗ (∗φ φ) , (17.40)

where π : F(M )/G2 → M is the natural basepoint projection. Let I be the dif-
ferential ideal on F (M )/G2 = Λ3+ (T ∗ M ) generated by dσ and dτ. The following
result is proved in Bryant (1987):

Theorem 17.3 The ideal I on Λ3+ (T ∗ M ) is involutive. A section σ ∈ Ω3+ (M )

is an integral of I only if it is gσ -parallel. Modulo diffeomorphisms, the general
I-integral σ depends on 6 functions of 6 variables.

17.4.1 Hypersurfaces
The group G2 acts transitively on S 6 ⊂ R7 , with stabilizer SU (3). Hence, an
oriented N 6 ⊂ M inherits a canonical SU (3)-structure, which is determined by
the (1, 1)-form ω and (3, 0)-form Ω = φ + i ψ defined by

ω=n σ and Ω = φ + i ψ = N ∗ σ − i (n ∗σ σ). (17.41)

In fact, if one defines f : R × N → M by

 
f (t, p) = expp t n(p) , (17.42)

then

f ∗ σ = dt ∧ ω + Re(Ω) and f ∗ (∗σ σ) = 1

2
ω 2 − dt ∧ Im(Ω), (17.43)

where, now, ω and Ω are forms on N that depend on t.

For each fixed t = t0 , the induced SU (3)-structure on N satisfies
 
d Re(Ω) = d(ft∗0 σ) = 0 and d( 12 ω 2 ) = d ft∗0 (∗σ σ) = 0, (17.44)

so these are necessary conditions on the SU (3)-structure on N that it be induced

by immersion into a G2 -holonomy manifold M .
360 Non-embedding and non-extension results in special holonomy

Theorem 17.4 A real-analytic SU (3)-structure on N 6 is induced by embedding

into a G2 -manifold if and only if its defining forms ω and Ω satisfy
 
d Re(Ω) = 0 and d 12 ω 2 = 0. (17.45)
Proof. The necessity of (17.45) has already been demonstrated, so I will just
prove the sufficiency.
Define a tautological two-form ω and three-form Ω on F(N )/SU (3) as
follows. For a coframe u : Tx N → C3 , define these forms at [u] = u · SU (3) ∈
F(N )/SU (3) by
   
ω[u] = π ∗ u∗ ( 2i (t dz ∧ dz̄)) and Ω[u] = π ∗ u∗ (dz 1 ∧ dz 2 ∧ dz 3 ) (17.46)
where π : F(N )/SU (3) → N is the basepoint projection.
On X = R × F(N )/SU (3), consider the three-form and four-form defined by
σ = dt ∧ ω + Re(Ω)
(17.47)
τ= 1
2 ω2 − dt ∧ Im(Ω).
Let I be the EDS generated by the closed four-form dσ and five-form dτ. Then
the calculation used to prove Theorem 17.3 (see Bryant 1987) shows that I is
involutive, with characters
(s0 , s1 , . . . , s7 ) = (0, 0, 0, 1, 4, 10, 13, 0). (17.48)
   1 2
Since d Re(Ω) = d 2 ω = 0, the given SU (3)-structure on N defines a
regular integral manifold L ⊂ X of I lying in the hypersurface t = 0.
Since the given SU (3)-structure is assumed to be real-analytic, the system I
and the integral manifold L ⊂ X are real-analytic by constuction. Hence the
Cartan–Kähler theorem can be applied to conclude that L lies in a unique I-
integral M 7 ⊂ X. The pullback of σ to M is then a closed definite three-form σ
on M while the pullback of τ to M is closed and equal to ∗σ σ. Thus, (M, σ) is
a G2 -manifold. By construction, N is imbedded into M as the locus t = 0 and
the SU (3)-structure on N induced by σ is the given one.

Theorem 17.5 There exist non-real-analytic SU (3)-structures on N 6 whose
associated forms (ω, Ω) satisfy (17.45) but that are not induced from an immer-
sion into a G2 -manifold (M, σ).
In fact, if a non-analytic SU (3)-structure satisfies (17.45) and
  
∗ ω ∧ d Im(Ω) = C (17.49)
for some constant C, then it cannot be G2 -immersed.
Non-analytic SU (3)-structures satisfying (17.45) and (17.49) do exist.
Proof. When an SU (3)-structure on N 6 with defining forms (ω, Ω) is induced
via a G2 -immersion N 6 → M 7 , the mean curvature H of N in M is given by
  
−12H = ∗ ω ∧ d Im(Ω) . (17.50)
 G2 -Manifolds 361

Thus, when the right-hand side of this equation is constant, it follows by elliptic
regularity that N 6 is a real-analytic submanifold of the real-analytic (M 7 , σ).
Thus, if the given SU (3)-structure on N satisfying (17.45) and (17.49) is
not real-analytic, it cannot be induced by an embedding into a G2 -holonomy
manifold M .
It remains to construct a non-analytic example satisfying (17.45)
 and (17.49).
Here is why it is somewhat delicate: Since dim GL(6, R)/SU (3) = 28, a choice
of an SU (3)-structure (ω, Ω) on N 6 depends on 28 functions of 6 variables.
Modulo diffeomorphisms, this leaves 22 functions of 6 variables. On the other
hand, the equations
    
d Re(Ω) = 0, d( 12 ω 2 ) = 0, and ∗ ω ∧ d Im(Ω) = C (17.51)
constitute 15 + 6 + 1 = 22 equations for the SU (3)-structure.
Thus, the equations to be solved are “more determined” than in the analogous
SU (2) case. Nevertheless, their diffeomorphism invariance still allows one to
construct the desired example, as will now be shown.
Say that a three-form
 1 φ ∈ Ω3 (N 6 ) is elliptic if, at each point, it is linearly
equivalent to Re dz ∧dz ∧dz 3 . This is a open pointwise condition on φ (i.e.
2

it is stable in Hitchin’s sense 2001): The elliptic three-forms are sections of an

open subbundle Λ3e (T ∗ N ) ⊂ Λ3 (T ∗ N ). I will denote the set of elliptic three-forms
on N by Ω3e (N ).
Fix an orientation of N 6 . An elliptic φ ∈ Ω3e (N ) then defines a unique,
orientation-preserving almost-complex structure Jφ on N 6 such that
Ωφ = φ + i Jφ∗ (φ) (17.52)
is of Jφ -type (3, 0).
Now assume that φ ∈ Ω3e (N ) is closed. Then dΩφ is purely imaginary and
yet must be a sum of terms of Jφ -type (3, 1) and (2, 2). Thus, dΩφ is purely of
Jφ -type (2, 2).
Let Λ1,1
+ (N, Jφ ) denote the set of real two-forms that are of Jφ -type (1, 1) and
that are positive on all Jφ -complex lines. The squaring map
σ : Λ1,1
+ (N, Jφ ) → Λ
2,2
(N, Jφ )
given by σ(ω) = ω2 is a diffeomorphism onto the open set Λ2,2 + (N, Jφ ) ⊂
Λ2,2 (N, Jφ ) that consists of the real four-forms of Jφ -type (2, 2) that are positive
on all Jφ -complex two-planes.
Now fix a constant C = 0. One sees from the above discussion that it is a
C 1 -open condition on φ that
dΩφ = 6i C (ωφ )2 for some ωφ = ωφ ∈ Ω1,1
+ (N, Jφ ). (17.53)
Now, the pair (ωφ , Ωφ ) are the defining forms of an SU (3)-structure on N if
and only if
1
6 (ωφ )
3
− 18 i Ωφ ∧ Ωφ = 0. (17.54)
362 Non-embedding and non-extension results in special holonomy

This is a single, first-order scalar equation on the closed three-form φ. It is easy

to see that there are non-analytic solutions. (For example, if one starts with
the standard structure induced on the six-sphere in flat R7 endowed with its
flat G2 -structure, then small perturbations of the corresponding closed φ can be
made that solve (17.54) but for which the induced SU (3)-structure has constant
curvature on a proper open subset of S 6 . Such an SU (3)-structure clearly cannot
be real-analytic everywhere.)
Assuming (17.54) is satisfied, we have
d(Re Ωφ ) = dφ = 0, (17.55)
and
1   
d 2 (ωφ )
2
= d −3i C1 dΩφ = 0, (17.56)
and finally,
   
∗φ ωφ ∧ d(Im Ωφ ) = ∗φ ωφ ∧ 16 C (ωφ )2 = C. (17.57)

17.4.2 Flow interpretation

On N 6 × R, with (ω, Ω) defining an SU (3)-structure on N 6 depending on t ∈ R,
consider the equations
   
d dt ∧ ω + Re(Ω) = 0 and d 12 ω 2 − dt ∧ Im(Ω) = 0, (17.58)
which assert that σ = dt∧ω + Re(Ω), which is a definite three-form on M =
N 6 × R, is both closed and co-closed (since ∗σ σ = 12 ω 2 − dt∧ Im(Ω)).
Think of Ω as φ + iJφ∗ (φ), so that the SU (3)-structure is determined by (ω, φ)
where φ = Re(Ω). The conditions (17.58) for fixed t are then
dφ = 0 and d(ω 2 ) = 0, (17.59)
while the G2 -evolution equations implied by (17.58) for such (ω, φ) are then
d d   
(φ) = dω and (ω) = −Lω −1 d Jφ∗ (φ) , (17.60)
dt dt
where Lω : Ω2 (N ) → Ω4 (N ) is the invertible map Lω (β) = ω ∧β.
Theorems 17.4 and 17.5 show that solutions to the “G2 -flow” (17.60) do exist
for analytic initial SU (3)-structures satisfying the closure conditions (17.59), but
need not exist for non-analytic initial SU (3)-structures satisfying these closure
conditions.

17.5 Spin(7)-manifolds
For background on the group Spin(7) ⊂ SO(8) and Spin(7)-manifolds, the reader
can consult Bryant (1987), Salamon (1989), and Joyce (2000). I will generally
follow the notation in Bryant (1987).
 Spin(7)-manifolds 363

The main point is that the group Spin(7) ⊂ SO(8) is the GL(8, R)-stabilizer
of the four-form Φ0 ∈ Λ4 (R8 ), defined by

Φ0 = e0 ∧ φ + ∗φ φ (17.61)

where φ is defined by (17.37) and e0 , . . . , e7 is a basis of linear forms on R8 =

R ⊕ R7 . The group Spin(7) is a connected Lie group of dimension 21 that
is the double cover of SO(7). It acts transitively on the unit sphere in R8
and the Spin(7)-stabilizer of e0 is G2 . The GL(8, R)-orbit of Φ0 in Λ4 (R8 )
will be denoted by Λ4s (R8 ). This orbit has dimension 43, so it is not open
in Λ4 (R8 ).
A Spin(7)-structure on M 8 is thus specified by a four-form Φ ∈ Ω4 (M ) that
is linearly equivalent to Φ0 at each point of M . The set of such four-forms will
be denoted Ω4s (M ). They are the sections of the bundle F (M )/Spin(7) → M ,
which has a natural embedding into Λ4 (T ∗ M ).
Given a four-form Φ ∈ Ω4s (M ), the corresponding Spin(7)-structure B = BΦ
is defined to be the set of coframings u : Tx M → R8 that satisfy u∗ Φ0 = Φx .
Conversely, every Spin(7)-structure B → M is of the form B = BΦ for a unique
four-form Φ ∈ Ω4s (M ).
In particular, each Φ ∈ Ω4s (M ) determines a metric gΦ and orientation ∗Φ
on M by requiring that the elements u ∈ BΦ be oriented isomorphisms.
It is a fact (Bryant 1987) that BΦ is torsion-free if and only if Φ is gΦ -parallel,
which, in turn, holds if and only if

dΦ = 0. (17.62)

Thus, a Spin(7)-manifold can be regarded as a pair (M 8 , Φ) where Φ ∈ Ω4s (M )

satisfies the non-linear system of PDE (17.62).
By a theorem of Bonan (see Besse 1987, chapter X), for any Spin(7)-
manifold (M, Φ), the associated metric gΦ has vanishing Ricci tensor. In par-
ticular, by a result of DeTurck and Kazdan (1981), the metric gΦ is real-analytic
in gΦ -harmonic coordinates. Since Φ is gΦ -parallel, it, too, must be real-analytic
in gΦ -harmonic coordinates.
Define a four-form Φ on F(M )/Spin(7) by the following rule: For u : Tx M →
R8 and [u] = u·Spin(7), set
 
Φ[u] = π ∗ u∗ Φ0 (17.63)

where π : F(M ) → M is the basepoint projection. Let I be the ideal generated

by dΦ on F (M )/Spin(7). The following result is proved in Bryant (1987):

Theorem 17.6 A section Φ ∈ Ω4s (M ) satisfies (17.62) if and only if it is an

integral of I. The ideal I is involutive. Modulo diffeomorphisms, the general
I-integral Φ depends on 12 functions of 7 variables.
364 Non-embedding and non-extension results in special holonomy

17.5.1 Hypersurfaces
Spin(7) acts transitively on S 7 and the stabilizer of a point is G2 . An oriented
hypersurface N 7 ⊂ M 8 inherits a G2 -structure σ ∈ Ω3+ (M ) that is defined by the
rule
σ=n Φ (17.64)
where n is the oriented normal vector field along N . It also satisfies
∗σ σ = N ∗ Φ. (17.65)
The structure equations show that
 
∗σ σ ∧ dσ = 28H (17.66)
where H is the mean curvature of N in (M, gΦ ).
Theorem 17.7 If σ ∈ Ω3+ (N 7 ) is real-analytic and satisfies d(∗σ σ) = 0, then σ
is induced by an immersion of N into a Spin(7)-manifold (M, Φ).
Proof. The argument in this case is entirely analogous to the SU (2) and G2
cases already treated.
Recall the definitions of σ and τ on F(N 7 )/G2 and, on X = R × F(N )/G2 ,
define
Φ = dt ∧ σ + τ. (17.67)
Let I be the ideal on R × F(N )/G2 generated by dΦ. The same calculation
used to prove Theorem 17.6 (see Bryant 1987) then yields that I is involutive,
with character sequence
(s0 , s1 , . . . , s8 ) = (0, 0, 0, 0, 1, 4, 10, 20, 0). (17.68)
Since d(∗σ σ) = 0, the G2 -structure σ defines a regular I-integral L ⊂ X within
the locus t = 0.
Since σ is assumed to be real-analytic, it follows that X, I, and L are
real-analytic with respect to the obvious induced analytic structures on the
appropriate underlying manifolds. Thus, the Cartan–Kähler theorem applies to
show that L lies in a unique I-integral M 8 ⊂ X. The form Φ then pulls back
to M to be a closed Φ ∈ Ω4s (M ) which induces the given σ on N  L ⊂ M defined
by t = 0.

Theorem 17.8 There exist non-real-analytic G2 -structures σ ∈ Ω3+ (N 7 ) that
satisfy
d(∗σ σ) = 0 (17.69)
but that are not induced from a Spin(7)-immersion.
In fact, if a non-analytic G2 -structure σ satisfies (17.69) and
 
∗σ σ ∧ dσ = C (17.70)
 Spin(7)-manifolds 365

where C is a constant, then it cannot be Spin(7)-immersed.

Non-analytic G2 -structures σ ∈ Ω3+ (N 7 ) satisfying (17.69) and (17.70) do
exist.
Proof. The first claim follows from the second and third.
If σ ∈ Ω3+ (N 7
 ) is induced from at Spin(7)-immersion (N, σ) → (M , Φ) and
8

has ∗σ σ ∧dσ equal to a constant, then the hypersurface N ⊂ M has constant

mean curvature. Since (M, gΦ ) is real-analytic, constant mean curvature hyper-
surfaces in M are also real-analytic, so it follows that σ must be real-analytic.
This proves the second claim.
It remains now to verify the third claim by explaining how to construct non-
analytic G2 -structures σ satisfying (17.69) and (17.70). To save space, I will only
sketch the argument, the details of which are somewhat involved, though the idea
is the same as for SU (2):
If such a σ is to be real-analytic, it will have to be real-analytic in gσ -harmonic
coordinates. Now, the system of first-order equations
 
d(∗σ σ) = 0, ∗σ σ ∧ dσ = C, and d(∗σ dx) = 0
for σ ∈ Ω3+ (R7 ) is only 21 + 1 + 7 = 29 equations for 35 unknowns. This under-
determined system is not elliptic, but its symbol mapping has constant rank and
it can be embedded into an appropriate sequence to show that it has non-real-
analytic solutions.

17.5.2 Flow interpretation

Finally, let us consider the “flow” interpretation. If f : N → M is an ori-
ented smooth hypersurface in a Spin(7)-manifold (M, Φ) with oriented unit
normal n : N → T M , then the normal exponential mapping can be used to
embed a neighborhood U of N in M into R × N in such a way that, on this
neighborhood U , one can write
Φ = dt ∧ σ + ∗σ σ (17.71)
where σ ∈ Ω3+ (N )
now depends on t (in the case that N is noncompact, the
domain of σ might depend on t). The closure of Φ implies that
d  
∗σ σ = dσ. (17.72)
dt
This equation is enough to determine the time derivative of σ as well because
of the following observation. For any real vector space V of dimension 7, the
map S : Λ3+ (V ∗ ) → Λ4 (V ∗ ) defined by S(φ) = ∗φ φ is a 2-to-1 smooth local dif-
feomorphism of Λ3+ (V ∗ ) onto an open cone Λ4+ (V ∗ ) ⊂ Λ4 (V ∗ ). In fact, if S(φ) =
S(ψ), then φ = ±ψ. Both φ and −φ are definite, but they each determine opposite
orientations, that is, ∗−φ 1 = − ∗φ 1. In particular, if one fixes an orientation
on V , then for any τ ∈ Λ4+ (V ∗ ), there is a unique element φ ∈ Λ3+ (V ∗ ) such
that ∗φ φ = τ and ∗φ 1 is a positive volume form on V . I will denote this
366 Non-embedding and non-extension results in special holonomy

element by S−1 (τ ) = φ. Using this notation and the assumption that N is

oriented, the above equation can be written in the more obviously “evolutionary”
form
d    
τ = d S−1 (τ ) (17.73)
dt
where

Φ = dt ∧ S−1 (τ ) + τ (17.74)

with τ ∈ Ω4+ (N ) depending on t.

The content of Theorems 17.7 and 17.8 is then that (17.73) has a solution
when the initial τ0 is closed and real-analytic, but need not have a solution for
a non-analytic closed τ0 .

Acknowledgments
Thanks to Duke University for its support via a research grant and to the NSF
for its support via DMS-0604195. Finally, I would like to thank the referee
for a careful reading of the manuscript and for helpful suggestions regarding
references.

References
Besse, A. (1987). Einstein manifolds. Ergebnisse der Mathematik und ihrer Grenzgebiete
(3) [Results in Mathematics and Related Areas (3)]. 10. Springer-Verlag, Berlin,
Germany. MR 0867684 (88f:53087).
Borel, A. and Lichnerowicz, A. (1952). Groupes d’holonomie des variétés riemanniennes.
C. R. Acad. Sci. Paris 234, 1835–37. MR 48133 (13,986b).
Bryant, R. (1987). Metrics with exceptional holonomy. Ann. Math. (2) 126, 525–76.
MR 916718 (1989b:53084).
Bryant, R. et al ., (1991). Exterior Differential Systems. MSRI Series 18, Springer-
Verlag, New York. MR 1083148 (1992h:58007).
Bryant, R. (2006). Some remarks on G2 -structures. Proceedings of Gökova Geometry-
Topology Conference 2005. pp. 75–109, Gökova Geometry/Topology Conference
(GGT), Gökova, MR 2282011 (2007k:53019) arXiv:math.DG/0305124.
Cartan, É. (1925). La géometrie des espaces de Riemann. Mémorial des Sciences
Mathematiques Fasc. IX.
Conti, D. and Salamon, S. (2007). Generalized Killing spinors in dimension 5. Trans.
Am. Math. Soc. 359, 5319–43.
DeTurck, D. and Kazdan, J. (1981). Some regularity theorems in Riemannian geometry.
Ann. Sci. Éc. Norm. Sup. 14, 249–60. MR 644518 (83f:53018).
Hitchin, N. (2001). Stable forms and special metrics. Global Differential Geometry: The
Mathematical Legacy of Alfred Gray (Bilbao, 2000). pp. 70–89, Contemporary Math-
ematics, Vol. 288, American Mathematical Society, Providence, RI. MR 1871001
(2003f:53065).
 References 367

Joyce, D. (2000). Compact Manifolds with Special Holonomy. Oxford Mathematical

Monographs, Oxford University Press, Oxford. MR 1787733 (2001k:53093).
Salamon, S. (1989). Riemannian Geometry and Holonomy Groups. Pitman Research
Notes in Mathematics, Vol. 201. Longman, Harlow, UK. MR 1004008 (90g:53058).
Schouten, J. (1918). On the number of degrees of freedom of the geodetically moving
systems. Proc. Kon. Acad. Wet. Amsterdam 21, 607–13.
 XVIII
BRANES ON POISSON VARIETIES

Marco Gualtieri
Dedicated to Nigel Hitchin on the occasion of his 60th birthday

18.1 Introduction
In this chapter we shall take a second look at a classical structure in differential
and algebraic geometry, that of a holomorphic Poisson structure, which is a
complex manifold with a holomorphic Poisson bracket on its sheaf of regular
functions. The structure is determined, on a real smooth manifold M , by the
choice of a pair (I, σI ), where I is an integrable complex structure tensor and
σI is a holomorphic Poisson tensor. We shall view (I, σI ) not as we normally do
but instead as a generalized complex structure, in the sense of Hitchin (2003). In
so doing, we shall obtain a new notion of equivalence between the pairs (I, σI )
which does not imply the holomorphic equivalence of the underlying complex
structures.
In studying this equivalence relation we are naturally led to an unexpected
connection to generalized Kähler geometry, as defined in Gualtieri (2004),
and to a method for constructing certain examples of these structures which
extends the recent work of Hitchin constructing bi-Hermitian metrics on Del
Pezzo surfaces (Hitchin 2007); in particular we obtain similar families of bi-
Hermitian metrics on all smooth Poisson Fano varieties, and in fact on any
smooth Poisson variety admitting a positive Poisson line bundle. We therefore
give an explicit construction of a subclass of the generalized Kähler struc-
tures proven to exist by the generalized Kähler stability theorem of Goto
(2007).
In both these efforts we shall find it useful to introduce an extension of the
notion of connection on a vector bundle, to allow differentiation not only in the
tangent but also in the cotangent directions; we call such a structure a generalized
connection. We also show that in the presence of a generalized metric, there is
a canonical connection D which plays the role of the Levi-Civita connection in
Kähler geometry: namely, we show that (J , G) is generalized Kähler if and only
if DJ = 0.
In the final section we make some speculative comments concerning the rela-
tionship between generalized Kähler geometry and non-commutative geometry,
a topic we hope to clarify in the future.
 Gerbe trivializations 369

18.2 Gerbe trivializations

Let M be a manifold equipped with a U (1) gerbe with connection (specifically, a
gerbe with connective structure in the sense of Brylinski 1993). This determines
canonically a Courant algebroid E over M , in the same way that a U (1)
principal bundle P determines an Atiyah Lie algebroid E = T P/U (1) over M .
See Ševera (1998–2000) and Hitchin (2006a) for details of this construction, and
see Courant (1990), Roytenberg (1999), and Liu, Weinstein, and Xu (1997) for
details concerning Courant algebroids; we review their main properties presently.
The Courant algebroid E is an extension of real vector bundles

π∗ π
0 / T∗ / E / T / 0, (18.1)

where T and T ∗ denote the tangent and cotangent bundles of M . Further, E is

equipped with a nondegenerate symmetric bilinear form ·, · of split signature,
such that π ∗ ξ, a = ξ(π(a)). Finally, there is a bilinear Courant bracket [·, ·] on
C ∞ (E) such that
r [[a, b], c] = [a, [b, c]] − [b, [a, c]] (Jacobi identity)
r [a, f b] = f [a, b] + (π(a)f )b (Leibniz rule)
r π(a)b, c = [a, b], c + b, [a, c] (invariance of bilinear form)
r [a, a] = π ∗ da, a (skew-symmetry anomaly)
The choice of an isotropic complement to T ∗ in E is a contractible one, and so
an isotropic splitting s : T −→ E of the sequence (18.1) always exists. Each such
splitting determines a closed three-form H ∈ Ω3 (M ), given by

(iX iY H)(Z) = [s(X), s(Y )], s(Z). (18.2)

The cohomology class [H]/2π ∈ H 3 (M, R) is independent of the choice of split-

ting, and coincides with the image of the Dixmier–Douady class of the gerbe
in real cohomology. Furthermore, [H] classifies the Courant algebroid up to
isomorphism, as shown by Ševera (1998–2000).
Courant algebroids may be naturally pulled back by the inclusion S ⊂ M of a
submanifold; as a bundle over S, the result is simply given by
π −1 (T S)
ES = , (18.3)
Ann(T S)
and its bracket and inner product are inherited in a straightforward manner.
A trivialization of the gerbe along S induces a Courant trivialization in the
following sense:

Definition 18.1 A Courant trivialization along S is an integrable isotropic

splitting s : T S −→ ES of the pullback Courant algebroid. Such trivializations
exist if and only if ι∗ [H] = 0 for ι : S → M is the inclusion.
370 Branes on Poisson varieties

Integrability is the requirement that the subbundle s(T S) ⊂ ES be closed

under the Courant bracket. Integrable maximal isotropic subbundles of a Courant
algebroid are called Dirac structures; therefore s(T S) is simply a Dirac structure
transverse to T ∗ S. As a result of a Courant trivialization, ES is canonically
isomorphic to T S ⊕ T ∗ S with its natural pairing and the bracket
[X + ξ, Y + η] = LX (Y + η) − iY dξ.
Now suppose that S0 , S1 ⊂ M are submanifolds with smooth intersection, and
suppose we have gerbe trivializations on each of them. Then on X = S0 ∩ S1 we
obtain a pair of gerbe trivializations, which must differ by a line bundle L01 with
U (1) connection ∇01 . Let s0 , s1 be the splittings of EX determined by the two
gerbe trivializations. Then s1 − s0 : T X −→ T ∗ X is given by X → iX F01 where
F01 ∈ Ω2 (M ) is the curvature of ∇01 .
The notion of Courant trivialization provides a convenient way of charac-
terizing isomorphisms of Courant algebroids, as in the following example. The
notation E denotes the Courant algebroid E, equipped with the opposite bilinear
form −·, ·.
Example 18.1 Let EM , EN be Courant algebroids over the manifolds M, N ,
respectively. They are isomorphic precisely if there is a Courant trivialization of
the product Courant algebroid E M × EN along the graph of a diffeomorphism
ϕ : M −→ N in the product M × N .

18.3 Generalized connections

Let E be a Courant algebroid as in the previous section. In keeping with the
notion that the Courant algebroid is an analogue of the tangent bundle, we have
the following generalization of the usual notion of connection:
Definition 18.2 A generalized connection on a vector bundle V is a first-order
linear differential operator
D : C ∞ (V ) −→ C ∞ (E ⊗ V )
such that D(f s) = f Ds + (π ∗ df ) ⊗ s. Furthermore, if V has a Hermitian metric
h, then D is unitary when
d(h(s, t)) = h(Ds, t) + h(s, Dt).
If s : T −→ E is any splitting (not necessarily isotropic) of the Courant algebroid,
then using the decomposition E = s(T ) ⊕ T ∗ ∼ = T ⊕ T ∗ , we have
D = ∇ + χ, (18.4)
where ∇ is a usual unitary connection and χ is a vector field with values in the
bundle of skew-adjoint endomorphisms of V , that is, χ ∈ C ∞ (T ⊗ u(V )). The
tensor χ is independent of the choice of splitting, and we note that if V is of
rank 1, χ is simply a vector field on the manifold.
 Generalized connections 371

With respect to a different splitting s , such that

s − s = θ : T −→ T ∗ ,
we obtain a different decomposition D = ∇ + χ, where ∇ = ∇ + θ(χ).
A generalized connection has a natural curvature operator: for a, b ∈ C ∞ (E),
we define
R(a, b) = [Da , Db ] − D[a,b] ∈ C ∞ (u(V )).
This becomes tensorial in a, b when restricted to a Dirac structure L ⊂ E:
R|L ∈ C ∞ (∧2 L∗ ⊗ u(V )).
If L = T ∗ , for example, we obtain a bivector with values in the skew-adjoint
endomorphisms, R|T ∗ = [χ, χ].
The tensorial curvatures Rs , Rs associated to integrable splittings s, s of E
with s − s = F ∈ Ω2cl (M ) may be compared by projection to T :
Rs = Rs + d∇ (a) + a ∧ a,
where a = F (χ). Therefore if V has rank 1, we have that χ = iX for a real vector
field X, and
Rs − Rs = diX F.
In the particular case that we have a generalized connection D on E itself, it
is natural to compare the connection derivative with the Courant bracket; we
therefore introduce the torsion of D, and leave it as an exercise to verify it is
well-defined.
Definition 18.3 The torsion TD ∈ C ∞ (∧2 E ⊗ E) of a generalized connection
D on E itself is defined by
T (a, b, c) = Da b − Db a − [a, b]sk , c + 12 (Dc a, b − Dc b, a),
where [a, b]sk = 12 ([a, b] − [b, a]). If D preserves the canonical bilinear form ·, ·
on E, then TD is totally skew, that is, TD ∈ C ∞ (∧3 E).
A generalized Riemannian metric on the Courant algebroid E is the choice of a
maximal positive-definite subbundle C+ ⊂ E; this reduces the O(n, n) structure
of E to O(n) × O(n), and defines a positive-definite metric on E:
G(·, ·) = ·, ·|C+ − ·, ·|C− ,
⊥
where C− = C+ is the orthogonal complement with respect to ·, ·. We now
describe a construction of a canonical connection associated to the choice of
such a metric, inspired by calculations in Gualtieri (2004) and Hitchin (2006a)
relating metric connections with skew torsion to the Courant bracket.
The G-orthogonal complement to T ∗ is an isotropic splitting C0 ⊂ E and we
identify it with T , so that G induces a splitting E = T ⊕ T ∗ . The Courant bracket
372 Branes on Poisson varieties

in this splitting is
[X + ξ, Y + η]H = [X, Y ] + LX η − iY dξ + iY iX H,
where H ∈ Ω3cl (M )
is defined by (18.2). The splitting also defines an anti-
orthogonal automorphism C : E −→ E defined by C(X + ξ) = X − ξ, which
satisfies C(C± ) = C∓ . It also has the property, for Z, W ∈ C ∞ (E):
[CZ, CW ]H = C([Z, W ]−H ).
Theorem 18.1 Let C+ ⊂ E be a maximal positive-definite subbundle, i.e. a
generalized metric, as above. Let C : E −→ E be the map defined above. Write
Z = Z+ + Z− for the orthogonal projections of Z ∈ C ∞ (E) to C± . Then the
operator
DZ (W ) = [Z− , W+ ]+ + [Z+ , W− ]− + [CZ−, W− ]− + [CZ+, W+ ]+ (18.5)
defines a generalized connection on E, preserving both ·, · and the positive-
definite metric G. We call this the generalized Bismut connection
Proof. Using the properties of the Courant bracket and the orthogonality C+ =
⊥
C− , we have immediately the property Df Z W = f DZ W , for f ∈ C ∞ (M ). We
also have
DZ (f W ) = f DZ (W ) + (Z− f )W+ + (Z+ f )W− + (Z− f )W− + (Z+ f )W+
= f DZ (W ) + (Zf )W,
proving that D is a generalized connection.
It is clear from (18.5) that C± are preserved by the connection, since DZ W has
nonzero component in C± if and only if W does. Hence we obtain a decomposition
D = D+ ⊕ D− ,
where D± are generalized connection on C± respectively.
To prove that D preserves the canonical metric ·, · as well as the metric G,
we show that D ± preserve the induced metrics on C± . Let V, W ∈ C ∞ (C+ ), and
Z ∈ C ∞ (E). Then
Z+ V, W  = (CZ+ )V, W 
= −[CZ+ , V ], W ] − V, [CZ+ , W ]
= DZ+ V, W  + V, DZ+ W 
Also, we have
Z− V, W  = [Z− , V ], W  + V, [Z− , W ]
= DZ− V, W  + V, DZ− W .
Summing these two results, we obtain that D + preserves the metric on C+ ; the
same argument holds for C− , completing the proof.

 Generalized connections 373

The generalized connections D± define tensors χ± ∈ C ∞ (T ⊗ C± ), via the

decomposition (18.4). We see now that these vanish, since for Z ∈ C ∞ (T ∗ ) and
W ∈ C ∞ (C± ), we have

χ±
Z W = DZ W = [Z∓ , W± ]± + [CZ± , W± ]±

= [Z∓ + (CZ)∓ , W± ]±
= 0,

where we use the fact that Z ∈ T ∗ if and only if CZ = −Z.

As a result of this, we conclude that D± may be viewed as the usual metric
connections ∇± on T , via the projection isomorphisms π± : C± −→ T , i.e.
−1 ±
D ± = π± ∇ π±

The connections ∇± may be described as follows:

∇± ±
X Y = 2π± DX Y±
±
= 4π± DX∓
Y±

= 4π± [X∓ , Y± ]± .

We may easily compute the torsion T ± of the connections ∇± , for vector fields
X, Y, Z:

2g(T + (X, Y ), Z) = T + (X, Y ), Z+ 

= 4π+ [X− , Y+ ]+ − 4π+ [Y− , X+ ]+ − π[X, Y ], Z+ 
= 2[X− , Y+ ]+ + 2[X+ , Y− ]+ − [X, Y ] + iY iX H, Z+ 
= 2H(X, Y, Z) − 2[X+ − X− , Y+ − Y− ], Z+ 
= 2H(X, Y, Z),

by the fact that [X+ − X− , Y+ − Y− ] = 0 since the Courant bracket vanishes on

1-forms. A similar calculation gives g(T − (X, Y ), Z) = −H(X, Y, Z).
The above calculation shows that ∇± coincide with the Bismut connections
with totally skew torsion ±H. In this way, we have essentially repeated the
observation of (Hitchin 2006a) that ∇± may be conveniently expressed in terms
of the Courant bracket. To summarize, the generalized Bismut connection is
essentially a usual connection on E which restricts to C± to give the Bismut
connections with torsion ±H.

Proposition 18.1 The torsion TD of the generalized Bismut connection lies

in ∧3 C+ ⊕ ∧3 C− ⊂ ∧3 E, and is given by
∗ ∗
TD = 2(π+ H + π− H).
374 Branes on Poisson varieties

Proof. First we show that T (C+ , C− , ·) = 0, so that T ∈ C ∞ (∧3 C+ ⊕ ∧3 C− ).

Let x ∈ C ∞ (C+ ), y ∈ C ∞ (C− ) and z ∈ C ∞ (E). Then
TD (x, y, z) = Dx y − Dy x − [x, y], z
= [x, y]− − [y, x]+ − [x, y], z = 0,
as required.
Now take x, y, z ∈ C ∞ (C+ ). Since χD = 0, we have the identity
Dz x, y − Dz y, x = DCz x, y − DCz y, x
= [Cz, x], y − [Cz, y], x
= [x, y] − [y, x], Cz.
Therefore the torsion is given by
1
T (x, y, z) = Dx y − Dy x − [x, y]sk , z + [x, y] − [y, x], Cz
2
= Dx y − Dy x, z + [x, y], Cz − z

X Y − ∇Y X − [X, Y ], Z)
= 2g(∇+ +

= 2H(X, Y, Z).
A similar calculation for x, y, z ∈ C ∞ (C− ) gives T (x, y, z) = 2H(x, y, z) as well,
yielding the result.

As we have explained, the generalized Bismut connection D is completely
determined by a usual connection on T ⊕ T ∗ . Using the fact that the Bismut
connections satisfy ∇± = ∇ ± 12 g −1 H for ∇ the Levi-Civita connection, we may
write D explicitly with respect to the splitting E = T ⊕ T ∗ , and for X ∈ C ∞ (T ),
as follows:

1 2 −1

∇ ∧ g (iX H)
DX = 1 X 2
2 iX H ∇∗X
The significance of this connection in the context of generalized geometry was
first understood and investigated by Ellwood. Here we simply view it as a
generalized connection 1 mainly for the purpose of highlighting its properties
and defining its torsion tensor.

18.4 Generalized holomorphic bundles and branes

Suppose now that we have a generalized complex structure J on (M, E), which
is an orthogonal almost complex structure J : E −→ E whose +i eigenbun-
dle L ⊂ E ⊗ C is closed under the Courant bracket (Hitchin 2003). We now
1 The author thanks Yicao Wang for pointing this out and correcting an error in the previous

version.
 Generalized holomorphic bundles and branes 375

describe how the structures in the previous two sections may be made compatible
with J .

18.4.1 Generalized holomorphic bundles

The integrability of J guarantees that L = ker(J − i1) is a complex Lie alge-
broid, with associated de Rham complex
dL
C ∞ (∧k L∗ ) / C ∞ (∧k+1 L∗ ) (18.6)

A complex vector bundle equipped with a flat L-connection is called a gener-

alized holomorphic bundle (Gualtieri 2007). Therefore, generalized holomorphic
bundles form a category of Lie algebroid representations in the sense of (Evens,
Lu, and Weinstein 1999).
In the case that J is a usual complex structure, for instance, a generalized
holomorphic bundle consists of a holomorphic bundle V , together with a holo-
morphic section Φ ∈ H 0 (M, T1,0 M ⊗ End(V )) satisfying
Φ ∧ Φ = 0 ∈ H 0 (M, ∧2 T1,0 M ⊗ End(V ).
∗
Note that if M is holomorphic symplectic, then T1,0 is isomorphic to T1,0 , and
Φ may be viewed as a Higgs bundle, in the sense of Simpson (1992).
In the case that J is a symplectic structure, a generalized holomorphic bundle
is simply a flat bundle.
Definition 18.4 A unitary generalized connection D on a complex vector
bundle V is compatible with J when its curvature along L = ker(J − i1) is zero.
It follows immediately that the restriction of D to L defines a flat L-module
structure on V , making V a generalized holomorphic bundle. Conversely, suppose
V is a J -holomorphic bundle, that is, it is equipped with an L-connection as
follows:
2
∂ : C ∞ (V ) −→ C ∞ (L∗ ⊗ V ), ∂ = 0. (18.7)
This operator has symbol sequence given by wedging with σ(ξ) = + iJ )ξ ∈
1
2
(1
L, where we identify L∗ = L using the metric on E.
Choosing a Hermitian metric h on the bundle V , so that V  V ∗ , we may
view the complex conjugate of (18.7),
∂ : C ∞ (V ) −→ C ∞ (L ⊗ V ),
as a L-connection on V ∗ ; we then form the dual ∂ ∗ of this partial connection.
Finally we form the sum
D = ∂ + ∂ ∗ : C ∞ (V ) −→ C ∞ ((L ⊕ L) ⊗ V ) = C ∞ (E ⊗ V ),
which has symbol σ + σ̄ = π ∗ . Hence it defines a generalized connection on V .
We summarize the above in the following:
376 Branes on Poisson varieties

Proposition 18.2 Let V be a complex vector bundle with J -holomorphic

structure given by ∂, and choose a Hermitian metric on V . Then the operator
D = ∂ + ∂ ∗ : C ∞ (V ) −→ C ∞ (E ⊗ V )
is the unique unitary generalized connection extending ∂.

When V is a line bundle, there is a useful formula for the generalized

connection one-form in terms of a holomorphic trivialization, analogous to the
Poincaré–Lelong formula for the Chern connection on a Hermitian holomorphic
line bundle.

Proposition 18.3 (Generalized Poincaré–Lelong formula) Let V be a

generalized holomorphic Hermitian line bundle, and let s ∈ C ∞ (V ) be a holo-
morphic section. Where it is nonzero, it defines a trivialization of the unitary
generalized connection, D = d + iA, where
A = J d log |s| ∈ C ∞ (E).

Proof. Whenever s is nonzero, we have

s s
D = iA .
|s| |s|

Taking the projection to L, we obtain

dL log |s| = iA0,1 ,
so that A = −i(dL − dL ) log |s| = J d log |s|, as required.

In particular, if s is nonzero on an open dense set, then the vector field

πJ d log |s| = X must extend to a smooth vector field on the whole of M ,
since π(iA) coincides with χ ∈ C ∞ (T ⊗ u(V )), which is globally defined for any
generalized connection. But the map πJ |T ∗ : T ∗ −→ T is actually a Poisson
structure Q ∈ C ∞ (∧2 T ) (see Gualtieri 2007 for a discussion of this fact), and
hence s vanishes only along the zero locus of the Poisson structure Q, which is
a strong constraint on any generalized holomorphic section.
The above proposition may be used, by invoking the local existence of non-
vanishing holomorphic sections near points for which J is regular (i.e. Q has
locally constant rank), to show that the vector field χ of any Hermitian J -
holomorphic line bundle must be a Poisson vector field. It therefore defines a
characteristic class in the Poisson cohomology of Lichnerowicz (1977), which
is the cohomology of the complex (C ∞ (∧• T ), dQ ), where dQ Π = [Q, Π] is the
Schouten bracket with Q.

Corollary 18.1 The real vector field X = −iχ associated to any Hermitian
generalized holomorphic line bundle preserves the Poisson structure Q, that is, it
is a Poisson vector field. Furthermore its Poisson cohomology class [X] ∈ HQ
1
(M )
is independent of the Hermitian metric.
 Generalized holomorphic bundles and branes 377

Proof. X is Poisson since, by the proposition, it is locally Hamiltonian on an

open dense subset of M . Hence LX Q = 0 everywhere. Rescaling the Hermitian
metric by a positive function ef , we obtain a new vector field X  = X + 12 Qdf ,
which differs from X by a global Hamiltonian vector field. Hence [X] ∈ HQ 1
(M )
is independent of the choice of Hermitian structure.

We may also deduce this result from the more general fact that the tensor
product of a L-module with a L module is a Poisson module for Q. (This is a
direct consequence of the fact that the tensor product of the Dirac structures
L, L is the Dirac structure associated to Q, shown in Gualtieri 2007.) For any
generalized holomorphic line bundle V , therefore, the trivial bundle V ⊗ V
acquires a Q-module structure, and therefore, as described in Evens, Lu, and
1
Weinstein (1999), a characteristic class in HQ (M ).
There are always two natural J -holomorphic line bundles on any generalized
complex manifold: the trivial bundle, for which χ = 0 (for the standard Hermitian
structure), and the canonical line bundle KJ of pure spinors associated to
the maximal isotropic subbundle L ⊂ E ⊗ C. Since KJ ⊗ KJ is naturally the
determinant line det T ∗ , it follows that [X] = [−iχ] is actually the modular class
of the Poisson structure Q, in the sense of Weinstein (1997).

18.4.2 Generalized complex branes

Suppose we have a submanifold ι : S → M equipped with a Courant trivializa-
tion s : T S −→ ES . The Dirac structure s(T S) ⊂ ES may be canonically lifted to
a maximal isotropic subbundle of ι∗ E; this operation is called the push-forward
of Dirac structures (Bursztyn and Radko 2003):
τS := ι∗ s(T S) = {e ∈ E : π(e) ∈ T S and e + Ann(T S) ∈ s(T S)}.
Note that τS is an extension of the tangent bundle of S by its conormal bundle:

0 / N ∗S / τS / TS / 0.

In the presence of the generalized complex structure, there is a natural compat-

ibility condition, as follows:
Definition 18.5 A generalized complex submanifold is a trivialization of the
Courant algebroid along a submanifold ι : S → M which is compatible with the
generalized complex structure J , in the sense that
J τS = τS , (18.8)
where τS = ι∗ s(T S) ⊂ ι∗ E.
As shown in Gualtieri (2007), in the complex case (and for the trivial
gerbe), generalized complex submanifolds correspond to holomorphic submani-
folds equipped with unitary holomorphic line bundles, whereas in the symplectic
case they correspond to Lagrangian submanifolds equipped with flat line bundles
or the co-isotropic A-branes of Kapustin and Orlov (2003). A useful general
378 Branes on Poisson varieties

example of a generalized complex submanifold is the graph of an isomorphism

of generalized complex manifolds, as follows. The notation J denotes the same
endomorphism as J but in the opposite Courant algebroid E.
Example 18.2 Let (M, JM ), (N, JN ) be generalized complex manifolds. They
are isomorphic when there is a Courant algebroid isomorphism in the sense
of Example 18.1 which is a generalized complex submanifold of the product
(M × N, J M × JN ).
In Gualtieri (2007), it is shown that in the eigenspace decomposition with
respect to J
¯
τS ⊗ C =  + ,
the +i eigenbundle  inherits a Lie bracket, by extending sections randomly
to sections over M which remain +i eigensections of J , taking their Courant
bracket and restricting to S. Thus  becomes an elliptic complex Lie algebroid
over S. Therefore there is a notion of flat -module, The resulting -modules are
called branes in analogy to the physics literature.
Definition 18.6 A generalized complex brane is a vector bundle with flat -
connection, supported over a generalized complex submanifold.
Remark 18.1 Just as for generalized holomorphic bundles, we may choose to
represent branes using unitary connections with values in τS∗ , that is, operators
D : C ∞ (V ) −→ C ∞ (τS∗ ⊗ V )
with symbol given by the inclusion T ∗ S ⊂ τS∗ , and with vanishing curvature along
 ⊂ τS ⊗ C.
For a usual complex structure, a brane consists of a holomorphic bundle V sup-
ported on a complex submanifold S ⊂ M together with a choice of holomorphic
section φ ∈ H 0 (S, N1,0 S ⊗ End(V )) satisfying
φ ∧ φ = 0 ∈ H 0 (S, ∧2 N1,0 S ⊗ End(V )),
where N1,0 S denotes the holomorphic normal bundle of S.
On the other hand, for a symplectic structure, branes are complex flat bundles
if they are supported on Lagrangian submanifolds; they may also be supported
on co-isotropic submanifolds with holomorphic structure transverse to the char-
acteristic foliation (Kapustin and Orlov 2003, Gualtieri 2007), in which case they
are transversally holomorphic bundles, flat along the leaves.
Example 18.3 (Higgs bundles) Let C ⊂ X be a curve in a complex surface
equipped with holomorphic symplectic form Ω ∈ H 0 (X, Ω2,0 ), for example, a K3
surface. Also, let V −→ C be a Higgs bundle in the sense of Hitchin (1987), that
is, a holomorphic bundle together with a Higgs field θ ∈ H 0 (C, Ω1 ⊗ End(V )).
Since C is Lagrangian with respect to Ω, we have an isomorphism T1,0 C −→
 Generalized holomorphic bundles and branes 379

∗
N1,0 C, so that we may form φ = Ω−1 θ ∈ H 0 (C, N1,0 ⊗ End(V )), making (V, φ)
into a brane for the complex structure.
On the other hand, if (V, θ) is a stable Higgs bundle, then by the existence
theorem (Hitchin 1987) for solutions to Hitchin’s equations we obtain a complex
flat connection ∇ on V , rendering (C, V, ∇) into a symplectic brane with respect
to either the real or imaginary parts of Ω.

Example 18.4 Let V be a generalized holomorphic bundle, that is, a complex

vector bundle equipped with a flat L-connection, where L = ker(J − i1) for J
a generalized complex structure. Then the pullback of V to any generalized
complex submanifold S ⊂ M defines a generalized complex brane, as can be
seen easily from the inclusion  ⊂ L.

Another simple example of a generalized complex brane occurs when it

is supported on an isomorphism of generalized complex manifolds, as in
Example 18.16.

Proposition 18.4 Let S ⊂ M0 × M1 define an isomorphism of the generalized

complex manifolds (M0 , J0 ), (M1 , J1 ). Then the Lie algebroid  is isomorphic
to both Li = ker(Ji − i1), so that branes on S may be identified with generalized
holomorphic bundles on either manifold.

Proof. Let πi : M0 × M1 −→ Mi be the usual projection maps. The subbundle

 ⊂ (π0∗ L0 ⊕ π1∗ L1 )|S is transverse to both πi∗ Li , as we now show. If x ∈  ∩ π0∗ L0 ,
for instance, then (π1 )∗ x = ϕ∗ ((π0 )∗ x), where ϕ is the diffeomorphism defining
S. This implies (πi )∗ x = 0 and x ∈ N ∗ S ⊗ C, which clearly is transverse to π0∗ L0 .
Hence x = 0, and similarly for L1 .
This transversality means that we have isomorphic bundle maps onto each
factor:
p0 p1
L0 o  / L1

(18.9)
 π0  π1 
M0 o S / M1

We now show that the projections p0 , p1 are isomorphisms of Lie algebroids.

Given Z ∈ C ∞ (S, ), let X = p0 (Z) and Y = p1 (Z). Then Z may be expressed
as Z = (π0∗ X + π1∗ Y )|S . Computing the bracket of Z, Z  , we may use the given
extensions to M0 × M1 and compute their Courant bracket:
[Z, Z  ] = [π0∗ X + π1∗ Y, π0∗ X  + π1∗ Y  ]|S
= [π0∗ X, π0∗ X  ]|S + [π1∗ Y, π1∗ Y  ]|S
= (π0∗ [X, X  ] + π1∗ [Y, Y  ])|S , (18.10)
380 Branes on Poisson varieties

where we use the fact that sections pulled back from opposite factors M0 , M1
Courant commute in the product. Applying the projections to the final formula,
we obtain
p0 ([Z, Z  ]) = [p0 (Z), p0 (Z  )] and p1 ([Z, Z  ]) = [p1 (Z), p1 (Z  )],
as required.

We now describe the general form of a generalized complex brane when it
is supported on the whole manifold M ; these are usually called “space-filling
branes.” We first observe that the requirement that M be a generalized complex
submanifold of itself places a very strong constraint on J .
Proposition 18.5 (M, J ) is a generalized complex submanifold of itself if and
only if there exists an integrable isotropic splitting E = T ⊕ T ∗ of the Courant
algebroid with respect to which J has the form


I Q
J = , (18.11)
−I ∗
where I is a usual complex structure on the manifold and σ = P + iQ, for P =
IQ, is a holomorphic Poisson structure, that is, it satisfies [σ, σ] = 0.
Proof. Compatibility of the splitting with J forces J T = T , which holds if
and only if J is upper triangular, and the orthogonality of J together with the
fact J 2 = −1 guarantees that I is an almost complex structure and that Q is
a bivector of type (2, 0) + (0, 2). The −i eigenbundle of J is then the direct
∗
sum of T0,1 with the graph of σ : T1,0 −→ T1,0 . This is closed (involutive) for the
Courant bracket if and only if T0,1 is integrable and [σ, σ] = 0, as required.

In the splitting E = T ⊕ T ∗ for which J has the form (18.11), we see that
τS = T M , and further that  = T1,0 , so that -modules are precisely holomorphic
bundles with respect to the complex structure I.

18.5 Multiple branes and holomorphic Poisson varieties

Suppose that we have a Courant trivialization s making (M, J ) a generalized
complex submanifold of itself, so that E = T ⊕ T ∗ and J has the form (18.11).
Now we investigate the consequences of having a second trivialization s which
is also compatible with J . Let F ∈ Ω2cl (M, R) be the two-form taking s to s .
By Proposition 18.5, and the fact that the Poisson structure Q is independent
of splitting, we have





1 I Q 1 J Q
= , (18.12)
−F 1 −I ∗ F 1 −J ∗
for a second complex structure J such that σ  = JQ + iQ is holomorphic Poisson.
In particular we note the important fact that a generalized complex structure
 Multiple branes and holomorphic Poisson varieties 381

may be expressed as a holomorphic Poisson structure in several different ways,

and with respect to different underlying complex structures, depending on the
choice of splitting. Equation (18.12) is equivalent to the conditions

J − I = QF,
(18.13)
F J + I ∗ F = 0.

Phrased as a single condition on F , we obtain the non-linear equation

F I + I ∗ F + F QF = 0, (18.14)

which may be viewed as a deformation of the usual condition F I + I ∗ F = 0 that

F be of type (1, 1) with respect to the complex structure. Equation (18.14) has
been studied by Kapustin (2005), who showed that it corresponds to a non-
commutative version of the (1, 1) condition via the Seiberg–Witten transform
on tori. We take a different approach here, focusing rather on a groupoid
interpretation of the equivalent system (18.13).
The set of compatible global Courant trivializations forms a groupoid; we may
label each trivialization by the complex structure it induces on the base, and we
see from (18.12) or (18.13) that if FIJ takes I to J and FJK takes J to another
trivialization K, then FIJ + FJK takes I to K.

Definition 18.7 Fix a real manifold M with real Poisson structure Q. Let G
be the groupoid whose objects are holomorphic Poisson structures (Ii , σi ) on M
with fixed imaginary part given by Im(σi ) = Q, and whose morphisms Hom(i, j)
consist of real closed two-forms Fij ∈ Ω2cl (M, R) such that

Ij − Ii = QFij ,
(18.15)
Fij Ij + Ii∗ Fij = 0.

The composition of morphisms is then simply addition of two-forms Fij + Fjk . In

keeping with the interpretation of Fij as differences between gerbe trivializations,
we could define Hom(i, j) to consist of unitary line bundles Lij with curvature
Fij , such that composition of morphisms would coincide with tensor product.

Automorphisms of the Courant algebroid which fix J give rise to automor-

phisms of the groupoid of trivializations defined above; we describe these now.
Orthogonal automorphisms of the standard Courant bracket on T ⊕ T ∗ consist
of pairs (ϕ, B) ∈ Diff(M ) × Ω2cl (M, R), which act on T ⊕ T ∗ via X + ξ → ϕ∗ X +
(ϕ−1 )∗ ξ + iϕ∗ X B. Since our generalized complex structure has the form (18.11),
we may easily determine its automorphism group.
382 Branes on Poisson varieties

Proposition 18.6 The automorphism group Aut(J ) of the generalized complex

structure (18.11) is the set of pairs (ϕ, B) ∈ Diff(M ) × Ω2cl (M, R) such that
Qϕ = Q
I ϕ − I = QB
BI ϕ + I ∗ B = 0,
where Qϕ = ϕ∗ Q and I ϕ = ϕ∗ Iϕ−1
∗ .

These automorphisms therefore act on the groupoid of global generalized

complex submanifolds (18.15), sending (Ii , σi ) → (Iiϕ , (ϕ−1 )∗ σi ) and Fij →
(ϕ−1 )∗ Fij + B. Of course, we may wish to interpret B as the curvature of a
unitary line bundle U , in which case it would act on the groupoid line bundles
Lij by tensor product Lij → (ϕ−1 )∗ Lij ⊗ U .
Instead of viewing Fij as the difference between two generalized complex
submanifolds of (M, J ), we may interpret (18.12) as giving an isomorphism
between two different generalized complex structures on T ⊕ T ∗ . This rephrasing
leads immediately to the following:
Proposition 18.7 Let (Ii , σi ) and (Ij , σj ) be holomorphic Poisson struc-
tures on M with associated generalized complex structures Ji , Jj on T ⊕ T ∗
via (18.11), let Im(σi ) = Im(σj ) = Q, and let Fij ∈ Ω2cl (M, R) satisfy (18.15).
Then the graph of Fij over the diagonal ∆ ⊂ M × M defines a generalized com-
plex submanifold of (M × M, J i × Jj ), yielding an isomorphism of generalized
complex manifolds
∼
=
(M, Ji ) −→ (M, Jj ). (18.16)
In view of Proposition 18.4, this result implies that a morphism Fij from
(Ii , σi ) to (Ij , σj ) induces an equivalence between the categories of generalized
holomorphic bundles associated to Ji , Jj . We now explain this equivalence
explicitly, and its significance for the holomorphic Poisson structures involved.
Proposition 18.8 Let J be of the form (18.11), for I is a complex structure
and σ = P + iQ is a holomorphic Poisson structure. Then a generalized holo-
morphic bundle is precisely a holomorphic Poisson module (Polishchuk 1997),
that is, a holomorphic bundle V with an additional action of the structure sheaf
on the sheaf of holomorphic sections, denoted {f, s}, satisfying
{f, gs} = {f, g}s + g{f, s}, (18.17)
{{f, g}, s} = {f, {g, s}} − {g, {f, s}}, (18.18)
where f, g ∈ O, s ∈ O(V ), and {f, g} denotes the Poisson bracket induced by σ.
Proof. Let L = ker(J + i1), so that for J as in (18.11), we have
L = T0,1 ⊕ Γσ , (18.19)
 Multiple branes and holomorphic Poisson varieties 383

∗
where Γσ = {ξ + σ(ξ) : ξ ∈ T1,0 }. Let ∂ V : C ∞ (V ) −→ C ∞ (L∗ ⊗ V ) be a gen-
eralized holomorphic structure. Decomposing using (18.19) and identifying
∗   
Γσ = T1,0 , we write ∂ V = ∂ + ∂ , where ∂ : C ∞ (V ) −→ C ∞ (T0,1 ∗
⊗ V ) is a

usual holomorphic structure and ∂ : C ∞ (V ) −→ C ∞ (T1,0 ⊗ V ) satisfies, for
f ∈ C ∞ (M, C) and s ∈ C ∞ (V ),
 
∂ (f s) = f ∂ s + Zf ⊗ s,
where Zf = σ(df ) is the Hamiltonian vector field of f . This is equivalent to
2 
condition (18.17). Furthermore ∂ V = 0 implies that ∂ is holomorphic and
defines a Poisson module structure via

{f, s} = ∂ ∂f s,
   
since (∂ )2 = 0 implies ∂ ∂{f,g} = [∂ ∂f , ∂ ∂g ], which is equivalent to (18.18), as
required.

Letting Lk = ker(Jk + i1), we see from (18.12) that exp(Fij ) takes Li to Lj .
Hence the map on generalized holomorphic bundles induced by the isomor-
phism (18.16) may be described as composition with exp(Fij )
∂ i → eFij ◦ ∂ i .
This map may be made more explicit in terms of the associated generalized
connections. Choose a Hermitian structure on the Ji -holomorphic bundle (i.e. σi -
Poisson module), and let D = ∇ + χ be the extension of ∂ i as in Proposition 18.2.
Then Fij acts on D via

D → D = ∇ + χ,
(18.20)
∇ = ∇ + Fij (χ).
which then defines a σj -Poisson module. It is important to note that the σi -
Poisson module, which is Ii -holomorphic, inherits via (18.20) a Ij -holomorphic
structure, without the presence of any holomorphic map between (M, Ij ) and
(M, Ii ).
Given this result, it is natural to ask how restrictive the condition of admitting
a Poisson module structure actually is. The following is a simple result describing
the complete obstruction to the existence of a Poisson module structure on a
holomorphic line bundle.
Proposition 18.9 Let M be a holomorphic Poisson manifold, and let V be a
∗
holomorphic line bundle on M . Then the Atiyah class of V , α ∈ H 1 (T1,0 ), com-
bines with the Poisson structure σ ∈ H 0 (∧2 T1,0 ) to give the class σα ∈ H 1 (T1,0 ).
If σα = 0, then there is a well-defined secondary characteristic class fα ∈ Hσ2 (M )
in Poisson cohomology. V admits a Poisson module structure if and only if
both classes {σα, fα } vanish. The space of Poisson module structures is affine,
modeled on Hσ1 (M ).
384 Branes on Poisson varieties

Proof. A Poisson module structure on V is a holomorphic differential oper-

ator ∂ : O(V ) −→ O(T1,0 ⊗ V ) satisfying ∂(f s) = f ∂s + Zf ⊗ s, for f ∈ O and
s ∈ O(V ), where Zf is the σ-Hamiltonian vector field of f , and such that the
curvature vanishes. Let {Ui } be an open cover of M and let {si ∈ O(Ui , V )} be a
local trivialization of V such that si = gij sj for holomorphic transition functions
gij ; then
∂si = Xi ⊗ si , (18.21)

where Xi are holomorphic Poisson vector fields (since ∂d{f,g} = [∂df , ∂dg ]) such
that
Xi − Xj = Zlog gij . (18.22)

The Hamiltonian vector fields Zlog gij = σ(d log gij ) are a Čech representative for
the image of the Atiyah class under σ. Therefore, (18.22) holds if and only if
σα = 0 ∈ H 1 (T1,0 ). If σα = 0, then we may solve (18.22) for some holomorphic
vector fields X̃i . We can modify these by a global holomorphic vector field so
that they are each Poisson if and only if the global bivector field fσ defined
by fσ |Ui = [X̃i , σ] vanishes in Poisson cohomology, that is, fσ = [Y, σ] for Y ∈
H 0 (T1,0 ), in which case Xi = X̃i − Y |Ui defines a Poisson module structure as
required.
Given any holomorphic Poisson vector field Z ∈ Hσ1 (M ) and Poisson module
structure ∂, the sum ∂ + Z defines a new Poisson module structure. Con-
versely, two Poisson module structures ∂  , ∂ must satisfy ∂  − ∂ ∈ Hσ1 (M ), as
claimed.

It is remarked in (Polishchuk 1997) that the canonical line bundle K always

admits a natural Poisson module structure for any holomorphic Poisson structure
σ via the action, for f ∈ O and ρ ∈ O(K),

{f, ρ} = LZf ρ.

Based on these considerations, we obtain the following example.

Example 18.5 Let M = CP 2 , equipped with a holomorphic Poisson struc-

ture σ ∈ H 0 (O(3)). Note that H 1 (T1,0 ) = 0. Then K = O(−3) is canonically a
Poisson module, and since O(1)−3 = K, we see that the obstruction fσ from
Proposition 18.9 must vanish for O(1) as well (note that dim Hσ2 (CP 2 ) = 2, so
the obstruction space is nonzero). Hence all holomorphic line bundles O(k) admit
Poisson module structures. If σ is generic, these Poisson module structures are
unique, since Hσ1 (M ) = 0, due to the fact that only the zero holomorphic vector
field on CP 2 stabilizes a smooth cubic curve.

We conclude this section with a simple example of a generalized complex man-

ifold admitting multiple trivializations with non-biholomorphic induced complex
structures.
 Relation to generalized Kähler geometry 385

Proposition 18.10 Let E0 = E × C, the trivial line bundle over an elliptic

curve E, and let Ec , for c ∈ R, be the alternative holomorphic structure on
E × C obtained by endowing the bundle E × C with the holomorphic structure
associated to the point ic ∈ H 1 (E, O) = C. Then E0 and Ec are diffeomorphic,
non-biholomorphic complex manifolds. They are equipped with canonical holo-
morphic Poisson structures σ0 , σc vanishing to first order on the zero section,
and furthermore (E0 , σ0 ) and (Ec , σc ) are isomorphic as generalized complex
manifolds ∀c ∈ R (and hence have equivalent categories of Poisson modules).
Proof. Represent E as C∗ /{z → λz} and let w be the linear coordinate on
the fiber of E × C. Then the holomorphic structure Ec is given by the complex
volume form


dz dz̄
Ωc = ∧ dw + icw ,
z z̄
and the holomorphic Poisson structure σc is given by


∂ ∂ ∂
σc = z + icw̄ ∧w .
∂z ∂ w̄ ∂w
The pure spinor corresponding to the generalized complex structure (Ec , σc ) is
ρc = eσc Ωc .
Now let Fc = ic dz∧dz̄
z z̄ be a real multiple of the volume form on E (which may be
viewed as a curvature when c ∈ 2πZ). Then we verify that
 
z ∧ dw + icw z̄
eFc eσ0 Ω0 = w + dz dz̄

= eσc Ωc ,
showing that (E0 , σ0 ) and (Ec , σc ) are isomorphic as generalized complex
manifolds.

18.6 Relation to generalized Kähler geometry

A generalized Kähler structure is a pair (JA , JB ) of commuting generalized
complex structures such that
G(·, ·) = JA ·, JB ·
is a generalized Riemannian metric.
In Gualtieri (2004) it is shown that the integrability of the pair (JA , JB ) is
equivalent to the fact that the induced decomposition of the definite subspaces
C± given by
C± ⊗ C = L± ⊕ L± ,
where L± = ker(JA − i1) ∩ ker(JB ∓ i1), satisfies the condition that L± are
each involutive. Using the canonical generalized connection D introduced in
386 Branes on Poisson varieties

Theorem 18.1, we provide the following equivalent description of generalized

Kähler geometry:

Theorem 18.2 Let G be a generalized metric and let J be a G-orthogonal

almost generalized complex structure. Then (J , G) defines a generalized Kähler
structure if and only if DJ = 0 and the torsion TD ∈ C ∞ (∧3 E) is of type (2, 1) +
(1, 2) with respect to J .

Proof. We leave the forward direction to the reader. We show that if DJ = 0

(where, as usual, (Dx J )y = Dx (J y) − J (Dx y)) and the torsion is as above,
then J is integrable as a generalized complex structure. Note that under these
assumptions, the complementary generalized complex structure J  = GJ would
also be covariant constant, and be compatible with the torsion as well, by
Proposition 18.1. Therefore by the following argument J  is also integrable,
and we obtain the result.
We compute the Nijenhuis tensor of J , for x, y, z ∈ C ∞ (E) (in the following,
[·, ·] refers to the skew-symmetrized Courant bracket):

NJ (x, y), z = [J x, J y] − J [J x, y] − J [x, J y] − [x, y], z

= Dz (J x), J y − Dz (J y), J x + DJ z (J x), y − DJ z y, J x
+DJ z x, J y − DJ z (J y), x − Dz x, y + Dz y, x
−TD (J x, J y, z) − TD (J x, y, J z) − TD (x, J y, J z) − T (x, y, z).

The first eight terms cancel since Dx (J y) = J Dx y, and the last four terms
cancel since TD is of type (2, 1) + (1, 2). Therefore J is integrable, as
claimed.

We now explain that a solution to the system (18.13), if positive in a certain

sense, gives rise to a generalized Kähler structure. When the Poisson structure
Q vanishes, this result specializes to the fact that a positive holomorphic line
bundle with Hermitian structure defines a Kähler structure.

Definition 18.8 Let (I, J, Q, F ) be a solution to the system (18.13), that is, it
defines two global Courant trivializations compatible with a generalized complex
structure, separated by the two-form F . Then

g = − 12 F (I + J)

is a symmetric tensor, and if it is positive-definite, we say that F is positive.

If F is positive, then (g, I), (g, J) are both Hermitian structures. Let ωI = gI,
ωJ = gJ be their associated two-forms. Then we have the following:
 Relation to generalized Kähler geometry 387

Theorem 18.3 Let (I, J, Q, F ) be as above, and let F be positive. Then the
pair


 

1 1 J + I −(ωJ−1 − ωI−1 )
1
JB = ,
2 b1 ωJ − ωI −(J ∗ + I ∗ )
−b 1



 (18.23)
1 1 J − I −(ωJ−1 + ωI−1 ) 1
JA =
2 b1 ωJ + ωI −(J ∗ − I ∗ ) −b 1
defines a generalized Kähler structure on the standard Courant algebroid (T ⊕
T ∗ , [·, ·]0 ), for b ∈ Ω2 (M, R) given by
b = − 12 F (J − I).
Proof. It is easily verified that JA2 = JB2 = −1 and that [JA , JB ] = 0. To show
integrability, we first observe that JA has the form of a pure symplectic structure;
indeed, with the definitions above,


−F −1
JA = .
F
We see therefore that JA is integrable since dF = 0.
The structure JB is also integrable, as follows. Let LB = ker(JB − i) and Let
LB = L+ ⊕ L− be its decomposition into ±i eigenspaces for JA . Then
L+ = {X + (b + g)X : X ∈ TJ1,0 }
L− = {X + (b − g)X : X ∈ TI1,0 }
It follows from the definitions of b, g that b + g = −F J whereas b − g = F I. As
a result we have
L+ = {X − iF X : X ∈ TJ1,0 }
L− = {X + iF X : X ∈ TI1,0 }

which are integrable precisely when iX iY dF = 0 for all X, Y in TI1,0 or TJ1,0 . Of

course this holds since F is closed.

We note that the converse of this argument also holds; using the result from
Gualtieri (2004) that any generalized Kähler structure has the form (18.23), we
may show that any generalized Kähler structure (JA , JB ) with the property that
JA is symplectic gives rise to a solution to the system (18.13). More explicitly,
given the bi-Hermitian data (g, I, J) we determine F via
F = −2g(I + J)−1 ,
where (I + J) is invertible by the assumption on JA , and the Poisson structure
Q is given by
Q = (J − I)F −1 = 12 [I, J]g −1 .
388 Branes on Poisson varieties

This is consistent with Hitchin’s general observation (Hitchin 2006b) that

[I, J]g −1 defines a holomorphic Poisson structure for both I and J, for any
generalized Kähler structure.
In fact, the interpretation of Fij in Proposition 18.7 as defining a morphism
between holomorphic Poisson structures allows us to view the generalized Kähler
structure as a morphism between the holomorphic Poisson structures (I, σI ),
(J, σJ ). This point of view is related to the approach in Lindström et al. (2007).
to defining a generalized Kähler potential, and may help to resolve the problems
encountered there at non-regular points.
Given the equivalence between certain generalized Kähler structures and
configurations of generalized complex submanifolds shown in this section, we may
apply it to produce new examples of generalized Kähler structures, or indeed of
configurations of branes. We do this in the following section.

18.7 Construction of generalized Kähler metrics

Given a generalized complex submanifold, it is natural to construct more by
deformation; this is a familiar construction in symplectic geometry, where new
Lagrangian submanifolds may be produced by applying Hamiltonian or sym-
plectic diffeomorphisms. Therefore we would like to deform a given general-
ized complex submanifold by an automorphism of the underlying geometry,
as described in Proposition 18.6. If the automorphism used is positive in the
sense of Definition 18.8, then we will have constructed a generalized Kähler
structure, by Theorem 18.3. This construction is inspired by a construction
of Joyce contained in Apostolov, Gauduchon, and Grantcharov (1999), and
its generalization by Hitchin (2006b) to the construction of generalized Kähler
structures on Del Pezzo surfaces.
To reiterate, the goal of the construction is as follows: given a holomor-
phic Poisson structure (I, σI ) on M , with real and imaginary parts σI =
P + iQ, find a second complex structure J and a two-form F solving the
system (18.13), that is,

J − I = QF,
(18.24)
F J + I ∗ F = 0.

We are particularly interested in the case where g = − 12 F (I + J) is positive-

definite, as this then defines a generalized Kähler structure, however the con-
struction does not depend on it.
In this construction, the complex structure J will be obtained from I by
flowing along a vector field; as a result, J will be biholomorphic to I. Also, we
shall describe the construction in the case that F is the curvature of a unitary
connection, although it will be clear that integrality of the form F is not required.
1. We begin with a Hermitian complex line bundle L over a compact com-
plex manifold M ; the two-form F solving (18.24) will be chosen from the
 Construction of generalized Kähler metrics 389

cohomology class c1 (L). We first assume that L admits a holomorphic

structure ∂ 0 with respect to the “initial” complex structure I = I0 . The
associated Chern connection will be called ∇0 , and its curvature denoted F0 .
Recall that ∇0 is the unique Hermitian connection on L such that ∇0,1 0 = ∂ 0.
2. We then assume that L admits the structure of a holomorphic Poisson
module with respect to a holomorphic Poisson structure σI on M , which
by Proposition 18.9 occurs if and only if [σI F0 ] ∈ HI1 (T1,0 ) vanishes and the
secondary characteristic class in Hσ2I (M ) also vanishes. By Proposition 18.2,
we construct the Hermitian generalized connection D associated to this gen-
eralized holomorphic structure, and decompose it according to the splitting
T ⊕ T ∗:

D = ∇0 + iX,

where X is a real Q-Poisson vector field such that ∂X 1,0 = σI F0 , giving rise
to the real equations

LX Q = 0,
(18.25)
LX I0 = QF0 .

3. Let ϕt be the time t flow of the vector field X. Then we may transport F0 by
the flow, yielding the cohomologous family of two-forms Ft = ϕ∗−t F0 , which
satisfies

Ḟt = LX Ft = diX Ft .

We may also transport I0 by the flow, obtaining a family It = I0ϕt satisfying

I˙t = LX It = QFt ,

by (18.25). Note that Ft is type (1, 1) with respect to It . Also note that Ft
is the curvature of the family of connections
 t
∇t = ∇0 + iX Fs ds,
0

which are therefore the Chern connections associated to a family of holo-

morphic structures ∂ t on L, each holomorphic with respect to It .
4. We then compute the difference
 t
It − I0 = QFs ds
0
 t
= tQ 1t Fs ds (18.26)
0

= tQF t ,
390 Branes on Poisson varieties

where F t is the curvature of the average Chern connection on L:

 t
∇t = t1
∇s ds.
0

Setting t = 1 we obtain a solution to the first part of (18.24):

I1 − I0 = QF 1 .
5. Observe that the second part of (18.24) is automatically satisfied:
from (18.26) we have It − I0 = QGt , where
 t
Gt = Fs ds.
0

For t = 0, the expression

Gt It + I0∗ Gt (18.27)
vanishes, since G0 = 0. Taking the time derivative, we obtain
Ġt It + Gt I˙t + I0∗ Ġt = Ft It + Gt QFt + I0∗ Ft
= −(It∗ − I0∗ )Ft + Gt QFt
= −Gt QFt + Gt QFt = 0.
Therefore (18.27) vanishes for all t; since Ft = t−1 Gt , we obtain the result.
6. Positivity: If F0 is positive, that is, if the original line bundle L is positive,
then F t is positive for sufficiently small t. By (18.26), this gives a solution
to the system (18.24) for the Poisson structure tσI replacing σI .
We summarize the main result of this construction in the following:
Theorem 18.4 Let L be a positive holomorphic line bundle with Poisson
module structure over a compact complex manifold with holomorphic Poisson
structure σ. Let (g0 , I0 ) be the original Kähler structure it determines. Then the
choice of Hermitian structure on L determines a canonical family of generalized
Kähler structures {(gt , It , I0 ) : − < t < } such that the complex structure It
coincides with I0 only along the vanishing locus of σ for t = 0.
Example 18.6 One case where the existence of a positive Poisson module is
guaranteed is in the case of a Fano Poisson manifold, since the anti-canonical
bundle, which always admits a Poisson module structure, is positive. This extends
the result of Hitchin (2007), who showed that all smooth Fano surfaces (the Del
Pezzo surfaces) admit the families of generalized Kähler structures described
here.
We remark finally upon the relation of our construction to Hitchin’s result
for Del Pezzo surfaces (Hitchin 2007). To obtain the family of generalized
Kähler structures, he used a flow generated by a Poisson vector field X which
 Relation to non-commutative algebraic geometry 391

he expressed as the Hamiltonian vector field of log |s|2 , for s a holomorphic

section of the anti-canonical bundle vanishing at the zero locus of the Poisson
structure. From our point of view, he was making use of the generalized Poincaré–
Lelong formula of Proposition (18.3), since in the two-dimensional case there
is always a non-trivial generalized holomorphic section of the anti-canonical
bundle of a Poisson surface, namely, the Poisson structure itself. However, in
higher dimension, there is a dearth of global generalized holomorphic sections;
indeed by Proposition (18.3), such a section (if generically nonzero) must vanish
only along the zero locus of σ, which has codimension greater than one in
general.

18.8 Relation to non-commutative algebraic geometry

Since the observation in Gualtieri (2004) that the deformation space of a complex
manifold as a generalized complex manifold includes the “non-commutative”
directions in H 0 (∧2 T1,0 ), it was hoped that there might be a more precise
relationship between generalized complex structures and non-commutativity.
The presence of an underlying Poisson structure, for example, lends credence
to this idea. In the realm of generalized Kähler four-manifolds, we have
even more evidence in this direction, since, as observed originally in Apos-
tolov, Gauduchon, and Grantcharov (1999), the locus where the bi-Hermitian
complex structures (I+ , I− ) coincide is an anti-canonical divisor for both
structures.
If smooth, each connected component of this coincidence locus is an ellip-
tic curve C, and we may view it as embedded in two different complex
manifolds X± = (M, I± ).

X− X+
`AA >
AA }}
AA }}}
AA }
}}
C

In the examples constructed in Section 18.7 and by Hitchin (2007), X± have

natural holomorphic line bundles sitting over them, which we called L0 , L1 .
Pulling them back to C, we obtain holomorphic line bundles L0 , L1 over C.
Furthermore, since the flow satisfied LX It = QFt , we know that the flow restricts
to a holomorphic flow on C, the vanishing locus of Q. As a result, L0 , L1 are
related by an automorphism of C. This data (C, L0 , L1 ) is precisely what is
used in the approach of Bondal and Polishchuk (1993) to the classification of
Z-algebras describing non-commutative projective surfaces.
In fact, in our construction, we produce an example of an automorphism
(ϕ, F ) = (ϕ1 , F 1 ) in the sense of Proposition 18.6. Therefore, we may apply it
successively, producing an infinite family of generalized complex submanifolds
ϕk
with induced complex structures {Ik = I0 1 }, each Ik separated from I0 by the
392 Branes on Poisson varieties

line bundle Lk with connection ∇k , and all coinciding on the vanishing locus C
of Q. As a result, we obtain an infinite family of embeddings
k
F Fϕ Fϕ
··· / (M, I0 ) / (M, I1 ) / ··· / (M, Ik+1 ) / ···
cFF O : gggggg3
FF t g
FF tt ggggg
FF tttt
gggg ggggg
FF tt gggggg
gtgggg
C

Where the arrows on the top row indicate morphisms in the sense of the groupoid
of Definition 18.7. This may provide an alternative interpretation of Van den
Bergh’s construction of the twisted homogeneous coordinate ring (see Stafford
and van den Bergh 2001): let L = L0 |C , and let Lϕ = (ϕ−1 )∗ L. Then define the
vector spaces
i i+1 j−1
Hom(i, j) = H 0 (C, Lϕ ⊗ Lϕ ⊗ · · · ⊗ Lϕ )
and define a Z>0 -graded algebra structure on
9
A• = Hom(0, k), (18.28)
k≥0

via the multiplication, for a ∈ Ap and b ∈ Aq :

p
a · b = a ⊗ bϕ ,
p
where we use the natural map b → bϕ taking Hom(0, q) −→ Hom(p, p + q), and
the tensor product is viewed as a composition of morphisms.
Of course this is nothing but a recasting of the Van den Bergh construction;
there is a sense in which it captures only certain morphisms between the gen-
eralized complex submanifolds, namely, those which are visible upon restriction
to C. Though rare, there are sometimes generalized holomorphic sections of
the bundles Lk supported over all of M . In some sense, these sections must be
included in the morphism spaces as well.
For instance, performing our construction for L = O(1) over CP 2 , equipped
with a holomorphic Poisson structure σ ∈ H 0 (CP 2 , O(3)) with smooth zero
locus ι : C → CP 2 , the graded algebra (18.28) has linear growth instead of the
quadratic growth needed to capture a full non-commutative deformation of the
coordinate ring of CP 2 (these are the Sklyanin algebras, classified by Artin, Tate
and Van den Bergh 1991). It fails to include an additional generator in degree
3, as can be seen from the fact that the restriction map H 0 (CP 2 , O(3)) −→
H 0 (C, ι∗ O(3)) has one-dimensional kernel. However, it is important to note that
neither O(1) nor O(2) has generalized holomorphic sections over CP 2 , while
O(3) has a one-dimensional space of them. We end with this vague indication
that the morphisms supported on C should be combined with those supported
on the whole holomorphic Poisson manifold.
 References 393

Acknowledgements
I would like to thank Nigel Hitchin for many illuminating discussions, and also
the organizers of his birthday conference, especially Luis Álvarez-Cónsul and
Oscar Garcı́a-Prada, for a stimulating event. I would also like to thank Mike
Artin, Gil Cavalcanti, Izzet Coskun, and Pavel Etingof for helpful discussions.
Finally, I thank Yicao Wang for pointing out an error in a previous draft.

References
Apostolov, V., Gauduchon, P., and Grantcharov, G. (1999). Bihermitian structures on
complex surfaces, Proc. London Math. Soc. 79(3), 414–28.
Artin, M., Tate, J., and Van den Bergh, M. (1991). Modules over regular algebras of
dimension 3, Invent. Math. 106(2), 335–88.
Bondal, A. I. and Polishchuk, A. E. (1993). Homological properties of associative
algebras: the method of helices, Izv. Ross. Akad. Nauk Ser. Mat. 57(2), 3–50.
Brylinski, J.-L. (1993). Loop Spaces, Characteristic Classes and Geometric Quantiza-
tion, Vol. 107 of Progress in Mathematics, Birkhäuser Boston Inc., Boston, MA.
Bursztyn, H. and Radko, O. (2003). Gauge equivalence of Dirac structures and sym-
plectic groupoids, Ann. Inst. Fourier (Grenoble) 53(1), 309–37.
Courant, T. (1990). Dirac manifolds, Trans. Am. Math. Soc. 319, 631–61.
Ellwood, I. T. (2007). NS-NS fluxes in Hitchin’s generalized geometry, J. High Energy
Phys. 12, 084, 1–24.
Evens, S., Lu, J.-H., and Weinstein, A. (1999). Transverse measures, the modular
class and a cohomology pairing for Lie algebroids, Q. J. Math. Oxford Ser. (2)
50(200), 417–36.
Goto, R. (2007). Deformations of generalized complex and generalized Kähler struc-
tures, arXiv:0705.2495v2.
Gualtieri, M. (2004). Generalized complex geometry, DPhil thesis, Oxford University,
Oxford, math.DG/0401221.
Gualtieri, M. (2007). Generalized complex geometry, arXiv:math/0703298v2.
Hitchin, N. (1987). The self-duality equations on a Riemann surface, Proc. London
Math. Soc. (3) 55(1), 59–126.
Hitchin, N. (2003). Generalized Calabi-Yau manifolds, Q. J. Math. 54, 281–308.
Hitchin, N. (2006a). Brackets, forms and invariant functionals, Asian J. Math.
10(3), 541–60.
Hitchin, N. (2006b). Instantons, Poisson structures and generalized Kähler geometry,
Commun. Math. Phys. 265(1), 131–64.
Hitchin, N. (2007). Bihermitian metrics on Del Pezzo surfaces, J. Symplectic Geom.
5, 1–7.
Kapustin, A. (2005). A-branes and noncommutative geometry, hep-th/0502212.
Kapustin, A. and Orlov, D. (2003). Remarks on A-branes, mirror symmetry, and the
Fukaya category, J. Geom. Phys. 48, 84.
Lichnerowicz, A. (1977). Les variétés de Poisson et leurs algèbres de Lie associées,
J. Diff. Geom. 12(2), 253–300.
Lindström, U., Roček, M., von Unge, R., and Zabzine, M. (2007). Generalized Kähler
manifolds and off-shell supersymmetry, Commun. Math. Phys. 269(3), 833–49.
394 Branes on Poisson varieties

Liu, Z.-J., Weinstein, A., and Xu, P. (1997). Manin triples for Lie bialgebroids, J. Diff.
Geom. 45, 547–74.
Polishchuk, A. (1997). Algebraic geometry of Poisson brackets, J. Math. Sci.
(New York) 84(5), 1413–44. Algebraic Geometry, 7.
Roytenberg, D. (1999). Courant algebroids, derived brackets and even symplectic
supermanifolds, PhD thesis, University of California, Berkeley, CA.
Simpson, C. T. (1992). Higgs bundles and local systems, Inst. Hautes Études Sci. Publ.
Math. (75), 5–95.
Stafford, J. T. and van den Bergh, M. (2001). Noncommutative curves and noncom-
mutative surfaces, Bull. Am. Math. Soc. (N.S.) 38(2), 171–216 (electronic).
Ševera, P. (1998–2000). Letters to A. Weinstein concerning Courant algebroids. Unpub-
lished.
Weinstein, A. (1997). The modular automorphism group of a Poisson manifold,
J. Geom. Phys. 23, 379–94.
 XIX
CONSISTENT ORIENTATION OF MODULI SPACES

Daniel S. Freed, Michael J. Hopkins, and Constantin Teleman

For Nigel, with admiration

In a series of papers by Freed, Hopkins, and Teleman (2003, 2005, 2007a)

we develop the relationship between positive energy representations of the
loop group of a compact Lie group G and the twisted equivariant K-theory
τ +dim G
KG (G). Here G acts on itself by conjugation. The loop group represen-
tations depend on a choice of “level,” and the twisting τ is derived from the
level. For all levels the main theorem is an isomorphism of abelian groups,
and for special transgressed levels it is an isomorphism of rings: the fusion
τ +dim G
ring of the loop group and KG (G) as a ring. For G connected with
π1 G torsionfree we prove in Freed et al. (2007a, section 4 and 2007b, section
τ +dim G
7) that the ring KG (G) is a quotient of the representation ring of G
and we calculate it explicitly. In these cases it agrees with the fusion ring
of the corresponding centrally extended loop group. We also treat G = SO3
in Freed et al. (2007b (A.10)). In this chapter we explicate the multiplication
on the twisted equivariant K-theory for an arbitrary compact Lie group G.
We work purely in topology; loop groups do not appear. In fact, we construct
τ +dim G
a Frobenius ring structure on KG (G). This is best expressed in the
language of topological quantum field theory: we construct a two-dimensional
TQFT over the integers in which the abelian group attached to the circle is
τ +dim G
KG (G).
At first glance the ring structure seems apparent. The multiplication map µ :
G × G → G induces a pushforward on K-theory: the Pontrjagin product. But in
K-cohomology the pushforward is the wrong-way, or umkehr, map. Thus to define
it we must K-orient the map µ. Furthermore, the twistings must be accounted
for in the orientations. Finally, to ensure associativity we must consistently K-
orient maps constructed from µ by iterated composition. For connected and
simply connected groups there is essentially a unique choice, but in general one
must work more. This orientation problem is neatly formulated in the language
of topological quantum field theory. Cartesian products of G then appear as
moduli spaces of flat connections on surfaces, and the maps along which we
push forward are restriction maps of the connections to the boundary. What is
required, then, is a consistent orientation of these moduli spaces and restriction
maps. The existence of consistent orientations, which we prove in Theorem 19.1,
396 Consistent orientation of moduli spaces

is in some sense due to the Narasimhan–Seshadri theorem which identifies moduli

spaces of flat connections with complex manifolds of stable bundles: complex
manifolds carry a canonical orientation in K-theory. Our proof, though, uses
only the much more simple linear statement that the symbol of the de Rham
complex on a surface is the complexification of the symbol of the Dolbeault
complex. As we explain in Section 19.1, which serves as a heuristic introduction
and motivation, “consistent orientations on moduli spaces” is the topological
analog of “consistent measures on spaces of fields” in quantum field theory.
The latter is what one would like to construct in the path integral approach
to quantum field theory.
Our topological construction, outlined in Section 19.3, proceeds via a universal
orientation (Definition 19.1). The main observation is that the problem of consis-
tent orientations is a bordism problem, and the relevant bordism groups are those
constructed by Madsen and Tillmann (2001) in their formulation of the Mumford
conjectures (see Madsen and Weiss 2005 and Galatius et al. 2009 for proofs and
generalizations). A universal orientation induces a level (Definition 19.2). The
map from universal orientations to levels is an isomorphism for simply connected
and connected compact Lie groups G, but in general it may fail to be injective,
surjective, or both. The theories we construct are parametrized by universal
orientations, not by levels. It is interesting to ask whether universal orientations
also appear in related topological and conformal field theories as a refinement of
the level.
The two-dimensional TQFT we construct here is the dimensional reduction
of three-dimensional Chern–Simons theory, refined to have base ring Z in place
of C. Our construction is a priori in the sense that the axioms of TQFT—the
topological invariance and gluing laws—are deduced directly from the definition.
By contrast, rigorous constructions of many other TQFTs, such as the Chern–
Simons theory, proceed via generators and relations. Such constructions are
based on general theorems which tell that these generators and relations generate
a TQFT: gluing laws and topological invariance are satisfied. One can ask if there
is an a priori topological construction of Chern–Simons theory using twisted K-
theory. We do not know of one. In another direction we can extend TQFTs to
lower dimension, so look for a theory in 0-1-2 dimensions which extends the 1-2
dimensional theory constructed here. Again, we do not know a construction of
that extended theory.
Section 19.2 of this chapter is an exposition of twistings and orientation,
beginning on familiar ground with densities in differential geometry. Section 19.4
briefly considers this TQFT for families of one- and two-manifolds. Our purpose
is to highlight an extra twist which occurs: that theory is “anomalous.”
As far as we know, the problem of consistently orienting moduli spaces first
arises in work of Donaldson (Donaldson 1987; Donaldson and Kronheimer 1990).
He works with anti-self-dual connections on a four-manifold and uses excision
in index theory to relate all of the different moduli spaces. In both his situation
 Push–pull construction of TQFT 397

and ours the moduli spaces in question sit inside infinite-dimensional function
spaces, and the virtual tangent bundle to the moduli space extends to a virtual
bundle on the function space. Thus it suffices to orient over the function space,
and this becomes a universal problem. Presumably our methods apply to his
situation as well, but we have not worked out the details.
It is a pleasure and an honor to dedicate this chapter to Nigel Hitchin. We
greatly admire his mathematical taste, style, and influence. ¡Feliz cumpleaños y
que cumplas muchos más!

19.1 Push–pull construction of TQFT

19.1.1 Quantum field theory
The basic structure of an n-dimensional Euclidean quantum field theory may
be axiomatized simply. Let BRiemn be the bordism category whose objects
are closed oriented (n − 1)-dimensional Riemannian manifolds. A morphism
X : Y0 → Y1 is a compact-oriented n-dimensional Riemannian manifold X
together with an orientation-preserving isometry of its boundary to the disjoint
union −Y0 ! Y1 , where −Y0 is the oppositely oriented manifold. We term Y0
the incoming boundary and Y1 the outgoing boundary. A quantum field theory
is a functor from BRiemn to the category of Hilbert spaces and trace class
maps. The functoriality encodes the gluing law; there is also a symmetric
monoidal structure which encodes the behavior under disjoint unions. There
are many details and subtleties (e.g. see Chapter 9), but our concern is a
simpler topological version. Thus we replace BRiemn by the bordism cate-
gory BSO n of smooth-oriented manifolds and consider orientation-preserving
diffeomorphisms in place of isometries. We define an n-dimensional topological
quantum field theory (TQFT) to be a functor from BSOn to the category of
complex vector spaces. The functor is required to be monoidal: disjoint unions
map to tensor products. The functoriality expresses the usual gluing law and
the structure of the domain category BSOn encodes the topological invariance.
The example of interest here has an integral structure: the codomain is the
category of abelian groups rather than complex vector spaces. The integral-
ity reflects that the theory is a dimensional reduction (see Freed 2009 for a
discussion).
Physicists often employ a path integral to construct a quantum field theory.
Here is a cartoon version. To each manifold M is attached a space FM of fields
and so to a bordism X : Y0 → Y1 a correspondence diagram:

FX (19.1)
DD
s zzz DD t
z DD
zz DD
z} z !
FY0 FY1
398 Consistent orientation of moduli spaces

in which s, t are restriction maps. The important property of fields is locality: in

the diagram

FX  ◦X (19.2)
u II
u II r
uu
r
u II
uu II
zuu I$
FX FX 
u II II
uu II t s uu II t
s
uuu II
II uuu II
uu II uu II
uz u $ uz u I$
FY0 FY1 FY2

the space of fields FX  ◦X on the composition of bordisms X : Y0 → Y1 and

X  : Y1 → Y2 is the fiber product of the maps t, s . Fields are really infinite-
dimensional stacks—for example, in gauge theories the gauge transformations
act as morphisms of fields—and the maps and fiber products must be understood
in that sense.
The backdrop for the path integral is measure theory. If there exist mea-
sures µX , µY on the spaces FX , FY with appropriate gluing properties, then one
can construct a quantum field theory. Namely, define the Hilbert space HY =
L2 (FY , µY ) and the linear map attached to a bordism X : Y0 → Y1 as the push–
pull
ZX = t∗ ◦ s∗ : HY0 −→ HY1 .
The pushforward t∗ is integration. Thus if f ∈ L2 (FY0 , µY0 ) and g ∈
L2 (FY1 , µY1 ), then

   
ḡ, ZX (f )HY = g t(Φ) f s(Φ) dµX (Φ).
1
FX

One usually postulates an action functional SX : FX → C and a measure µ̃X

such that µX = e−SX µ̃X and the action satisfies the gluing law
   
SX  ◦X (Φ) = SX r(Φ) + SX  r (Φ)
in (19.2). These measures have not been constructed in most examples of
geometric interest.

19.1.2 Topological construction

Our idea is to replace the infinite-dimensional stack FX by a finite-dimensional
stack MX ⊂ FX of solutions to a first-order partial differential equation and to
shift from measure theory to algebraic topology. Examples of finite-dimensional
moduli spaces MX in supersymmetric field theory include anti-self-dual con-
nections in four dimensions and holomorphic maps in two dimensions. From
 Push–pull construction of TQFT 399

the physical point of view the differential equations are the BPS (Bogomolńyi-
Prasad-Sommerfield) equations of supersymmetry; from a mathematical point
of view they define the minima of a calculus of variations functional. In this
chapter we consider pure gauge theories. Fix a compact Lie group G and for any
manifold M let FM denote the stack of G-connections on M . Define MM as the
stack of flat G-connections on M . If we choose a set {mi } ⊂ M of “basepoints,”
one for each
M Ncomponent
 of M , then
 OMM is represented by the product of
groupoids i Hom π1 (M, mi ), G //G . A basic property of flat connections is
the gluing law (see (19.2)).
Lemma 19.1 Suppose X : Y0 → Y1 and X  : Y1 → Y2 are bordisms of smooth
manifolds. Then MX  ◦X is the fiber product of

MX MX  (19.3)
GG
GG t 
s www
GG w
GG ww
# w{ w
MY

Roughly speaking, this says that given flat connections on X, X  and an isomor-
phism of their restrictions to Y , one can construct a flat connection on X  ◦ X
and every flat connection on X  ◦ X comes this way.
Replace the infinite-dimensional correspondence diagram (19.1) with the finite-
dimensional correspondence diagram of flat connections:

MX (19.4)
FF
xx FF t
xx
s
x FF
x FF
x| x "
MY0 MY1

Whereas the path integral linearizes (19.1) using measure theory, we propose
instead to linearize (19.4) using algebraic topology. Let E be a generalized
cohomology theory. To every closed (n − 1)-manifold we assign the abelian group
AY = E • (MY ).
To a morphism X : Y0 → Y1 we would like to attach a homomorphism ZX :
AY0 → AY1 defined as the push–pull
ZX := t∗ ◦ s∗ : E • (MY0 ) −→ E • (MY1 ) (19.5)
in E-cohomology. Whereas the path integral requires measures consistent under
gluing to define integration t∗ , in our topological setting we require orientations
of t consistent with gluing to define pushforward t∗ . The consistency of orienta-
tions under gluing ensures that (19.5) defines a TQFT which satisfies the gluing
law (functoriality).
400 Consistent orientation of moduli spaces

This, then, is the goal of the chapter: we formulate the algebro-topological

home for consistent orientations and study a particular example. Namely,
specialize to n = 2 and require that the one-manifolds Y and two-manifolds X
be oriented. In other words, the domain category of our TQFT is BSO2 . For
Y = S 1 the moduli stack of flat connections is the global quotient

MY ∼
= G//G

of G by its adjoint action; the isomorphism is the holonomy of a flat connection

around the circle. Take the cohomology theory E to be complex K-theory.
The resulting two-dimensional TQFT on oriented manifolds is the dimensional
reduction of three-dimensional Chern–Simons theory for the group G. In this case
there is a map from consistent orientations to “levels” on G; the level is what
is usually used to describe Chern–Simons theory. A two-dimensional TQFT on
oriented manifolds determines a Frobenius ring and conversely. The Frobenius
ring constructed here is the Verlinde ring attached to the loop group of G. The
abelian group AS 1 is a twisted form of K(G//G) = KG (G) and its relation to
positive energy representations of the loop group is developed in Freed et al.
(2003, 2005, 2007a). In this chapter we describe a topological construction of the
ring structure.

19.1.3 Remarks
r Let X be the “pair of pants” with the two legs incoming and the single waist
outgoing. Then restriction to the outgoing boundary is the map t : (G ×
G)//G → G//G induced by multiplication µ : G × G → G. So ZX = t∗ ◦ s∗ ,
which defines the ring structure in a two-dimensional TQFT, is pushforward
by multiplication on G. Therefore, we do construct the Pontrjagin product
τ +dim G
on KG (G)—here τ is the twisting and there is a degree shift as well—
and have implicitly used an isomorphism of twistings µ∗ τ → τ ⊗ 1 + 1 ⊗ τ
which, since the TQFT guarantees an associative product, satisfies a com-
patibility condition for triple products. This isomorphism and compatibility
are embedded in our consistent orientation construction.
r We do not use the theorem (Abrams 1996, Moore and Segal 2006) which
constructs a two-dimensional TQFT from a Frobenius ring. Rather, our a
priori construction manifestly produces a TQFT which satisfies the gluing
law, and we deduce the Frobenius ring as a derived quantity.
r Three-dimensional Chern–Simons theory is defined on a bordism category of
manifolds which carry an extra topological structure. For oriented manifolds
this extra structure is described as a trivialization of p1 or signature, or a
certain sort of framing. (For spin manifolds it is described as a string struc-
ture or, since we are in sufficiently low dimensions, an ordinary framing.)
The two-dimensional reduction constructed here factors through the bordism
categories of oriented manifolds.
 Orientation and twisting 401

r The topological push–pull construction extends to families of bordisms

parametrized by a base manifold S. A choice of consistent orientation
determines this extension to a theory for families of manifolds, albeit an
“anomalous” theory (see Section 19.4 for a discussion).
r The pushforward t∗ is only defined if t : MX → MY is a representable
1
map of stacks, that is, only if the fibers of t are spaces—no automorphisms
allowed. This happens only if each component of X has a nonempty outgoing
boundary. Therefore, the push–pull construction only gives a partial TQFT.
We complete to a full TQFT using the nondegeneracy of a certain bi-additive
form (see Freed et al. 2003, section 17).
r As mentioned earlier, a standard TQFT is defined over the ring C whereas
this theory, being a dimensional reduction of a three-dimensional theory,
is defined over Z. It is possible to go further and refine the push–pull
construction to obtain a theory over K, where K is the K-theory ring
spectrum (see Freed 2009 for further discussion).
r The theory constructed here has two tiers—it concerns one- and two-
manifolds—so could be termed a “1-2 theory.” Extensions to 0-1-2 theories,
which have three tiers, are of great interest. The general structure of such
theories has been much studied recently in various guises (Moore and Segal
2006; Costello 2007; Hopkins and Lurie in preparation). A theory defined
down to points is completely local, and so ultimately has a simpler structure
than less local theories. We do not know if the push–pull construction here
can be extended to construct a 0-1-2 theory.

19.2 Orientation and twisting

19.2.1 Ordinary cohomology
The first example for a differential geometer is de Rham theory. Let M be
a smooth manifold and suppose it has a dimension equal to n. An orienta-
tion on M , which is an orientation of the tangent bundle T M , enables inte-
gration

: Ωnc (M ) −→ R
M

on forms of compact support. Absent an orientation we can integrate twisted

forms, or densities. The twisting is defined as follows. For any real vector space V
of dimension r let B(V ) denote the GLr R-torsor of bases of V . There is an
associated Z-graded real line o(V ) of functions f : B(V ) → R which satisfy f (b ·
A) = sign det A · f (b) for b ∈ B(V ), A ∈ GLr R; the degree of o(V ) is r. Applied
fiberwise this construction yields a flat Z-graded line bundle o(V ) → M for a
real vector bundle V → M . There is a twisted de Rham complex
d d d
0 −→ Ωo(V )−r (M ) −−→ Ωo(V )−r+1 (M ) −−→ · · · −−→ Ωo(V ) (M ) −→ 0 (19.6)
402 Consistent orientation of moduli spaces

where Ωo(V )+q (M ) is the space of smooth sections of the ungraded vector bundle
P r+q ∗
T M ⊗ o(V ). The cohomology of (19.6) is the twisted de Rham cohomology
o(V )+•
HdR (M ). Let o(M ) = o(T M ). Then integration is a map

: Ωo(M
c
)
(M ) −→ R. (19.7)
M

Notice this formulation notation works if M has several components of vary-

ing dimension: the degree of o(M ) is then the locally constant function
dim M : M → Z.
A similar construction works in integer cohomology. If π : V → M is a real
vector bundle over a space M (which need not be a manifold) we define o(V ) as
the orientation double cover of M determined by V and introduce a Z-grading
according to the rank of V . (Note rank V : M → Z is a locally constant function.)
There is an o(V )-twisted singular complex analogous to (19.6): cochains in this
complex are cochains on the double cover which change sign under the deck
transformation. The equivalence class of the twisting o(V ) is
N O  
o(V ) = rank V, w1 (V ) ∈ H 0 (M ; Z) × H 1 (M ; Z/2Z),

where w1 is the Stiefel–Whitney class. The relationship of the twisting to

π ∗ o(V )
integration occurs in the Thom isomorphism. The Thom class U ∈ Hcv (V )
lies in twisted cohomology with compact vertical support. Let B(V ) and S(V )
be the ball and sphere bundles
 relative to a metric on V , respectively. The
Thom space M V is the pair B(V ), S(V ) or equivalently, assuming M is a CW
complex, the quotient B(V )/S(V ). The composite
π∗ ∗
H −o(V )+• (M ) −−−→ H −π (V ) −−−→ H • (M V )
o(V )+• U

is an isomorphism—the Thom isomorphism—a generalization of the suspension

isomorphism. If M is a compact manifold and i : M → S n a (Whitney) embed-
ding with normal bundle ν → M , then the Pontrjagin–Thom collapse is the
map c : S n → M ν defined by identifying ν with a tubular neighborhood of M
and sending the complement of B(ν) in S n to the basepoint of M ν . Integration
is then the composite
c∗ suspension
H −o(ν)+n (M ) −−−
Thom
∼
−→ H n (M ν ) −−−→ H n (S n ) −−−−∼−−−→ Z. (19.8)
= =

Twistings obey a Whitney sum formula: there is a natural isomorphism

∼
=
o(V1 ⊕ V2 ) −−−→ o(V1 ) + o(V2 ).

Applied to T M ⊕ ν = n, where n is the trivial bundle of rank n, we conclude

that integration (19.8) is a map (compare (19.7))

H o(M ) (M ) −→ Z.
 Orientation and twisting 403

More generally, if p : M → N is a proper map there is a pushforward

p∗ : H o(p)+• (M ) −→ H • (N ), (19.9)
where o(p) = o(M ) − p∗ o(N ).

19.2.2 K-theory
This discussion applies to any multiplicative cohomology theory. 1 The only issue
is to determine the twisting of a real vector bundle in that theory. For complex
K-theory there are many possible models for the twisting τV of a vector bundle
V → M . In the Donovan–Karoubi picture (Donovan and Karoubi 1970) τV is
represented by the bundle of complex Z/2Z-graded Clifford algebras defined
by V . A bundle of algebras A → M of this type is considered trivial if A =
End(W ) for a Z/2Z-graded complex vector bundle W → M , that is, if A is
Morita equivalent to the trivial bundle of algebras M × C. The equivalence class
of τV is
 
[τV ] = rank V, w1 (V ), W3 (V ) ∈ H 0 (M ; Z/2Z) × H 1 (M ; Z/2Z) × H 3 (M ; Z).
(19.10)
Only torsion classes in H 3 (M ; Z) are realized by bundles of finite-dimensional
algebras, but we have in mind a larger model which includes nontorsion
classes. (Such models are developed in Atiyah and Segal 2006, Freed et al.
2007a, and Murray 1996 among other works.) There is a Whitney sum
isomorphism
∼
=
τV1 ⊕V2 −−−→ τV1 + τV2 ; (19.11)
the sum of twistings is realized by the tensor product of algebras. A spinc
structure on V induces an orientation, that is, a Morita equivalence
∼
=
τrank V −−−→ τV . (19.12)
An A-twisted vector bundle is a vector bundle with an A-module structure; it
represents an element of twisted K-theory.
The Whitney formula (19.11) allows us to attach a twisting to any virtual real
vector bundle: set
τ−V = −τV . (19.13)

Since the Thom space satisfies the stability condition X V ⊕n ∼

= Σn X V , where
Σ denotes suspension, there is also a Thom spectrum attached to any virtual
vector bundle and a corresponding Thom isomorphism theorem. An orientation,
which is an isomorphism as in (19.12), is equivalently a trivialization of the
twisting attached to the reduced bundle (V − rank V ).

1 Also to a cohomology theory defined by a module over a ring spectrum.

404 Consistent orientation of moduli spaces

Suppose τ is any twisting on a manifold N . We can put that extra twist-

ing into the pushforward in K-theory associated to a proper map p : M → N
(compare (19.9)):
∗
p∗ : K (τp +p τ )+•
(M ) −→ K τ +• (N ). (19.14)
∗
Here τp is the twisting τp = τM − p τN of the relative tangent bundle. In the next
section we encounter a situation in which τp + p∗ τ is trivialized, and so construct
a pushforward from untwisted K-theory to twisted K-theory (see (19.38)).
Twistings of K • (pt) form a symmetric monoidal two-groupoid; its classifying
space Picg K is thus an infinite loop space. The notation: Pic K is the classify-
ing space of invertible K-modules and Picg K the subspace classifying certain
“geometric” invertible K-modules including twisted forms of K-theory defined
by real vector bundles. As a space there is a homotopy equivalence
Picg K ∼ K(Z/2Z, 0) × K(Z/2Z, 1) × K(Z, 3) (19.15)
with a product of Eilenberg–MacLane spaces, but the group structure on Picg K
is not a product. The group of equivalence classes of twistings on M is the
group of homotopy classes of maps [M, Picg K], which as a set is the product of
cohomology groups in (19.10).
Let picg K denote the spectrum whose 0-space is Picg K and which is a
Postnikov section of the real KO-theory spectrum: the “Z/2, Z/2, 0, Z” bit
of the “. . . , Z/2, Z/2, 0, Z, 0, 0, 0, Z, . . . ” song. Thus the one-space B Picg K
of the spectrum picg K is a Postnikov section of BO. Also, let ko denote the
connective KO-theory spectrum. Its zero-space is the group completion of the
classifying space of the symmetric monoidal category of real vector spaces of
finite dimension (Segal 1977). The map which attaches a twisting of K • (pt) to
a real vector space, say via the Clifford algebra, induces a spectrum map
τ : ko −→ picg K. (19.16)
Remark 19.1 If M is a smooth stack, presented as the groupoid of a Lie
group acting on a smooth manifold, then its tangent space is presented as
a graded vector bundle. An orientation of M is then an orientation of this
graded bundle, viewed as a virtual bundle by taking the alternating sum of the
homogeneous terms. The virtual tangent bundle is classified by a map M → ko
whose composition with (19.16) gives the induced twisting of complex K-theory.
We apply this in next section 19.3 to the moduli space of flat connections on a
fixed oriented two-manifold. In that case the virtual tangent bundle is the index
of an elliptic complex and the map M → ko is computed from the Atiyah–Singer
index theorem (Atiyah and Singer 1971).
The Z part of Picg K, the Eilenberg–MacLane space K(Z, 3), has an attractive
geometric model: gerbes. Nigel Hitchin (2001) has developed beautiful appli-
cations of gerbes in differential geometry. There is a gerbe model for Picg K
too—one need only add Z/2Z-gradings.
 Universal orientations and consistent orientations 405

19.3 Universal orientations and consistent orientations

19.3.1 Overview
In this section, the heart of the chapter, we define universal orientations
(Definition 19.1) and prove that they exist (Theorem 19.1). A universal orien-
tation simultaneously orients the maps t in (19.4) along which we pushforward
classes in twisted K-theory (see (19.40) for the precise push–pull maps in the
theory). Universal orientations form a torsor for an abelian group (19.23). A
universal orientation determines a level (Definition 19.2), which is the quantity
typically used to label theories. The map (19.35) from universal orientations to
levels is not an isomorphism in general.
We begin with a closed oriented surface X. The virtual tangent space to
the stack MX of flat G-connections on X is the index of a twisted de Rham
complex (19.17), and we construct a universal symbol (19.20)—whence universal
index—for this operator. A trivialization of the universal twisting (19.22) is a
universal orientation, and it simultaneously orients the moduli stacks MX for
all closed oriented X.
For a surface X with (outgoing) boundary we must orient the restriction map
t : MX → M∂X on flat connections. It turns out that a universal orientation
does this, simultaneously and coherently for all X, as expressed in the isomor-
phism (19.37). An important step in the argument is Lemma 19.4, which uses
work of Atiyah and Bott (1964) to interpret the universal symbol in terms of
standard local boundary conditions for the de Rham complex.

19.3.2 Closed surfaces

Fix a compact Lie group G with Lie algebra g. Let X be a closed oriented two-
manifold and MX the moduli stack of flat G-connections on X. A point of MX
is represented by a flat connection A on a principal bundle P → X, and the
tangent space to MX at A by the deformation complex

d d
0 −→ Ω0X (gP ) −−−
A
→ Ω1X (gP ) −−−
A
→ Ω2X (gP ) −→ 0, (19.17)

the de Rham complex with coefficients in the adjoint bundle associated to P . This
is an elliptic complex. Its symbol σ satisfies the reality condition σ(−ξ) = σ(ξ)
for ξ ∈ T X, since (19.17) is a complex of real differential operators (Atiyah and
Singer 1971). Recall that the symbol of any complex differential operator lies
in Kcv0
(T X) ∼= K 0 (X T X ). The reality condition gives a lift σ ∈ KR0 (X iT X ),
where the imaginary tangent bundle iT X carries the involution of complex
conjugation (Atiyah 1966). Bott periodicity asserts that V ⊕ iV is canonically
KR-oriented for any real vector bundle V with no degree shift. In the language
of twistings of KR this means

(KR) (KR)
τV + τiV = 0. (19.18)
406 Consistent orientation of moduli spaces

Therefore
(KR) (19.18) (KR) (19.13)
−→ KR−τiT X (X) −−−−
Thom
KR0 (X iT X ) −−−
∼ ∼
−−→ KRτT X (X) −−−−
∼
−−→
= = =
(KR)
−τ−T X
KO0 (X −T X )
Thom
KR (X) −−−
∼
−→ (19.19)
=

from which we locate the symbol σ ∈ KO0 (X −T X ). Note that by Atiyah duality
KO0 (X −T X ) ∼
= KO0 (X) and the KO-homology group is well-known to be the
home of real elliptic operators.
Now (19.17) is a universal operator: its symbol is constructed from the exterior
algebra of T X and the adjoint representation of G; it does not depend on details
of the manifold X. Thus it is pulled back from a universal symbol. Let Vn →
BSOn denote the universal oriented n-plane bundle. The universal symbol lives
on the bundle V2 → BSO2 × BG, and by (19.19) we identify it as an element
σuniv ∈ KO0 (BSO2−V2 ∧ BG+ ). (19.20)
Here BG+ is the space BG with a disjoint basepoint adjoined and “∧” is the
smash product. Introduce the notation
M T SOn = BSOn−Vn
for this Thom spectrum and so write
σuniv ∈ KO 0 (M T SO2 ∧ BG+ ).
If f : X → BSO2 × BG is a classifying map for T X and P , and f˜ : X −T X →
MTSO2 ∧ BG+ the induced map on Thom spectra, then σ = f˜∗ σuniv . It is in
this sense that (19.20) is a universal symbol.
Remark 19.2 We digress to explain M T SOn in more detail. Let Grn+ (RN ) be
the Grassmannian of oriented n-planes in RN and
0 → Vn → N → QN −n → 0
the exact sequence of real vector bundles over Grn+ (RN ) in which Vn is the
universal subbundle and QN −n the universal quotient bundle. Denote the Thom
space of the latter as
ZN := Grn+ (RN )QN −n .
Then the suspension ΣZN is the Thom space of QN −n ⊕ 1 → Grn+ (RN ). But
QN −n ⊕ 1 is the pullback of QN +1−n → Grn+ (RN +1 ) under the natural inclusion
Grn+ (RN ) → Grn+ (RN +1 ), and in this manner we produce a map ΣZN → ZN +1 .
Whence the spectrum M T SOn = {ZN }N ≥0 . The notation identifies M T SOn
as an unstable version of the Thom spectrum M SO and also alludes to its
appearance in the work of Madsen and Tillmann (2001). There are analogous
spectra M T On , M T Spinn , M T Stringn , etc. If F : S N → ZN is transverse to
the 0-section, then X := F −1 (0-section) is an n-manifold and the pullback
 Universal orientations and consistent orientations 407

of QN −n − N is stably isomorphic to −T X. Thus a map 2 S → M T SOn classifies

a map M → S of relative dimension n together with a rank n bundle W → M
and a stable isomorphism 3 T (M/S) ∼= W . An important theorem of Madsen and
Weiss (2005) and Galatius et al. (2009) identifies M T SOn as a bordism theory
of fiber bundles rather than a bordism theory of arbitrary maps.
If a smooth manifold M parametrizes a family of flat G-connections on X—
that is, P → M × X is a G-bundle with a partial flat connection along X—then
there is a classifying map M → MX and the pullback of the stable tangent
bundle of MX to M is the index of the family of elliptic complexes (19.17).
Note that if we replace the adjoint bundle gP in (19.17) by the trivial bundle
of rank dim G then the resulting elliptic complex does not vary over M ; its
index is a trivializable bundle. Hence the reduced stable tangent bundle to MX
is computed by the de Rham complex coupled to the reduced adjoint bundle
ḡP = gP − dim G.
There is a corresponding reduced universal symbol (compare (19.20))
σ̄univ ∈ KO 0 (M T SO2 ∧ BG). (19.21)
It induces a universal twisting in K-theory and a consistent orientation is
constructed by trivializing this twisting.
Definition 19.1 A universal orientation is a null homotopy of the composition
σ̄ τ
M T SO2 ∧ BG −−univ
−−→ ko −−→ picg K. (19.22)
Two universal orientations are said to be equivalent if the null homotopies are
homotopy equivalent.
The set of equivalence classes of universal orientations is a torsor for the abelian
group
O(G) := [ΣM T SO2 ∧ BG, picg K]. (19.23)
We prove in Theorem 19.1 below that universal orientations exist. In fact, there
is a canonical universal orientation, so the torsor of universal orientations may
be naturally identified with the abelian group (19.23). Definition 19.1 is designed
to orient the moduli spaces attached to closed surfaces. In an equivalent form
it leads to the pushforward maps we need for surfaces with boundary and to
twisted K-theory of moduli spaces attached to the boundary (see the discussion
preceding (19.31)).
Return now to the family of partial G-connections on P → M × X. The bundle
P → M × X is classified by a map f : M → M T SO2 × BG and the Atiyah–
Singer index theorem (Atiyah and Singer 1971) implies that the index of the
2That is, a stable map from the suspension spectrum of S to M T SOn .
3Thom bordism theories, such as M SO, retain the information of the stable normal bundle.
Madsen–Tillmann theories, such as M T SOn , track the stable tangent bundle, which is one
more justification for the “T” in the notation.
408 Consistent orientation of moduli spaces

family of operators (19.17) is f ∗ σuniv . Therefore, a universal orientation pulls

back to an orientation of the index of (19.17) (c.f. Remark 19.1). It follows that
a universal orientation simultaneously orients MX for all closed oriented two-
manifolds X.

19.3.3 Surfaces with boundary

As a preliminary we observe two topological facts about M T SOn .
Lemma 19.2
(1) M T SO1  S −1 , the desuspension of the sphere spectrum.
(2) There is a cofibration

 b
/ M T SOn / / (BSOn )+ .
Σ−1 M T SOn−1 (19.24)

Proof. For (1) simply observe

M T SO1  BSO1−V1  pt−R  S −1 . (19.25)
For (2) begin with the cofibration built from the sphere and ball bundles of the
universal bundle:
  
S(Vn )+ / B(Vn )+ / / B(Vn ), S(Vn ) , (19.26)

Then identify BSOn−1 as the unit sphere bundle S(Vn ) and write (19.26) in
terms of Thom spaces:

0  / BSO0 / / BSOVn . (19.27)

BSOn−1 n n

Here 0 is the vector bundle of rank zero. Now add −Vn to each of the vector
bundles in (19.27) and note that the restriction of Vn to BSOn−1 is Vn−1 ⊕ 1.
Consider the diagram

b q
Σ−1 M T SO1 ∧ BG / M T SO2 ∧ BG / (M T SO2 , Σ−1 M T SO1 ) ∧ BG
TTTT
TTTT σ̄univ
TTTT 
TTTT σ̄univ
TTTT 
*
ko
(19.28)

The top row is a cofibration. From (19.24) we can replace (M T SO2 , Σ−1 M T SO1 )
with (BSO2 )+ .


Lemma 19.3 Define σ̄univ in (19.28) to be the map (BSO2 )+ ∧ BG → ko
induced by the standard representation of SO2 smash with the reduced adjoint
 Universal orientations and consistent orientations 409

representation of G. Then the triangle in (19.28) commutes and the diagram

gives a canonical null homotopy of the composite σ̄univ ◦ b.

Proof. Recalling the isomorphisms in (19.19), and using the fact that the
universal symbol σ̄univ is canonically associated to a representation of SO2 × G,
we locate σ̄univ ∈ KRSO 0
2 ×G
(−R2 )c ∼ = KRSO 0
2 ×G
(iR2 )c . (Recall that the involu-
2 ∼ 2
tion on C = R ⊕ iR is complex conjugation and the subscript “c” denotes
2

compact support.) Similarly, σ̄univ

∈ KRSO0
2 ×G
(pt) ∼
= KRSO 0
2
(C2 )c . Using the

ThomP isomorphism P we identify σ̄univ as the difference of classes represented
by • C2 ⊗ gC and • C2 ⊗ Cdim G , where in both summands θ ∈ C2 acts as
(θ) ⊗ id. Exterior multiplication (θ) is exact for θ = 0, so this does represent
a class with compact support. Also, as commutes with complex conjugation it
is Real in the sense of Atiyah (1966). It remains to observe that its restriction
under KRSO 0
2 ×G
(C2 )c → KRSO 0
2 ×G
(iR2 )c is the universal symbol σ̄univ of the
de Rham complex coupled to the reduced adjoint bundle.

For any n a point of the 0-space of the pair of spectra (M T SOn ,

Σ−1 M T SOn−1 ) is represented by a map from (B N , S N −1 ) into the pair of Thom
spaces attached to

QN −n / QN −n

 
 / Gr+ (RN )
+
Grn−1 (RN −1 ) n

for N sufficiently large. A map of (B N , S N −1 ) into this pair which is transverse to

the 0-section gives a compact oriented n-manifold M with boundary embedded
in (B N , S N −1 ), a rank n bundle W → M equipped with a stable isomorphism
TM ∼ = W , and a splitting of W ∂M as the direct sum of a rank (n − 1) bundle
and a trivial line bundle. The composition with the boundary map

r : (M T SOn , Σ−1 M T SOn−1 ) → M T SOn−1 (19.29)

is the restriction of this data to ∂M .

Now let M be a smooth manifold, X a compact oriented two-manifold with
boundary, and P → M × X a principal G-bundle with partial flat connection
along X. This data induces a classifying map M → MX and, forgetting the
connection, a classifying map

f : M −→ (M T SO2 , Σ−1 M T SO1 ) ∧ BG. (19.30)

There are induced classifying maps M → M∂X and

f˙ : M −→ M T SO1 ∧ BG
410 Consistent orientation of moduli spaces

for the boundary data; here f˙ = r ◦ f . View X as a bordism X : ∅ → ∂X; later

we incorporate incoming boundary components. The following key result relates
the universal topology above to surfaces with boundary.

Lemma 19.4 The map σ̄univ ◦ f : M → ko classifies the reduced tangent
bundle of the restriction map t : MX → M∂X —the bundle T MX − t∗ T M∂X
reduced to rank zero—pulled back to M .

Proof. At a point A ∈ M there is a flat connection on P {A}×X . The tangent

space to MX at that point is computed by the twisted de Rham complex (19.17),
•
so is represented by the twisted de Rham cohomology HA (X). Similarly, the
•
tangent space to M∂X at the restriction t(A) of A to the boundary is Ht(A) (∂X).
From the long exact sequence of the pair (X, ∂X) we deduce that the difference
T MX − t∗ T M∂X at A is the twisted relative de Rham cohomology HA •
(X, ∂X).
Now the twisted relative de Rham cohomology is the index of the deformation
complex (19.17) with relative boundary conditions (Gilkey 1995, section 4.1). In
other words, we consider the subcomplex of forms which vanish when restricted
to ∂X. This is an example of a local elliptic boundary value problem. Atiyah
and Bott (1964) interpret local boundary  conditions in K-theory and
 prove an
index formula. More precisely, the triple B(T X), ∂B(T X), S(T X) leads to the
exact sequence
   
KR−1 B(T X) ∂X
, S(T X) ∂X
−→ KR0 B(T X), ∂B(T X)
   
−→ KR0 B(T X), S(T X) −→ KR0 B(T X) ∂X
, S(T X) ∂X

The symbol σ of an elliptic operator lives in the third group, and Atiyah and Bott
construct a lift to the second group from a local boundary condition. (The image
of σ in the last group is an obstruction to the existence of local boundary condi-
tions; the image of the first group in the second measures differences of boundary
conditions.) The relative boundary conditions on the twisted de Rham complex
are universal, so the corresponding lift of the symbol occurs in the universal
setting. Recall from the proof of Lemma 19.3 the exact sequence (19.28), now
extended one step to the left:
0
KRG (iR)c −→ KRSO
0
2 ×G
(C2 )c −→ KRSO
0
2 ×G
(iR2 )c −→ KRG
0
(iR2 )c

The group G acts trivially in all cases. The Atiyah–Bott procedure applied
to the relative boundary conditions gives a lift of σ̄univ ∈ KRSO 0
2 ×G
(iR2 ))c
0 2 
to KRSO2 ×G (C )c . Recall that σ̄univ , constructed in the proof of Lemma 19.3,
is also a lift of σ̄univ . But by periodicity we find KRG 0
(iR)c ∼
= KRG 0
(−R)c ∼=
0 ∼ 1
KOG (−R)c = KOG (pt) which vanishes by Anderson (1964). Thus the lift

of σ̄univ is unique and σ̄univ computes the relative twisted de Rham cohomology.
This completes the proof.

 Universal orientations and consistent orientations 411

19.3.4 The level

A universal orientation induces a level, which is commonly used to identify the
theory. One observation arising from this study is that the level is a derived
quantity, and it is the universal orientation which determines the theory. We
explain the relationship, and deduce the existence of universal orientations, in
this subsection.
To begin we recast Definition 19.1 of a universal orientation in a form suited
for surfaces with boundary. Consider the diagram

q r
M T SO2 ∧ BG / (M T SO2 , Σ−1 M T SO1 ) ∧ BG / M T SO1 ∧ BG
TTTT
TTTT σ̄univ
TTTT 
−λ
TTTT σ̄univ
TTTT  
) τ
/ picg K
ko
(19.31)

The top row is a cofibration, the continuation of the top row of (19.28) in the
Puppe sequence. Recall that a universal orientation is a null homotopy of τ ◦

σ̄univ = τ ◦ σ̄univ ◦ q.

Lemma 19.5 A universal orientation is equivalent to a map −λ in (19.31)


and a homotopy from τ ◦ σ̄univ to −λ ◦ r.

The proof is immediate. In view of (19.25) and adjunction we can write

λ : Σ∞ BG → Σpicg K as a map from the suspension spectrum of BG to the
spectrum picg K, or equivalently as a map

λ : BG −→ B Picg K (19.32)

on spaces.

Definition 19.2 The map (19.32) is the level induced by a universal

orientation.

There is a map K(Z, 4) → B Picg K (see (19.15)) and in some important

cases, for example, if G is connected and simply connected, the level factors
through K(Z, 4), that is, the level is a class λ ∈ H 4 (BG).
A universal orientation is more than a level: it is a map −λ : M T SO1 ∧ BG →

picg K together with a homotopy of −λ ◦ r and τ ◦ σ̄univ in (19.31). Our next
result proves that universal orientations exist.

Theorem 19.1 There is a canonical universal orientation µ. The correspond-

ing level h is the negative of BG → BO → B Picg K, where the first map is
induced from the reduced adjoint representation ḡ and the second is projection to
a Postnikov section.
412 Consistent orientation of moduli spaces

Proof. Since complex vector spaces have a canonical K-theory orientation,

τ
the composite map k → ko → picg K is null, where k is the connective K-theory
spectrum (see the text preceding (19.16)). Therefore, a universal orientation is
given by filling in the left dotted arrow in the diagram:
q r
M T SO2 ∧ BG / (M T SO2 , Σ−1 M T SO1 ) ∧ BG / M T SO1 ∧ BG


σ̄univ −λ
  τ

k / ko / picg K
(19.33)

and specifying a homotopy which makes the left square commute. There is
a natural choice: smash the K-theory Thom class U : M T U1  M T SO2 → k
with the complexified reduced adjoint representation ḡC . This is the universal
rewriting of de Rham on a Riemann surface in terms of Dolbeault, at least on the
symbolic level. In terms of the proof of LemmaP 19.3, the map0 σ̄univ , 2restricted
to M T SO2 , is the exterior algebra complex ( • C2 , ) in KRSO 2
(iR )c . Write
R2 = LR for the complex line L = C and C2 ∼= R2 ⊗ C ∼ = L ⊕ L. Then the symbol
complex at θ ∈ iR2 ,
(θ) (θ)
C / L⊕L / L⊗L ,

is the realification of the complex

(θ)
C / L

which defines the K-theory Thom class. Tensor with the complexified reduced
adjoint representation ḡC to complete the argument.
η
To compute the level of µ we factorize τ as ko → Σ−1 ko → picg K, where the
first map is multiplication by η : S 0 → S −1 and the second is projection to a
Postnikov section. The map η fits into the Bott sequence k → ko → Σ−1 ko, and
so we extend (19.33) to the left:

(19.34)
The homotopy which expresses the commutativity of the right-hand square
induces the map α in this diagram, and the map −λ induced in (19.33) is the
suspension of α. We claim that there is a unique α, up to homotopy, which makes
the left square in (19.34) commute. For the difference of any two choices for α is a
map Σ−1 M T SO1 ∧ BG → Σ−1 ko, and the homotopy classes of such maps form
 Universal orientations and consistent orientations 413

the group KO1 (BG) which vanishes (Anderson 1964). It is easy to find a map α
as follows. Since Σ−1 M T SO1  S −2 (Lemma 19.2(1)), the upper left map is
the inclusion of the bottom cell of M T SO2 ∧ BG. The composite Σ−1 M T SO1 ∧
BG  Σ−2 BG → k factors as Σ−2 BG → Σ−2 k → k, where the first map is the
double desuspension of ḡC and the second Bott periodicity. Choose α to be the
double desuspension of ḡ, the real reduced adjoint representation. This completes
the proof.

Since equivalence classes of universal orientations form a torsor for the
group O(G) in (19.23), the canonical universal orientation identifies the torsor
of universal orientations with O(G). Notice the natural map
 : O(G) = [ΣM T SO2 ∧ BG, picg K] −→ [M T SO1 ∧ BG, picg K]
∼
= [BG, B Picg K] (19.35)
from universal orientations to levels. If g ∈ O(G), then the level of µ + g is
(g) − h. If G is connected, simply connected, and simple, then [BG, B Picg K] ∼
=
H 4 (BG; Z) ∼= Z and h is the dual Coxeter number of G times a generator. Then
g → (g) − h is a version of the ubiquitous “adjoint shift.”
Remark 19.3 For any G the top homotopy group of Map(ΣM T SO2 , picg K)
and of B Picg K is π4 , which is infinite cyclic. So there is a homomorphism
of H 4 (BG; Z) into the domain and codomain of (19.35), and on these subspaces
 is an isomorphism. This means that we can change a consistent orientation by
an element of H 4 (BG; Z), and the level changes by the same amount.

19.3.5 Pushforward maps

Suppose we have chosen a universal orientation with level λ. Let X be a compact-
oriented two-manifold with boundary. We can work as before with a family of
flat connections on X parametrized by a smooth manifold M , but instead for
simplicity we work universally on MX . As in (19.30) fix a classifying map
f : MX −→ (M T SO2 , Σ−1 M T SO1 ) ∧ BG; (19.36)
then there is an induced classifying map
f˙ = r ◦ f : M∂X −→ M T SO1 ∧ BG
for r the map (19.29). Set τ = f˙∗ (λ). Let t : MX → M∂X be the restriction map

on flat connections. According to Lemma 19.4 the composition τ ◦ σ̄univ ◦ f is the
reduced twisting τt − (dim MX − t∗ dim M∂X ). The homotopy which expresses
commutativity of the square in (19.31) gives an isomorphism
∼
τt − (dim MX − t∗ dim M∂X ) −→ −t∗ τ.
=
(19.37)
In principle, dim MX and dim M∂X are locally constant functions which vary
over the moduli space. However, the Euler characteristic of the deformation
414 Consistent orientation of moduli spaces

complex (19.17) is independent of the connection, and we easily deduce

dim MX − t∗ dim M∂X ≡ (dim G) b0 (∂X) (mod 2),

where b0 is the number of connected components. (We only track degrees in

K-theory modulo 2 [see (19.10)].) According to the discussion in Section 19.2
(especially (19.14)), there then is an induced pushforward

t∗ : K 0 (MX ) −→ K τ + (dim G) b0 (∂X) (M∂X ). (19.38)

This is the pushforward (19.5) associated to the bordism X : Y0 → Y1 with

Y0 = ∅ and Y1 = ∂X. The invariant (19.5) is then t∗ (1).
Note the special case ∂X = S 1 . Then MS 1 = G//G is the global quotient stack
of the action of G on G by conjugation. The codomain of (19.38) is thus

K τ +dim G (MS 1 ) ∼ τ +dim G

= K τ +dim G (G//G) = KG (G).

This is the basic space of the two-dimensional TQFT we construct (see

Section 19.1).
Observe that the universal orientation, in the form described around (19.31),
leads to the twist τ in the codomain of (19.38). This is the mechanism which was
envisioned in (19.14) when we discussed twistings in general: we have constructed
a pushforward from untwisted K-theory to twisted K-theory. The universal
orientation neatly accounts for the construction of a Frobenius ring structure
on twisted K-theory.
To treat an arbitrary bordism X : Y0 → Y1 we note that the deformation
complex at a flat connection a on a principal G-bundle Q → Y0 is

da
0 / Ω0 (g ) / Ω1 (g ) / 0 (19.39)
Y0 Q Y0 Q

The operator da is skew-adjoint. Therefore, there is a canonical trivialization of

the K-theory class of (19.39); for example, a canonical isomorphism ker da ∼ =
coker da . Suppose a classifying map f is given as in (19.36) and let f˙0 , f˙1 denote
its restriction to the boundary connections on Y0 , Y1 . Set τi = f˙i∗ (λ). Then
(19.37) and the canonical trivialization of (19.39) on the incoming boundary
lead to the desired push–pull map
s∗ ∗
K τ0 + (dim G) b0 (Y0 ) (MY0 ) −→ K s τ0 + (dim G) b0 (Y0 )
(MX )
t
∗
−→ K τ1 + (dim G) b0 (Y1 ) (MY1 ) (19.40)

This is the map (19.5) with the twistings induced from the universal
orientation.
A universal orientation induces consistent orientations on the outgoing restric-
tion maps of bordisms. In other words, if X : Y0 → Y1 and X  : Y1 → Y2 are
 Families of surfaces, twistings, and anomalies 415

bordisms, then the push–pull maps derived from the diagram

MX  ◦X
s KK
r ss KK r 
ss KK
ss KK
sy s K%
MX MX 
JJ JJ
s ttt JJ t tt
 t JJ t
tt JJ s
tt JJ
tt JJ tt JJ
ty t J% ty t J%
MY0 M Y1 MY2

satisfy
∗
(t r )∗ ◦ (sr)∗ = [t∗ ◦ s ] ◦ [t∗ ◦ s∗ ].
This follows from Lemma 19.1 and the “Fubini property”
(t r  )∗ = t∗ r∗ (19.41)
of pushforward. The orientation of t induces orientations of r  and t r  , since
the diamond is a fiber product. At stake in (19.41) is the consistency of the
orientations, which is ensured by the use of a universal orientation. The details
of this argument 4 will be given on another occasion.
One caveat: since MX , M∂X are stacks we can only pushforward along repre-
sentable maps, and this forces every component of X to have a nonempty out-
going boundary. As mentioned at the end of section 19.1, the partial topological
quantum field theory obtained from the push–pull construction extends to a full
theory using the invertibility of the (co)pairing attached to the cylinder (Freed
et al. 2003, section 17).

19.4 Families of surfaces, twistings, and anomalies

We begin with a general discussion of topological quantum field theories for fam-
ilies. Let F be an n-dimensional TQFT in the most naive sense. Thus F assigns
a finite-dimensional complex vector space F (Y ) to a closed oriented (n − 1)-
manifold Y and a linear map F (X) : F (Y0 ) → F (Y1 ) to a bordism X : Y0 → Y1 .
In particular, F (X) ∈ C if X is closed. Suppose that Y → S is a fiber bundle
with a closed oriented (n − 1)-manifold. Then the vector spaces assigned to each
fiber fit together into a complex vector bundle F (Y/S) → S. If γ : [0, 1] → S is
a path, then γ ∗ Y → [0, 1] is a bordism from the fiber Yγ(0) to the fiber Yγ(1) .
The topological invariance of F shows that F (γ ∗ Y) : F (Yγ(0) ) → F (Yγ(1) ) is
unchanged by a homotopy of γ, and so F (Y/S) → S carries a natural flat
connection. Then a family of bordisms X → S from Y0 → S to Y1 → S produces
4 The main point is that consistent orientations themselves form an invertible topological

field theory, and these field theories factor through the group completion of bordism, that is,
the Madsen–Tillmann space.
416 Consistent orientation of moduli spaces

 
a section F (X/S) of Hom F (Y0 /S), F (Y1 /S) → S; the topological invariance
and gluing law of the TQFT imply thatthis section is flat. In other words,
F (X/S) ∈ H 0 S; Hom F (Y0 /S), F (Y1 /S) . It is natural, then, to postulate that
a TQFT in families gives more, namely, classes of all degrees:
  
F (X/S) ∈ H • S; Hom F (Y0 /S), F (Y1 /S) . (19.42)
These classes are required to satisfy gluing laws and topological invariance as
well as naturality under base change.
The idea of a TQFT in families—at least in two dimensions—was introduced
in the mid-1990s. In two dimensions it is often formulated in a holomorphic
language (e.g. Kontsevich and Manin 1994), and classes are required to extend
to the Deligne–Mumford compactification of the moduli space of Riemann
surfaces.
Our push–pull construction works for families of surfaces—with a twist. The
purpose of this section is to alert the reader to the twist. 5
Suppose X → S is a family of bordisms from Y0 → S to Y1 → S, where Yi →
S are fiber bundles of oriented one-manifolds. Then the moduli stacks of flat
connections form a correspondence diagram over S:

MX/S
II
s uuu II t
uu II
uu II
uz u I$
MY0 /S MY1 /S
JJ t
JJ tt
JJ tt
JJ tt π1
JJ t
tz t
π0
$
S

The push–pull constructs a map from twisted K(MY0 /S ) to twisted K(MY1 /S ).

We can also work locally over S: the K-theory of the fibers of πi form bundles
of spectra over S and the push–pull gives a map of these spectra. But for our
purposes the global push–pull suffices. This construction is a variation of (19.42):
we use K-theory rather than cohomology.
The discussion of Section 19.3 goes through with one important change. It
comes in the paragraph preceding (19.21). For simplicity suppose Y0 = ∅ so that
the boundary ∂X = Y1 is entirely outgoing. Fix A ∈ MXs a flat connection on a
principal G-bundle P → Xs . Then the KO-theory class of the de Rham complex
coupled to the reduced adjoint bundle ḡP = gP − dim G computes the difference
TA MXs − (dim G) H • (Xs ), where H • (Xs ) is the real cohomology of Xs viewed
as a class in KO-theory. In Section 19.3 we treat H • (Xs ) as a trivial vector
space (there S = pt), but now H • (Xs ) varies with s ∈ S and so can give rise
5 We thank Veronique Godin for the perspicacious sign question which prompted this

exposition.
 Families of surfaces, twistings, and anomalies 417

to a non-trivial twisting. More precisely, H • (Xs ) is the fiber at s ∈ S of a flat

vector bundle H• (X/S) → S. Let τX/S denote the twisting of complex K-theory
attached to this bundle. This twisting replaces the degree shift in (19.37) and
the pushforward (19.38) is modified to include that extra twist:
∗
t∗ : K 0 (MX/S ) −→ K τ + (dim G) π1 τX/S (MY1 /S ). (19.43)

The degree shift is now incorporated into the twist τX/S , and there may be non-
trivial contributions to the twist from w1 and W3 of H• (X/S) → S as well. (The
degree shift and twistings vanish canonically if dim G is even.)

Example 19.1 Consider the disjoint union X of two 2-disks. The boundary
circles are outgoing, as above. Suppose that dim G is odd. For a single disk
τ +1
the pushforward t∗ (1) in (19.38) lands in KG (G) and is the unit 1 in the
Verlinde ring. Thus for the disjoint union of two disks, t∗ (1) is the image of 1 ⊗ 1
(τ,τ )
under the external product KG τ +1
(G) ⊗ KGτ +1
(G) → KG×G (G × G). Now con-
sider the family X → S with fiber X and base S = S in which the monodromy
1

exchanges the two disks. The flat bundle H• (X/S) → S has rank 2 and non-
trivial w1 . According to (19.43), then, t∗ (1) for the family lives in the twisted
∗
group K (τ,τ )+π τX/S (M∂X/S ). On each fiber of π : M∂X/S → S we recover the
class 1 ⊗ 1 above. But upon circling the loop S = S 1 this class changes sign in the
π ∗ τX/S -twisted K-group. Said differently, the diffeomorphism which exchanges
the disks acts by a sign on 1 ⊗ 1. Of course, one might predict this from the
τ +1
sign rule in graded algebra: the Verlinde ring KG (G) is in odd degree, so upon
exchanging the factors of 1 ⊗ 1 one picks up a sign. It shows up here as an extra
twisting.

This extra twisting is a topological analog of what is usually called an anomaly

in quantum field theory. In an anomalous theory in n dimensions the partition
function on a closed n-manifold, rather than being a complex-valued function
on a space of fields, is a section of a complex line bundle over that space of
fields. Furthermore, there is a gerbe over the space of fields on a closed (n − 1)-
manifold, and for a bordism the partition function is a map of the gerbes attached
to the boundary. In the homological version described at the beginning of this
section, the parameter space S plays the role of the space of fields and for a
family of closed n-manifolds the partition function in a non-anomalous theory is
an element of H • (S; R). An anomalous theory would assign a flat complex line
bundle L → S to the family, and the partition function would live in the twisted
cohomology H • (S; L) = H L+• (S). In the two-dimensional TQFT we construct
using push–pull on K-theory, the extra K-theory twist τX/S is the anomaly
(see (19.43)). Notice that there is no gerbe attached to a family of one-manifolds
(better: it is canonically trivial). We remark that the anomaly is itself a particular
example of an invertible topological quantum field theory.
418 Consistent orientation of moduli spaces

Acknowledgments
The work of Daniel S. Freed, Michael J. Hopkins, and Constantin Tele-
man was supported by NSF grant DMS-0603964, DMS-0306519, and EPSRC
GR/SO6165/01, respectively.
We thank Veronique Godin, Jacob Lurie, Ib Madsen, and Ulrike Tillmann for
enlightening conversations.

References
Abrams, L. (1996). Two-dimensional topological quantum field theories and Frobenius
algebras. J. Knot Theory Ramifications 5(5), 569–87.
Anderson, D. W. (1964). The real K-theory of classifying spaces. Proc. Nat. Acad. Sci.
51(4), 634–6.
Atiyah, M. F. (1966). K-theory and reality. Q. J. Math. Oxford Ser. (2) 17, 367–86.
Atiyah, M. F. and Bott, R. (1964). The index problem for manifolds with boundary.
Differential Analysis, Bombay Colloquium, 1964. Oxford University Press, London,
pp. 175–86.
Atiyah, M. and Segal, G. (2006). Twisted K-Theory and Cohomology, Inspired by
S. S. Chern, Nankai Tracts Mathematics, Vol. 11, World Scientific Publication,
Hackensack, NJ, pp. 5–43, arXiv:math/0510674.
Atiyah, M. F. and Singer, I. M. (1971). The index of elliptic operators. V, Ann. Math.
(2) 93, 139–49.
Costello, K. (2007). Topological conformal field theories and Calabi-Yau categories.
Adv. Math. 210(1), 165–214, arXiv:math/0412149.
Donaldson, S. K. (1987). The orientation of Yang-Mills moduli spaces and 4-manifold
topology. J. Differential Geom. 26(3), 397–428.
Donaldson, S. K. and Kronheimer, P. B. (1990). The geometry of four-manifolds. Oxford
Mathematical Monographs. Oxford Science Publications, New York.
Donovan, P. and Karoubi, M. (1970). Graded Brauer groups and K-theory with local
coefficients. Inst. Hautes Études Sci. Publ. Math. No. 38, 5–25.
Freed, D. S. (2009). Remarks on Chern–Simons theory. Bull. Amer. Math. Soc. 46,
221–254, arXiv:0808.2507.
Freed, D. S., Hopkins, M. J. and Teleman, C. Loop groups and twisted K-theory I,
arXiv:0711.1906.
Freed, D. S., Hopkins, M. J. and Teleman, C. Loop groups and twisted K-theory II,
arXiv:math/0511232.
Freed, D. S., Hopkins, M. J. and Teleman, C. Loop groups and twisted K-theory III,
arXiv:math/0312155.
Freed, D. S., Hopkins, M. J. and Teleman, C. (2007b). Twisted equivariant K-theory
with complex coefficients. J. Topol. 1, arXiv:math/0206257.
Galatius, S., Madsen, I., Tillmann, U. and Weiss, M. (2009). The homotopy type of
the cobordism category. Acta Math. 202, 195–239, arXiv:math/0605249.
Gilkey, P. B. (1995). Invariance Theory, the Heat Equation, and the Atiyah-Singer Index
Theorem, 2nd edn., Studies in Advanced Mathematics. CRC Press, Boca Raton, FL.
Hitchin, N. (2001). Lectures on special Lagrangian submanifolds. Winter School on
Mirror Symmetry, Vector Bundles and Lagrangian Submanifolds (Cambridge, MA,
 References 419

1999). AMS/IP Studies in Advanced Mathematics, Vol. 23, American Mathematical

Society, Providence, RI, pp. 151–82, arXiv:math/9907034.
Hopkins, M. J. and Lurie, J. in preparation.
Kontsevich, M. and Manin, Yu. (1994). Gromov-Witten classes, quantum cohomol-
ogy, and enumerative geometry. Commun. Math. Phys. 164(3), 525–62, arXiv:hep-
th/9402147.
Madsen, I. and Tillmann, U. (2001). The stable mapping class group and Q(CP∞ + ).
Invent. Math. 145(3), 509–44.
Madsen, I. and Weiss, M. (2005). The stable mapping class group and stable homotopy
theory. European Congress of Mathematics. European Mathematical Society, Zürich,
pp. 283–307.
Moore, G. W. and Segal, G. B. (2006). D-branes and K-theory in 2D topological field
theory, arXiv:hep-th/0609042.
Murray, M. K. (1996). Bundle gerbes. J. London Math. Soc. (2) 54(2), 403–16.
Segal, G. B. (1977). K-homology theory and algebraic K-theory. K-theory and Operator
Algebras (Proceedings of the Conference, University of Georgia, Athens, G.A, 1975).
Lectures Notes in Mathematics, Vol. 575, Springer, Berlin, Germany, pp. 113–27.
This page intentionally left blank
 INDEX

α particle, 286–294 Israel–Wilson, 30

A-model, 113–127, 299 rational map, 288–296
Abelian Area form, 88
group, 133, 405–407 Asymptotic expansion, 197
Lie group, 240 Asymptotically
variety, 92, 100, 304–305, faithful, 181, 186, 204–206
315–321 flat, 188, 201–204
Action functional, 398 Atiyah
ADE singularity, 14 –Singer theorem, 1, 404–407
ADHM construction, 11, 34 class, 383–384
Adjoint duality, 406
bundle, 113, 405–409 Automorphism
orbit, 87 group, 125, 132
representation, 99, 105, 412–413 of a Riemann surface, 185
Affine outer, 212–213
Lie algebra, 180 Axiom
manifold, 308–316 of concatenation, 172
sphere, 154 of tensoring, 172
Albanese TQFT, 179
image, 103
variety, 165 B-field, 8, 238
ALE manifold, 14, 29, 325, 330 B-model, 113–127, 299
Alexander Banach algebra, 266
–Wermer theorem, 261, 271–275 Baryon number, 281, 286, 293–295
duality, 267, 277 Betti
Algebra cohomology, 336
C ∗ , 166 number, 28, 324
Banach, 266 Bihermitian metric, 368, 387
Clifford, 403–404 Bimodule, 168
cluster, 156 Binary icosahedral group, 220,
exterior, 406 224–228
Kac–Moody, 40, 342 Bipolar theorem, 272
Lie, 86 Bockstein map, 253
Poisson, 101 Bogomolny equations, 3, 12, 118, 283
Sklyanin, 392 Bordism
Algebraic category, 400
curve, 219 group, 396–397
cycle, 275–278 theory, 407
hypersurface, 263, 278 Borel subgroup, 41
Algebroid, 369–392 Bott periodicity, 405, 410–413
Almost complex structure, 380 Boundary, 270, 397
Ample line bundle, 94 condition, 48, 405
Anomaly, 415–417 of a chain, 275
Faddeev–Mickelsson, 256 of a surface, 407
Anosov, 186–187 outer, 274
Ansatz regularity, 261
Gibbons–Hawking, 89 value problem, 71–90, 410
Hitchin, 120 BPS equations, 398
422 Index

Bracket Fukaya–Floer, 300

Courant, 370–392 linear, 168, 173
Lie, 76 of U (1) bundles, 240
Poisson, 89 of Lie algebroid representations, 375
Schouten, 376 of manifolds, 239
Braid group, 122 tensor, 97, 180
Branch point, 139, 145 Cauchy
Branched –Kovalewskaya theorem, 352
cover, 46, 230 –Riemann equations, 348
hyperbolic structure, 142, 151 Cayley cubic surface, 218
Brane, 8, 117, 174, 377–392 Čech complex, 243
Bundle Cell, 131
adjoint, 113, 405–409 big, 42
canonical, 47, 95, 113, 377 Chain, holomorphic, 261
circle, 4 Character, 92
conormal, 310 Cheeger–Simons, 269
cotangent, 5, 13, 99, 106, 134–135, multiplicative, 133
308 sequence, 353–363
dual, 238–240 variety, 341–342
flat, 13, 113, 118, 147, 307 Characteristic class, 144, 239, 245, 258, 376,
gerbe, 237–258 383
Higgs, 5, 14, 92–110, 114, 119, 129–158, Cheeger–Simons character, 269
375, 379 Chern
Hilbert, 253, 256–257 –Simons theory, 5, 13, 178, 258, 396, 400
holomorphic, 115, 117, 120, 151, 174, 210, class, 23, 27, 36, 192, 239, 247, 264, 325
217, 383 connection, 97, 376, 389–390
line, 7, 92, 129, 165, 188–189, 238, 264, formula, 268
305, 388, 417 Chevalley theorem, 106
loop, 257 Classifying
P1 , 102 map, 407–409
parabolic, 342 space, 404
Poincaré, 304–305, 317–320 Clifford algebra, 403–404
principal, 42, 92, 99, 105, 167, 241, 242, Cluster algebra, 156
306, 326, 369 Clutching construction, 244
spin, 24 Cobordism, 172–175
stable, see Stable Cofibration, 408–411
tangent, 15, 40, 105, 196–197, 369, 397, Coframing, 354–358
404–407 Coherent sheaf, 168, 301–304
U (1), 238, 242, 258 Cohomology
vector, 13, 35, 67, 92, 135, 212, 213, 3-dimensional, 237–239
408 class, 15, 17, 27, 42, 57, 61, 64, 150, 241,
Verlinde, 180 389
de Rham, 200, 328, 401–402
C ∗ algebra, 166 Dolbeault, 336
Calabi integer, 402
–Yau hypersurface, 315 intersection, 330
–Yau manifold, 7–8, 300–321, 337 middle-dimensional, 324, 339–340
–Yau metric, 331 non-abelian, 216, 336
conjecture, 1–2 Poisson, 384
inequality, 26 rational, 331
Calculus of variations, 399 theory, 399
Caloron, 4, 34–68 vanishing, 47
Canonical line bundle, 13, 47, 95, 113, 122, Coisotropic submanifold, 307–312, 378
300, 377 Commutative
Cartan–Kähler theorem, 346–349 diagram, 53
Category, 166–168, 177 differential operators, 101
bordism, 397, 400 Compactification, 416
derived, 300, 304–306 Companion matrix, 51
 Index 423

Completely integrable, see Integrable function, 78

system subset, 266, 272
Complex structure, 29, 92, 118–123, Cotangent bundle, 5, 13, 99, 134–135, 308,
143, 328 369
generalized, 8, 15, 114, 368–392 Courant
Complexification algebroid, 369–392
of a Lie group, 113, 126 bracket, 370–392
Concatenation, 172 trivialization, 369–392
of paths, 349 Covariant
Cone, 21, 270, 365 constant section, 181–183, 199, 204
hyperbolic, 139–145 derivative, 35
Kähler, 29 Covering
of homology classes, 264 Galois, 94
Configuration space, 171 map, 167
Conformal universal, 88
field theory, 180 Coxeter number, dual, 413
invariance, 17 Crepant resolution, 331
map, 155 Critical point, 15, 18, 171
structure, 134, 143, 148–150, 155 Cube, 286–294
Conjecture Cubic
Calabi, 1–2 Cayley, 218
Kontsevich, 301 curve, 305
mirror symmetry, 300–301 curl, 75
Mumford, 396 Current, 268–277
Poincaré, 188 Curvature, 246, 351, 370–371, 378, 381–382,
Sen, 329–330 385, 390
SYZ, 7 Hermitian, 269
Yamabe, 1, 18 mean, 355, 365
Connection, 12, 210, 239, 255 of a connection, 239
bundle gerbe, 247 of a line bundle, 189
Chern, 97, 376, 389–390 operator, 24
flat, 5, 96, 115, 120, 179–181, 213, 217, Riemann, 20
319, 333, 375–379, 399–400, 416 scalar, 18, 357–358
formal, 199–200 sectional, 76
Gauss–Manin, 213–217 tensor, 75
generalized, 368–374 three-, 247, 250
Hitchin, 188–193, 198–204 two-form, 123–124
hypergeometric, 212 Weyl, 24
Klein, 226 Curve
Levi-Civita, 191, 346, 349, 356, algebraic, 219
368–374 cubic, 305
logarithmic, 215–225 elliptic, 213, 231, 302–305, 385
meromorphic, 121 holomorphic, 300
one-form, 351 Jordan, 146
projectively flat, 177–206 operator, 184–185
quotient, 226 projective, 92
self-dual, 2–3, 38, 56 rational, 216, 301
torsion-free, 350 real analytic, 261–278
U (1), 301–302 simple closed, 178
unitary, 370, 388 spectral, 3, 7, 11, 34, 42, 100, 130
Connective structure, 240–242, 246–247, Cycle
250 algebraic, 275–278
Conormal bundle, 310 Hecke, 95
Constant mean curvature, 365 Lagrangian, 307–308
Continuity method, 83 Cylinder, 57
Convex, 60–62
domain, 154 ∂¯ operator, 115, 126, 190–193
foliated structure, 158 Darboux theorem, 306
424 Index

Data Dolbeault
principal part, 43 cohomology, 336
spectral, 35, 45 complex, 396
de Rham Donaldson theory, 124
class, 18, 241 Double coset space, 167
cohomology, 200, 328, 401–402 Dual
complex, 8, 243, 375, 396, 405–410 bundle, 238–240
functor, 216 group, 114, 165
Deck transformation, 402 Duality
Deformation Alexander, 267, 277
complex, 2, 405, 414 Atiyah, 406
infinitesimal, 92 Langlands, 113–127, 174
isomonodromic, 122–123, 215–222 Poincaré, 329
of complex structures, 303 Serre, 150
of geometric structures, 130–132 strong-weak, 324
quantization, 197–198 T–, 116
space, 142, 155 Dubrovin–Mazzocco solution, 231
Degree, 402
of a map, 241, 281, 292 E polynomial, 332–333, 338
Del Pezzo surface, 388–390 Eigenfunction, 90
Deligne–Mumford compactification, 416 Eilenberg–MacLane space, 404
Derived category, 300, 304–306 Einstein
Descent data, 249 –Hilbert action, 23
Desuspension, 408, 413 –Maxwell equations, 3, 17–30
Determinant line bundle, 93 –Weyl manifolds, 3
Developing map, 138 equations, 1
Diffeomorphism, 75, 148, 186, 353, 359, 370, manifold, 1
379, 388, 397, 417 metric, 2, 28
area-preserving, 71 universe, 158
Differential Elliptic
form, 14, 76, 80, 134, 241, 267, 328, complex, 56
351–356, 363, 369–372, 386, curve, 213, 231, 302–305, 385
388–389 fibration, 299–321
Hopf, 148, 151 function, 3, 7
ideal, 357 integral, 6, 215
operator, 75, 101, 117, 181–206 operator, 406–407
quadratic, 116 regularity, 355
Dihedral Embedding, 348–366, 392
group, 6 Energy, 74, 281–282, 293, 326, 395
solution, 220, 233 of a harmonic map, 148–149
symmetry, 285–287 Equation(s)
Dimensional reduction, 13, 215, 327, Bogomolny, 3, 12, 118, 283
397 BPS, 398
Dirac Cauchy–Riemann, 348
equation, 34, 55 Dirac, 34, 55
operator, 1–2, 25, 37, 256 Einstein, 1
structure, 370–371, 377 Einstein–Maxwell, 3, 17–30
Dirichlet problem, 71, 83 Euler–Lagrange, 74, 78, 87
Discriminant Fuchsian, 213, 218, 228–229
Hitchin, 103 heat, 6
locus, 303 Hitchin, 113–127
Distribution, 271 hypergeometric, 210–213, 228–230
Divisor, 48, 66, 164, 267, 276, 303 KdV, 12
at infinity, 52 Lamé, 213
class, 92 Lax, 46
normal crossing, 332 Maxwell, 1, 17
Dixmier–Douady class, 245–254, 258 monad, 67
Dodecahedron, 286, 293 Monge–Ampère, 82, 352
 Index 425

Nahm, 3–5, 7, 36–39, 46, 67, 71–90 Form, see Differential form
Painlevé, 6, 214–234 Formal
Riccati, 215 connection, 199–200
Schlesinger, 217 trivialization, 188
Seiberg–Witten, 3, 17–30 Fourier
self-duality, 2, 34, 327 –Laplace transform, 224–225
Skyrme, 281 –Mukai
Yang–Mills, 2–6, 13, 215, 327 transform, 299–321
Equivariant K-theory, 395 Framing, 400
Euler Fredholm operator, 37, 56
φ function, 325 Fricke space, 142–148
–Lagrange equations, 74, 78, 87 Frobenius
characteristic, 329, 413 manifold, 213
class, 104, 144–145 ring, 395, 400, 414
Exact sequence, 54 Fubini property, 415
Exponential mapping, 365 Fuchsian
Exterior equation, 213, 218, 228–229
algebra, 406 representation, 129, 137–148, 154
differential system, 8, 348–366 Fukaya–Floer category, 300
Extremal Kähler metric, 17–30 Fullerene, 283
Function
F4 , 227–228 Green, 59, 271
Faddeev–Mickelsson anomaly, 256 harmonic, 274
Family Morse, 171
of elliptic complexes, 407 plurisubharmonic, 269, 273
of surfaces, 415–417 Poisson commuting, 6, 101
Fano spaces, 58, 61
manifold, 303–305, 368 theta, 6–7
Poisson manifold, 390 Fundamental
Federer theorem, 274 form, second, 347
Fibration group, 5, 92, 98, 113, 122, 129–130, 138,
elliptic, 299–321 158, 215–218, 230, 327, 395
Lagrangian, 299–321 Fusion ring, 395
of groupoids, 237
special Lagrangian, 337 G2 , 348
Fibre product, 240–242, 399 manifold, 15, 358–365
Filtration, 49 structure, 8
weight, 332 Galois group, 94
Finite field, 331–332, 340 Gauge
Flag, 41 group, 86, 114, 327
Flat theory, 123, 188, 398–399
bundle, 13, 113, 118, 147, 307 Gauss
connection, 5, 96, 115, 179–181, 213, 217, –Bonnet formula, 24
319, 333, 375–379, 399–400, 416 –Manin connection, 213–217
line bundle, 133 theorem, 72
structure, 213, 347 Gelfand transformation, 266
Floer homology group, 306 Generalized
Flop, 304 cohomology theory, 399
Flow complex structure, 8, 15, 114, 368–392
geodesic, 158 connection, 368–374
of a vector field, 389 holomorphic bundle, 374–377
of hypersurfaces, 8 Kähler manifold, 4, 368
of line bundles, 49 submanifold, 392
of maps, 171 Genus, 13, 180
of sheaves, 46 geometric, 30
of structures, 348 of Painlevé curve, 219
Fock space, 256 Geodesic, 3, 78
Foliated structure, 158 flow, 158
426 Index

Geometric Half-form, 193

genus, 30 Hamiltonian, 87
invariant theory, 93, 136, 166 vector field, 383–384
Langlands, see Langlands Hard Lefschetz, 303
quotient, 67 Harmonic
Gerbe, 6, 237–258, 404 coordinates, 347, 355
U (1), 369–370 curvature, 19
Gibbons–Hawking form, 324–342
ansatz, 89 function, 71, 88, 89, 274
instanton, 14 L2 form, 7
GIT quotient, 337, 340 map, 96, 148
Gluing metric, 150
between bundles, 51 representative, 61
law, 397–400 spinor, 38
theorems, 29 two-form, 17, 21, 25
Grassmannian, 406 Hausdorff
Lagrangian, 158 dimension, 265
Green’s function, 59, 271 measure, 270
Grothendieck Heat
dessins d’enfants, 230 equation, 6
fundamental group, 98 specific, 71–72
Group Hecke
abelian, 133, 405–407 correspondence, 5, 94
action on a space, 167 cycle, 95
automorphism, 125, 132 operator, 117–118
bordism, 396–397 transformation, 102
braid, 122 Hermitian
complexified, 87 line bundle, 238, 376
dihedral, 6 metric, 269, 275–276, 370, 375–377,
dual, 114, 165 390
fundamental, 5, 92, 98, 113, 122, 129–130, structure, 180, 189, 195–202
138, 158, 215–218, 230, 327, 395 symmetric space, 130, 146, 156–158
Galois, 94 Yang–Mills, 301
gauge, 86, 114, 327 Hessian group, 222
Hessian, 222 Higgs
holonomy, 74 bundle, 5, 14, 92–110, 114, 119,
icosahedral, 210–220, 232 129–158, 375, 379
Kac–Moody, 254 field, 12, 40, 115, 136, 151,
Lie, see Lie group 327–328
loop, 66, 395 pair, see Higgs bundle
mapping class, 122, 132–134, 178–218 Hilbert
monodromy, 210–234 bundle, 253, 256–257
of holomorphic maps, 167 polynomial, 302
of Lie type, 338 space, 398
orthogonal, 75 space representation, 182–183
Picard, 164–165 Hirzebruch surface, 304
quantum, 180 Hitchin
reductive, 105–106 –Chatterjee gerbe, 244–251
reflection, 221–226 connection, 188–193, 198–204
surface, 130 discriminant, 103
symmetry, 210 equations, 113–127
symplectic, 157 fibration, 116
topological, 182 moduli space, 166, 328, 336
triangle, 229–232 morphism, 99–101
unitary, 156 representation, 157–158
Valentiner, 222, 224, 227 Hodge
Weyl, 106, 215, 219, 227–228 mixed structure, 331–342
Groupoid, 166–168, 237–238, 245, 381, 399 number, 338–339
 Index 427

star operator, 12, 17, 73, 114, 134, 328, Calabi–Yau, 315
354 complex, 218
structure, 130
theory, 98, 134, 328–330 Icosahedral
theory, hyperkähler, 324–342 group, 210–220, 232
Holomorphic solution, 6, 232–234
p chain, 261 symmetry, 282, 286
bundle, 115, 117, 120, 164–217 Immersion, 347
curve, 300 Index theorem, 1, 404–407
differential form, 134 Infinite-dimensional
line bundle, 129, 264, 302, 305, 383 Lie group, 13
map, 124, 165–167, 190 manifold, 74
Poisson structure, 368–392 Instanton, 2–4, 14, 34, 325, 337
Poisson variety, 380–388 bubbling, 37
representation, 165 Integer cohomology, 402
section, 263, 267, 276, 376, 391–392 Integrable system, 5–6, 100, 116, 130, 337
vector bundle, 151, 175 Integral manifold, 353, 358
Holonomy, 239–241, 247, 250–256, 300, 337, Intersection
346–366, 400 cohomology, 330
function, 187 form, 340–342
group, 74 pairing, 113, 267
representation, 138, 145, 154 Intrinsic torsion, 346–366
Homogeneous space, 353 Invariant
Homology Seiberg–Witten, 27
class, 264 Toledo, 146–147
Floer, 306 Involution, 54, 405
Homotopy, 241 anti-holomorphic, 110
class, 186–187 hyperelliptic, 181
equivalence, 404 Involutive
null, 407–411 ideal, 354–357
theory, 168–175 tableau, 351
Hopf differential, 148, 151 Isomonodromic
Hull deformation, 122–123, 215–222
λ, 277 family, 229
projective, 262–265, 269 Isothermal migration, 85
Hurewicz theorem, 133 Israel–Wilson ansatz, 30
Hyperbolic
affine sphere, 154 Jacobian, 101, 305
cone, 139–145 variety, 92, 129, 134
geometry, 129 Jimbo’s formula, 222–223
space, 137 Jones polynomial, 5, 178, 188
structure, 137–144, 151 Jordan curve, 146
surface, 137 Jump
triangle group, 229 of holomorphic structure, 49
Hypercohomology, 100 of isomorphism class, 165
Hyperconvex curves, 155–156
Hyperelliptic involution, 181 K-theory, 156, 168, 395–417
Hypergeometric twisted, 8
connection, 212 K3 surface, 1–2, 314
equation, 210–213, 228–230 Kähler
Hyperkähler class, 20, 119
Hodge theory, 324–342 cone, 29
manifold, 4–8, 13, 119, 314, 351–358 form, 147, 267, 326
metric, 324–342 manifold, 126, 130, 189, 197, 265,
quotient, 13, 127, 326–330 269–270
structure, 89, 115, 135–136 metric, 89, 188–194, 275–276
Hypersurface, 8, 73, 347–366 potential, 74, 87
algebraic, 263, 275, 278 surface, 18
428 Index

Kac–Moody isomorphism, 43
algebra, 40, 342 simplicial, 244
extension, 257 Linear
group, 254 category, 168, 173
monopole, 66 system, 41
KdV equation, 12 Link, 177
Killing–Coxeter matrix, 224–225 Linking number, 261–263, 267–269,
Klein 275–276
connection, 226 affine, 268
polynomial, 286 projective, 262
theorem, 229 reduced, 262
Knot, 188 Local
Kodaira dimension, 29 bundle gerbe, 244–246
Kontsevich conjecture, 301 Locality, 164–175
Locally product metric, 350
L2 Logarithmic connection, 215–225
cohomology, 341 Loop, 293
harmonic form, 7, 324–342 bundle, 257
Hermitian structure, 202 group, 66, 395
inner product, 74 space, 171–175, 238, 254, 404
norm, 88 Lorentzian space, 84
norm of curvature, 20, 36
section, 38, 58, 180–181 Möbius transformation, 284
solution, 12, 44, 64 Macdonald polynomial, 341
Lagrangian, 85 Magnetic monopole, see Monopole
cycle, 307–308 Manifold, see also Variety
fibration, 299–321 affine, 308–316
fibre, 100, 116 ALE, 14, 29, 325, 330
Poincaré cycle, 313–319 Calabi–Yau, 7–8, 300–321, 337
submanifold, 378–379, 388 Einstein, 1
subspace, 158 Einstein–Weyl, 3
torus, 116 Fano, 303–305, 368
Lamé equation, 213 Frobenius, 213
Langlands G2 , 15, 358–365
correspondence, 168 generalized Kähler, 4, 368
dual group, 165 hyperkähler, 4–8, 13, 119, 314, 351–358
duality, 113–127, 174 infinite-dimensional, 74
programme, 5, 13, 164, 338–339 integral, 353–358
Laplacian, 71, 90, 192 Kähler, 126, 130, 189, 197, 265, 269–270
Large complex structure limit, 301–304, 315 non-commutative, 174
Lattice, 308 one-, 400–401, 416
Lax equations, 46 projective, 261–278
Legendre transform, 81 quasi-projective, 264
Level, 411 Riemannian, 172, 397
Levi-Civita connection, 191, 346, 349, 356, seven-, 358
368–374 Stein, 265
Lie SU (2), 348–357
algebra, 86, 180, 185 symplectic, 113, 118–122, 129, 134,
bracket, 76 188–189
derivative, 75, 196 three-, 177, 188, 255–256
group, 13, 113, 125, 129, 130, 178, 240, two-, 113–127, 400–401, 405–409, 413
254–255, 258, 326, 346, 395–417 Map
Line bundle, 7, 92, 129, 165, 188–189, 238, conformal, 155
305, 388, 417 covering, 167
canonical, 13, 122, 300 developing, 138
flat, 133 harmonic, 148
Hermitian, 266, 376 holomorphic, 124, 165–167, 190
holomorphic, 302, 305 meromorphic, 171
 Index 429

Mapping class, 185–187 of flat bundles, 113

group, 122, 132–134, 178–218 of flat connections, 5, 129, 180, 301, 396,
Marked 404–405, 416
conformal structure, 148–150 of geometric structures, 129
point, 177 of Higgs bundles, 7, 13, 95, 324, 331, 333
Marking, 138–144 of holomorphic bundles, 168, 174
Mass-norm, 272 of holomorphic maps, 398
Matrix of instantons, 327
companion, 51 of Lagrangian subspaces, 7
Pauli, 37, 56, 281 of monopoles, 327
S-, 179, 185 of parabolic bundles, 342
Maximal of self-dual connections, 396–398
representation, 156–158 of stable bundle, 5
torus, 41, 119 of stable bundles, 306
Maximum-modulus principle, 265 of symplectic structures, 301
Maxwell equations, 1, 17 SL(n, C), 187
McKay correspondence, 304 Moment map, 4, 314, 326–328
Mean curvature, 355, 365 Monad, 67
Measure Monge–Ampère equation, 82, 352
Hausdorff, 270 Monodromy, 120–122, 210–234, 303, 314, 321
Poisson–Jensen, 270 Monopole, 3–4, 7, 11, 34, 39, 283–284, 325,
theory, 398 337
Meromorphic Kac–Moody, 66
connection, 121 metric, 4, 6
map, 171 moduli space, 327
Metaplectic correction, 193 U (1), 55
Metric Morita equivalence, 167–168, 403
bihermitian, 368, 387 Morphism, Hitchin, 99–101
Calabi–Yau, 331 Morse function, 171
Einstein, 2, 28 Multiplicative character, 133
extremal Kähler, 17–30 Multiplicative cohomology theory, 403
harmonic, 150 Mumford conjecture, 396
Hermitian, 269, 275–276, 301, 370,
375–377, 390 Nahm
hyperkähler, 324–342 equations, 3–5, 7, 36–39, 46, 67, 71–90
Kähler, 89, 188–194, 275–276 transform, 34–68
locally Euclidean, 79 Neilsen–Thurston classification, 185–187
locally product, 350 Nerve, 242
monopole, 4, 6 Newlander–Nirenberg theorem, 352
real-analytic, 346–366 Newstead relation, 339
Riemannian, 164, 346, 385 Nijenhius tensor, 386
self-dual, 350 Non-abelian
Taub–NUT, 313–314 cohomology, 216, 336
Middle-dimensional cohomology, 324, Hodge theory, 98, 134
339–340 Non-commutative
Minimizer, 86 geometry, 8, 165–171, 368, 391–392
Minkowski space, 2 manifold, 174
Mirror symmetry, 7–8, 113–127, 299–321, Norm, mass, 272
337–341 Normal crossing divisor, 332
Mixed Hodge theory, 331–342 Null homotopy, 407–411
Modular
form, 325 O(n, n), 371
functor, 177–179 Octahedral solution, 220, 227, 233
Moduli space, 92, 395–417 Octahedron, 287–290
Hitchin, 166, 328, 336 Octic curve, 233
of coherent sheaves, 302 Operator
of complex structures, 8, 300–301 ¯ 115, 126, 190–193
∂,
of elliptic curves, 302 curve, 184–185
430 Index

Operator (cont.) Jones, 5, 178, 188

differential, 181–206 Macdonald, 341
Dirac, 1–2, 37 Polystable, 93, 105
Fredholm, 37, 56 Higgs pair, 136
Laplace, 192 Poncelet
positive, 73 problem, 233
prequantum, 195 theorem, 6
product expansion, 173 Pontrjagin
shift, 43 –Thom collapse, 402
Toeplitz, 181–206 class, 258, 400
trace-class, 172 product, 395, 400
Orbifold, 337 Porous-dam problem, 84
Orbit, adjoint, 87 Positive
Ordinary differential equation, 210–234 operator, 73
Orientation structure, 156
of moduli spaces, 395–417 Postnikov section, 404, 411–412
universal, 407–417 Potential, 78
Kähler, 74
P1 bundle, 102 Pre-Hilbert space, 195
Painlevé equation, 6, 214–234 Principal
Pair of pants, 400 bundle, 42, 92, 99, 167, 242, 326,
Pairing, Kronecker, 267 369
Parabolic G bundle, 306
bundle, 342 part data, 43
subgroup, 105 Z-bundle, 241
Parallel Product of gerbes, 239
form, 353, 359 Proj, 266
transport, 239, 349 Projective
Partition function, 417 hull, 262–265, 269
Path integral, 178, 396–399 linking number, 262
Pauli matrices, 37, 56, 281 manifold, 261–278
Periodic function, 72 variety, 94
Periodicity, Bott, 405, 410–413 winding number, 263
Picard group, 164–165 Projectively flat connection, 177–206
Platonic solid, 294 Prym variety, 13, 101
Pluripolar, 265 Pseudo
Plurisubharmonic convex function, 352
function, 269, 273 differential operator, 196
quasi, 263 Puiseau expansion, 226
Poincaré Pullback of gerbes, 245
–Lelong formula, 268, 276, 376, 391
bundle, 304–305, 317–320 Quadratic differential, 116
conjecture, 188 Quantization, 101, 156, 178–180, 193,
duality, 329 293
Poisson deformation, 197–198
–Jensen measure, 270 Quantum
bracket, 89 field theory, 164–175, 177–187, 238, 397,
cohomology, 376–377, 384 see also TQFT
commuting, 6, 101 group, 180
module, 383 mechanics, 171
structure, 217 Quasi-projective variety, 264, 332
variety, 368–392 Quaternionic
Polar, 272 Lie group, 127
Polarization, 6, 185 vector space, 13
Pole, 49, 57, 120–121, 212–234 Quiver variety, 327, 333–336
Polynomial Quotient
E, 332–333, 338 G3 /G, 218–220
Hilbert, 302 bundle, 406
 Index 431

by finite group, 7 Riemannian

by translation, 35 functional, 20
connection, 226 manifold, 172, 397
discrete, 140–145 metric, 164, 346, 385
geometric, 67 symmetric space, 71
GIT, 337, 340 Rigidity
hyperkähler, 13, 127, 326–330 of a connection, 211–213
interpretation of stack, 125 of geometric structures, 130
of groups, 132
stack, 414 S 1 , 89, 400
symplectic, 4 S-duality, 324–342
S-matrix, 179, 185
Ramification, 118–123 Scalar
Rational curvature, 18, 357–358
curve, 216, 301 flat, 21
map, 4, 66, 229–231, 283–296 Schlesinger equations, 217
Real structure, 43 Schouten bracket, 376
Real-analytic metric, 346–366 Schubert variety, 45
Reduction Schwarz list, 210–234
dimensional, 327, 397 Second fundamental form, 347
symplectic, 307 Seiberg–Witten
Reductive group, 105–106 equations, 3
Reflection group, 221–226 invariant, 27
Regularity, boundary, 261 theory, 17–30
Relative boundary condition, 410 transform, 381
Representation Self-dual
adjoint, 99, 105 connection, 2–3, 35–68
completely reducible, 96 equations, 2, 34, 327
Fuchsian, 129, 137–148, 154 four-manifold, 6
Hilbert space, 182–183 metric, 350
Hitchin, 157–158 two-form, 19, 25, 351–356
holomorphic, 165 Semisimple coefficient, 121–123
holonomy, 138, 145 Semistable, 93, 105, 115
into SL(2, C), 104 vector bundle, 135
into SU (2), 65 Sen’s conjecture, 329–330
irreducible, 49, 178, 351 Sequence
monodromy, 210–234 exact, 52
of fundamental group, 129 short exact, 44
of mapping class group, 179 Serre duality, 150
unitary, 92 Sheaf
variety, 130 coherent, 168, 301–304
Residue, 43, 50, 120, 215–217 divisorial, 305, 317
Resolution, crepant, 331 of groupoids, 238, 257
Riccati of holomorphic maps, 42
equation, 215 of regular functions, 368
solution, 233 torsion-free, 305
Ricci Shift operator, 43
form, 18, 191 Short exact sequence, 44
potential, 191–193 Sigma model, 124–127
tensor, 18, 22, 350, 359 Signature, 400
Riemann Simple closed curve, 178
–Hilbert correspondence, 216, 224 Simplicial space, 244
–Hurwicz formula, 219, 222–223 Singularity, 90, 119–125, 142, 212–234, 307,
–Roch theorem, 63, 203 315
curvature, 20 ADE, 14
sphere, 40, 284 orbifold, 105
surface, 12, 88, 92, 124, 129–158, 164, 175, Skein theory, 179
273, 324, 330, 340, 412 Sklyanin algebra, 392
432 Index

Skyrmion, 7, 281–296 Stable

Slope, 93 λ, 277
Smash, 408 bundle, 5, 166, 396
Soliton, 282 Higgs pair, 136
Space holomorphic vector bundle, 301
classifying, 404 isomorphism, 249, 407
configuration, 171 principal bundle, 105
deformation, 142, 155 subset, 262, 266
double coset, 167 tangent bundle, 407
Eilenberg–MacLane, 404 three-form, 14
Fock, 256 vector bundle, 93, 135
Fricke, 142–148 Stack, 118, 123–127, 166, 398–417
Hermitian symmetric, 146, 156–158 quotient, 414
Hilbert, 398 Star-product, 200–201
homogeneous, 353 Stein manifold, 265
hyperbolic, 137 Stereographic projection, 247, 284
loop, 171–175, 254, 404 Stiefel–Whitney class, 146, 194, 402
Lorentzian, 84 Stokes
Minkowski, 2 data, 224
moduli, see Moduli space matrix, 121
of connections, 126 theorem, 76, 224, 247, 261
of continuous maps, 171 String structure, 257
of maps, 127 Strong-weak duality, 324
projective, 261–278 SU (2) manifold, 348–357
simplicial, 244 SU (3) structure, 359
Stein, 265 Subgroup
symmetric, 77 Borel, 41
Teichmüller, 5–6, 129, 144, 148–150, discrete, 99
187 maximal compact, 105
Thom, 402, 406, 408 parabolic, 44, 105
Space-filling brane, 380 unipotent, 41
Spark, 267 Submanifold, 171, 239–241, 250, 261–278,
Special 347–366, 369, 377–378
Kähler geometry, 7–8 codimension three, 6
Lagrangian fibration, 337 coisotropic, 307–312
Lagrangian submanifold, 7, 301–321 complex, 300
Spectral generalized, 392
cover, 299, 305–306 Lagrangian, 100
curve, 3, 7, 11, 34, 42, 100, 130 one-dimensional, 185
data, 35, 45 special Lagrangian, 7, 301–321
line, 12 Submersion, 242
Spectrum, 168 Subspace, Lagrangian, 158
K-theory ring, 401 Subvariety, 261–278
Thom, 403, 406 Superpotential, 299, 303, 321
Sphere, 244, 347 Surface
affine, 154 cubic, 218
punctured, 210–222 Del Pezzo, 388–390
Riemann, 40, 284 group, 130
six-, 358 Hirzebruch, 304
spectrum, 408 hyperbolic, 137
two-, 314 K3, 1–2, 314
Spin bundle, 24 Kähler, 18
spinc structure, 24–27 Kodair-fibred, 26
Spinning top, 3 of genus g, 178
Spinor, 256, 377, 385 of genus 2, 131, 139–142
harmonic, 38 of genus 3, 142
Square integrable, see L2 of Kähler type, 30
 Index 433

rational or ruled, 23 Darboux, 306

Riemann, 12, 124, 129–158, 164, 175, 273, Federer, 274
324, 330, 340, 412 Gauss, 72
Suspension, 402, 406 Gauss–Bonnet, 24
Symbol, 101, 378, 410 Hurewicz, 133
sequence, 375 Klein, 229
universal, 405–407 Newlander–Nirenberg, 352
Symmetric Poncelet, 6
power, 151 Riemann–Roch, 63, 203
space, 71, 77, 130, 146, 156–158 rigidity, 99
tensor field, 22 Stokes, 76, 224, 247, 261
Symmetry group, 210 Torelli, 103, 134
Symplectic uniformiaztion, 143, 148
form, 100, 195 universal coefficient, 133
geometry, 299–321 vanishing, 65
group, 157 Theta
manifold, 113, 118–122, 129, 134, characteristic, 96
188–189 function, 6–7
quotient, 4 Thickening, 353
reduction, 307 Thom
structure, 158, 375, 378, 387 isomorphism, 402–403
surface, 217 space, 408
transform, 300 spectrum, 403, 406
vector space, 133 Three-form, 369–372
Symplectomorphism, 134 Three-tier theory, 173
SYZ Toeplitz operator, 181–206
conjecture, 7 Toledo invariant, 146–147
construction, 337 Topological
transform, 299–321 field theory, 113
Szegö group, 182
kernel, 2 quantum field theory, see TQFT
projection, 194 solution, 220
vector space, 172
T-duality, 116 Torelli theorem, 103, 134
Tangent bundle, 15, 40, 196–197, 369, 397, Torsion, 372, 386
404–407 -free connection, 350
unit, 158 Torsor, 405–407, 413
Tannaka completion, 98 Torus, 73, 301, 381
Taub–NUT metric, 313–314 2g-dimensional, 92, 129, 134
Teichmüller Lagrangian, 116
component, 104, 153–156 maximal, 41, 119
space, 5–6, 129, 144, 148–150, 187 symplectic, 312
Tensor two-, 88
category, 97, 180 TQFT, 177–187, 395–417
Nijenhuis, 386 Trace-class operator, 172
Ricci, 350, 359 Transform(ation)
Tesselation, 154 Fourier–Laplace, 224–225
Tetrahedral Fourier–Mukai, 299–321
solution, 220, 227, 233 Gelfand, 266
symmetry, 282, 289–292 Legendre, 81
Theorem Möbius, 284
Alexander–Wermer, 261, 271–275 Nahm, 34–68
Atiyah–Singer, 1, 404–407 Seiberg–Witten, 381
Bipolar, 272 SYZ, 299–321
Cartan–Kähler, 346–349 twistor, 39
Cauchy–Kovalewskaya, 352 Transgression, 238
Chevalley, 106 Triangle group, 229–232
434 Index

Triangulation, 251 Prym, 13, 101

Trivialization, Courant, 369–392 quiver, 327, 333–336
Twin Lagrangian fibration, 299–321 representation, 130
Twisted Schubert, 45
de Rham cohomology, 410 Vector
form, 401–404 bundle, 92, 135, 408
K-theory, 403–407 field, 75, 191–193, 201, 370–376,
Twistor 388–390
space, 6, 63 Verlinde
theory, 2–3, 14 bundle, 180
transform, 39 formula, 203
Two-form, 351–356, 381, 386, ring, 400, 417
388–389 Volume form, 73, 385

Uniformization theorem, 143, 148 Weierstrass, cubic, 305

Unipotent subgroup, 41 Weight filration, 332
Unit sphere bundle, 408 Weitzenböck formula, 25, 37
Unitary Wess–Zumino–Witten term, 255
connection, 388 Weyl
group, 156 curvature, 24
representation, 92 group, 106, 215, 219, 227–228
Universal Winding number, 263, 267–269, 276
coefficient theorem, 133 projective, 263
covering, 88
orientation, 407–417 Yamabe
symbol, 405–407 conjecture, 1, 18
constant, 23
Vacuum expectation value, 172 Yang–Mills
Valentiner group, 222, 224, 227 equations, 2–6, 13, 215, 327
Variational problem, 17, 21, 85 functional, 168
Variety, see also Manifold theory, 124–125, 174, 325
abelian, 92, 100, 304–305, 315–321 Young diagram, 178
Albanese, 165
Jacobian, 92, 129, 134 Zariski
Poisson, 368–392 -dense, 158
Projective, 94 topology, 131