Asymptotic Analysis in General Relativity

L O N D O N M AT H E M AT I C A L S O C I E T Y L E C T U R E N OT E S E R I E S
Managing Editor:
Professor M. Reid, Mathematics Institute,
University of Warwick, Coventry CV4 7AL, United Kingdom
The titles below are available from booksellers, or from Cambridge University Press at
www.cambridge.org/mathematics
325 Lectures on the Ricci flow, P. TOPPING
326 Modular representations of finite groups of Lie type, J.E. HUMPHREYS
327 Surveys in combinatorics 2005, B.S. WEBB (ed)
328 Fundamentals of hyperbolic manifolds, R. CANARY, D. EPSTEIN & A. MARDEN (eds)
329 Spaces of Kleinian groups, Y. MINSKY, M. SAKUMA & C. SERIES (eds)
330 Noncommutative localization in algebra and topology, A. RANICKI (ed)
331 Foundations of computational mathematics, Santander 2005, L.M PARDO, A. PINKUS, E. SÜLI
& M.J. TODD (eds)
332 Handbook of tilting theory, L. ANGELERI HÜGEL, D. HAPPEL & H. KRAUSE (eds)
333 Synthetic differential geometry (2nd Edition), A. KOCK
334 The Navier–Stokes equations, N. RILEY & P. DRAZIN
335 Lectures on the combinatorics of free probability, A. NICA & R. SPEICHER
336 Integral closure of ideals, rings, and modules, I. SWANSON & C. HUNEKE
337 Methods in Banach space theory, J.M.F. CASTILLO & W.B. JOHNSON (eds)
338 Surveys in geometry and number theory, N. YOUNG (ed)
339 Groups St Andrews 2005 I, C.M. CAMPBELL, M.R. QUICK, E.F. ROBERTSON & G.C. SMITH (eds)
340 Groups St Andrews 2005 II, C.M. CAMPBELL, M.R. QUICK, E.F. ROBERTSON & G.C. SMITH (eds)
341 Ranks of elliptic curves and random matrix theory, J.B. CONREY, D.W. FARMER, F. MEZZADRI
& N.C. SNAITH (eds)
342 Elliptic cohomology, H.R. MILLER & D.C. RAVENEL (eds)
343 Algebraic cycles and motives I, J. NAGEL & C. PETERS (eds)
344 Algebraic cycles and motives II, J. NAGEL & C. PETERS (eds)
345 Algebraic and analytic geometry, A. NEEMAN
346 Surveys in combinatorics 2007, A. HILTON & J. TALBOT (eds)
347 Surveys in contemporary mathematics, N. YOUNG & Y. CHOI (eds)
348 Transcendental dynamics and complex analysis, P.J. RIPPON & G.M. STALLARD (eds)
349 Model theory with applications to algebra and analysis I, Z. CHATZIDAKIS, D. MACPHERSON, A. PILLAY
& A. WILKIE (eds)
350 Model theory with applications to algebra and analysis II, Z. CHATZIDAKIS, D. MACPHERSON, A. PILLAY
& A. WILKIE (eds)
351 Finite von Neumann algebras and masas, A.M. SINCLAIR & R.R. SMITH
352 Number theory and polynomials, J. MCKEE & C. SMYTH (eds)
353 Trends in stochastic analysis, J. BLATH, P. MÖRTERS & M. SCHEUTZOW (eds)
354 Groups and analysis, K. TENT (ed)
355 Non-equilibrium statistical mechanics and turbulence, J. CARDY, G. FALKOVICH & K. GAWEDZKI
356 Elliptic curves and big Galois representations, D. DELBOURGO
357 Algebraic theory of differential equations, M.A.H. MACCALLUM & A.V. MIKHAILOV (eds)
358 Geometric and cohomological methods in group theory, M.R. BRIDSON, P.H. KROPHOLLER
& I.J. LEARY (eds)
359 Moduli spaces and vector bundles, L. BRAMBILA-PAZ, S.B. BRADLOW, O. GARCÍA-PRADA &
S. RAMANAN (eds)
360 Zariski geometries, B. ZILBER
361 Words: Notes on verbal width in groups, D. SEGAL
362 Differential tensor algebras and their module categories, R. BAUTISTA, L. SALMERÓN & R. ZUAZUA
363 Foundations of computational mathematics, Hong Kong 2008, F. CUCKER, A. PINKUS & M.J. TODD (eds)
364 Partial differential equations and fluid mechanics, J.C. ROBINSON & J.L. RODRIGO (eds)
365 Surveys in combinatorics 2009, S. HUCZYNSKA, J.D. MITCHELL & C.M. RONEY-DOUGAL (eds)
366 Highly oscillatory problems, B. ENGQUIST, A. FOKAS, E. HAIRER & A. ISERLES (eds)
367 Random matrices: High dimensional phenomena, G. BLOWER
368 Geometry of Riemann surfaces, F.P. GARDINER, G. GONZÁLEZ-DIEZ & C. KOUROUNIOTIS (eds)
369 Epidemics and rumours in complex networks, M. DRAIEF & L. MASSOULIÉ
370 Theory of p-adic distributions, S. ALBEVERIO, A.YU. KHRENNIKOV & V.M. SHELKOVICH
371 Conformal fractals, F. PRZYTYCKI & M. URBAŃSKI
372 Moonshine: The first quarter century and beyond, J. LEPOWSKY, J. MCKAY & M.P. TUITE (eds)
373 Smoothness, regularity and complete intersection, J. MAJADAS & A. G. RODICIO
374 Geometric analysis of hyperbolic differential equations: An introduction, S. ALINHAC
375 Triangulated categories, T. HOLM, P. JØRGENSEN & R. ROUQUIER (eds)
376 Permutation patterns, S. LINTON, N. RUŠKUC & V. VATTER (eds)
377 An introduction to Galois cohomology and its applications, G. BERHUY
378 Probability and mathematical genetics, N. H. BINGHAM & C. M. GOLDIE (eds)
379 Finite and algorithmic model theory, J. ESPARZA, C. MICHAUX & C. STEINHORN (eds)
380 Real and complex singularities, M. MANOEL, M.C. ROMERO FUSTER & C.T.C WALL (eds)
381 Symmetries and integrability of difference equations, D. LEVI, P. OLVER, Z. THOMOVA
& P. WINTERNITZ (eds)
Downloaded from https://www.cambridge.org/core. University of New England, on 04 Jan 2018 at 16:45:21, subject to the Cambridge Core
terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781108186612
382 Forcing with random variables and proof complexity, J. KRAJÍČEK
383 Motivic integration and its interactions with model theory and non-Archimedean geometry I, R. CLUCKERS,
J. NICAISE & J. SEBAG (eds)
384 Motivic integration and its interactions with model theory and non-Archimedean geometry II, R. CLUCKERS,
J. NICAISE & J. SEBAG (eds)
385 Entropy of hidden Markov processes and connections to dynamical systems, B. MARCUS, K. PETERSEN
& T. WEISSMAN (eds)
386 Independence-friendly logic, A.L. MANN, G. SANDU & M. SEVENSTER
387 Groups St Andrews 2009 in Bath I, C.M. CAMPBELL et al. (eds)
388 Groups St Andrews 2009 in Bath II, C.M. CAMPBELL et al. (eds)
389 Random fields on the sphere, D. MARINUCCI & G. PECCATI
390 Localization in periodic potentials, D.E. PELINOVSKY
391 Fusion systems in algebra and topology, M. ASCHBACHER, R. KESSAR & B. OLIVER
392 Surveys in combinatorics 2011, R. CHAPMAN (ed)
393 Non-abelian fundamental groups and Iwasawa theory, J. COATES et al. (eds)
394 Variational problems in differential geometry, R. BIELAWSKI, K. HOUSTON & M. SPEIGHT (eds)
395 How groups grow, A. MANN
396 Arithmetic differential operators over the p-adic integers, C.C. RALPH & S.R. SIMANCA
397 Hyperbolic geometry and applications in quantum chaos and cosmology, J. BOLTE & F. STEINER (eds)
398 Mathematical models in contact mechanics, M. SOFONEA & A. MATEI
399 Circuit double cover of graphs, C.-Q. ZHANG
400 Dense sphere packings: a blueprint for formal proofs, T. HALES
401 A double Hall algebra approach to affine quantum Schur–Weyl theory, B. DENG, J. DU & Q. FU
402 Mathematical aspects of fluid mechanics, J.C. ROBINSON, J.L. RODRIGO & W. SADOWSKI (eds)
403 Foundations of computational mathematics, Budapest 2011, F. CUCKER, T. KRICK, A. PINKUS
& A. SZANTO (eds)
404 Operator methods for boundary value problems, S. HASSI, H.S.V. DE SNOO & F.H. SZAFRANIEC (eds)
405 Torsors, étale homotopy and applications to rational points, A.N. SKOROBOGATOV (ed)
406 Appalachian set theory, J. CUMMINGS & E. SCHIMMERLING (eds)
407 The maximal subgroups of the low-dimensional finite classical groups, J.N. BRAY, D.F. HOLT
& C.M. RONEY-DOUGAL
408 Complexity science: the Warwick master’s course, R. BALL, V. KOLOKOLTSOV & R.S. MACKAY (eds)
409 Surveys in combinatorics 2013, S.R. BLACKBURN, S. GERKE & M. WILDON (eds)
410 Representation theory and harmonic analysis of wreath products of finite groups,
T. CECCHERINI-SILBERSTEIN, F. SCARABOTTI & F. TOLLI
411 Moduli spaces, L. BRAMBILA-PAZ, O. GARCÍA-PRADA, P. NEWSTEAD & R.P. THOMAS (eds)
412 Automorphisms and equivalence relations in topological dynamics, D.B. ELLIS & R. ELLIS
413 Optimal transportation, Y. OLLIVIER, H. PAJOT & C. VILLANI (eds)
414 Automorphic forms and Galois representations I, F. DIAMOND, P.L. KASSAEI & M. KIM (eds)
415 Automorphic forms and Galois representations II, F. DIAMOND, P.L. KASSAEI & M. KIM (eds)
416 Reversibility in dynamics and group theory, A.G. O’FARRELL & I. SHORT
417 Recent advances in algebraic geometry, C.D. HACON, M. MUSTAŢǍ & M. POPA (eds)
418 The Bloch–Kato conjecture for the Riemann zeta function, J. COATES, A. RAGHURAM, A. SAIKIA
& R. SUJATHA (eds)
419 The Cauchy problem for non-Lipschitz semi-linear parabolic partial differential equations, J.C. MEYER
& D.J. NEEDHAM
420 Arithmetic and geometry, L. DIEULEFAIT et al. (eds)
421 O-minimality and Diophantine geometry, G.O. JONES & A.J. WILKIE (eds)
422 Groups St Andrews 2013, C.M. CAMPBELL et al. (eds)
423 Inequalities for graph eigenvalues, Z. STANIĆ
424 Surveys in combinatorics 2015, A. CZUMAJ et al. (eds)
425 Geometry, topology and dynamics in negative curvature, C.S. ARAVINDA, F.T. FARRELL & J.-F. LAFONT (eds)
426 Lectures on the theory of water waves, T. BRIDGES, M. GROVES & D. NICHOLLS (eds)
427 Recent advances in Hodge theory, M. KERR & G. PEARLSTEIN (eds)
428 Geometry in a Fréchet context, C. T. J. DODSON, G. GALANIS & E. VASSILIOU
429 Sheaves and functions modulo p, L. TAELMAN
430 Recent progress in the theory of the Euler and Navier–Stokes equations, J.C. ROBINSON, J.L. RODRIGO,
W. SADOWSKI & A. VIDAL-LÓPEZ (eds)
431 Harmonic and subharmonic function theory on the real hyperbolic ball, M. STOLL
432 Topics in graph automorphisms and reconstruction (2nd Edition), J. LAURI & R. SCAPELLATO
433 Regular and irregular holonomic D-modules, M. KASHIWARA & P. SCHAPIRA
434 Analytic semigroups and semilinear initial boundary value problems (2nd Edition), K. TAIRA
435 Graded rings and graded Grothendieck groups, R. HAZRAT
436 Groups, graphs and random walks, T. CECCHERINI-SILBERSTEIN, M. SALVATORI & E. SAVA-HUSS (eds)
437 Dynamics and analytic number theory, D. BADZIAHIN, A. GORODNIK & N. PEYERIMHOFF (eds)
438 Random walks and heat kernels on graphs, M.T. BARLOW
439 Evolution equations, K. AMMARI & S. GERBI (eds)
440 Surveys in combinatorics 2017, A. CLAESSON et al. (eds)
441 Polynomials and the mod 2 Steenrod algebra I, G. WALKER & R.M.W. WOOD
442 Polynomials and the mod 2 Steenrod algebra II, G. WALKER & R.M.W. WOOD
443 Asymptotic analysis in general relativity, T. DAUDÉ, D. HÄFNER & J.-P. NICOLAS (eds)
444 Geometric and cohomological group theory, P.H. KROPHOLLER, I.J. LEARY, C. MARTÍNEZ-PÉREZ &
B.E.A. NUCINKIS (eds)
London Mathematical Society Lecture Note Series: 443
Asymptotic Analysis in General Relativity
Edited by
T H I E R RY DAU D É
Université de Cergy-Pontoise, France
D I E T R I C H H Ä F N E R
Université Grenoble Alpes, France
JEAN-PHILIPPE NICOLAS
Université de Bretagne Occidentale, France
University Printing House, Cambridge CB2 8BS, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre,
New Delhi – 110025, India
79 Anson Road, #06–04/06, Singapore 079906
Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of
education, learning, and research at the highest international levels of excellence.
www.cambridge.org
Information on this title: www.cambridge.org/9781316649404
DOI: 10.1017/9781108186612
© Cambridge University Press 2018
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2018
Printed in the United Kingdom by Clays, St Ives plc
A catalogue record for this publication is available from the British Library.
Library of Congress Cataloguing in Publication data
Names: Daudé, Thierry, 1977– editor. | Häfner, Dietrich, editor. |
Nicolas, J.-P. (Jean-Philippe), editor.
Title: Asymptotic analysis in general relativity / edited by Thierry Daudé
(Université de Cergy-Pontoise), Dietrich Häfner (Université Grenoble Alpes),
Jean-Philippe Nicolas (Université de Bretagne Occidentale).
Other titles: London Mathematical Society lecture note series ; 443.
Description: Cambridge, United Kingdom ; New York, NY :
Cambridge University Press, 2017. |
Series: London Mathematical Society lecture note series ; 443 |
Includes bibliographical references.
Identifiers: LCCN 2017023160 | ISBN 9781316649404 (pbk.) | ISBN 1316649407 (pbk.)
Subjects: LCSH: General relativity (Physics)–Mathematics.
Classification: LCC QC173.6 .A83 2017 | DDC 530.11–dc23 LC record available at
https://lccn.loc.gov/2017023160
ISBN 978-1-316-64940-4 Paperback
Cambridge University Press has no responsibility for the persistence or accuracy of
URLs for external or third-party internet websites referred to in this publication
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
Contents
1 Introduction to Modern Methods for Classical and

Quantum Fields in General Relativity page 1
Thierry Daudé, Dietrich Häfner and Jean-Philippe Nicolas
1.1 Geometry of Black Hole Spacetimes 2
1.2 Quantum Field Theory on Curved Spacetimes 3
1.3 Conformal Geometry and Conformal Tractor Calculus 4
1.4 A Minicourse in Microlocal Analysis and Wave Propagation 5
References 6
2 Geometry of Black Hole Spacetimes 9

Lars Andersson, Thomas Bäckdahl and Pieter Blue
2.1 Introduction 9
2.2 Background 13
2.3 Black Holes 28
2.4 Spin Geometry 43
2.5 The Kerr Spacetime 53
2.6 Monotonicity and Dispersion 55
2.7 Symmetry Operators 65
2.8 Conservation Laws for the Teukolsky System 70
2.9 A Morawetz Estimate for the Maxwell Field on Schwarzschild 75
References 79
3 An Introduction to Conformal Geometry and Tractor

Calculus, with a view to Applications in General Relativity 86
Sean N. Curry and A. Rod Gover
3.1 Introduction 86
3.2 Lecture 1: Riemannian Invariants and Invariant Operators 91
v
vi Contents
3.3 Lecture 2: Conformal Transformations and Conformal Covariance 94

3.4 Lecture 3: Prolongation and the Tractor Connection 104
3.5 Lecture 4: The Tractor Curvature, Conformal Invariants and
Invariant Operators 120
3.6 Lecture 5: Conformal Compactification of Pseudo-Riemannian
Manifolds 128
3.7 Lecture 6: Conformal Hypersurfaces 145
3.8 Lecture 7: Geometry of Conformal Infinity 151
3.9 Lecture 8: Boundary Calculus and Asymptotic Analysis 154
Appendix: Conformal Killing Vector Fields and Adjoint Tractors 160
References 168
4 An Introduction to Quantum Field Theory on

Curved Spacetimes 171
Christian Gérard
4.1 Introduction 171
4.2 A Quick Introduction to Quantum Mechanics 175
4.3 Notation 178
4.4 CCR and CAR Algebras 179
4.5 States on CCR/CAR Algebras 183
4.6 Lorentzian Manifolds 190
4.7 Klein–Gordon Fields on Lorentzian Manifolds 192
4.8 Free Dirac Fields on Lorentzian Manifolds 197
4.9 Microlocal Analysis of Klein–Gordon Quasi-Free States 201
4.10 Construction of Hadamard States 209
References 217
5 A Minicourse on Microlocal Analysis for Wave Propagation 219

András Vasy
5.1 Introduction 219
5.2 The Overview 221
5.3 The Basics of Microlocal Analysis 230
5.4 Propagation Phenomena 285
5.5 Conformally Compact Spaces 315
5.6 Microlocal Analysis in the b-Setting 349
References 371
1
Introduction to Modern Methods for Classical
and Quantum Fields in General Relativity
Thierry Daudé, Dietrich Häfner and Jean-Philippe Nicolas
The last few decades have seen major developments in asymptotic analysis
in the framework of general relativity, with the emergence of methods that,
until recently, were not applied to curved Lorentzian geometries. This has led
notably to the proof of the stability of the Kerr–de Sitter spacetime by P. Hintz
and A. Vasy [17]. An essential feature of many recent works in the field is the
use of dispersive estimates; they are at the core of most stability results and are
also crucial for the construction of quantum states in quantum field theory,
domains that have a priori little in common. Such estimates are in general
obtained through geometric energy estimates (also referred to as vector field
methods) or via microlocal/spectral analysis. In our minds, the two approaches
should be regarded as complementary, and this is a message we hope this
volume will convey succesfully. More generally than dispersive estimates,
asymptotic analysis is concerned with establishing scattering-type results.
Another fundamental example of such results is asymptotic completeness,
which, in many cases, can be translated in terms of conformal geometry as
the well-posedness of a characteristic Cauchy problem (Goursat problem) at
null infinity. This has been used to develop alternative approaches to scattering
theory via conformal compactifications (see for instance F. G. Friedlander
[11] and L. Mason and J.-P. Nicolas [22]). The presence of symmetries in the
geometrical background can be a tremendous help in proving scattering results,
dispersive estimates in particular. What we mean by symmetry is generally the
existence of an isometry associated with the flow of a Killing vector field,
though there exists a more subtle type of symmetry, described sometimes as
hidden, corresponding to the presence of Killing spinors for instance. Recently,
the vector field method has been adapted to take such generalized symmetries
into account by L. Andersson and P. Blue in [2].
This volume compiles notes from the eight-hour mini-courses given at the
summer school on asymptotic analysis in general relativity, held at the Institut
1
Downloaded from https://www.cambridge.org/core. University of Exeter, on 30 Dec 2017 at 23:36:17, subject to the Cambridge Core terms
of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781108186612.001
2 T. Daudé, D. Häfner and J.-P. Nicolas
Fourier in Grenoble, France, from 16 June to 4 July 2014. The purpose of the
summer school was to draw an up-to-date panorama of the new techniques
that have influenced the asymptotic analysis of classical and quantum fields in
general relativity in recent years. It consisted of five mini-courses:
• “Geometry of black hole spacetimes” by Lars Andersson, Albert Einstein
Institut, Golm, Germany;
• “An introduction to quantum field theory on curved spacetimes” by Christian
Gérard, Paris 11 University, Orsay, France;
• “An introduction to conformal geometry and tractor calculus, with a view to
applications in general relativity” by Rod Gover, Auckland University, New
Zealand;
• “The bounded L2 conjecture” by Jérémie Szeftel, Paris 6 University, France;
• “A minicourse on microlocal analysis for wave propagation” by András Vasy,
Stanford University, United States of America.
Among these, only four are featured in this book. The proof of the bounded
L2 conjecture having already appeared in two different forms [20, 21], Jérémie
Szeftel preferred not to add yet another version of this result; his lecture notes
are therefore not included in the present volume.
1.1. Geometry of Black Hole Spacetimes

The notion of a black hole dates back to the 18th century with the works of
Simpson and Laplace, but it found its modern description within the framework
of general relativity. In fact the year after the publication of the general
theory of relativity by Einstein, Karl Schwarzschild [30] found an explicit
non-trivial solution of the Einstein equations that was later understood to
describe a universe containing nothing but an eternal spherical black hole.
The Kerr solution appeared in 1963 [19] and, with the singularity theorems
of Hawking and Penrose [15], black holes were eventually understood as
inevitable dynamical features of the evolution of the universe rather than mere
mathematical oddities. The way exact black hole solutions of the Einstein
equations were discovered was by imposing symmetries. First Schwarzschild
looked for spherically symmetric and static solutions in four spacetime
dimensions, which reduces the Einstein equations to a non-linear ordinary
differential equation (ODE). The Kerr solution appears when one relaxes one
of the symmetries and looks for stationary and axially symmetric solutions.
Roy Kerr obtained his solution by imposing on the metric the so-called “Kerr–
Schild” ansatz that corresponds to assuming a special algebraic property for
Modern Methods for Classical and Quantum Fields in General Relativity 3
the Weyl tensor, namely that it has Petrov-type D, which is similar to the
condition for a polynomial to have two double roots. This algebraic speciality
of the Weyl tensor can be understood as another type of symmetry assumption
about spacetime. This is a generalized symmetry that does not correspond
to an isometry generated by the flow of a vector field, but is related to the
existence of a Killing spinor. The Kerr family, which contains Schwarzschild’s
spacetime as the zero angular momentum case, is expected to be the unique
family of asymptotically flat and stationary (perhaps pseudo-stationary, or
locally stationary, would be more appropriate) black hole solutions of the
Einstein vacuum equations (there is a vast literature on this topic, see for
example the original paper by D. Robinson [27], his review article [28] and the
recent analytic approach by S. Alexakis, A. D. Ionescu, and S. Klainerman [1]).
Moreover it is believed to be stable (there is also an important literature on this
question, the stability of Kerr–de Sitter black holes was established recently in
[17], though the stability of the Kerr metric is still an open problem). These
two conjectures play a crucial role in physics where it is commonly assumed
that the long term dynamics of a black hole stabilizes to a Kerr solution. The
extended lecture notes by Lars Andersson, Thomas Bäckdahl, and Pieter Blue
take us through the many topics that are relevant to the questions of stability
and uniqueness of the Kerr metric, including the geometry of stationary and
dynamical black holes with a particular emphasis on the special features of
the Kerr metric, spin geometry, dispersive estimates for hyperbolic equations
and generalized symmetry operators. The type D structure is an essential focus
of the course, with the intimate links between the principal null directions,
the Killing spinor, Killing vectors and tensors, Killing–Yano tensors and
symmetry operators. All these notions are used in the final sections where
some conservation laws are derived for the Teukolsky system governing the
evolution of spin n/2 zero rest-mass fields, and a new proof of a Morawetz
estimate for Maxwell fields on the Schwarzschild metric is given.
1.2. Quantum Field Theory on Curved Spacetimes

In the 1980s, Dimock and Kay started a research program concerning scatter-
ing theory for classical and quantum fields on the Schwarzschild spacetime;
see [9]. Their work was then pushed further by Bachelot, Häfner, and others,
leading in particular to a mathematically rigorous description of the Hawking
effect on Schwarzschild and Kerr spacetimes, see e.g. [4], [14]. In the
Schwarzschild case there exists a global timelike Killing vector field in the
exterior of the black hole that can be used to define vacuum and thermal states.
However, it is not clear how to extend these states to the whole spacetime.
From a more conceptual point of view this is also quite unsatisfactory because
the construction of vacuum states on the Minkowski spacetime uses the
full Poincaré group. In addition general spacetimes will not even be locally
stationary. On a curved spacetime, vacuum states are therefore replaced by
so-called Hadamard states. These Hadamard states were first characterized
by properties of their two-point functions, which had to have a specific
asymptotic expansion near the diagonal. In 1995 Radzikowski reformulated
the old Hadamard condition in terms of the wave front set of the two-point
function; see [26]. Since then, microlocal analysis has played an important
role in quantum field theory in curved spacetime, see e.g. the construction
of Hadamard states using pseudodifferential calculus by Gérard and Wrochna
[13]. The lectures given by Christian Gérard give an introduction to quantum
field theory on curved spacetimes and in particular to the construction of
Hadamard states.
1.3. Conformal Geometry and Conformal Tractor Calculus

Conformal compactifications were initially used in general relativity by André
Lichnerowicz for the study of the constraints. It is Roger Penrose who
started applying this technique to Lorentzian manifolds, more specifically
to asymptotically flat spacetimes, in the early 1960s (see Penrose [25]). The
purpose was to replace complicated asymptotic analysis by simple and natural
geometrical constructions. To be precise, a conformal compactification allows
one to describe infinity for a spacetime (M, g) as a finite boundary for the
manifold M equipped with a well-chosen metric ĝ that is conformally related
to g. Provided a field equation has a suitably simple transformation law under
conformal rescalings, ideally conformal invariance or at least some conformal
covariance, the asymptotic behavior of the field on (M, g) can be inferred
from the local properties at the boundary of the conformally rescaled field
on (M, ĝ). Penrose’s immediate goal was to give a simple reformulation of
the Sachs peeling property as the continuity at the conformal boundary of the
rescaled field. But he had a longer term motivation which was to construct a
conformal scattering theory for general relativity, allowing the setting of data
for the spacetime at its past null conformal boundary and to propagate the
associated solution of the Einstein equations right up to its future null confor-
mal boundary. Since its introduction, the conformal technique has been used
to prove global existence for the Einstein equations, or other non-linear hyper-
bolic equations, for sufficiently small data (see for example Y. Choquet-Bruhat
and J. W. York [8]), to construct scattering theories for linear and non-linear test
fields, initially on static backgrounds and, in recent years, in time dependent
situations and on black hole spacetimes (see L. Mason and J.-P. Nicolas [22]
and Nicolas [24] and references therein). It has also been applied to spacetimes
with a non-zero cosmological constant. There is an important literature from
the schools of R. Mazzeo and R. Melrose and more recently numerous studies
using the tractor calculus approach by A. R. Gover and his collaborators.
Tractor calculus in its conformal version started from the notion of a local
twistor bundle on four-dimensional spin-manifolds as an associated bundle to
the Cartan conformal connection, though it in fact dates back to T. Y. Thomas’s
work [31]. The theory in its modern form first appeared in the founding paper
by T. Bailey, M. Eastwood, and Gover [6] where its origins are also thoroughly
detailed. The extended lecture notes by Sean Curry and Rod Gover give an
up-to-date presentation of the conformal tractor calculus: the first four lectures
are mainly focused on the search for invariants; the second half of the course
uses tractor calculus to study conformally compact manifolds with application
to general relativity as its main motivation.
1.4. A Minicourse in Microlocal Analysis

and Wave Propagation
One of the central questions in mathematical relativity is the stability of
the Kerr or the Kerr–de Sitter spacetime. As mentioned above, stability has
been established by Hintz and Vasy for the Kerr–de Sitter metric, and the
question remains open for the Kerr metric. The advantage of the Kerr–de Sitter
case is that the inverse of the Fourier transformed d’Alembert operator has a
meromorphic extension across the real axis in appropriate weighted spaces.
The poles of this extension are then called resonances. Resonances in general
relativity were first studied from a mathematical point of view by Bachelot
and Motet-Bachelot in [5]. Bony and Häfner gave a resonance expansion of
the local propagator for the wave equation on the Schwarzschild–de Sitter
metric [7] using the localization of resonances by Sá Barreto-Zworski [29].
Then Dyatlov, Hintz, Vasy, Wunsch, and Zworski made new progress leading
eventually to a resonance expansion for the wave equation on spacetimes which
are perturbations of the Kerr–de Sitter metric; see the work of Vasy [32]. The
whole program culminated in the proof of the non-linear stability of the Kerr–
de Sitter metric by Hintz and Vasy [17]. Many aspects come into this study.
The first is trapping. Trapping situations were studied in the 1980s for the wave
equation outside two obstacles by Ikawa who obtained local energy decay with
loss of derivatives in this situation; see [18]. The trapping that appears on the
Kerr (or the Kerr–de Sitter) metric is r-normally hyperbolic at least for small
angular momentum. Suitable resolvent estimates for this kind of situation have
been shown by Wunsch–Zworski [33] and Dyatlov [10]. Another important
aspect is the presence of supperradiance due to the fact that there is no globally
timelike Killing field outside a Kerr–de Sitter black hole. Whereas the cut-off
resolvent can nevertheless be extended meromorphically across the real axis
using the work of Mazzeo–Melrose [23] and several different Killing fields
(see [12]), a more powerful tool to obtain suitable estimates is the Fredholm
theory for non-elliptic settings developed by Vasy [32]. Microlocal analysis
was first developed for linear problems. Nevertheless, as the work of Hintz–
Vasy shows strikingly enough, it is also well adapted to quasilinear problems.
In this context one needs to generalize some of the important theorems (such
as the propagation of singularities) to very rough metrics. This program has
been achieved by Hintz; see [16]. The last important aspect in the proof of
the non-linear stability of the Kerr–de Sitter metric is the issue of the gauge
freedom in the Einstein equations. Roughly speaking, a linearization of the
Einstein equations can create resonances whose imaginary parts have the “bad
sign,” leading to exponentially growing modes. These resonances turn out to be
“pure gauge” and can therefore be eliminated by an adequate choice of gauge;
see [17]. The lectures notes by András Vasy introduce the essential tools used
in the proof of the non-linear stability of the Kerr–de Sitter metric.
References
[1] S. Alexakis, A. D. Ionescu, S. Klainerman, Rigidity of stationary black holes
with small angular momentum on the horizon, Duke Math. J. 163 (2014), 14,
2603–2615.
[2] L. Andersson, P. Blue, Hidden symmetries and decay for the wave equation on the
Kerr spacetime, Ann. of Math. (2) 182 (2015), 3, 787–853.
[3] L. Andersson, P. Blue, Uniform energy bound and asymptotics for the Maxwell
field on a slowly rotating Kerr black hole exterior, J. Hyperbolic Differ. Equ., 12
(2015), 4, 689–743.
[4] A. Bachelot, The Hawking effect, Ann. Inst. H. Poincaré Phys. Théor. 70 (1999),
1, 41–99.
[5] A. Bachelot, A. Motet-Bachelot, Les résonances d’un trou noir de Schwarzschild,
Ann. Inst. H. Poincaré Phys. Théor. 59 (1993), 1, 3–68.
[6] T. N. Bailey, M. G. Eastwood, A. R. Gover, Thomas’s structure bundle for
conformal, projective and related structures, Rocky Mountains J. Math. 24
(1994), 4, 1191–1217.
[7] J.-F. Bony, D. Häfner, Decay and non-decay of the local energy for the wave
equation on the de Sitter–Schwarzschild metric, Comm. Math. Phys. 282 (2008),
3, 697–719.
[8] Y. Choquet-Bruhat, J. W. York, The Cauchy problem. In A. Held, editor, General
relativity and gravitation, Vol. 1, 99–172, Plenum, New York and London, 1980.
[9] J. Dimock, B. S. Kay, Classical and quantum scattering theory for linear scalar
fields on the Schwarzschild metric, Ann. Physics 175 (1987), 2, 366–426.
[10] S. Dyatlov, Spectral gaps for normally hyperbolic trapping, Ann. Inst. Fourier
(Grenoble) 66 (2016), 1, 55–82.
[11] F.G. Friedlander, Radiation fields and hyperbolic scattering theory, Math. Proc.
Camb. Phil. Soc. 88 (1980), 483–515.
[12] V. Georgescu, C. Gérard, D. Häfner, Asymptotic completeness for superradiant
Klein–Gordon equations and applications to the De Sitter Kerr metric, J. Eur.
Math. Soc. 19 (2017), 2371–2444.
[13] C. Gérard, M. Wrochna, Construction of Hadamard states by pseudo-differential
calculus, Comm. Math. Phys. 325 (2014), 2, 713–755.
[14] D. Häfner, Creation of fermions by rotating charged black holes, Mém. Soc. Math.
Fr. (N.S.) 117 (2009), 158 pp.
[15] S. Hawking, R. Penrose, The singularities of gravitational collapse and cosmol-
ogy, Proc. Roy. Soc. London Series A, Mathematical and Physical Sciences, 314
(1970), 1519, 529–548.
[16] P. Hintz, Global analysis of quasilinear wave equations on asymptotically de
Sitter spaces, Ann. Inst. Fourier (Grenoble) 66 (2016), 4, 1285–1408.
[17] P. Hintz, A. Vasy, The global non-linear stability of the Kerr–de Sitter family of
black holes, arXiv:1606.04014.
[18] M. Ikawa, Decay of solutions of the wave equation in the exterior of two convex
obstacles, Osaka J. Math. 19 (1982), 3, 459–509.
[19] R. P. Kerr, Gravitational field of a spinning mass as an example of algebraically
special metrics, Phys. Rev. Letters 11 (1963), 5, 237–238.
[20] S. Klainerman, I. Rodnianski, J. Szeftel, The bounded L2 curvature conjecture,
Invent. Math. 202 (2015), 1, 91–216.
[21] S. Klainerman, I. Rodnianski, J. Szeftel, Overview of the proof of the bounded L2
curvature conjecture, arXiv:1204.1772v2.
[22] L. J. Mason, J.-P. Nicolas, Conformal scattering and the Goursat problem, J.
Hyperbolic Differ. Equ., 1 (2) (2004), 197–233.
[23] R. Mazzeo, R. Melrose, Meromorphic extension of the resolvent on complete
spaces with asymptotically constant negative curvature, J. Funct. Anal. 75 (1987),
2, 260–310.
[24] J.-P. Nicolas, Conformal scattering on the Schwarzschild metric, Ann. Inst.
Fourier (Grenoble) 66 (2016), 3, 1175–1216.
[25] R. Penrose, Zero rest-mass fields including gravitation: asymptotic behaviour,
Proc. Roy. Soc. London A284 (1965), 159–203.
[26] M. Radzikowski, Micro-local approach to the Hadamard condition in quantum
field theory on curved space–time, Comm. Math. Phys. 179 (1996), 3, 529–553.
[27] D. C. Robinson, Uniqueness of the Kerr black hole, Phys. Rev. Lett. 34 (1975),
905.
[28] D. C. Robinson, Four decades of black hole uniqueness theorems. In D.L.

Wiltshire, M. Visser and S.M. Scott, editors, The Kerr space–time, 115–143,
Cambridge University Press, 2009.
[29] A. Sá Barreto, M. Zworski, Distribution of resonances for spherical black holes,
Math. Res. Lett. 4 (1997), 1, 103–121.
[30] K. Schwarzschild, Über der Gravitationsfeld eines Massenpunktes nach der
Einsteinschen Theorie, K. Preus. Akad. Wiss. Sitz. 424 (1916).
[31] T. Y. Thomas, On conformal geometry, Proc. Nat. Acad. Sci. 12 (1926), 352–359.
[32] A. Vasy, Microlocal analysis of asymptotically hyperbolic and Kerr–de Sitter
spaces (with an appendix by Semyon Dyatlov), Invent. Math. 194 (2013), 2,
381–513.
[33] J. Wunsch, M. Zworski, Resolvent estimates for normally hyperbolic trapped sets,
Ann. Henri Poincaré 12 (2011), 7, 1349–1385.
Laboratoire AGM, Département de Mathématiques, Université de Cergy-Pontoise,

95302 Cergy-Pontoise cedex
E-mail address: [email protected]
Université Grenoble Alpes, Institut Fourier, UMR 5582 du CNRS, 100, rue des maths,
38610 Gières, France
LMBA, Université de Brest, 6 avenue Victor Le Gorgeu, 29238 Brest Cedex 3, France
2
Geometry of Black Hole Spacetimes
Lars Andersson, Thomas Bäckdahl and Pieter Blue
Abstract. These notes, based on lectures given at the summer school on

Asymptotic Analysis in General Relativity, collect material on the Einstein
equations, the geometry of black hole spacetimes, and the analysis of fields
on black hole backgrounds. The Kerr model of a rotating black hole in a
vacuum is expected to be unique and stable. The problem of proving these
fundamental facts provides the background for the material presented in these
notes.
Among the many topics which are relevant to the uniqueness and stability
problems are the theory of fields on black hole spacetimes, in particular for
gravitational perturbations of the Kerr black hole and, more generally, the
study of nonlinear field equations in the presence of trapping. The study
of these questions requires tools from several different fields, including
Lorentzian geometry, hyperbolic differential equations, and spin geometry,
which are all relevant to the black hole stability problem.
2.1. Introduction
A short time after Einstein published his field equations for general relativity
in 1915, Karl Schwarzschild discovered an exact and explicit solution of the
Einstein vacuum equations describing the gravitational field of a spherical body
at rest. In analyzing Schwarzschild’s solution, one finds that if the central
body is sufficiently concentrated, light emitted from its surface cannot reach
an observer at infinity. It was not until the 1950s that the global structure of
the Schwarzschild spacetime was understood. By this time causality theory
and the Cauchy problem for the Einstein equations were firmly established,
although many important problems remained open. Observations of highly
energetic phenomena occurring within small spacetime regions, eg. quasars,
made it plausible that black holes played a significant role in astrophysics, and
by the late 1960s these objects were part of mainstream astronomy. The term
9
10 L. Andersson, T. Bäckdahl and P. Blue
“black hole” for this type of object came into use in the 1960s. According
to our current understanding, black holes are ubiquitous in the universe, in
particular most galaxies have a supermassive black hole at their center, and
these play an important role in the life of the galaxy. Our galaxy also has at
its center a very compact object, Sagittarius A*, with a diameter of less than
one astronomical unit, and a mass estimated to be 106 M . Evidence for this
includes observations of the orbits of stars in its vicinity.
Recall that a solution to the Einstein vacuum equations is a Lorentzian
spacetime (M, gab ), satisfying Rab = 0, where Rab is the Ricci tensor of gab .
The Einstein equation is the Euler–Lagrange equation of the diffeomorphism
invariant Einstein–Hilbert action functional, given by the integral of the scalar
curvature of (M, gab ),

Rdμg .
M
The diffeomorphism invariance, or general covariance, of the action has the
consequence that Cauchy data for the Einstein equation must satisfy a set
of constraint equations, and that the principal symbol of the Euler–Lagrange
equation is degenerate.1 After introducing suitable gauge conditions, the
Einstein equations can be reduced to a hyperbolic system of evolution equa-
tions. It is known that, for any set of sufficiently regular Cauchy data satisfying
the constraints, the Cauchy problem for the Einstein equation has a unique
solution which is maximal among all regular, vacuum Cauchy developments.
This general result, however, does not give any detailed information about the
properties of the maximal development.
There are two main conjectures about the maximal development. The strong
cosmic censorship conjecture (SCC) states that a generic maximal develop-
ment is inextendible, as a regular vacuum spacetime. There are examples where
the maximal development is extendible, and has non-unique extensions, which
furthermore may contain closed timelike curves. In these cases, predictability
fails for the Einstein equations, but if SCC holds, they are non-generic. At
present, SCC is only known to hold in the context of families of spacetimes
with symmetry restrictions; see [98, 7] and references therein. Further, some
1
From the perspective of hyperbolic partial differential equations, the Einstein equations are both
over and under-determined. Contracting the Einstein equation against the normal to a smooth
spacelike hypersurface gives elliptic equations that must be satisfied on the hypersurface;
these are called the constraint equations. After introducing suitable gauge conditions, the
combination of the gauge conditions and the remaining Einstein equations form a hyperbolic
system of evolution equations. Furthermore, if the initial data satisfies the constraint equations,
then the solution to this hyperbolic system, when restricted to any spacelike hypersurface, also
satisfies the constraint equations. If the initial hypersurface is null, the situation becomes more
complicated to summarize but simpler to treat in full detail.
Geometry of Black Hole Spacetimes 11
non-linear stability results without symmetry assumptions, including the sta-

bility of Minkowski space and the stability of quotients of the Milne model
(also known as Löbell spacetimes, see [53, 18] and references therein), can
be viewed as giving support to SCC. The weak cosmic censorship conjecture
states that for a generic isolated system (i.e. an asymptotically flat solution of
the Einstein equations), any singularity is hidden from observers at infinity. In
this case, the spacetime contains a black hole region, i.e. the complement of
the part of the spacetime visible to observers at infinity. The black hole region
is bounded by the event horizon, the boundary of the region of spacetime
which can be seen by observers at future infinity. Both of these conjectures
remain wide open, although there has been limited progress on some problems
related to them. The weak cosmic censorship conjecture is most relevant for
the purpose of these notes; see [110].
The Schwarzschild solution is static, spherically symmetric, asymptotically
flat, and has a single free parameter M which represents the mass of the
black hole. By Birkhoff’s theorem it is the unique solution of the vacuum
Einstein equations with these properties. In 1963 Roy Kerr [68] discovered
a new, explicit family of asymptotically flat solutions of the vacuum Einstein
equations which are stationary, axisymmetric, and rotating. Shortly after this,
a charged, rotating black hole solution to the Einstein–Maxwell equations,
known as the Kerr–Newman solution, was found, cf. [87, 88]. Recall that a
vector field ν a is Killing if ∇(a νb) = 0. A Kerr spacetime admits two Killing
fields, the stationary Killing field (∂t )a which is timelike at infinity, and the
axial Killing field (∂φ )a . The Kerr family of solutions is parametrized by the
mass M and the azimuthal angular momentum per unit mass a. In the limit
a = 0, the Kerr solution reduces to the spherically symmetric Schwarzschild
solution.
If |a| ≤ M, the Kerr spacetime contains a black hole, while if |a| > M,
there is a ringlike singularity which is naked, in the sense that it fails to be
hidden from observers at infinity. This situation would violate the weak cosmic
censorship conjecture, and one therefore expects that an overextreme Kerr
spacetime is unstable and, in particular, that it cannot arise through a dynamical
process from regular Cauchy data.
For a geodesic γ a (λ) with velocity γ̇ a = dγ a /dλ, in a stationary
axisymmetric spacetime,2 there are three conserved quantities, the mass
μ2 = γ̇ a γ̇b , energy e = γ̇ a (∂t )a , and angular momentum z = γ̇ a (∂φ )a .
In a general axisymmetric spacetime, geodesic motion is chaotic. However, as
was discovered by Brandon Carter in 1968, there is a fourth conserved quantity
2
We use the signature + − −−; in particular timelike vectors have a positive norm.
for geodesics in the Kerr spacetime, the Carter constant k; see Section 2.5 for
details. By Liouville’s theorem, this allows one to integrate the geodesic
equations by quadratures, and thus geodesics in the Kerr spacetime do not
exhibit a chaotic behavior.
The Carter constant is a manifestation of the fact that the Kerr spacetime is
algebraically special, of Petrov type {2, 2}, also known as type D. In particular,
there are two repeated principal null directions for the Weyl tensor. As shown
by Walker and Penrose [112] a vacuum spacetime of Petrov type {2, 2} admits
an object satisfying a generalization of Killing’s equation, namely a Killing
spinor κAB , satisfying ∇A (A κBC) = 0. As shown in the just cited paper, this
leads to the presence of four conserved quantities for null geodesics.
Assuming some technical conditions, any asymptotically flat, stationary
black hole spacetime is expected to belong to the Kerr family, a fact which is
known to hold in the real-analytic case. Further, the Kerr black hole is expected
to be stable in the sense that a small perturbation of the Kerr spacetime settles
down asymptotically to a member of the Kerr family.
There is much observational evidence pointing to the fact that black holes
exist in large numbers in the universe, and that they play a role in many
astrophysically significant processes. For example, most galaxies, including
our own galaxy, are believed to contain a supermassive black hole at their
center. Further, dynamical processes involving black holes, such as mergers,
are expected to be important sources of gravitational wave radiation, which
could be observed by existing and planned gravitational wave observatories.3
Thus, black holes play a central role in astrophysics.
Due to its conjectured uniqueness and stability properties, these black holes
are expected to be modelled by the Kerr or Kerr–Newman solutions. However,
in order to establish the astrophysical relevance of the Kerr solution, it is vital
to find rigorous proofs of both of these conjectures, which can be referred
to as the black hole uniqueness and stability problems, respectively. A great
deal of work has been devoted to these and related problems, and although
progress has been made, both remain open at present. The stability problem
for the analog of the Kerr solution in the presence of a positive cosmological
constant, the Kerr–de Sitter solution, has recently been solved for the case of
small angular momenta [117].
Overview
Section 2.2 introduces a range of background material on general relativity,
including a discussion of the Cauchy problem for the Einstein equations.
3
At the time of writing, the first such observation has just been announced [1].
The discussion of black hole spacetimes is started in Section 2.3 with a

detailed discussion of the global geometry of the extended Schwarzschild
spacetime, followed by some background on marginally outer trapped surfaces
and dynamical black holes. Section 2.4 introduces some concepts from
spin geometry and the related Geroch–Held–Penrose (GHP) formalism. The
Petrov classification is introduced and some properties of its consequential
algebraically special spacetimes are presented. In Section 2.5 the geometry
of the Kerr black hole spacetime is introduced.
Section 2.6 contains a discussion of null geodesics in the Kerr spacetime.
A construction of monotone quantities for null geodesics based on vector
fields with coefficients depending on conserved quantities, is introduced. In
Section 2.7, symmetry operators for fields on the Kerr spacetime are discussed.
Dispersive estimates for fields are the analog of monotone quantities for null
geodesics, and in constructing these, symmetry operators play a role analogous
to the conserved quantities for the case of geodesics.
2.2. Background
2.2.1. Minkowski Space
Minkowski space M is R4 with metric which in a Cartesian coordinate system
(xa ) = (t, xi ) takes the form4
2
dτM = dt2 − (dx1 )2 − (dx2 )2 − (dx3 )2 .
Introducing the spherical coordinates r, θ , φ we can write the metric in the form
−dt2 + dr2 + r2 d
2S2 , where d
2S2 is the line element on the standard S2 ,
d
2S2 = (gS2 )ab dxa dxb = dθ 2 + sin2 θ dφ 2 . (2.1)
A tangent vector ν a is timelike, null, or spacelike when gab ν a ν b > 0, = 0, or
< 0, respectively. Vectors with gab ν a ν b ≥ 0 are called causal. Let p, q ∈ M.
We say that p is in the causal (timelike) future of q
if p − q is causal (timelike). The causal and timelike
futures J + (p) and I + (p) of p ∈ M are the sets of
points which are in the causal and timelike futures
of p, respectively. The corresponding past notions
are defined analogously.
4 2 = (g ) dxa dxb , and metrics, eg. (g ) ,

Here and below we shall use line elements, eg. dτM M ab M ab
interchangeably.
Let u, v be given by
u = t − r, v = t + r.
In terms of these coordinates the line element takes the form
2
dτM = dudv − r2 d
2S2 . (2.2)
We see that there are no terms du2 , dv2 , which correspond to the fact that
both u, v are null coordinates. In particular, the vectors (∂u )a , (∂v )a are null.
A complex null tetrad is given by
√ 1
la = 2(∂u )a = √ ∂t )a + (∂r )a , (2.3a)
2
√ 1
na = 2(∂v )a = √ (∂t )a − (∂r )a , (2.3b)
2

1 i
ma = √ (∂θ )a + (∂φ )a (2.3c)
2r sin θ
normalized so that na la = 1 = −ma m̄a , with all other inner products of tetrad
legs zero. Complex null tetrads with this normalization play a central role in the
Newman–Penrose (NP) and GHP formalisms; see Section 2.4. In these notes
we will use such tetrads unless otherwise stated.
In terms of a null tetrad, we have
gab = 2(l(a nb) − m(a m̄b) ). (2.4)
Introduce compactified null coordinates U, V, given by
U = arctan u, V = arctan v.
These take values in {(−π/2, π/2) × (−π/2, π/2)} ∩ {V ≥ U}, and we
can thus present Minkowski space in a causal diagram; see Figure 2.1. Here
each point represents an S2 and we have drawn null vectors at 45◦ angles.
A compactification of Minkowski space is now given by adding the null
boundaries5 I ± , spatial infinity i0 , and timelike infinity i± as indicated in the
figure. Explicitly,
I + = {V = π/2}
I − = {U = −π/2}
i0 = {V = π/2, U = −π/2}
i± = {(V, U) = ±(π/2, π/2)}.
5
Here I is pronounced “Scri” for “script I.”
i+
I+
r=0
∂U
∂V
i0
I−
{t = constant}
i−
Figure 2.1. Causal diagram of Minkowski space
In Figure 2.1, we have also indicated schematically the t-level sets which
approach spatial infinity i0 . Causal diagrams are a useful tool which, if applied
with proper care, can be used to understand the structure of quite general
spacetimes. Such diagrams are often referred to as Penrose or Carter–Penrose
diagrams.
In particular, as can be seen from Figure 2.1, we have M = I − (I + )∩I + (I − ),
i.e. any point in M is in the past of I + and in the future of I − . This is
related to the fact that M is asymptotically simple, in the sense that it admits a
conformal compactification with a regular null boundary, and has the property
that any inextendible null geodesic hits the null boundary. For massless fields
on Minkowski space, this means that it makes sense to formulate a scattering
map which takes data on I − to data on I + ; see [93].
Let
T = V + U, R = V − U. (2.5)
Then, with 2 = 2 cos U cos V, the conformally transformed metric g̃ab =

2 gab takes the form
g̃M
ab = dT − dR − sin Rd
S2
2 2 2 2
= dT 2 − d
2S3
which we recognize as the metric on the cylinder R × S3 . This spacetime is

known as the Einstein cylinder, and can be viewed as a static solution of the
Einstein equations with dust matter and a positive cosmological constant [50].
2.2.2. Lorentzian Geometry and Causality

We now consider a smooth Lorentzian four-manifold (M, gab ) with signature
+ − −−. Each tangent space in a four-dimensional spacetime is isometric to
Minkowski space M, and we can carry intuitive notions of causality over from
M to M. We say that a smooth curve γ a (λ) is causal if the velocity vector
γ̇ a = dγ a /dλ is causal. Two points in M are causally related if they can be
connected by a piecewise smooth causal curve. The concept of causal curves
is most naturally defined for C0 curves. A C0 curve γ a is said to be causal if
each pair of points on γ a are causally related. We may define a timelike curve
and timelike related points in an analogous manner.
We now assume that M is time oriented, i.e. that there is a globally
defined timelike vector field on M. This allows us to distinguish between
future and past directed causal curves, and to introduce a notion of the causal
and timelike future of a spacetime point. The corresponding past notions are
defined analogously. If q is in the causal future of p, we write p q. This
introduces a partial order on M. The causal future J + (p) of p is defined as
J + (p) = {q : p q} while the timelike future I + (p) is defined in an analogous
manner, with timelike replacing causal. A subset ⊂ M is achronal
I+(p) if there is no pair p, q ∈ M such that q ∈
I + (p), i.e. does not intersect its timelike future
or past. The domain of dependence D(S) of S ⊂ M
S
is the set of points p such that any inextendible
p
causal curve starting at p must intersect S.
Definition 2.1 A spacetime M is globally hyperbolic if there is a closed,
achronal ⊂ M such that M = D( ). In this case, is called a Cauchy
surface.
D(S)
Due to the results of Bernal and Sanchez
[28], global hyperbolicity is characterized by the S
existence of a smooth, Cauchy time function
τ : M → R. A function τ on M is a time
function if ∇ a τ is timelike everywhere, and it is Cauchy if the level sets
t = τ −1 (t) are Cauchy surfaces. If τ is smooth, its level sets are then
smooth and spacelike. It follows that a globally hyperbolic spacetime M is
globally foliated by Cauchy surfaces, and in particular is diffeomorphic to a
product × R. In the following, unless otherwise stated, we shall consider

only globally hyperbolic spacetimes.
If a globally hyperbolic spacetime M is a subset of a spacetime M , then
the boundary ∂M of M in M is called the Cauchy horizon.
Example 2.1 Let O be the origin in Minkowski space, and let M = I + (O) =
{t > r} be its timelike
√ future. Then M is globally hyperbolic with Cauchy
time function τ = t2 − r2 . Further, M is a subset of Minkowski space M,
which is a globally hyperbolic space with Cauchy time function t. Minkowski
space is geodesically complete and hence inextendible. The boundary {t = r}
is the Cauchy horizon ∂M of M. Past inextendible causal geodesics (i.e. past
causal rays) in M end on ∂M. In particular, M is incomplete. However, M is
extendible, as a smooth flat spacetime, with many inequivalent extensions.
We remark that for a globally hyperbolic spacetime, which is extendible, the

extension is in general non-unique. In the particular case considered in example
2.1, M is an extension of M, which also happens to be maximal and globally
hyperbolic. In the vacuum case, there is a unique maximal globally hyperbolic
extension, cf. Section 2.2.5 below. However, a maximal extension is in general
non-unique, and may fail to be globally hyperbolic.
2.2.3. Conventions and Notation

We will mostly use abstract indices, cf. [94], but will sometimes work with
coordinate indices, and unless confusion arises we will not be too specific about
this. We raise and lower indices with gab , e.g. ξ a = gab ξb , with gab gbc = δ a c ,
where δ a c is the Kronecker delta, i.e. the tensor with the property that δ a c ξ c =
ξ a for any ξ a .
Let a···d be the Levi-Civita symbol, i.e. the skew symmetric expression
which in any coordinate system has the property that 1···n = 1. The volume
√
form of gab is (μg )abcd = |g|abcd . Given (M, gab ) we have the canonically
defined Levi-Civita covariant derivative ∇a . For a vector ν a , this is of the form
∇a ν b = ∂a ν b + ac
b c
ν
where ac b = 1 gbd (∂ g + ∂ g − ∂ g ) is the Christoffel symbol. In order

2 a dc c db d ac
to fix the conventions used here, we recall that the Riemann curvature tensor is
defined by
(∇a ∇b − ∇b ∇a )ξc = Rabc d ξd .
The Riemann tensor Rabcd is skew symmetric in the pairs of indices ab, cd,
Rabcd = R[ab]cd = Rab[cd] , is pairwise symmetric Rabcd = Rcdab , and satisfies
the first Bianchi identity R[abc]d = 0. Here square brackets [· · · ] denote

antisymmetrization. We shall similarly use round brackets (· · · ) to denote
symmetrization. Further, we have ∇[a Rbc]de = 0, the second Bianchi identity. A
contraction gives ∇ a Rabcd = 0. The Ricci tensor is Rab = Rc acb and the scalar
curvature R = Ra a . We further let Sab = Rab − 14 Rgab denote the tracefree part
of the Ricci tensor. The Riemann tensor can be decomposed as follows,
Rabcd = − 12 gad gbc R + 12 gac gbd R + 2 gbd Sac

1 1 1
− 12 gbc Sad
− 2 gad Sbc + 2 gac Sbd + Cabcd .
1 1
(2.6)
This defines the Weyl tensor Cabcd which is a tensor with the symmetries of
the Riemann tensor, and vanishing traces, Cc acb = 0. Recall that (M, gab )
is locally conformally flat if and only if Cabcd = 0. It follows from the
contracted second Bianchi identity that the Einstein tensor Gab = Rab − 12 Rgab
is conserved, ∇ a Gab = 0.
2.2.4. Einstein Equation

The Einstein equation in geometrized units with G = c = 1, where G, c denote
Newton’s constant and the speed of light, respectively, cf. [109, Appendix F],
is the system
Gab = 8πTab . (2.7)
This equation relates geometry, expressed in the Einstein tensor Gab on the
left-hand side, to matter, expressed via the energy momentum tensor Tab on the
right-hand side. For example, for a self-gravitating Maxwell field Fab , Fab =
F[ab] , we have

1 1
Tab = Fac Fbc − Fcd F gab .
cd
4π 4
The source-free Maxwell field equations
∇ a Fab = 0, ∇[a Fbc] = 0
imply that Tab is conserved, ∇ a Tab = 0. The contracted second Bianchi

identity implies that ∇ a Gab = 0, and hence the conservation property of Tab
is implied by the coupling of the Maxwell field to gravity. These facts can be
seen to follow from the variational formulation of Einstein gravity, given by
the action

R
I= dμg − Lmatter dμg
M 16π M
where Lmatter is the Lagrangian describing the matter content in spacetime. In

the case of Maxwell theory, this is given by
1
LMaxwell =
Fcd F cd .
4π
Recall that in order to derive the Maxwell field equation, as an Euler–Lagrange
equation, from this action, it is necessary to introduce a vector potential for Fab ,
by setting Fab = 2∇[a Ab] , and to carry out the variation with respect to Aa . It
is a general fact that for generally covariant (i.e. diffeomorphism invariant)
Lagrangian field theories which depend on the spacetime location only via the
metric and its derivatives, the symmetric energy momentum tensor
1 ∂Lmatter
Tab = √
g ∂gab
is conserved when evaluated on solutions of the Euler–Lagrange equations.
As a further example of a matter field, we consider the scalar field, with
action
Lscalar = 12 ∇ c ψ∇c ψ
where ψ is a function on M. The corresponding energy-momentum tensor is
Tab = ∇a ψ∇b ψ − 12 ∇ c ψ∇c ψgab
and the Euler–Lagrange equation is the free scalar wave equation
∇ a ∇a ψ = 0. (2.8)
As (2.8) is another example of a field equation derived from a covariant
action which depends on the spacetime location only via the metric gab or its
derivatives, the symmetric energy-momentum tensor is conserved for solutions
of the field equation.
In both of the just mentioned cases, the energy momentum tensor satisfies
the dominant energy condition Tab ν a ζ b ≥ 0 for future directed causal vectors
ν a , ζ a . This implies the null energy condition
Rab ν a ν b ≥ 0 if νa ν a = 0. (2.9)
These energy conditions hold for most classical matter models.
There are many interesting matter systems which are worthy of considera-
tion, such as fluids, elasticity, kinetic matter models including Vlasov, as well
as fundamental fields such as Yang–Mills, to name just a few. We consider only
spacetimes which satisfy the null energy condition, and for the most part we
shall in these notes be concerned with the vacuum Einstein equations,
Rab = 0. (2.10)
2.2.5. The Cauchy Problem

Given a spacelike hypersurface6 in M with timelike normal T a , induced
metric hab , and second fundamental form kab , defined by kab X a Y b =
∇a Tb X a Y b for X a , Y b tangent to , the Gauss and Gauss–Codazzi equations
imply the constraint equations
R[h] + (kab hab )2 − kab kab = 16πTab T a T b (2.11a)
∇[h]a (kbc h ) − ∇[h] kab = Tab T .
bc b b
(2.11b)
A three-manifold together with tensor fields hab , kab on solving the
constraint equations is called a Cauchy data set. The constraint equations
for general relativity are analogs of the constraint equations in Maxwell and
Yang–Mills theory, in that they lead to Hamiltonians which generate gauge
transformations.
Consider a 3+1 split of M, i.e. a one-parameter family of Cauchy surfaces
t , with a coordinate system (xa ) = (t, xi ), and let
(∂t )a = NT a + X a
be the split of (∂t )a into a normal and tangential piece. The fields (N, X a ) are
called lapse and shift. The definition of the second fundamental form implies
the equation
L∂t hab = −2Nkab + LX hab .
In the vacuum case, the Hamiltonian for gravity can be written in the form

NH + X a Ja + boundary terms
where H and J are the densitized left-hand sides of (2.11). If we consider

only compactly supported perturbations in deriving the Hamiltonian evolution
equation, the boundary terms mentioned above can be ignored. However,
for (N, X a ) not tending to zero at infinity, and considering perturbations
compatible with asymptotic flatness, the boundary term becomes significant,
cf. Section 2.2.6.4.
The resulting Hamiltonian evolution
√ equations, written in terms of hab and
its canonical conjugate π ab = h(kab − (hcd kcd hab )), are usually called the
ADM (for Arnowitt–Deser–Misner) evolution equations.
Let ⊂ M be a Cauchy surface. Given functions φ0 , φ1 on and F on M,
the Cauchy problem for the wave equation is that of finding solutions to
6
Where there is no likelihood of confusion, we shall denote abstract indices for objects on by
a, b, c, . . . .

∇ a ∇a ψ = F, ψ = φ 0 , L∂t ψ = φ1 .
Assuming suitable regularity conditions, the solution is unique and stable with
respect to the initial data. This fact extends to a wide class of non-linear hyper-
bolic PDEs including quasilinear wave equations, i.e. equations of the form
Aab [ψ]∂a ∂b ψ + B[ψ, ∂ψ] = 0
with Aab a Lorentzian metric depending on the field ψ.
Given a vacuum Cauchy data set, ( , hab , kab ), a solution of the Cauchy
problem for the Einstein vacuum equations is a spacetime metric gab with
Rab = 0, such that (hab , kab ) coincides with the metric and second fundamental
form induced on from gab . Such a solution is called a vacuum extension of
( , hab , kab ).
Due to the fact that Rab is covariant, the symbol of Rab is degenerate. In
order to get a well-posed Cauchy problem, it is necessary either to impose
gauge conditions or to introduce new variables. A standard choice of gauge
condition is the harmonic coordinate condition. Let gab be a given metric on
M. The identity map i : M → M is harmonic if and only if the vector field
V a = gbc (bc
a
−a
bc )
vanishes. Here bca , a are the Christoffel symbols of the metrics g ,
bc ab gab .
Then V is the tension field of the identity map i : (M, gab ) → (M,
a gab ). This
is harmonic if and only if
V a = 0. (2.12)
Since harmonic maps with a Lorentzian domain are often called wave maps,
the gauge condition (2.12) is sometimes called a wave map gauge condition.
A particular case of this construction, which can be carried out if M admits a
global coordinate system (xa ), is given by letting
gab be the Minkowski metric
defined with respect to (xa ). Then a = 0 and (2.12) is simply
bc
∇ b ∇b xa = 0, (2.13)
which is usually called the wave coordinate gauge condition.
Going back to the general case, let ∇ be the Levi-Civita covariant derivative
defined with respect to
gab . We have the identity
1 √ ab + ∇(a Vb)
Rab = − 12 √ ∇ a gg ∇b gab + Sab [g, ∇g] (2.14)
g
a gcd . Setting
where Sab is an expression which is quadratic in first derivatives ∇
V = 0 in (2.14) yields Rab , and (2.10) becomes a quasilinear wave equation
a harm
ab = 0.
Rharm (2.15)
By standard results, the equation (2.15) has a locally well-posed Cauchy

problem in Sobolev spaces H s for s > 5/2. Using more sophisticated
techniques, well-posedness can shown to hold for any s > 2 [71]. Recently
a local existence has been proved under the assumption of curvature bounded
in L2 [73]. Given a Cauchy data set ( , hab , kab ), together with initial values
for lapse and shift N, X a on , it is possible to find Lt N, Lt X a on such
that the V a are zero on . A calculation now shows that, due to the constraint
equations, L∂t V a is zero on . Given a solution to the reduced Einstein vacuum
equation (2.15), one finds that V a solves a wave equation. This follows from
∇ a Gab = 0, due to the Bianchi identity. Hence, due to the fact that the Cauchy
data for V a is trivial, it holds that V a = 0 on the domain of the solution. Thus,
the solution to (2.15) is a solution to the full vacuum Einstein equation (2.10).
This proves local well-posedness for the Cauchy problem for the Einstein
vacuum equation. This fact was first proved by Yvonne Choquet-Bruhat [54];
see [99] for background and history.
Global uniqueness for the Einstein vacuum equations was proved by
Choquet-Bruhat and Geroch [35]. The proof relies on the local existence
theorem sketched above, patching together local solutions. A partial order is
defined on the collection of vacuum extensions, making use of the notion of
a common domain. The common domain U of two extensions M, M is the
maximal subset in M which is isometric to a subset in M . We can then define
a partial order by saying that M ≤ M if the maximal common domain is M.
Given a partially ordered set, a maximal element exists by Zorn’s lemma. This
is proven to be unique by an application of the local well-posedness theorem
for the Cauchy problem sketched above. For a contradiction, let M, M be two
inequivalent extensions, and let U be the maximal common domain. Due to the
Haussdorff property of spacetimes, this leads to a contradiction. By finding a
partial Cauchy surface which touches the boundary of U (see Figure 2.2 and
making use of local uniqueness) one finds a contradiction to the maximality of
U. It should be noted that here uniqueness holds up to isometry, in keeping with
the general covariance of the Einstein vacuum equations. These facts extend
to the Einstein equations coupled to hyperbolic matter equations. See [101]
∂U
Figure 2.2. A partial Cauchy surface which touches the boundary of ∂U
for a construction of the maximal globally hyperbolic extension which does

not rely on Zorn’s lemma; see also [114]. The global uniqueness result can
be generalized to Einstein-matter systems, provided the matter field equation
is hyperbolic and that its solutions do not break down. General results on
this topic are lacking, see however [92] and references therein. The minimal
regularity needed for global uniqueness is a subtle issue, which has not been
fully addressed. In particular, results on local well-posedness are known; see
e.g. [72] and references therein, which require less regularity than the best
results on global uniqueness.
2.2.6. Remarks
We shall now make several remarks relating to the above discussion.
2.2.6.1. Bianchi Identities as a Hyperbolic System

The vacuum Einstein equation Rab = 0 implies that the Weyl tensor Cabcd sat-
isfies the Bianchi identity ∇ a Cabcd = 0. This is the massless spin-2 equation.
In particular, this is a first order hyperbolic system for the Weyl tensor.
The spin-2 equation (i.e. the equation ∇ a Wabcd for a Weyl test field Wabcd ,
a tensor field with the symmetries and trace properties of the Weyl tensor)
implies algebraic conditions relating the field and the curvature. In particular,
in a sufficiently general background a Weyl test field must be proportional to
the Weyl tensor Cabcd of the spacetime. This holds in particular for spacetimes
of Petrov type D (cf. Section 2.4.6 below for the definition of Petrov types; see
[10, §2.3] and references therein).
One may view the Bianchi identity for the Weyl tensor as the main gravita-
tional field equation, and the vacuum Einstein equation as a type of “constraint”
equation, which allows one to relate the Weyl tensor to the Riemann curvature
of spacetime. The first-order system for the Weyl tensor can be extended to a
first-order system including the first and second Cartan structure equations. A
hyperbolic system can be extracted by introducing suitable gauge conditions;
see Section 2.2.6.3.
2.2.6.2. Null Condition

Consider the Cauchy problem for the semilinear wave equation on Minkowski
space,
∇ a ∇a ψ = Qab ∇a ψ∇b ψ

with data ψ t=0 = ψ0 , ∂t ψ t=0 = ψ1 , where > 0 and ψ1 , ψ2 are suitably
regular functions. Solutions exist globally for small data (i.e. for sufficiently
small > 0) if and only if Qab satisfies the null condition, Qab ξa ξb = 0,
for any null vector ξ a . An example given by Fritz John shows that the equation
∇ a ∇a ψ = |∂t ψ|2 , for which the null condition fails, can have blowup for small
data, cf. [104]. Similar results hold also for quasilinear equations; in particular
for quasilinear wave equations satisfying a suitable null condition, one has
stability of the trivial solution.
For the vacuum Einstein equation in harmonic coordinates, we have
ab = − 2 g ∂c ∂d gab + Sab (g, ∂g)

Rharm 1 cd
where the lower order term Sab contains terms of the form ∂a gcd ∂b gef gce gdf ,
and hence the null condition fails to hold for the Einstein vacuum equation in
harmonic coordinates. For this reason the problem of stability of Minkowski
space in Einstein gravity is subtle. The stability of Minkowski space was first
proved by Christodoulou and Klainerman [37]. Later a proof using harmonic
coordinates was given by Lindblad and Rodnianski [76]. This exploits the
fact that the equation Rharm
ab = 0 satisfies a weak form of the null condition.
Consider the system
∇ a ∇a ψ = |∂t φ|2 (2.16a)
∇ ∇a φ = Q ∇a φ∇b φ
a ab
(2.16b)
on Minkowski space, where Qab has a null structure. For this system, the null
condition fails to hold. However, φ satisfies an equation with null structure and
therefore has good dispersion. The equation for ψ has a source defined in terms
of φ but no bad self-interaction. One finds therefore that the solution to (2.16)
exists globally for small data, but with slightly slower falloff than a solution of
an equation satisfying the null condition.
2.2.6.3. Gauge Source Functions

As has been pointed out by Helmut Friedrich, see [55] for discussion, one
may introduce gauge source functions V a = F a (xb , gcd ) without affecting
the reduction procedure. The gauge source functions can be designed to yield
damping effects or to control the evolution of the lapse and shift. This has
frequently been used in numerical relativity. A related strategy is to add terms
involving factors of the constraints Ca . Such terms vanish for a solution of the
field equations, but may provide improved behavior for the reduced system.
It is often convenient to introduce a suitably normalized tetrad ea a .
Important examples are orthonormal tetrads, satisfying ea a eb b gab =
diag(+1, −1, −1, −1), and the null tetrads (la , na , ma , m̄a ) with la na = 1,
ma m̄a = −1, all other inner products being zero. Such tetrads appear naturally
when working with spinors; see Section 2.4.
The field equations can be written as a system of equations for tetrad

components, connection coefficients, and curvature. Introducing tetrad gauge
source functions Vab = (∇ c ∇c eaa )ebb gab it is possible to extract a first-order
symmetric hyperbolic system with V a , Vab taking values involving tetrad,
connection coefficients and curvature. This opens up a lot of interesting
possibilities, but has not been widely used. The phantom gauge introduced by
Chandrasekhar [34, p. 240] was shown in [3] to correspond to a tetrad gauge
condition of the above type, and is therefore compatible with a well-posed
Cauchy problem.
Let (M, gab ) be a vacuum spacetime. Let g(s)ab be a one-parameter family
of vacuum metrics and let

d
hab = g(s)ab .
ds s=0
Then hab solves the linearized Einstein equation DRab = 0, where DRab is the
Frechet derivative of the Ricci tensor at gab in the direction hab . A calculation,
cf. [3], shows that if we impose the linearized wave map gauge condition, then
hab satisfies the Lichnerowicz wave equation
∇ c ∇c hab + 2Racbd hcd = 0.
2.2.6.4. Asymptotically Flat Data

The Kerr black hole represents an isolated system, and the appropriate data
for the black hole stability problem should therefore be asymptotically flat.
To make this precise we suppose there is a compact set K in M and a map
: M \ K → R3 \ B(R, 0), where B(R, 0) is a Euclidean ball. This defines a
Cartesian coordinate system on the end M \ K so that hab − δab falls off to zero
at infinity, at a suitable rate. Here δab is the Euclidean metric in the Cartesian
coordinate system constructed above. Similarly, we require that kab falls off
to zero.
Let xa be the chosen Euclidean coordinate system and let r be the Euclidean
radius r = (δab xa xb )1/2 . Following Regge and Teitelboim [97], see also [25],
we assume that gab = δab + hab with
hab = O(1/r), ∂a hbc = O(1/r2 ),
kab = O(1/r2 ).
Further, we impose the parity conditions
hab (x) = hab (−x), kab (x) = −kab (−x). (2.17)
These falloff and parity conditions guarantee that the ADM 4-momentum and
angular momentum are well defined. It was shown in [63] that data satisfying
the parity condition conditions (2.17) are dense among data which satisfy an
asymptotic flatness condition in terms of weighted Sobolev spaces.
Let ξ a be an element of the Poincare Lie algebra and assume that NT a + X a
tends in a suitable sense to ξ a at infinity. Then the action for Einstein gravity
can be written in the form

Rdμg = Pa ξ a + π ij ḣij − NH + X i Ji .
M
Here we may view Pa as a map to the dual of the Poincare Lie algebra, i.e.
a momentum map. Evaluating Pa ξ a on a particular element of the Poincare
Lie algebra gives the corresponding momentum. These can also be viewed as
charges at infinity. We have

1
P0 = lim (∂i gji − ∂j gii )dσ i (2.18a)
16π r→∞ Sr

1
Pi = lim πij dσ j (2.18b)
8π r→∞ Sr
where dσ i denotes the hypersurface area element of a family of spheres (which

can be taken to be coordinate spheres) Sr foliating a neighborhood of infinity.
See [82] and references therein for a recent discussion of the conditions under
which these expressions are well-defined.
The energy and linear momentum (P0 , Pi ) provide the components of a
4-vector Pa , the ADM 4-momentum. Assuming the dominant energy condition
then under the above asymptotic conditions, Pa is future causal, and timelike
unless the maximal development (M, gab ) is isometric to Minkowski space.
Further, Pa transforms as a Minkowski 4-vector, and the ADM mass is given
√
by M = Pa Pa . The boost theorem [38] implies, given an asymptotically
flat Cauchy data set, that one may find a boosted Cauchy surface in its
development such that the data is in the rest frame with respect to , i.e.
Pa = M(∂t )a .
Since the constraint quantities H, Ji vanish for solutions of the Einstein
equations, the gravitational Hamiltonian takes the value Pa ξ a , and hence the
ADM mass and momenta defined by (2.18) are conserved for an evolution
with lapse and shift (N, X i ) → (1, 0) at infinity. If we consider the analog of
the above definitions for a hyperboloidal slice which meets I, then the ADM
mass and momentum are replaced by the Bondi mass and momentum. An
example of a hyperboloidal slice in Minkowski space is given by a level set
of the time function T , cf. (2.5), in the compactification of Minkowski space.
For the Bondi 4-momentum, one has the important feature that gravitational
energy is radiated through I, which means that it is not conserved. See [40]
and references therein for further details.
2.2.6.5. Killing Initial Data

A Killing initial data set is a Cauchy data set ( , hab , kab ) such that the
development (M, gab ) is a spacetime with a Killing field ν a , i.e.
Lν gab = 2∇(a νb) = 0.
Now let ν a be a solution to the wave equation ∇ a ∇a νb = 0, but not necessarily
a Killing field. In a vacuum spacetime, we then have
∇ d ∇d (∇(a νb) ) = 2Rc (ab) d ∇(c νd) .
This implies that the tensor Lν gab satisfies a wave equation, so if it has trivial
Cauchy data on , then ν a is a Killing field in the domain of dependence of
. This allows us to characterize Lie symmetries of a development (M, gab )
purely in terms of the Cauchy data. Another way to formulate this statement is
that Lie symmetries propagate. This fact, which is closely related to the global
uniqueness for the Cauchy problem, allows one to study symmetry restrictions
of the Einstein equations. Much work has been done to study consistent
subsystems of the Einstein equation, implied by imposing symmetries on
the initial data. Examples include Bianchi (spatially homogenous), T 2 (two
commuting spatial Killing fields with torus orbits), U 1 (one spatial Killing
field generating a circle action). Note, however, there are also the so-called
surface symmetric spacetimes, which arise in a somewhat different manner.
In addition, there are consistent subsystems which are not given by symmetry
restrictions. Examples are the polarized Gowdy spacetimes and half-polarized
T 2 -symmetric spacetimes. See [7] and references therein for further details.
The analog of the principle that symmetries propagate is also valid for
spinors. This leads to the notion of Killing spinor initial data, which is relevant
to the problem of Kerr characterization; see [20] for further details.
2.2.6.6. Komar Integrals

Assume that ν a is a Killing vector field. Then we have ∇a νb = ∇[a νb] . A
calculation shows
∇ a (∇a ξb − ∇b ξa ) = −2Rbc ξ c .
Hence, in a vacuum,

eabcd ∇ c ξ d
S
depends only on the homology class of the two-surface S. The analogous fact
have ∇ Fab =cd0, ∇[a Fbc] = 0,
for the source free Maxwell equation, where we a
is the conservation of the charge integrals S Fab , S abcd F , which again

depend only on the homology class of S. These statements are immediate
consequences of Stokes’ theorem.
If we consider asymptotically flat spacetimes, we have in the stationary case,

with ξ a = (∂t )a ,

1
P ξa = −
a
abcd ∇ c ξ d ,
8π S
where on the left-hand side we have the ADM 4-momentum evaluated at
infinity. Similarly, in the axially symmetric case, with ηa = (∂φ )a ,

1
J=− abcd ∇ c ηd .
16π S
These integrals again depend only on the homology class of S. See [66, §6] for
background to these facts. For a non-symmetric, but asymptotically flat, space-
time, letting S tend to infinity through a sequence of suitably round spheres
yields the linkage integrals, which again reproduce the ADM momenta [113].
2.3. Black Holes

2.3.1. The Schwarzschild Solution
Before introducing the Kerr solution, we will discuss the spherically sym-
metric, static Schwarzschild black hole spacetime. This exhibits some of the
features of the Kerr solution and has the advantage that the algebraic form of
the line element is much simpler. However, it must be noted that, due to the
fact that the Schwarzschild spacetime is static and spherically symmetric, the
essential difficulties in analyzing fields on the Kerr background stemming from
the complicated trapping (i.e. the fact null geodesics orbiting the black hole fill
an open spacetime region, cf. section 2.3.7) and superradiance are not seen
in the Schwarzschild case. Therefore, one should be careful in generalizing
notions from Schwarzschild to Kerr.
In Schwarzschild coordinates (t, r, θ , φ), the Schwarzschild metric takes
the form
gab dxa dxb = fdt2 − f −1 dr2 − r2 d
2S2 (2.19)
with f = 1 − 2M/r. Here d

2S2 = dθ 2 + sin2 θ dφ 2 is the line element on the
unit 2-sphere. The coordinate r is the area radius, defined by 4π r2 = A(S(r, t)),
where S(r, t) is the 2-sphere with constant t, r. The line element given in
equation (2.19) is valid for r > 0, but has a coordinate singularity at r = 2M,
which is also the location of the event horizon. Historically, this fact caused
some confusion, and was only fully cleared up in the 1950s due to the work
of Kruskal and Szekeres, see e.g. [84, Chapter 31] and references therein.
The metric is in fact regular, and the line element given above is valid also
for 0 < r < 2M. At r = 0, there is a curvature singularity, where spacetime
curvature diverges as 1/r3 . The Schwarzschild metric is asymptotically flat and

the parameter M coincides with the ADM mass.
We remark that by setting f = 1 − 2M/r + Q2 /r2 , the Schwarzschild
line element becomes that of Reissner–Nordström, a spherically symmetric
solution to the Einstein–Maxwell
ab equations, with field strength of the form
F tr = Q/r2 . Here Q = 8π1
S F dσab .
In order to get a better understanding of the Schwarzschild spacetime, it
is instructive to consider its maximal extension. In order to do this, we first
introduce the tortoise coordinate r∗ ,

r
r∗ = r + 2M log −1 . (2.20)
2M
r∗
r
r+
This solves dr∗ = f −1 dr, r∗ (4M) = 4M. As r 2M, r∗ diverges

logarithmically to −∞, and for large r, r∗ ∼ r. Inverting (2.20) yields

r∗
r = 2MW e 2M −1 + 2M (2.21)
where W is the principal branch of the Lambert W function.7 We can now

introduce null coordinates
u = t − r∗ , v = t + r∗ .
A null tetrad is given by

2 a
la = ∂ ,
f v

2 a
n =
a
∂ ,
f u

1 i a
m =√
a
∂θ +
a
∂ .
2r sin θ φ
7
The Lambert W function, or product logarithm, is defined as the solution of W(x)eW(x) = x for
x > 0. It satisfies W (x) = W(x)/((W(x) + 1)x). The principal branch is analytic at x = 0 and is
real valued in the range (−e−1 , ∞) with values in (−1, ∞). In particular, W(0) = 0. See [44].
On the exterior region in Schwarzschild, (u, v) take values in the range

(−∞, ∞) × (−∞, ∞). Let U , V be a pair of coordinates taking values in
(−π/2, π/2), and related to u, v by
u = − 4M log(− tan U ), U ∈ (−π/2, 0)

v = 4M log(tan V), V ∈ (0, π/2).
We have
t = 12 (v + u) = 4M log (− tan V tan U )

tan V
r∗ = 12 (v − u) = 4M log − .
tan U
In terms of U , V we have
r = 2MW(−e−1 tan U tan V) + 2M (2.22)
and r > 0 thus corresponds to tan U tan V < 1. The line element now takes the
form
dUdV 32M 3 − r
gab dxa dxb = e 2M − r2 d
2S2 . (2.23)
cos2 U cos2 V r
The form (2.23) of the Schwarzschild line element is non-degenerate in the
range
(U, V) ∈ (−π/2, π/2) × (−π/2, π/2) ∩ {−π/2 < U + V < π/2}. (2.24)
In particular, the location r = 2M of the coordinate singularity in the

line element (2.19) corresponds to U V = 0. The line element (2.23) has
a coordinate singularity, which is also a curvature singularity, at r = 0
(corresponding to tan U tan V = 1), and at U = ±π/2, V = ±π/2
(corresponding to u, v taking unbounded values). Figure 2.3 shows the region
given in (2.24), with lines of constant t, r indicated. Using the causal diagram
for the extended Schwarzschild solution, one can easily find the null infinities
I ± , spatial infinity i0 , timelike infinities i± , and the horizons H± at r = 2M,
which are indicated. Region I is the domain of outer communication, i.e.
I − (I + ) ∩ I + (I − ), while region II is the future trapped (or black hole) region.
The level sets of t hit the bifurcation sphere B located at U = V = 0,
where ∂t = 0. In particular, we see that the Schwarzschild coordinates are
degenerate, since the level sets of t do not foliate the extended Schwarzschild
spacetime. On the other hand, a global Cauchy foliation of the maximally
extended Schwarzschild spacetime is given by the level sets of the Kruskal
time function T = 12 (V + U ).
i+
I+
2M
3M
III II I
r=
r=
B i0
IV I−
i−
{t = constant}
Figure 2.3. Causal diagram for the extended Schwarzschild solution
ka Given a null vector ka , perpendicular to a spacelike two-

surface S, we may define the null expansion with respect to ka by
ka = 12 δka log(A(S)) (2.25)

S
where δka denotes the variation in the direction ka .
Then ka is the expansion of the area element of S, along the null geodesic
with velocity ka . If we let ka = (∂V )a , we have
⎧
⎨ > 0 in region I,
k a = 0 on H+ ,
⎩
< 0 in region II.
Thus, the area of a bundle of null rays in region I is expanding with respect to
a future, outgoing null vector like ∂V , while in region II they are contracting.
Actually, in region II, we find that the expansion with respect to any future null
vector is negative.
Although null vectors are conventionally drawn at 45◦ angles, due to the fact
that each point in the causal diagram represents a sphere, this does not give a
complete description. From the causal diagram it is clear that from each point
in the DOC there are null curves which escape through I ± or fall in through
the horizons H± . By continuity, it is clear that there must be null curves which
neither escape through I nor fall in through the horizons H. We refer to these as
orbiting or trapped null geodesics. In the Schwarzschild spacetime, the trapped
null geodesics are located at r = 3M; see Figure 2.3. The presence of trapped
null geodesics is a robust feature of black hole spacetimes.
Although the region covered by the null coordinates U , V is compact, the
line element (2.23) is of course isometric to the form given in (2.19). A
conformal factor = cos U cos V may now be introduced, which brings I ±
to a finite distance. Letting g̃ab = 2 gab , and adding these boundary pieces to
(M, g̃ab ), provides a conformal compactification8 of the maximally extended
Schwarzschild spacetime.
2.3.1.1. Gravitational Redshift

A robust fact about black hole spacetimes is that radiation emanating from near
the event horizon is strongly red shifted before reaching infinity. In the limit
as the source approaches the horizon, the redshift tends to infinity. Let γ̇ a be
a null geodesic. The observed frequency of a plane fronted wave with wave
plane perpendicular to γ̇ a is
ξ a γ̇a
ω=
(ξ a ξ a)
1/2
where ξ a γ̇a is conserved along the null geodesic. In Schwarzschild, ξ a ξa =

f = (1 − 2M/r). If we let ω1 , ω2 be the observed frequency at r1 , r2 , we find
ω2 1 − 2M/r1
= 0 as r1 2M.
ω1 1 − 2M/r2
2.3.1.2. Orbiting Null Geodesics

Consider a null geodesic γ a in the Schwarzschild spacetime. Due to the
spherical symmetry of the Schwarzschild spacetime, we may assume without
loss of generality that θ̇ = 0 and set θ = π/2, so that γ a moves in the
equatorial plane. Hence the geodesic energy and azimuthal angular momentum
e = −ξ a γ̇a and z = ηa γ̇a are conserved. We have
z = ηa γ̇ b gab = r2 φ̇
In fact the same is true for the momenta corresponding to each of the three
rotational Killing fields. Thus, we may consider the total squared angular
momentum L2 given by
L2 = 2r2 m(a m̄b) γ̇ a γ̇ b = r4 (gS2 )ab γ̇ a γ̇ b . (2.26)
8
There are subtleties concerning the regularity of the conformal boundary of Schwarzschild; and
the naive choice of conformal factor mentioned above does not lead to an analytic
compactification. See [60] for recent developments.
For geodesics moving in the equatorial plane, we have L2 = z 2 . Rewriting

gab γ̇ a γ̇ b = 0 using (2.4) and these definitions gives
ṙ2 + V = e2 (2.27)
where
f 2
L . V=
r2
Equation 2.27 can be viewed as the equation for a particle moving in a
potential V.
r
r+ 3M
Figure 2.4. The potential for radial motion of null geodesics in Schwarzschild.
An analysis shows that V has a unique critical point at r = 3M, and hence
a null geodesic with ṙ = 0 in the Schwarzschild spacetime must orbit at r =
3M. We call such null geodesics trapped. The critical point r = 3M is a local
maximum for V and hence the orbiting null geodesics are unstable. The sphere
r = 3M is called the photon sphere. A similar analysis can be performed for
massive particles orbiting the Schwarzschild black hole; see [109, Chapter 6]
for further details.
The geometric optics correspondence between waves packets and null
geodesics indicates that the phenomenon of trapped null geodesics is an
obstacle to dispersion, i.e. the tendency for waves to leave every stationary
region. For waves of finite energy, the fact that the trapped orbits are unstable
can be used to show that such waves in fact disperse. This is a manifestation of
the uncertainty principle.
The close relation between the equation for radial motion of null geodesics
and the wave equation ∇ a ∇a ψ = 0 can be seen as follows. Equation (2.27)
can be written in the form
r4 ṙ2 + R(r, e, L) = 0, (2.28)
where
R = −r4 e2 + r2 fL. (2.29)
On the other hand, the wave equation in the Schwarzschild exterior spacetime
takes the form
R
r2 ∇ a ∇a ψ = ∂r (r2 f )∂r + 2 .
r f
Here R = R(r, ∂t , ) where

is the spherical Laplacian. This is the same
expression as in the equation for the radial motion of null geodesics, but with
, using the correspondence e ↔ i∂t ,
e, L2 replaced by symmetry operators ∂t ,
L2 ↔ − . If we perform separation of variables, the angular Laplacian
is replaced by its eigenvalues −( + 1). This relation between the potential
for radial motion of null geodesics and the term R in the d’Alembertian is a
curious and interesting fact, and importantly this relation holds also in Kerr
(see Figure 2.4).
2.3.2. Raychaudhouri Equation and Comparison Theory

Assume that ka is a null vector field which generates affinely parametrized
geodesics, kb ∇b ka = 0. Let
= 12 ∇a ka (2.30)
be the divergence, or null expansion,9 of the null congruence generated by ka .
For any ka as above, we have
ka ∇a + 2 + σ σ̄ + 12 Rab ka kb = 0 (2.31)
where σ σ̄ = 12 (∇(a kb) ∇ (a kb) − 12 (∇a ka )2 is the squared shear. Equation (2.31)
describes the evolution of the null expansion along null geodesics γ a (λ)
generated by ka . Assuming the null energy condition (2.9), we have
ka ∇a + 2 ≤ 0. (2.32)

Hence, if S < c0 < 0, we find that −∞ along γ a at some finite affine
time λ0 .
Recall that a geodesic in a Riemannian manifold ceases to be minimizing
at its first conjugate point. This can be shown by “rounding off the corner,”
which decreases length. In the Lorentzian case, “rounding off the corner,” see
Figure 2.5, increases Lorentzian length, and one finds that points along null
geodesic γ a past γ a (λ0 ) are timelike related to γ a (0). This means that the
Figure 2.5. Rounding off the corner
9
The definition of in (2.30) agrees with (2.25); we have dropped the subindex on to avoid
clutter. The null expansion is often defined as ∇a ka , however we shall here use the
normalization as in (2.30).
Ta
ηa
la
Figure 2.6. The normals to S
geodesic in particular leaves the boundary of the causal future of S. It is known

that any p ∈ ∂J + (S) is connected to S by a null geodesic without conjugate
points. Combining this argument with the inequality (2.32) shows that if θka <
c0 for some c0 < 0, we find that the boundary of the causal future of S can
extend only for a finite affine parameter range.
Now let be a spacelike Cauchy surface with future timelike normal T a .
For a two-sided surface S ⊂ , we say that a null normal ka to S is outward
pointing if the projection of ka to points into the exterior of , i.e. the
component of \ connected to E. Let ηa be the outward pointing normal
to S in . Then ka = T a + ηa is future directed and outward pointing. Let
H = ∇a ηa be the mean curvature of S in . Then ka = 12 (trS k + H), where
trS k = hij kij − kij ηi ηj is the trace of kij restricted to S. See Figure 2.6. If
the outgoing null expansion ka satisfies ka = 0 (< 0, > 0), we call S a
marginally outer trapped (trapped, untrapped) surface.
Consider the Schwarzschild spacetime; see 2.19. If we designate the null
vector (∂V )a as outgoing, then the coordinate spheres St,r are outer untrapped
in regions I, IV, outer trapped in regions II, III, and marginally trapped on H.
Due to their importance, we introduce the acronym MOTS for “marginally
outer trapped surface.” These are analogs of minimal surfaces in Riemannian
geometry. In particular, a MOTS is critical with respect to variation of area
along the outgoing null directions. For a stationary black hole spacetime, the
event horizon is foliated by MOTS.
As an application of the above remarks, we have the following incomplete-
ness result.
Theorem 2.1 ([15, §7]) Let (M, gab ) be a globally hyperbolic spacetime
satisfying the null energy conditon, and let ( , hij , kij ) be a Cauchy surface
in (M, gab ) with a non-compact exterior. Assume that S is outer trapped in the
sense that the outgoing null expansion θ of S satisfies θ < c0 < 0 for some
c0 < 0. Then (M, gab ) is causally geodesically incomplete.
Remark 2.1 Results similar to theorem 2.1 are usually referred to as “sin-
gularity theorems,” but they actually demonstrate that the spacetime M has
a non-trivial Cauchy horizon ∂M, without giving any information about its
properties. Versions of such results were originally proved by Hawking and
Penrose, see [62]. Motivated by the strong cosmic censorship conjecture, one
expects that for a generic spacetime, the spacetime metric becomes irregular
as one approaches ∂M, and hence that a regular extension beyond ∂M is
impossible. For example, in the Schwarzschild spacetime, curvature diverges
as 1/r3 as one approaches the Cauchy horizon at r = 0. This can be seen by
looking at the invariantly defined Kretschmann scalar Rabcd Rabcd = 48M/r6 .
The detailed behavior of the geometry at the Cauchy horizon in generic
situations is subtle and far from understood; see however [78] and references
therein for recent developments. For cosmological singularities, strong cosmic
censorship including curvature blowup for generic data has been established in
some symmetric situations, see [99, §5.2] and references therein.
By the weak cosmic censorship conjecture, one expects that in a generic
asymptotically flat spacetime, ∂M is hidden from observers at infinity, and
hence that the domain of outer communication has a non-trivial boundary,
the event horizon. This motivates the idea that MOTS may be viewed as
representing the apparent horizon of a black hole; see Section 2.3.3. Due to
the fact that the MOTS can be understood in terms of Cauchy data, this point
of view is important in considering dynamical black holes.
2.3.3. The Apparent Horizon

Consider the Vaidya line element, cf. [96, §5.1.8]
ds2 = fdv2 − 2dvdr − r2 d
sS2 (2.33)
with f = 1 − 2M(v)/r, where the mass aspect function M(v) is an increasing
function of the retarded time coordinate v. The matter in the Vaidya spacetime
is infalling null dust. Those regions where dM/dv = 0 are empty. We see that
there is no dr2 term in (2.33), so r is a null coordinate. Setting M(v) ≡ M, gives
the Schwarzschild line element in ingoing Eddington–Finkelstein coordinates.
A calculation shows that there are MOTS located at r = 2M(v). Hence, if
M(v) varies from M1 to M2 in an interval (v1 , v2 ) and is constant elsewhere
(see Figure 2.7) we find that the MOTS move outwards, to the event horizon,
located at r = 2M2 .
In general, the spacetime tube swept out by the MOTS might, provided it
exists, be termed a marginally outer trapped tube (MOTT). By known stability
results for MOTS, this exists locally in generic situations, see [16], see also
Section 2.3.4. Thus, heuristically the MOTS and MOTT represent the apparent
horizon, and the fact that the apparent horizon moves outward corresponds to
M2 M(v)
M1
Figure 2.7. Mass aspect function M(v) varying from M1 to M2
apparent horizon
event horizon null dust
Figure 2.8. Event and apparent horizons in the Vaidya spacetime
the growth of mass of the black hole due to the stress-energy or gravitational
energy crossing the horizon; see Figure 2.8.
Remark 2.2 1. The event horizon is teleological, in the sense that determin-
ing its location requires complete knowledge of spacetime. In particular, it is
not possible to compute its location from Cauchy data without constructing
the complete spacetime evolution. On the other hand, the notion of MOTS
and apparent horizon are quasilocal notions, which can be determined
directly from Cauchy data.
2. The location of MOTSs is not a spacetime concept but depends on the
choice of Cauchy slicing. See [26] for results on the region of spacetime
containing trapped surfaces. It was shown by Wald and Iyer [111] that
there are Cauchy surfaces in the extended Schwarzschild spacetime which
approach the singularity arbitrarily closely and such that the past of these
Cauchy surfaces does not contain any outer trapped surfaces.
3. The interior of the outermost MOTS is called the trapped region (a
notion which depends on the Cauchy slicing). Based on the weak cosmic
censorship conjecture, and the above remarks, one expects this to be in the
black hole region, which is bounded by the event horizon. See [39, Theorem
6.1] for a result in this direction.
2.3.4. Results on MOTS and the Trapped Region

Several theorems about MOTS have been proved in the last decade. In
particular, if a Cauchy surface contains a MOTS, then there is an outermost
MOTS.
If we conside a Cauchy slicing ( t ), then if t0 contains a MOTS, then
for t > t0 , t contains a MOTS. However, the location of the MOTS may
jump, e.g. due to the formation of a MOTS surrounding the previous one; see
Figure 2.9. This phenomenon is seen in numerical simulations of colliding
black holes, cf. [86]. There, examples with two merging black holes are consid-
ered. When the apparent horizons of the two black holes are sufficiently close
together, a new apparent horizon surrounding both is formed, in accordance
with the results in [17, 15].
If the null energy condition (NEC) holds, then in a generic situation the
MOTT is spacelike [15], and hence from the point of view of the exterior part
of M it represents an outflow boundary. This means that it is not necessary
to impose any boundary condition on the MOTT in order to get a well-posed
Cauchy problem. This leads to the exterior Cauchy problem. As mentioned
above, cf. Figure 2.10, in strong field situations, it can happen that the MOTS
jumps out. In this case, one must then restart the solving of the exterior Cauchy
problem at the jump time. This corresponds closely to what one sees in a
0 0
Figure 2.9. The location of the MOTS jumps outward when a new one forms
outside the previous one
Figure 2.10. The exterior Cauchy problem
numerical evolution of strong field situations, e.g. of merging black holes,

when using horizon trackers to determine the location of MOTS.
2.3.5. Formation of Black Holes

The first example of a dynamically forming black hole through the collapse
of a cloud of dust was constructed by Oppenheimer and Snyder [91] in 1939.
Examples of the formation of a black hole by concentration of gravitational
radiation were constructed by Christodoulou [36]. There has been much recent
work refining and extending these results, see [70] and references therein.
In order to understand the formation of black holes, it is important to have
good conditions for the existence of marginally outer trapped surfaces in a
given Cauchy surface. Such results have been proved by Schoen and Yau [102],
see also [41]. The result in [102] makes use of Jang’s equation to show that
MOTS form if a sufficiently dense concentration of matter is present. A related
result for the vacuum case is given in [47]; see also [115].
2.3.6. Black Hole Stability

Taking the trapped region as representing a dynamical black hole, the above
discussion leads to a picture of the evolution of dynamical black holes, as
well as their formation. Based on these general considerations, we can now
give a heuristic formulation of the black hole stability problem and related
conjectures. Recall that the Kerr black hole spacetime, which we shall study
in detail below, is conjectured to be the unique rotating vacuum black hole
spacetime, and further to be dynamically stable.
The black hole stability conjecture is that Cauchy data sufficiently close, in
a suitable sense, to Kerr Cauchy data10 have a maximal development which
is future asymptotic to a Kerr spacetime; see Figure 2.11. In approaching
this problem, one may use the results on the evolution of MOTS mentioned
above, cf. Section 2.3.4 to consider only the exterior Cauchy problem. It is
important to note that the parameters of the “limiting” Kerr spacetime cannot
be determined in any effective manner from the initial data.
As discussed above, cf. Section 2.2.6.6, if we are restricted to axial
symmetry, then angular momentum is quasilocally conserved. This means that
if we further restrict to zero angular momentum, the end state of the evolution
must be a Schwarzschild black hole.
10
See [10], see also e.g. [21, 81] for discussions of the problem of characterizing Cauchy data as
Kerr data.
i+
H
I+
i0
trapping I−
Figure 2.11. Expected causal diagram for the exterior of a solution arising from a
perturbation of the Kerr solution
Thus, the black hole stability conjecture for the axially symmetric case is that
the maximal development of sufficiently small (in a suitable sense), axially
symmetric, deformations of Schwarzschild Cauchy data with zero angular
momentum is asymptotic in the future to a Schwarzschild spacetime.11 In
this case, due to the loss of energy through I + , the mass of the “limiting”
Schwarzschild black hole cannot be determined directly from the Cauchy data.
A conjecture related to the black hole stability conjecture, but which is even
more far reaching, may be termed the end state conjecture. Here the idea is
that the maximal evolution of generic asymptotically flat vacuum initial data is
asymptotic in a suitable sense, to a collection of black holes moving apart, with
the near region of each black hole approaching a Kerr geometry. No smallness
condition is implied.
The heuristic ideas relating to weak cosmic censorship and Kerr as the final
state of the evolution of an isolated system, together with Hawking’s area
theorem, was used by Penrose to motivate the Penrose inequality,

Amin
≤ MADM
16π
where Amin is the minimal area of any surface surrounding all past and future
trapped regions in a given Cauchy surface, and MADM is the ADM mass at
infinity. The Riemannian version of the Penrose inequality has been proved
by Bray [31], and Huisken and Ilmanen [65]. The spacetime version of the
11
The zero angular momentum, axially symmetric case of the black hole stability problem could
also be called the polarized case.
Penrose inequality remains open. It should be stressed that the formulation

of the inequality given above may have to be adjusted. Interesting possible
approaches to the problem have been developed by Bray and Khuri, see [61]
and references therein.
2.3.7. The Kerr Metric

In this section we shall discuss the Kerr metric, which is the main object
of our considerations. Although many features of the geometry and analysis
on black hole spacetimes are seen in the Schwarzschild case, there are many
new and fundamental phenomena present in the Kerr case. Among these are
complicated trapping, i.e. the fact that trapped null geodesics fill up an open
spacetime region, the fact that the Kerr metric admits only two Killing fields,
but a hidden symmetry manifested in the Carter constant, and the fact that the
stationary Killing vector field ξ a fails to be timelike in the whole domain of
outer communications, which leads to a lack of a positive conserved energy
for waves in the Kerr spacetimes. This fact is the origin of superradiance and
the Penrose process. See [108] for a recent survey.
The Kerr metric describes a family of stationary, axisymmetric, asymptot-
ically flat vacuum spacetimes, parametrized by ADM mass M and angular
momentum per unit mass a. The expressions for mass and angular momentum
introduced in Section 2.2.6 when applied in Kerr geometry yield M and J =
aM. In Boyer–Lindquist coordinates (t, r, θ, φ), the Kerr metric takes the form
( − a2 sin2 θ )dta dtb 2a sin2 θ (a2 + r2 − )dt(a dφb)
gab = + (2.34)

dra drb sin2 θ (a2 + r2 )2 − a2 sin2 θ dφa dφb
− − dθa dθb − ,

where = a2 − 2Mr + r2 and = a2 cos2 θ + r2 . The volume element is

| det gab | = sin θ (2.35)
There is a ring-shaped singularity at r = 0, θ = π/2. For |a| ≤ M, the
Kerr √ spacetime contains a black hole, with event horizon at r = r+ ≡
M + M 2 − a2 , while for |a| > M, the singularity is naked in the sense that
it is causally connected to observers at infinity. The area of the horizon is
AHor = 4π(r+ 2 + a2 ). This achieves its maximum of 16π M 2 when a = 0,
which provides one of the ingredients in the heuristic argument for the Penrose
inequality; see Section 2.3.6. The case |a| = M is called extreme. We shall
here be interested only in the subextreme case, |a| < M, as this is the only case
where we expect black hole stability to hold.
Figure 2.12. The ergoregion
The Boyer–Lindquist coordinates are analogous to the Schwarzschild coor-

dinates in Section 2.3.1 and, upon setting a = 0, (2.34) reduces to (2.19).
The line element takes a simple form in Boyer-Lindquist coordinates, but,
similarly to the Schwarzschild coordinates, they have the drawback that they
are not regular at the horizon.
The Kerr metric admits two Killing vector fields ξ a = (∂t )a (stationary)
and ηa = (∂φ )a (axial). Although the stationary Killing field ξ a is timelike
near infinity, since gab ξ a ξ b → 1 as r → ∞, ξ a becomes spacelike for r
sufficiently small, when 1 − 2M/ < 0. In the Schwarzschild case a = 0,
this occurs at the event horizon r = 2M. However, for a rotating Kerr black
hole with 0 < |a| ≤ M, there is a region, called the ergoregion, outside the
event √horizon, where ξ a is spacelike. The ergoregion is bounded by the surface
M + M 2 − a2 cos2 θ which touches the horizon at the poles θ = 0, π ; see
Figure 2.12. In the ergoregion, null and timelike geodesics can have negative
energy with respect to ξ a . The fact that there is no globally timelike vector
field in the Kerr exterior is the origin of superradiance, i.e. the fact that waves
which scatter off the black hole can leave the ergoregion with more energy (as
measured by a stationary observer at infinity) than was sent in. This effect was
originally found by an analysis based on separation of variables, but it can be
demonstrated rigorously, see [52]. However, it is a subtle effect and not easy to
demonstrate numerically, see [75].
If we consider a dynamical spacetime containing a rotating black hole, then
the presence of the ergoregion allows for the Penrose process, which extracts
rotational energy from the black hole, see [74], see also [48] for a numerical
study of superradiance of gravitational waves in a dynamical spacetime.
Let ωH = a/(r+ 2 + a2 ) be the rotation speed of the black hole. The Killing
field χ = ξ + ωH ηa is null on the event horizon in Kerr, which is therefore

a a
a Killing horizon. For |a| < M, there is a neighborhood of the horizon in the
black hole exterior where χ a is timelike. The surface gravity κ, defined by
κ 2 = − 12 (∇ a χ b )(∇a χb ) takes the value κ = (r+ − M)/(r+ 2 + a2 ), and is posi-
tive in the subextreme case |a| < M. By general results, a Killing horizon with
non-vanishing surface gravity is bifurcate, i.e. there is a cross-section where
the null generator vanishes. In the Schwarzschild case, this is the 2-sphere
U = V = 0. See [90, 96] for background on the geometry of the Kerr
spacetime; see also [89].
2.4. Spin Geometry

The 2-spinor formalism, and the closely related GHP formalism, are important
tools in Lorentzian geometry and the analysis of black hole spacetimes; we
will introduce them here. A detailed account of this material is given by
Penrose and Rindler [94]. Following the conventions there, we use the abstract
index notation with lower case Latin letters a, b, c, . . . for tensor indices, and
unprimed and primed upper case Latin letters A, B, C, . . . , A , B , C , . . . for
spinor indices. Tetrad and dyad indices are boldface Latin letters following the
same scheme, a, b, c, . . . , A, B, C, . . . , A , B , C , . . . . For coordinate indices
we use Greek letters α, β, γ , . . . .
2.4.1. Spinors on Minkowski Space

Consider Minkowski space M, i.e. R4 with coordinates (xα ) = (t, x, y, z) and
metric
gαβ dxα dxβ = dt2 − dx2 − dy2 − dz2 .
Define a complex null tetrad (i.e. frame) (ga a )a=0,··· ,3 = (la , na , ma , m̄a ), as in
(2.3) above, normalized so that la na = 1, ma m̄a = −1, so that
gab = 2(l(a nb) − m(a m̄b) ). (2.36)
Similarly, let A A be a dyad (i.e. frame) in C2 , with dual frame A A . The

complex conjugates will be denoted ¯A A , ¯A A and again form a basis in
another two-dimensional complex space denoted C̄2 , and its dual. We can
identify the space of complex 2 × 2 matrices with C2 ⊗ C̄2 . By construction,

the tensor products A A ¯A A and A A ¯A A form a basis in C2 ⊗ C̄2 and its dual.
Now, with xa = xa ga a , writing
0 2
x x
xa ga AA ≡ 3 1 (2.37)
x x
defines the soldering forms, also known as Infeld–van der Waerden symbols

ga AA , (and analogously gAA a ). By a slight abuse of notation we may write

xAA = xa instead of xAA = xa ga AA or, dropping reference to the tetrad,12

xAA = xa ga AA . In particular, xa ∈ M corresponds to a 2×2 complex Hermitian
12
The Infeld–van der Waerden symbols are sometimes denoted σa AA

matrix xAA ∈ C2 ⊗ C̄2 . Taking the complex conjugate of both sides of (2.37)
gives

x̄a = x̄A A = (xAA )∗ ,
where ∗ denotes Hermitian conjugation. This extends to a correspondence
C4 ↔ C2 ⊗ C̄2 with complex conjugation corresponding to Hermitian con-
jugation.
Note that

det(xAA ) = x0 x1 − x2 x3 = xa xa /2. (2.38)
We see from the above that the group

a b
SL(2, C) = A = , a, b, c, d ∈ C, ad − bc = 1
c d
acts on X ∈ C2 ⊗ C̄2 by
X → AXA∗ .
In view of (2.38) this exhibits SL(2, C) as a double cover of the identity
component of the Lorentz group SO0 (1, 3), the group of linear isometries of
M. In particular, SL(2, C) is the spin group of M. The canonical action
(A, v) ∈ SL(2, C) × C2 → Av ∈ C2
of SL(2, C) on C2 is the spinor representation. Elements of C2 are called
(Weyl) spinors. The conjugate representation given by
(A, v) ∈ SL(2, C) × C2 → Āv ∈ C2
is denoted C̄2 .

Spinors13 of the form xAA = α A β A correspond to matrices of rank one, and
hence to complex null vectors. Denoting oA = 0 A , ιA = 1 A , we have from
the above that

la = oA oA , na = ιA ιA , ma = oA ιA , m̄a = ιA oA . (2.39)
This gives a correspondence between a null frame in M and a dyad in C2 .
The action of SL(2, C) on C2 leaves invariant a complex area element,
a skew-symmetric bispinor. Such a unique spinor AB is determined by the
normalization
gab = AB ¯A B .
13
It is conventional to refer to spin-tensors e.g. of the form xAA or ψABA simply as spinors.
The inverse AB of AB is defined by AB CB = δA C , AB AC = δC B . As with

gab and its inverse gab , the spin-metric AB and its inverse AB is used to lower
and raise spinor indices,
λB = λA AB , λA = AB λB .
We have
AB = oA ιB − ιA oB .
In particular,
oA ιA = 1. (2.40)
k 2 l 2
An element φA···DA ···D of C C̄ is called a spinor of valence (k, l).
The space of totally symmetric14 spinors φA···DA ···D = φ(A···D)(A ···D ) is
denoted Sk,l . The spaces Sk,l for k, l non-negative integers yield all irreducible
representations of SL(2, C). In fact, one can decompose any spinor into
“irreducible pieces,” i.e. as a linear combination of totally symmetric spinors in
Sk,l with factors of AB . The above mentioned correspondence between vectors
and spinors extends to tensors of any type, and hence the just mentioned
decomposition of spinors into irreducible pieces carries over to tensors as
well. Examples are given by Fab = φAB A B , a complex anti-self-dual 2-
form, and − Cabcd = ABCD A B C D , a complex anti-self-dual tensor with
the symmetries of the Weyl tensor. Here, φAB and ABCD are symmetric.
2.4.2. Spinors on Spacetime

Now let (M, gab ) be a Lorentzian 3+1 dimensional spin manifold with metric
of signature + − −−. The spacetimes we are interested in here are spin ones,
in particular any orientable, globally hyperbolic 3+1 dimensional spacetime
is spin, cf. [58, page 346]. If M is spin, then the orthonormal frame bundle
SO(M) admits a lift to Spin(M), a principal SL(2, C)-bundle. The associated
bundle construction now gives vector bundles over M corresponding to the
representations of SL(2, C), in particular we have bundles of valence (k, l)
spinors with sections φA···DA ···D . The Levi-Civita connection lifts to act on
sections of the spinor bundles,
∇AA : ϕB···DB ···D → ∇AA ϕB···DB ···D (2.41)
where we have used the tensor–spinor correspondence to replace the index a
by AA . We shall denote the totally symmetric spinor bundles by Sk,l and their
spaces of sections by Sk,l .
14
The ordering between primed and unprimed indices is irrelevant.
The above mentioned correspondence between spinors and tensors, and the
decomposition into irreducible pieces, can be applied to the Riemann curvature
tensor. In this case, the irreducible pieces correspond to the scalar curvature,
traceless Ricci tensor, and the Weyl tensor, denoted by R, Sab , and Cabcd ,
respectively. The Riemann tensor then takes the form
Rabcd = − 12 gad gbc R + 12 gac gbd R + 2 gbd Sac

1 1 1
− 12 gbc Sad
− 2 gad Sbc + 2 gac Sbd + Cabcd .
1 1
(2.42)
The spinor equivalents of these tensors are

¯ A B C D AB CD ,
Cabcd = ABCD ¯A B ¯C D + (2.43a)
Sab = − 2ABA B , (2.43b)
R = 24 . (2.43c)
2.4.3. Fundamental Operators

Projecting (2.41) on its irreducible pieces gives the following four fundamental
operators, introduced in [9].
Definition 2.2 The differential operators
Dk,l : Sk,l → Sk−1,l−1 , Ck,l : Sk,l → Sk+1,l−1 ,
†
Ck,l : Sk,l → Sk−1,l+1 , Tk,l : Sk,l → Sk+1,l+1
are defined as

(Dk,l ϕ)A1 ...Ak−1 A1 ...Al−1 ≡ ∇ BB ϕA1 ...Ak−1 B A1 ...Al−1 B , (2.44a)
A1 ...Al−1 B A ...Al−1
(Ck,l ϕ)A1 ...Ak+1 ≡ ∇(A1 ϕA2 ...Ak+1 ) 1 B , (2.44b)
A2 ...Al+1 )
† B(A1
(Ck,l ϕ)A1 ...Ak−1 A1 ...Al+1 ≡∇ ϕA1 ...Ak−1 B , (2.44c)
A2 ...Al+1 )
(A1
(Tk,l ϕ)A1 ...Ak+1 A1 ...Al+1 ≡ ∇(A1 ϕA2 ...Ak+1 ) . (2.44d)
The operators are called respectively the divergence, curl, curl-dagger, and
twistor operators.
With respect to complex conjugation, the operators D, T satisfy Dk,l =

† †
Dl,k , Tk,l = Tl,k , while Ck,l = Cl,k , Ck,l = Cl,k .
Denoting the adjoint of an operator by A with respect to the bilinear pairing

(φA1 ···Ak A ···A , ψA1 ···Ak A ···A ) = φA1 ···Ak A ···A ψ A1 ···Ak A1 ···Al dμ
1 l 1 l 1 l
by A† , and the adjoint with respect to the sesquilinear pairing

φA1 ···Ak A ···A , ψA1 ···Al A ···A = φA1 ···Ak A ···A ψ̄ A1 ···Ak A1 ···Al dμ
1 l 1 k 1 l
by A! , we have
(Dk,l )† = −Tk−1,l−1 , (Tk,l )† = −Dk+1,l+1 ,
† † †
(Ck,l )† = Ck+1,l−1 , (Ck,l ) = Ck−1,l+1 ,
and
(Dk,l )! = −Tl−1,k−1 , (Tk,l )! = −Dl+1,k+1 ,
† ! †
(Ck,l )! = Cl−1,k+1 , (Ck,l ) = Cl+1,k−1 .
†
As we will see in Section 2.4.4, the kernels of C2s,0 and C0,2s are the massless
spin-s fields. The kernels of Tk,l are the valence (k, l) Killing spinors, which
we will discuss further in Section 2.4.5 and Section 2.4.7. A complete set of
commutator properties of these operators can be found in [9].
2.4.4. Massless Spin-s Fields

†
For s ∈ 12 N, ϕA···D ∈ ker C2s,0 is a totally symmetric spinor ϕA···D = ϕ(A···D) of
valence (2s, 0) which solves the massless spin-s equation
†
(C2s,0 ϕ)A···BD = 0.
For s = 1/2, this is the Dirac–Weyl equation ∇A A ϕA = 0, for s = 1, we

have the left and right Maxwell equation ∇A B φAB = 0 and ∇A B ϕA B = 0, i.e.
†
(C2,0 φ)AA = 0, (C0,2 ϕ)AA = 0.
An important example is the Coulomb Maxwell field on Kerr,
2
φAB = − o(A ιB) . (2.45)
(r − ia cos θ )2
This is a non-trivial sourceless solution of the Maxwell equation on the Kerr
background. We note that the scalars components, see Section 2.4.8 below, of
the Coulomb field φ1 = (r − ia cos θ )−2 while φ0 = φ2 = 0.
For s > 1, the existence of a non-trivial solution to the spin-s equation
implies curvature conditions, a fact known as the Buchdahl constraint [32],
0 = (A DEF φB...C)DEF . (2.46)
This is easily obtained by commuting the operators in
†
0 = (D2s−1,1 C2s,0 φ)A...C . (2.47)
For the case s = 2, the equation ∇A D ABCD = 0 is the Bianchi equation,
which holds for the Weyl spinor in any vacuum spacetime. Due to the Buchdahl
constraint, it holds that in any sufficiently general spacetime, a solution of the
spin-2 equation is proportional to the Weyl spinor of the spacetime.
2.4.5. Killing Spinors

A1 ···Al
Spinors "A1 ···Ak ∈ Sk,l satisfying

(Tk,l ")A1 ···Ak+1 A1 ···Al+1 = 0,
are called Killing spinors of valence (k, l). We denote the space of Killing
spinors of valence (k, l) by KS k,l . The Killing spinor equation is an over-
determined system. The space of Killing spinors is a finite dimensional space,
and the existence of Killing spinors imposes strong restrictions on M; see
Section 2.4.7 below. Killing spinors νAA ∈ KS 1,1 are simply conformal
Killing vector fields, satisfying ∇(a νb) − 12 ∇ c νc gab . A Killing spinor κAB ∈
KS 2,0 corresponds to a complex anti-self-dual conformal Killing–Yano 2-form
YABA B = κAB A B satisfying the equation
∇(a Yb)c − 2ζc gab + ζ(a gb)c = 0, (2.48)
where in the four-dimensional case, ζa = 13 ∇b Y b a .

In the mathematics literature, Killing spinors of valence (1, 0) are known
as twistor spinors. The terms “conformal Killing–Yano form” or “twistor
form” is used also for the real 2-forms corresponding to Killing spinors of
valence (2, 0), as well as for forms of higher degree and on higher dimensional
manifolds, appearing as kernel elements of an analogous Stein–Weiss operator.
Further, we may mention that Killing spinors LABA B ∈ KS 2,2 are traceless
symmetric conformal Killing tensors Lab , satisfying the equation
∇(a Lbc) − 13 g(ab ∇ d Lc)d = 0. (2.49)
In particular, any tensor of the form ζ gab for some scalar field ζ is a conformal
Killing tensor. If γ a is a null geodesic and Lab is a conformal Killing tensor,
then Lab γ̇ a γ̇ b is conserved along γ a . For any κAB ∈ KS 2,0 we have LABA B =
κAB κ̄A B ∈ KS 2,2 . See Section 2.4.7 below for further details.
2.4.6. Algebraically Special Spacetimes

Let ϕA···D ∈ Sk,0 . A spinor αA is a principal spinor of ϕA···D if
ϕA···D α A · · · α D = 0.
Table 2.1. The Petrov classification
I {1, 1, 1, 1} ABCD = α(A βB γC δD) I

II {2, 1, 1} ABCD = α(A αB γC δD)
II
D {2, 2} ABCD = α(A αB βC βD)
D III
III {3, 1} ABCD = α(A αB αC βD)
N {4} ABCD = αA αB αC αD N
O {−} ABCD = 0 O
An application of the fundamental theorem of algebra shows that any ϕA···D ∈

Sk,0 has exactly k principal spinors αA , . . . , δA , and hence is of the form
ϕA···D = α(A · · · δD) .
(i)
If ϕA···D ∈ Sk,0 has n distinct principal spinors αA , repeated mi times, then
ϕA···D is said to have algebraic type {m1 , . . . , mn }. Applying this to the Weyl
tensor leads to the Petrov classification; see Table 2.1.15
A principal spinor oA determines a principal null direction la = oA ōA . The
Goldberg–Sachs theorem states that, in a vacuum spacetime, the congruence
generated by a null field la is geodetic and shear free16 if and only if la is a
repeated principal null direction of the Weyl tensor Cabcd (or equivalently oA
is a repeated principal spinor of the Weyl spinor ABCD ).
2.4.6.1. Petrov Type D

The Kerr metric is of Petrov type D, and many of its important properties
follow from this fact. The vacuum type D spacetimes have been classified
by Kinnersley [69], see also Edgar et al. [49]. The family of Petrov type D
spacetimes includes the Kerr–NUT family and the boost-rotation symmetric
C-metrics. The only Petrov type D vacuum spacetime which is asymptotically
flat and has positive mass is the Kerr metric; see theorem 2.3.
A Petrov type D spacetime has two repeated principal spinors oA , ιA , and
correspondingly there are two repeated principal null directions la , na , for the
Weyl tensor. We can without loss of generality assume that la na = 1 and
define a null tetrad by adding complex null vectors ma , m̄a normalized such
that ma m̄a = −1. By the Goldberg–Sachs theorem both la , na are geodetic
and shear free, and only one of the five independent complex Weyl scalars is
non-zero, namely
15
The Petrov classification is exclusive, so a spacetime belongs at each point to exactly one
16
Petrov class.
If la is geodetic and shear then the spin coefficients σ , κ, cf. (2.61) below, satisfy σ = κ = 0.
2 = − la mb m̄d nc Cabcd . (2.50)
In this case, the Weyl spinor takes the form

1
ABCD = 2 o(A oB ιC ιD) .
6
See (2.64) below for the explicit form of 2 in the Kerr spacetime.
The following result is a consequence of the Bianchi identity.
Theorem 2.2 ([112]) Assume (M, gab ) is a vacuum spacetime of Petrov type
D. Then (M, gab ) admits a one-dimensional space of Killing spinors κAB of
the form
κAB = −2κ1 o(A ιB) (2.51)
−1/3
where oA , ιA are the principal spinors of ABCD and κ1 ∝ 2 .
Remark 2.3 Since the Petrov classes are exclusive, we have 2 = 0 for a
Petrov type D space.
2.4.7. Spacetimes Admitting a Killing Spinor

Differentiating the Killing spinor equation (Tk,l φ)A···DA ···D = 0, and com-
muting derivatives, yields an algebraic relation between the curvature, Killing
spinor, and their covariant derivatives which restrict the curvature spinor, see
[9, §2.3], see also [10, §3.2]. In particular, for a Killing spinor κA···D of valence
(k, 0), k ≥ 1, the condition
(ABC F κD···E)F = 0 (2.52)
must hold, which restricts the algebraic type of the Weyl spinor. For a valence
(2, 0) Killing spinor κAB , the condition takes the form
(ABC E κD)E = 0. (2.53)
It follows from (2.53) that a spacetime admitting a valence (2, 0) Killing

spinor is of type D, N, or O. The space of Killing spinors of valence (2, 0)
on Minkowski space (or any space of Petrov type O) has complex dimension

10. The explicit form in Cartesian coordinates xAA is

κ AB = U AB + 2xA (A V B) A + xAA xBB WA B ,

where U AB , V B A , W A B are arbitrary constant symmetric spinors, see[2,
Eq. (4.5)]. One of these corresponds to the spinor in (2.51), in spheroidal
coordinates it takes the form given in (2.65) below.
A further application of the commutation properties of the fundamental

operators yields the fact that the 1-form
†
ξAA = (C2,0 κ)AA (2.54)
is a Killing field, ∇(a ξb) = 0, provided M is a vacuum. Clearly the real and
imaginary parts of ξa are also Killing fields. If ξa is proportional to a real
Killing field,17 we can without loss of generality assume that ξa is real. In
this case, the 2-form
Yab = 32 i(κAB ¯A B − κ̄A B AB ) (2.55)
is a Killing–Yano tensor, ∇(a Yb)c = 0, and the symmetric 2-tensor
Kab = Ya c Ycb (2.56)
is a Killing tensor,
∇(a Kbc) = 0. (2.57)
Further, in this case,
ζa = ξ b Kab (2.58)
is a Killing field, see [64, 43]. Recall that the quantity Lab γ̇ a γ̇ b is conserved
along null geodesics if Lab is a conformal Killing tensor. For Killing tensors,
this fact extends to all geodesics, so that if Kab is a Killing tensor, then Kab γ̇ a γ̇ b
is conserved along a geodesic γ a . See [10] for further details and references.
2.4.8. GHP Formalism

Taking the point of view that the null tetrad components of tensors are sections
of complex line bundles with the action of the non-vanishing complex scalars
corresponding to the rescalings of the tetrad, respecting the normalization,
leads to the GHP formalism [59].
Given a null tetrad la , na , ma , m̄a we have a spin dyad oA , ιA as discussed
above. For a spinor ϕA···D ∈ Sk,0 , it is convenient to introduce the Newman–
Penrose scalars
ϕi = ϕA1 ···Ai Ai+1 ···Ak ιA1 · · · ιAi oAi+1 · · · oAk . (2.59)
In particular, ABCD corresponds to the five complex Weyl scalars i , i =
0, . . . 4. The definition ϕi extends in a natural way to the scalar components
of spinors of valence (k, l).
17
We say that such spacetimes are of the generalized Kerr–NUT class, see [19] and references
therein.
The normalization (2.40) is left invariant under rescalings oA → λoA , ιA →

λ−1 ιA where λ is a non-vanishing complex scalar field on M. Under such
rescalings, the scalars defined by projecting on the dyad, such as ϕi given by
(2.59), transform as sections of complex line bundles. A scalar ϕ is said to have
type {p, q} if ϕ → λp λ̄q ϕ under such a rescaling. Such fields are called properly
weighted. The lift of the Levi-Civita connection ∇AA to these bundles gives
a covariant derivative denoted a . Projecting on the null tetrad la , na , ma , m̄a
gives the GHP operators
þ = l a a , þ = na a , ð = ma a , ð = m̄a a .
The GHP operators are properly weighted, in the sense that they take properly
weighted fields to properly weighted fields, for example if ϕ has type {p, q},
then þ ϕ has type {p + 1, q + 1}. This can be seen from the fact that la =

oA ōA has type {1, 1}. There are 12 connection coefficients in a null frame, up
to complex conjugation. Of these, eight are properly weighted, the GHP spin
coefficients. The other connection coefficients enter in the connection 1-form
for the connection a .
The following formal operations take weighted quantities to weighted
quantities,
−
(bar) : la → la , na → na , ma → m̄a , m̄a → ma , {p, q} → {q, p},

(prime) : l → n , n → l , m → m̄ , m̄ → m ,
a a a a a a a a
{p, q} → {−p, −q},
∗
(star) : l → m , n → −m̄ , m → −l , m̄ → n ,
a a a a a a
{p, q} → {p, −q}.
a a
(2.60)
The properly weighted spin coefficients can be represented as
κ = m b l a ∇a l b , σ = mb ma ∇a lb , ρ = mb m̄a ∇a lb , τ = mb na ∇a lb , (2.61)
together with their primes κ , σ , ρ , τ .
A systematic application of the above formalism allows one to write the
tetrad projection of the geometric field equations in a compact form. For
example, the Maxwell equation corresponds to the four scalar equations
given by
(þ −2ρ)φ1 − (ð − τ )φ0 = −κφ2 , (2.62)
with its primed and starred versions.
Working in a spacetime of Petrov type D gives drastic simplifications, in
view of the fact that choosing the null tedrad so that la , na are aligned with
principal null directions of the Weyl tensor (or equivalently choosing the spin
dyad so that oA , ιA are principal spinors of the Weyl spinor); as has already
been mentioned, the Weyl scalars are zero with the exception of 2 , and the
only non-zero spin coefficients are ρ, τ and their primed versions.
2.5. The Kerr Spacetime

Taking into account the background material given in Section 2.4, we can now
state some further properties of the Kerr spacetime. As mentioned above, the
Kerr metric is algebraically special, of Petrov type D. An explicit principal null
tetrad (la , na , ma , m̄a ) is given by the Carter tetrad [116]
a(∂φ )a (a2 + r2 )(∂t )a 1/2 (∂r )a
la = √ + √ + √ , (2.63a)
21/2 1/2 21/2 1/2 2 1/2
a(∂φ )a (a2 + r2 )(∂t )a 1/2 (∂r )a
na = √ + √ − √ , (2.63b)
21/2 1/2 21/2 1/2 2 1/2
(∂θ )a i csc θ (∂φ )a ia sin θ (∂t )a
ma = √ + √ + √ . (2.63c)
2 1/2 2 1/2 2 1/2
In view of the normalization of the tetrad, the metric takes the form gab =
2(l(a nb) − m(a m̄b) ). We may remark that the choice of la , na to be aligned with
the principal null directions of the Weyl tensor, together with the normalization
of the tetrad, fixes the tetrad up to rescalings.
We have
M
2 = − . (2.64)
(r − ia cos θ )3
κAB = 23 (r − ia cos θ )o(A ιB) . (2.65)
With κAB as in (2.65), equation (2.54) yields
ξ a = (∂t )a , (2.66)
and from (2.55) we get
Yab = a cos θl[a nb] − irm[a m̄b] . (2.67)
With the normalizations above, the Killing tensor (2.56) takes the form
Kab = 14 (2 l(a nb) − r2 gab ) (2.68)
and (2.58) gives
ζ a = a2 (∂t )a + a(∂φ )a . (2.69)
Recall that for a geodesic γ , the quantity k = 4Kab γ̇ a γ̇ b , known as Carter’s
constant, is conserved. Explicitly,
k = γ̇θ2 + a2 sin2 θ e2 + 2aez + a2 cos2 θμ2 (2.70)
where γ̇θ = γ̇ a (∂θ )a . For a = 0, the tensor Kab cannot be expressed as a
tensor product of Killing fields [112], and similarly Carter’s constant k cannot
be expressed in terms of the constants of motion associated with Killing fields.
In this sense Kab and k manifest a hidden symmetry of the Kerr spacetime.
As we shall see in Section 2.7, these structures are also related to symmetry
operators and separability properties, as well as conservation laws, for field
equations on Kerr, and more generally in spacetimes admitting Killing spinors
satisfying certain auxiliary conditions.
2.5.1. Characterizations of Kerr

Consider a vacuum Cauchy data set ( , hij , kij ). We say that ( , hij , kij ) is
asymptotically flat if has an end R3 \ B(0, R) with a coordinate system (xi )
such that
hij = δij + O∞ (rα ), kij = O∞ (rα−1 ) (2.71)
for some α < −1/2. The Cauchy data set ( , hij , kij ) is asymptotically
Schwarzschildean if

2A α 2xi xj
hij = − 1 + δij − − δij + o∞ (r−3/2 ), (2.72a)
r r r2

β 2xi xj
kij = 2 − δij + o∞ (r−5/2 ), (2.72b)
r r2
where A is a constant, and α, β are functions on S2 , see [20, §6.5] for details.
Here, the symbols o∞ (rα ) are defined in terms of weighted Sobolev spaces,
see [20, §6.2] for details.
If (M, gab ) is a vacuum and contains a Cauchy surface ( , hij , kij ) sat-
isfying (2.71) or (2.72), then (M, gab ) is asymptotically flat, respectively
asymptotically Schwarzschildean, at spatial infinity. In this case there is a
spacetime coordinate system (xα ) such that gαβ is asymptotic to the Minkowski
line element with asymptotic conditions compatible with (2.72). For such
spacetimes, the ADM 4-momentum Pμ is well defined. The positive mass
theorem states that Pμ is future directed causal, Pμ Pμ ≥ 0 (where the
contraction is in the asymptotic Minkowski line element), P0 ≥ 0, and gives
conditions under which Pμ is strictly timelike. This holds in particular if
contains an apparent horizon.
Mars [80] has given a characterization of the Kerr spacetime as an asymp-
totically flat vacuum spacetime with a Killing field ξ a asymptotic to a time
translation, positive mass, and an additional condition on the Killing form
FAB = (C1,1 ξ )AB ,
ABCD F CD ∝ FAB .
A characterization in terms of algebraic invariants of the Weyl tensor has been
given by Ferrando and Saez [51]. The just mentioned characterizations are in
terms of spacetime quantities. As mentioned in Section 2.2.6.5 Killing spinor

initial data propagates, which can be used to formulate a characterization of
Kerr in terms of Cauchy data, see [19, 20, 21, 22].
We here give a characterization in terms of spacetimes admitting a Killing
spinor of valence (2, 0).
Theorem 2.3 Assume that (M, gab ) is vacuum, asymptotically Schwarzs-
childean at spacelike infinity and contains a Cauchy slice bounded by an
apparent horizon. Assume further that (M, gab ) admits a non-vanishing
Killing spinor κAB of valence (2, 0). Then (M, gab ) is locally isometric to the
Kerr spacetime.
Proof Let Pμ be the ADM 4-momentum vector for M. By the positive mass
theorem, Pμ Pμ ≥ 0. In the case where M contains a Cauchy surface bounded
by an apparent horizon, then Pμ Pμ > 0 by [23, Remark 11.5].18
Recall that a spacetime with a Killing spinor of valence (2, 0) is of Petrov
type D, N, or O. From asymptotic flatness and the positive mass theorem,
we have Cabcd Cabcd = O(1/r6 ), and hence there is a neighborhood of
spatial infinity where M is Petrov type D. It follows that near spatial infinity,
−1/3
κAB = −2κ1 o(A ιB) , with κ1 ∝ 2 = O(r). It follows from our asymptotic
†
conditions that the Killing field ξAA = (C2,0 κ)AB is O(1) and hence asymptotic
to a translation, ξ → A as r → ∞, for some constant vector Aμ . It follows
μ μ
from the discussion in [4, §4] that Aμ is non-vanishing. Now, by [24, §III], it
follows that in the case Pμ Pμ > 0, then Aμ is proportional to Pμ , see also [25].
We are now in the situation considered in the work by Bäckdahl and Valiente
Kroon, see [21, Theorem B.3], and hence we can conclude that (M, gab ) is
locally isometric to the Kerr spacetime.
Remark 2.4 1. This result can be turned into a characterization in terms of
Cauchy data along the lines in [20].
2. Theorem 2.3 can be viewed as a variation on the Kerr characterization given
in [21, Theorem B.3]. In the version given here, the asymptotic conditions
on the Killing spinor have been removed.
2.6. Monotonicity and Dispersion

The dispersive properties of fields, i.e. the tendency of the energy density
contained within any stationary region to decrease asymptotically to the future,
18
Section 11 appears only in the ArXiv version of [23].
is a crucial property for solutions of field equations on spacetimes, and any

proof of stability must exploit this phenomenon. In view of the geometric
optics approximation, the dispersive property of fields can be seen in an
analogous dispersive property of null geodesics, i.e. the fact that null geodesics
in the Kerr spacetime which do not orbit the black hole at a fixed radius must
leave any stationary region in at least one of the past or future directions. In
Section 2.6.1 we give an explanation for this fact using tools which can readily
be adapted to the case of field equations, while in Section 2.6.2 we outline how
these ideas apply to fields.
We begin by a discussion of conservation laws. For a null geodesic γ a , we
define the energy associated with a vector field X and evaluated on a Cauchy
hypersurface to be
eX [γ ]( ) = gab X a γ̇ b | .
Since γ̇ b ∇b γ̇ a = 0 for a geodesic, integrating the derivative of the energy gives
λ2
eX [γ ]( 2 ) − eX [γ ]( 1 ) = (γ̇a γ̇b )∇ (a X b) dλ, (2.73)
λ1
where λi is the unique value of λ such that γ (λ) is the intersection of γ with
i . Formula (2.73) is particularly easy to work with, if one recalls that
1
∇ (a X b) = − LX gab .
2
The tensor ∇ (a X b) is commonly called the “deformation tensor.” In the
following, unless there is the possibility of confusion, we will drop reference
to γ and in referring to eX .
Conserved quantities play a crucial role in understanding the behavior of
geodesics as well as fields. By (2.73), the energy eX is conserved if X a is a
Killing field. In the Kerr spacetime we have the Killing fields ξ a = (∂t )a , ηa =
(∂φ )a with the corresponding conserved quantities energy e = (∂t )a γ̇a and
azimuthal angular momentum z = (∂φ )a γ̇a . In addition, the squared particle
mass μ2 = gab γ̇ a γ̇ b and the Carter constant k = Kab γ̇ a γ̇ b are conserved along
any geodesic γ a in the Kerr spacetime. The presence of the extra conserved
quantity allows one to integrate the equations of geodesic motion.19
For a covariant field equation derived from an action principle which
depends on the background geometry only via the metric and its derivatives, the
19
In general, the geodesic equation in a four-dimensional stationary and axi-symmetric
spacetime cannot be integrated, and the dynamics of particles may in fact be chaotic; see
[57, 79] and references therein. Note however that the geodesic equations are not separable in
the Boyer–Lindquist coordinates. On the other hand, the Darboux coordinates have this
property, cf. [56].
symmetric stress-energy tensor Tab is conserved. As an example, we consider

the wave equation
∇ a ∇a ψ = 0 (2.74)
which has the stress-energy tensor
Tab = ∇(a ψ∇b) ψ̄ − 12 ∇ c ψ∇c ψ̄gab . (2.75)
Let ψ be a solution to (2.74). Then Tab is conserved, ∇ a Tab
= 0. For a vector
field X we have ∇ (Tab X ) given in terms of the deformation tensor,
a a b
∇ a (Tab X b ) = Tab ∇ (a X b) .
Let (JX )a = Tab X b be the current corresponding to X a . By the above, we have
conserved currents Jξ and Jη corresponding to the Killing fields ξ a , ηa .
An application of Gauss’s law gives the analog of (2.73),

(JX )a dσ a − (JX )a dσ a = Tab ∇ (a X b)
2 1
where
is a spacetime region bounded by 1 , 2 .
2.6.1. Monotonicity for Null Geodesics

We shall consider only null geodesics, i.e. μ = 0. In this case we have
k = Kab γ̇ a γ̇ b
= 2 l(a nb) γ̇ a γ̇ b
= 2 m(a m̄b) γ̇ a γ̇ b . (2.76)
We note that the tensors 2 l(a nb) and 2 m(a m̄b) are conformal Killing tensors;
see Section 2.4.5. From (2.76) it is clear that k is non-negative. A calculation
using (2.63) gives
1 2
2 l(a nb) ∂a ∂b = [(r + a2 )∂t + a∂φ ]2 − ∂r2

1
2 m(a m̄b) ∂a ∂b = ∂θ2 + ∂φ2 + a2 sin2 θ ∂t2 + 2a∂t ∂φ .
sin2 θ
Let Z = (r2 + a2 )e + az . Recall that ṙ = γ̇ r = grr γ̇r where grr = −/ .
Now we can write 0 = gab γ̇ a γ̇ b in the form
2 ṙ2 + R(r; e, z , k) = 0 (2.77)
where
R = −Z 2 + k (2.78)
Equation (2.77) is the exact analog of (2.28) for the Schwarzschild case. It is
clear from (2.76) that, for null geodesics, k corresponds in the Schwarzschild
case with a = 0, to L2 , the squared total angular momentum. It is possible to
derive equations similar to (2.77) for the other coordinates t, θ, φ, which allows
the solution of the geodesic equations by quadratures, see e.g. [105] for details.
Equation (2.77) allows one to make a qualitative analysis of the motion of
null geodesics in the Kerr spacetime. In particular, we find that the location
of orbiting null geodesics is determined by R = 0, ∂r R = 0. Due to the
form of R, the location of orbiting null geodesics depends only on the ratios
k/z 2 , e/z . One finds that orbiting null geodesics exist for a range of radii
r1 ≤ r ≤ r2 , with r+ < r1 < 3M < r2 . Here r1 , r2 depend on a, M and
as |a| M, r1 r+ , and r2 4M. The orbits at r1 , r2 are restricted to
the equatorial plane, those at r1 are co-rotating, while those at r2 are counter-
rotating. For r1 < r < r2 , the range of θ depends on r (see Figure 2.13). There
is r3 = r3 (a, M), r1 < r3 < r2 such that the orbits at r3 reach the poles, i.e.
θ = 0, θ = π; see Figure 2.13. For such geodesics, it holds that z = 0.
(a) (b)
Figure 2.13. The Kerr photon region. In subfigure (a), |a| M and the
ergoregion, see Section 2.3.7, is well separated from the photon region (bordered
in black). The radius r3 where geodesics reach the poles is indicated by a gray,
dashed line. In subfigure (b), |a| is close to M and the ergoregion overlaps the
photon region
(a) (b)
Figure 2.14. Examples of orbiting null geodesics in Kerr with a = M/2. In

subfigure (a), k/z 2 is small, while in subfigure (b), this constant is larger
For the following discussion, it is convenient to introduce

q = k − 2aez − z 2 = Qab γ̇a γ̇b ,
where
cos2 θ
Qab = (∂θ )a (∂θ )b + (∂φ )a (∂φ )b + a2 sin2 θ (∂t )a (∂t )b . (2.79)
sin2 θ
By construction, q is a sum of conserved quantities, and is therefore conserved.
Further, it is non-negative, since it is a sum of non-negative terms. In the
following we use (e, z , q) as parameters for null geodesics. Since we are
considering only null geodesics, there is no loss of generality compared to
using (e, z , k) as parameters.
For a null geodesic with given parameters (e, z , q), a simple turning point
analysis shows that there is a number ro ∈ (r+ , ∞) such that the quantity
(r − ro )γ̇ r increases overall. This quantity corresponds to the energy eA for
the vector field A = −(r − ro )∂r . Following this idea, we may now look for
a function F which will play the role of −(r − ro ), so that for A = F∂r , the
energy eA is non-decreasing for all λ and not merely non-decreasing overall.
For a = 0, both ro and F will necessarily depend on both the Kerr parameters
(M, a) and the constants of motion (e, z , q); the function F will also depend
on r, but no other variables.
We define Aa = F(∂r )a with
F = F(r; M, a, e, z , q).
It is important to note that this is a map from the tangent bundle to the tangent
bundle, and hence Aa = F(∂r )a cannot be viewed as a standard vector field,
which is a map from the manifold to the tangent bundle.
To derive a monotonicity formula, we wish to choose F so that eA has a
non-negative derivative. We define the covariant derivative of A by holding
the values of (e, z , q) fixed and computing the covariant derivative as if A
were a regular vector field. Similarly, we define LA gab by fixing the values
of the constants of geodesic motion. Since the constants of motion have zero
derivative along null geodesics, equation (2.73) remains valid.
Recall that null geodesics are conformally invariant up to reparameteriza-
tion. Hence, it is sufficient to work with the conformally rescaled metric gab .
Furthermore, since γ is a null geodesic, for any function qreduced , we may
subtract qreduced gab γ̇a γ̇a wherever it is convenient. Thus, the change in eA
is given as the integral of

1
γ̇a γ̇b ∇ A = − LA ( g ) − qreduced g
(a b) ab ab
γ̇a γ̇b .
2
The Kerr metric can be written as

1 ab
gab = −(∂r )a (∂r )b − R , (2.80)

where the tensorial form of Rab can be read off from the earlier definitions.
We now calculate −LA gab γ̇a γ̇b using (2.80). Ignoring distracting factors of ,
, the most important terms are
−2(∂r F)γ̇r γ̇r + F(∂r Rab )γ̇a γ̇b = −2(∂r F)γ̇r γ̇r + F(∂r R).
The second term in this sum will be non-negative if F = ∂r R(r; M, a; e, z , q).
Recall that the vanishing of ∂r R(r; M, a; e, z , q) is one of the two conditions
for orbiting null geodesics. With this choice of F, the instability of the null
geodesic orbits ensures that, for these null geodesics, the coefficient in the first
term, −2(∂r F), will be positive. These observations motivate the form of F
which yields non-negativity for all null geodesics.
It remains to make explicit the choices of F and qreduced . Once these choices
are made, the necessary calculations are straightforward but rather lengthy. Let
z and w be smooth functions of r and the Kerr parameters (M, a). Let R̃ denote
∂r ( z R(r; M, a; e, z , q)) and choose F = zwR̃ and qreduced = (1/2)(∂r z)wR̃ .
In terms of these functions,

2 z1/2
γ̇a γ̇b ∇ A = 2 w(R̃ ) − z
(a b) 1 1/2 3/2
∂r w 1/2 R̃ γ̇r2 . (2.81)

If z and w are chosen to be positive, then the first term on the right-hand side of
(2.81) which contains a square (R̃ )2 is non-negative. If we now take z = z1 =
(r2 + a2 )−2 and w = w1 = (r2 + a2 )4 /(3r2 − a2 ), then20

z1/2 3r4 + a4 3r4 − 6a2 r2 − a4
−∂r w 1/2 R̃ =2 z
2
+ 2 q. (2.82)
(3r2 − a2 )2 (3r2 − a2 )2
The coefficient of q is positive for r > r+ when |a| < 31/4 2−1/2 M =∼ 0.93M.
Since q is non-negative, the right-hand side of (2.82) is non-negative, and
hence also the right-hand side of equation (2.81) is non-negative, for this
range of a. Since equation (2.81) gives the rate of change, the energy eA is
monotone.
These calculations reveal useful information about the geodesic motion. The
positivity of the term on the right-hand side of (2.82) shows that R̃ can have
at most one root, which must be simple. In turn, this shows that R can have at
20
Equation (2.82) corrects a misprint in [12, Eq. (1.15b)].
most two roots. For orbiting null geodesics R must have a double root, which
must coincide with the root of R̃ . It is convenient to think of the corresponding
value of r as being ro .
The first term in (2.81) vanishes at the root of R̃ , as it must, so that eA can
be constantly zero on the orbiting null geodesics. When a = 0, the quantity R̃
reduces to −2(r − 3M)r−4 (z 2 + q), so that the orbits occur at r = 3M. The
continuity in a of R̃ guarantees that its root converges to 3M as a → 0 for
fixed (e, z , q).
From the geometrics optics approximation, it is natural to imagine that
the monotone quantity constructed in this section for null geodesics might
imply the existence of monotone quantities for fields, which would imply
some form of dispersion. For the wave equation, this is true. In fact, the
above discussion, when carried over to the case of the wave equation, closely
parallels the proof of the Morawetz estimate for the wave equation given in
[12], see Section 2.6.2 below. The quantity (γ̇α γ̇β )(∇ (α X β) ) corresponds to
the Morawetz density, i.e. the divergence of the momentum corresponding to
the Morawetz vector field. The role of the conserved quantities (e, z , q) for
geodesics is played, in the case of fields, by the energy fluxes defined via
second-order symmetry operators corresponding to these conserved quantities.
The fact that the quantity R vanishes quadratically on the trapped orbits is
reflected in the Morawetz estimate for fields, by a quadratic degeneracy of the
Morawetz density at the trapped orbits.
2.6.2. Dispersive Estimates for Fields

As discussed in Section 2.6.1, one may construct a suitable function of
the conserved quantities for null geodesics in the Kerr spacetime which is
monotone along the geodesic flow. This function may be viewed as arising
from a generalized vector field on phase space. The monotonicity property
implies, as discussed there, that non-trapped null geodesics disperse, in the
sense that they leave any stationary region in the Kerr spacetime. As mentioned
in Section 2.6.1, in view of the geometric optics approximation for the wave
equation, such a monotonicity property for null geodesics reflects the tendency
for waves in the Kerr spacetime to disperse.
At the level of the wave equation, the analog of the just mentioned
monotonicity estimate is called the Morawetz estimate. For the wave equation
∇ a ∇a ψ = 0, a Morawetz estimate provides a current Ja defined in terms of ψ
and some of its derivatives, with the property that ∇ a Ja has suitable positivity
properties, and that the flux of Ja can be controlled by a suitable energy defined
in terms of the field.
Let ψ be a solution of the wave equation ∇ a ∇a ψ = 0. Define the current

Ja by
Ja = Tab Ab + 12 q(ψ̄∇a ψ + ψ∇a ψ̄) − 12 (∇a q)ψ ψ̄,
where Tab is the stress-energy tensor given by (2.75). We have
∇ a Ja = Tab ∇ (a Ab) + q∇ c ψ∇c ψ̄ − 12 (∇ c ∇c q)ψ ψ̄. (2.83)
We now specialize to Minkowski space, with the line element gab dxa dxb =
dt2 − dr2 − dθ 2 − r2 sin2 θ dφ 2 . Let

E(τ ) = Ttt d3 x
{t=τ }
be the energy of the field at time τ , where Ttt is the energy density. The energy
is conserved, so that E(t) is independent of t.
Setting Aa = r(∂r )a , we have
∇ (a Ab) = gab − (∂t )a (∂t )b . (2.84)
With q = 1, we get
∇ a Ja = −Ttt .
With the above choices, the bulk term ∇ a Ja has a sign. This method can be
used to prove dispersion for solutions of the wave equation. In particular, by
introducing suitable cutoffs, one finds that for any R0 > 0, there is a constant
C, so that
t1
Ttt d3 xdt ≤ C(E(t0 ) + E(t1 )) ≤ 2CE(t0 ), (2.85)
t0 |r|≤R0

see [85]. The local energy, |r|≤R0 Ttt d3 x, is a function of time. By (2.85) it is
integrable in t, and hence it must decay to zero as t → ∞, at least sequentially.
This shows that the field disperses. Estimates of this type are called Morawetz
or integrated local energy decay estimates.
†
For a solution φAB of the Maxwell equation (C2,0 φ)AA = 0, the stress-
energy tensor Tab given by
Tab = φAB φ̄A B
is conserved, ∇ a Tab = 0. Further, Tab has trace zero, with T a a = 0.

Restricting to Minkowski space and setting Ja = Tab Ab , with Aa = r(∂r )a
we have
∇ a Ja = −Ttt
which again gives local energy decay for the Maxwell field on Minkowski
space.
For the wave equation on Schwarzschild we can choose

(r − 3M)(r − 2M)
Aa = (∂r )a , (2.86a)
3r2
6M 2 − 7Mr + 2r2
q= . (2.86b)
6r3
This gives
Mgab (r − 3M) M(r − 2M)2 (∂r )a (∂r )b

−∇ (a Ab) = − +
3r3 r4
(r − 3M) ((∂θ ) (∂θ ) + csc θ (∂φ )a (∂φ )b )
2 a b 2
+ 5
, (2.87a)
3r

M|∂r ψ|2 (r − 2M)2 |∂θ ψ|2 + |∂φ ψ|2 csc2 θ (r − 3M)2
−∇a J a = +
r4 3r5
M|ψ| (54M − 46Mr + 9r )
2 2 2
+ . (2.87b)
6r6
Here, Aa was chosen so that the last two terms (2.87a) have good signs. The
form of q given here was chosen to eliminate the |∂t ψ|2 term in (2.87b). The
first terms in (2.87b) are clearly non-negative, while the last is of a lower-order
and can be estimated using a Hardy estimate [12]. The effect of trapping in
Schwarzschild at r = 3M is manifested in the fact that the angular derivative
term vanishes at r = 3M.
In the case of the wave equation on Kerr, the above argument using a
classical vector field cannot work due to the complicated structure of the
trapping. However, making use of higher-order currents constructed using
second-order symmetry operators for the wave equation, we can construct a
generalized Morawetz vector field analogous to the vector field Aa . This is
discussed in Section 2.6.1 and has been carried out in detail in [12].
If we apply the same idea for the Maxwell field on Schwarzschild, there is
no reason to expect that local energy decay should hold, in view of the fact that
the Coulomb solution is a time-independent solution of the Maxwell equation
which does not disperse. In fact, with

2M
Aa = F(r) 1 − (∂r )a , (2.88)
r
we have

−Tab ∇ (a Ab) = − φ AB φ̄ A B (T1,1 A)ABA B (2.89)
(r − 2M)
= |φ0 |2 + |φ2 |2 F (r)
2r
|φ1 |2 r(r − 2M)F (r) − 2F(r)(r − 3M)
− . (2.90)
r2
If F is chosen to be positive, then the coefficient of the extreme components in

(2.90) is positive. However, at r = 3M, the coefficient of the middle component
is necessarily of the opposite sign. It is possible to show that no choice of F
will give positive coefficients for all components in (2.90).
The dominant energy condition, that Tab V a W b ≥ 0 for all causal vectors
V , W a , is a common and important condition on stress energy tensors. In
a
Riemannian geometry, a natural condition on a symmetric 2-tensor Tab would

be non-negativity, i.e. the condition that for all X a , one has Tab X a X b ≥ 0.
However, in order to prove dispersive estimates for null geodesics and the
wave equation, the dominant energy condition on its own is not sufficient and
non-negativity cannot be expected for stress energy tensors. Instead, a useful
condition to consider is non-negativity modulo trace terms, i.e. the condition
that for every X a there is a q such that Tab X a X b + qT a a ≥ 0. For null geodesics
and the wave equation, the tensors γ̇a γ̇b and ∇a u∇b u = Tab + T γ γ gab are both
non-negative, so γ̇a γ̇b and Tab are non-negative modulo trace terms.
From equation (2.87a), we see that −∇ (a Ab) is of the form f1 gab + f2 ∂ra ∂rb +
f3 ∂θa ∂θb + f4 ∂φa ∂φb where f2 , f3 , and f4 are non-negative functions. That is
−∇ (a Ab) is a sum of a multiple of the metric plus a sum of terms of the
form of a non-negative coefficient times a vector tensored with itself. Thus,
from the non-negativity modulo trace terms, for null geodesics and the wave
equation respectively, there are functions q such that γ̇a γ̇b ∇ a Ab = γ̇a γ̇b ∇ a Ab +
qgab γ̇a γ̇b ≤ 0 and Tab ∇ a Ab +qT a a ≤ 0. For null geodesics, since gab γ̇a γ̇b = 0,
the q term can be ignored. For the wave equation, one can use the terms
involving q in equations (2.83), to cancel the T a a term in ∇ a Ja . For the wave
equation, this gives non-negativity for the first-order terms in −∇ a Ja , and one
can then hope to use a Hardy estimate to control the zeroth-order terms.
If we now consider the Maxwell equation, we have the fact that the
Maxwell stress energy tensor is traceless, T a a = 0 and does not satisfy the
non-negativity condition. Therefore it also does not satisfy the condition of
non-negativity modulo trace. This appears to be the fundamental underlying
obstruction to proving a Morawetz estimate using Tab . This can be seen
as a manifestation of the fact that the Coulomb solution does not disperse.
In fact, it is immediately clear that the Maxwell stress
energy cannot be used directly to prove dispersive
estimates since it does not vanish for the Coulomb
field (2.45) on the Kerr spacetime. We may remark
that the existence of the Coulomb solution on the Kerr
spacetime is a consequence of the fact that the exterior
of the black hole contains non-trivial 2-spheres and the existence of two
conserved charge integrals S Fab dσ ab , S (∗F)ab dσ ab . Hence this is valid also
for dynamical black hole spacetimes.
2.7. Symmetry Operators

A symmetry operator for a field equation is an operator which takes solutions
to solutions. In order to analyze higher spin fields on the Kerr spacetime, it is
important to gain an understanding of the symmetry operators for this case. In
the paper [9] we have given a complete characterization of those spacetimes
admitting symmetry operators of a second order for the field equations of spins
0, 1/2, 1, i.e. the conformal wave equation, the Dirac–Weyl equation and the
Maxwell equation, respectively, and given the general form of the symmetry
operators, up to equivalence. In order to simplify the presentation here, we shall
discuss only the spin-1 case, and restrict ourselves to spacetimes admitting a
valence (2, 0) Killing spinor κAB . We first give some background on the wave
equation.
2.7.1. Symmetry Operators for the Kerr Wave Equation

As shown by Carter [33], if Kab is a Killing tensor in a Ricci flat spacetime, the
operator
K = ∇a K ab ∇b (2.91)
is a commuting symmetry operator for the d’Alembertian,
[∇ a ∇a , K] = 0.
In particular there is a second-order symmetry operator for the wave equation,

i.e. an operator which maps solutions to solutions,
∇ a ∇a ψ = 0 ⇒ ∇ a ∇a Kψ = 0.
Due to the form of the Carter Killing tensor, Kab , cf. (2.68), the operator K
defined by (2.91) contains derivatives with respect to all coordinates.
√
Recall that ∇ a ∇a = μ1g ∂a μg gab ∂b , where μg = det(gab ) is the volume
element. For Kerr in Boyer–Lindquist coordinates, we have from (2.35) μg =
μ, with μ = sin θ. After rescaling the d’Alembertian by , and using the just
mentioned facts, one finds
R(r; ∂t , ∂φ , Q)
∇ a ∇a = − ∂r ∂r + (2.92)

where
1
Q= ∂a μQab ∂b .
μ
In view of the form of Qab given in (2.79), we see that Q contains derivatives
only with respect to θ, φ, t, but not with respect to r. Thus, it is clear from
(2.92) that Q is a commuting symmetry operator for the rescaled d’Alembertian

∇ a ∇a ,
[ ∇ a ∇a , Q] = 0.
In addition to the symmetry operator Q related to the Carter constant, we have

the second-order symmetry operators generated by the Killing fields ξ a ∇a =
∂t , ηa ∇a = ∂φ . The operator Q can be termed a hidden symmetry, since it
cannot be represented in terms of operators generated by the Killing fields.
The above shows that we can write
∇ a ∇a = R + S
where the operators R, S commute, [R, S] = 0, and R contains derivatives with

respect to the non-symmetry coordinate r, and the two symmetry coordinates
t, φ, while S contains derivatives with respect to the non-symmetry coordinate
θ , and with respect to t, φ.
By making a separated ansatz
ψω,,m (t, r, θ , φ) = e−iωt eimφ Rω,,m (r)Sω,,m (θ )
the equation ∇ a ∇a ψ = 0 becomes a pair of scalar ordinary differential

equations
RR + λR = 0 (2.93a)
SS = λS (2.93b)
where λ = λω,,m . Here it should be noted that equation (2.93b) is to be

considered as a boundary value problem on [0, π ] with boundary conditions
determined by the requirement that φ be smooth. In the Schwarzschild case
a = 0, we can take S =, the angular Laplacian. The eigenfunctions of
are the spherical harmonics Y,m (θ, φ) = eimφ Y (θ ). The eigenvalues of
are
λ,m = −( + 1).
The solutions to the eigenproblem SS = λS are the spheroidal harmonics;
the eigenvalues in this case are not known in closed form, but depend on the
time frequency ω, and are indexed by , m. For real ω, it is known that the
eigensystem is complete, but for general ω this is not known.
One may now apply a Fourier transform and represent a typical solution ψ
to the wave equation in the form

ψ = dω e−iωt eimφ Rω,,m Sω,,m ,
,m
analyze the behavior of the separated modes ψω,,m , and recoved estimates for
ψ after inverting the Fourier transform. In order to do this, one must show a
priori that the Fourier transform can be applied. This can be done by applying
cutoffs, and removing these after estimates have been proved using Fourier
techniques. This approach has been followed in e.g. [45, 14, 13]. In recent
work by Dafermos, Rodnianski, and Shlapentokh-Rothman, see [46], proving
boundedness and decay for the wave equation on Kerr for the whole range
|a| < M, makes use of the technical condition of time integrability, i.e. that the
solution to the wave equation and its derivatives to a sufficiently high order is
bounded in L2 on time lines,
∞
dt|∂ α ψ(t, r, θ , φ)|2 .
−∞
This condition is consistent with integrated local energy decay and is removed
at the end of the argument.
However, by working directly with currents defined in terms of second-order
symmetry operators, one may prove a Morawetz estimate directly for the wave
equation on the Kerr spacetime. This was carried out for the case |a| M
in [12]. This involves introducing a generalization of the vector field method
to allow for currents defined in terms of generalized, operator valued, vector
fields. These are operator analogs of the generalized vector field Aa introduced
in Section 2.6.1.
Fundamental for either of the above mentioned approaches, is that the
analysis of the wave equation on the Kerr spacetime is based on the hidden
symmetry manifested in the existence of the Carter constant, or the conserved
quantity q, and its corresponding symmetry operator Q.
2.7.2. Symmetry Operators for the Maxwell Field

There are two spin-1 equations (left and right) depending on the helicity of the
spinor. These are
†
(C2,0 φ)AA = 0 (left), and (C0,2 ϕ)AA = 0 (right).
The real Maxwell equation ∇ a Fab = 0, ∇[a Fbc] = 0 for a real two form
Fab = F[ab] is equivalent to either the right or the left Maxwell equations.
Henceforth we will always assume that φAB solves the left Maxwell equation.

Given a conformal Killing vector ν AA , we follow [6, Equations (2) and
(15)], see also [5], and define a conformally weighted Lie derivative acting
on a symmetric valence (2s, 0) spinor field as follows.

Definition 2.3 For ν AA ∈ ker T1,1 , and ϕA1 ...A2s ∈ S2s,0 , we define

L̂ν ϕA1 ...A2s ≡ ν BB ∇BB ϕA1 ...A2s + sϕB(A2 ...A2s ∇A1 )B ν BB
CC
+ 4 ϕA1 ...A2s ∇
1−s
νCC . (2.94)
† †
If ν a is a conformal Killing field, then (C2,0 L̂ν ϕ)AA = L̂ν (C2,0 ϕ)AA . It
follows that the first-order operator ϕ → L̂ν ϕ defines a symmetry operator
of the first order, which is also of the first kind. For the equations of spins 0
and 1, the only first-order symmetry operators are given by conformal Killing
fields. For the spin-1 equation, we may have symmetry operators of the first
kind, taking left fields to left, i.e. ker C † → ker C † and of the second kind,
taking left fields to right, ker C † → ker C . Observe that symmetry operators of
the first kind are linear symmetry operators in the usual sense, while symmetry
operators of the second kind followed by complex conjugation gives anti-linear
symmetry operators in the usual sense.
Recall that the Kerr spacetime admits a constant of motion for geodesics
q which is not reducible to the conserved quantities defined in terms of
Killing fields, but rather is defined in terms of a Killing tensor. Similarly, in
a spacetime with Killing spinors, the geometric field equations may admit
symmetry operators of order greater than one, not expressible in terms of the
symmetry operators defined in terms of (conformal) Killing fields. We refer to
such symmetry operators as “hidden symmetries.”
In general, the existence of symmetry operators of the second order implies
the existence of Killing spinors (of valence (2, 2) for the conformal wave
equation and for Maxwell symmetry operators of the first kind for Maxwell, or
(4, 0) for Maxwell symmetry operators of the second kind) satisfying certain
auxiliary conditions. The conditions given in [9] are valid in arbitrary four-
dimensional spacetimes, with no additional conditions on the curvature. As
shown in [9], the existence of a valence (2, 0) Killing spinor is a sufficient
condition for the existence of second-order symmetry operators for the spin-s
equations, for s = 0, 1/2, 1.
Remark 2.5 1. If κAB is a Killing spinor of valence (2, 0), then LABA B =
κAB κ̄A B and LABCD = κ(AB κCD) are Killing spinors of valence (2, 2) and
(4, 0), respectively, satisfying the auxiliary conditions given in [9].
2. In the case of aligned matter with respect to ABCD , any valence (4, 0)
Killing spinor LABCD factorizes, i.e. LABCD = κ(AB κCD) for some Killing
spinor κAB of valence (2, 0) [9, Theorem 8]. An example of a spacetime
with aligned matter which admits a valence (2, 2) Killing spinor that does
not factorize is given in [9, §6.3], see also [83].
Proposition 2.1 ([9]) 1. The general symmetry operator of the first kind for
the Maxwell field, of order at most two, is of the form
χAB = QφAB + (C1,1 A)AB , (2.95)
where φAB is a Maxwell field, and AAA is a linear concomitant21 of the first
†
order, such that AAA ∈ ker C1,1 and Q ∈ ker T0,0 , i.e. is locally constant.
2. The general symmetry operator of the second kind for the Maxwell field is
of the form
†
ωA B = (C1,1 B)A B , (2.96)
where BAA is a first-order linear concomitant of φAB such that BAA ∈
ker C1,1 .
†
Remark 2.6 The operators C1,1 and C1,1 are the adjoints of the left and right
†
Maxwell operators C2,0 and C0,2 . The conserved currents for the Maxwell field
can be characterized in terms of solutions of the adjoint Maxwell equations
†
(C1,1 A)A B = 0 (2.97a)
(C1,1 B)AB = 0. (2.97b)
Definition 2.4 Given a spinor κAB ∈ S2,0 we define the operators E2,0 : S2,0 →
S2,0 and E¯0,2 : S0,2 → S0,2 by
(E2,0 ϕ)AB = − 2κ(A C ϕB)C , (2.98a)
C
(E¯0,2 φ)A B = − 2κ̄(A φB )C . (2.98b)
Let κi be the Newman–Penrose scalars for κAB . If κAB is of algebraic type
{1, 1} then κ0 = κ2 = 0, in which case κAB = −2κ1 o(A ιB) . A direct calculation
gives the following result.
Lemma 2.1 Let κAB ∈ S2,0 and assume that κAB is of algebraic type {1, 1}.
Then the operators E2,0 , E¯2,0 remove the middle component and rescale the
extreme components as
(E2,0 ϕ)0 = − 2κ1 ϕ0 , (E2,0 ϕ)1 = 0, (E2,0 ϕ)2 = 2κ1 ϕ2 , (2.99a)
(E¯0,2 φ)0 = − 2κ̄1 φ0 , (E¯0,2 φ)1 = 0, (E¯0,2 φ)2 = 2κ̄1 φ2 . (2.99b)
Remark 2.7 If κAB is a Killing spinor in a Petrov type D spacetime, then κAB
is of algebraic type {1, 1}.
21
A concomitant is a covariant, local partial differential operator.
Definition 2.5 Define the first-order 1-form linear concomitants AAA , BAA by
†
AAA [κAB , φAB ] = − 13 (E2,0 φ)AB (C0,2 κ̄)B A + κ̄A B (C2,0 E2,0 φ)A B , (2.100a)
AAA [νAA , φAB ] = νBA φA B , (2.100b)
† †
BAA [κAB , φAB ] = κAB (C2,0 E2,0 φ)B A + 1 B
3 (E2,0 φ)AB (C2,0 κ) A . (2.100c)
When there is no room for confusion, we suppress the arguments and write
simply AAA , BAA . The following result shows that AAA , BAA solves the adjoint
Maxwell equations, provided φAB solves the Maxwell equation.
Lemma 2.2 ([9, §7]) Assume that κAB is a Killing spinor of valence (2, 0), that
νAA is a conformal Killing field, and that φAB is a Maxwell field. Then, with
AAA , BAA given by (2.100) it holds that AAA [κAB , φAB ] and AAA [νAA , φAB ]
†
satisfy (C1,1 A)A B = 0, and BAA [κAB , φAB ] satisfies (C1,1 B)AB = 0.
Remark 2.8 Proposition 2.1 together with Lemma 2.2 show that the existence
of a valence (2, 0) Killing spinor implies that there are non-trivial second-order
symmetry operators of the first and second kind for the Maxwell equation.
2.8. Conservation Laws for the Teukolsky System

Recall that the operators C and C † are adjoints, and hence their composition
yields a wave operator. We have the identities (valid in a general spacetime)
†
ϕAB + 8 ϕAB − 2ABCD ϕ CD = − 2(C1,1 C2,0 ϕ)AB , (2.101a)
†
ϕABCD − 6(AB FH ϕCD)FH = − 2(C3,1 C4,0 ϕ)ABCD . (2.101b)
Here ϕAB and ϕABCD are elements of S2,0 and S4,0 , respectively. This means
†
that the Maxwell equation (C2,0 φ)AA = 0 in a vacuum spacetime implies the
wave equation
φAB − 2ABCD φ CD = 0. (2.102)

†
Similarly, in a vacuum spacetime, the Bianchi system (C4,0 )A ABC = 0 holds
for the Weyl spinor, and we arrive at the Penrose wave equation
ABCD − 6(AB FH CD)FH = 0. (2.103)
Restricting to a vacuum type D spacetime, and projecting the Maxwell wave

equation (2.102) and the linearized Penrose wave equation (2.103) on the
principal spin dyad, one obtains wave equations for the extreme Maxwell
˙ 0,
scalars φ0 , φ2 and the extreme linearized Weyl scalars ˙ 4.
−2/3
Letting ψ (s) denote φ0 , 2 φ2 for s = 1, −1, respectively, and
˙ 0 , −4/3
˙ 4 for s = 2, −2, respectively, one finds that these fields satisfy the
2
system
T 2s − 4s 2 ]ψ = 0,
2 (s)
[ (2.104)
see [3, §3], where, in GHP notation

T p = 2(þ −pρ − ρ̄)(þ − ρ ) − 2(ð −pτ − τ̄ )(ð − τ ) + (3p − 2)2 . (2.105)
Equation (2.104) was first derived by Teukolsky [106, 107] for massless spin-s
fields and linearized gravity on Kerr, and is referred to as the Teukolsky Master
Equation (TME). It was shown by Ryan [100] that the tetrad projection of the
linearized Penrose wave equation yields the TME, see also Bini et al. [29, 30].
In the Kerr case, The TME admits a commuting symmetry operator, and hence
allows separation of variables; the equation is valid for fields of all half-integer
spins between 0 and 2.
As discussed above, the TME is a wave equation for the weighted field ψ (s) .
It is derived from the spin-s field equation by applying a first-order operator
and hence is valid for the extreme scalar components of the field, rescaled as
explained above. It is important to emphasize that there is a loss of information
in deriving the TME from the spin-s equation. For example, if we consider two
independent solutions of the TME with spin weights s = ±1, these will not in
general be components of a single Maxwell field. If indeed this is the case, the
Teukolsky–Starobinsky identities (TSI) (also referred to as Teukolsky–Press
relations), see [67] and references therein, hold.
The TME admits commuting symmetry operators Ss , Rs , so that

T 2s − 4s 2 = Rs + Ss
2
with [Rs , Ss ] = 0, and such that, as in the case for the wave equation discussed
in Section 2.7.1, the operators Rs , Ss involve derivatives with respect to r and
θ , respectively, in addition to derivatives in the symmetry directions t, φ. This
shows that one may make a consistent separated ansatz
ψ (s) (t, r, θ , φ) = e−iωt eimφ R(s) (r)S(s) (θ )
where R(s) solves the radial TME
(Rs + λs,ω,,m )R(s) = 0
where λs,ω,,m is an eigenvalue for the angular Teukolsky equation Ss S(s) =
λS(s) , which is the equation for a spin-weighted spheroidal harmonic.
Although the TSI are usually discussed in terms of separated forms of ψ (s) ,
we are here interested in the TSI as differential relations between the scalars of
extreme spin weights. From this point of view, the TSI expresses the fact that
the Debye potential construction starting from the different Maxwell scalars
for a given Maxwell field φAB yields scalars of the same Maxwell field. The
equations for the Maxwell scalars in terms of Debye potentials can be found
in Newman–Penrose notation in [42]. These expressions correspond to the
components of a symmetry operator of the second kind. See [2, §5.4.2] for
further discussion, where also the GHP version of the formulas can be found.
An analogous situation obtains for the case of linearized gravity, see [77]. In
this case, the TSI are of the fourth order. Thus, for a Maxwell field, or a solution
of the linearized Einstein equations on a Kerr, or more generally a vacuum type
D background, the pair of Newman–Penrose scalars of extreme spin weights
for the field satisfy a system of differential equations consisting of both the
TME and the TSI.
Although the TME is derived from an equation governed by a variational
principle, it has been argued by Anco, see the discussion in [95], that the
Teukolsky system admits no real variational principle, due to the fact that the
operator T p defined by the above fails to be formally self-adjoint. Hence, the
issue of real conserved currents for the Teukolsky system, which appear to be
necessary for estimates of the solutions, appears to be open. However, as we
shall demonstrate here, if we consider the combined TME and TSI in the spin-
1 or Maxwell case, as a system of equations for both of the extreme Maxwell
scalars φ0 , φ2 , this system does admit both a conserved current and a conserved
stress-energy-like tensor.
2.8.1. A New Conserved Tensor for Maxwell

Let φAB → χAB be the second order symmetry operator of the first kind given
by (2.95) with Q = 0 and AAA given by (2.100a), and let ξAA be given by
(2.54). Then the current

AA = 12 ξ BB χAB φ̄A B + 12 ξ BB φAB χ̄A B ,
is conserved. In fact, as discussed in [10, section 6], −AA is equivalent to a
current Vab ξ b defined in terms of a symmetric tensor Vab which we shall now
introduce. Let
†
ηAA ≡ (C2,0 E2,0 φ)AA . (2.106)
where (E2,0 φ)AB is given by (2.98a) and define the symmetric tensor Vab by
VABA B ≡ 12 ηAB η̄A B + 12 ηBA η̄B A + 13 (E2,0 φ)AB (L̂ξ̄ φ̄)A B
+ 13 (E¯2,0 φ̄)A B (L̂ξ φ)AB . (2.107)
Then, as we shall now show, Vab is itself conserved,
∇ a Vab = 0,
and hence may be viewed as a higher-order stress-energy tensor for the

Maxwell field. The tensor Vab has several important properties. First of all, if
M is of Petrov type D, it depends only on the extreme Maxwell scalars φ0 , φ2 ,
and hence cancels the static Coulomb Maxwell field (2.45) on Kerr which has
only the middle scalar non-vanishing. This can be proved using Lemma 2.1,
cf. [10, Corollary 6.2]. Further, the tensor
UAA BB = 12 ηAB η̄A B + 12 ηBA η̄B A
is a superenergy tensor for the 1-form field ηAA , and hence satisfies the
dominant energy condition, cf. [27, 103]. Note that the notion of a superenergy
tensor extends to spinors of arbitrary valence. Similarly to the wave equation
stress energy, Vab has a non-vanishing trace, V a a = U a a = −η̄a ηa .
In order to analyze Vab , we first collect some properties of the one-form ηAA
as defined in (2.106).
Lemma 2.3 ([8, Lemma 2.4]) Let κAB ∈ KS 2,0 , and assume the aligned matter
condition holds with respect to κAB . Let ξAA be given by (2.54). Further, let φAB
be a Maxwell field, and let ηAA be given by (2.106). Then we have
(D1,1 η) = 0, (2.108a)
(C1,1 η)AB = 3 (L̂ξ φ)AB ,
2
(2.108b)
†
(C1,1 η)A B = 0, (2.108c)
AA
ηAA ξ = κ AB (L̂ξ φ)AB . (2.108d)
The following lemma gives a general condition, without assumptions on

the spacetime geometry, for a tensor constructed along the lines of Vab to be
conserved. The proof is a straightforward computation.
Lemma 2.4 ([8, Lemma 3.1]) Assume that ϕAB ∈ S2,0 satisfies the system
† †
(C1,1 C2,0 ϕ)A B = 0, (2.109a)
†
(C1,1 C2,0 ϕ)AB = $AB , (2.109b)
for some $AB ∈ S2,0 . Let

†
ςAA = (C2,0 ϕ)AA , (2.110)
and define the symmetric tensor XABA B by
XABA B = 12 ςAB ς̄A B + 12 ςBA ς̄B A + 12 $̄A B ϕAB + 12 $AB ϕ̄A B . (2.111)
Then

∇ BB XABA B = 0. (2.112)
We now have the following result, which follows directly from Lemma 2.4
and the identities for ηAA given in Lemma 2.3 together with the above remarks.
Theorem 2.4 ([8, Theorem 1.1]) Assume that (M, gab ) admits a valence (2, 0)
Killing spinor κAB and assume that the aligned matter condition holds with
respect to κAB . Let φAB be a solution of the Maxwell equation. Then the tensor
VABA B given by (2.107) is conserved, i.e.

∇ AA VABA B = 0
If in addition (M, gab ) is of Petrov type D, then Vab depends only on the
extreme components of φAB .
The properties of Vab indicate that Vab , rather than the Maxwell stress-
energy Tab , may be used in proving dispersive estimates for the Maxwell
field. In Section 2.9 we shall outline the proof of a Morawetz estimate for
the Maxwell field on the Schwarzschild background, making use of a related
approach.
2.8.2. Teukolsky Equation and Conservation Laws

We end this section by pointing out the relation between the fact that Vab is
conserved and the TME and TSI which follow from the Maxwell equation in a
Petrov type D spacetime.
A computation shows that the identities
† †
(C1,1 C2,0 E2,0 φ)A B = 0 (2.113a)
†
(E2,0 C1,1 C2,0 E2,0 φ)AB = 3 (L̂ξ E2,0 φ)AB
2
(2.113b)
follow from the Maxwell equations, cf. [8, Eq. (3.5)]. We see that this system
is equivalent to (2.109a), (2.109b), with ϕAB = E2,0 φ and $AB = 23 (L̂ξ φ)AB .
This shows that the fact that Vab is conserved is a direct consequence of (2.113),
which in fact are the covariant versions of the TME and TSI. In order to
make this clear for the case of the TME, we project (2.113b)on the dyad. A
calculation shows that
0 = − þ þ ϕ0 + ρ þ ϕ0 + ρ̄ þ ϕ0 + ð ð ϕ0 − τ ð ϕ0 − τ ð ϕ0 , (2.114a)
0 = − ρ þ ϕ2 − ρ þ ϕ2 + þ þ ϕ2 + τ̄ ð ϕ2 + τ ð ϕ2 − ð ð ϕ2 (2.114b)
where ϕ0 = −2κ1 φ0 and ϕ2 = 2κ1 φ2 . We see from this that (2.113b) is

equivalent to the scalar form of TME for Maxwell given in (2.104) above.
Further, one can show along the same lines that (2.113a) is equivalent to the
TSI for Maxwell given in scalar form in [2, §5.4.2], cf. [8, §3.1].
2.9. A Morawetz Estimate for the Maxwell Field

on Schwarzschild
In this section, we shall outline the proof of the Morawetz estimate for the
Maxwell field on the Schwarzschild spacetime given recently in [11].
Assume that κAB is a valence (2, 0) Killing spinor, such that κCD κ CD = 0.
In the case of the Schwarzschild spacetime, κAB is given by (2.65). Define the

Killing fields ξ AA , ηAA in terms of κAB by (2.54) and (2.106), respectively.
Now let φAB be a solution to the source-free Maxwell equation
†
(C2,0 φ)AA = 0 and define
UAA = − 12 ∇AA log(−κCD κ CD ), (2.115a)

ϒ = κ AB φAB , (2.115b)
AB = (E2,0 φ)AB , (2.115c)
βAA = ηAA + U B
A AB . (2.115d)
In the Schwarzschild case, we have

ξ AA = (∂t )AA , UAA = − r−1 ∇AA r. (2.116)
Analogously to Lemma 2.3, we have
Lemma 2.5 ([11, Lemma 8])
βAA = − UAA ϒ + (T0,0 ϒ)AA , (2.117a)
AA
(D1,1 β) = − U βAA , (2.117b)
A
(C1,1 β)AB = U(A βB)A , (2.117c)
†
(C1,1 β)A B =U A
(A β|A|B ) . (2.117d)
The superenergy tensors for βAA and AB are given by
HABA B = 12 βAB β A B + 12 βBA β B A , (2.118a)
WABA B = AB A B . (2.118b)
Choosing the principal tetrad in Schwarzschild given by specializing (2.63) to
a = 0 gives in a standard manner an orthonormal frame,

T AA ≡ √1 (oA ōA + ιA ῑA ),
X AA ≡ √1 (ōA ιA + oA ῑA ),
2 2

Y AA ≡ √i (−ōA ιA + oA ῑA ),
Z AA ≡ √1 (oA ōA − ιA ῑA ).
2 2
The tensor HABA B , which agrees up to lower order terms with the conserved
tensor VABA B introduced in Section 2.8.1, is not itself conserved, it yields a
conserved energy current.
Lemma 2.6 ([11, Lemma 11]) For the Schwarzschild spacetime we have

∇ BB HABA B = − UAA β BB β̄B B , (2.119a)
AA BB
ξ ∇ HABA B = 0. (2.119b)

In particular, ξ BB HABA B is a future causal conserved current.
This result makes use of the fact that the Schwarzschild spacetime is non-
rotating. For the Kerr spacetime with non-vanishing angular momentum, the
1-form UAA fails to be real and the current Hab ξ b is not conserved.
For a vector field Aa and a scalar q, define the Morawetz current Pa by

PAA = HABA B ABB − 12 qβ̄A B AB − 12 qβA B A B

+ 12 A B A B (T0,0 q)BB . (2.120)
For any spacelike hypersurface , we define the energy integrals

Eξ ( ) = Hab ξ b N a dμ , (2.121)

Eξ +A,q ( ) = Hab ξ b + Pa N a dμ . (2.122)

In view of Lemma 2.6, Eξ ( ) is non-negative and conserved.
We shall make the following explicit choices of the Aa and q,
(r − 3M)(r − 2M)
Aa = (∂r )a , (2.123a)
2r2
9M 2 (r − 2M)(2r − 3M)
q= . (2.123b)
4r5
2.9.1. Positive Energy

Before proving the integrated decay estimate, we shall verify that the energy
(2.122) is non-negative and uniformly equivalent to the energy (2.121).
From the properties of spin-weighted spherical harmonics, one derives the
inequalities

|ϕ0 |2 dμSr ≤ r2 | ð ϕ0 |2 dμSr , (2.124a)
Sr Sr

|ϕ2 |2 dμSr ≤ r2 | ð ϕ2 |2 dμSr (2.124b)
Sr Sr
for the extreme scalars ϕ0 , ϕ2 of a smooth symmetric spinor field ϕAB , cf.
[11, Lemma 6]. Here Sr is a sphere with constant t, r in the Schwarzschild
spacetime.
By making use of the Cauchy–Schwarz inequality, and the Hardy type

inequalities (2.124), we get

Theorem 2.5 ([11, Theorem 13, Corollary 14]) Let AAA and q be given by
(2.123a) and (2.123b).
1. For any constant |c1 | ≤ 10/9 and any spherically symmetric slice with

future pointing timelike normal N AA such that N AA NAA = 1 we have a
positive energy

N AA (HABA B ξ BB + c1 PAA )dμ i ≥ 0. (2.125)

2. For any spherically symmetric slice with future pointing timelike normal

N AA such that N AA NAA = 1 the energies Eξ ( ) and Eξ +A,q ( ) are
uniformly equivalent,
1
10 Eξ ( ) ≤ Eξ +A,q ( ) ≤ 19
10 Eξ ( ). (2.126)
In particular, we find that using Theorem 2.5, we can dominate the integral
of the bulk term for the Morawetz current over a spacetime domain bounded
by Cauchy surfaces 1 , 2 , in terms of the energies Eξ ( 1 ), Eξ ( 2 ). This
is the essential step in the proof of an integrated energy decay (or Morawetz)
estimate.
We shall apply (2.124) to AB . We have

Z | + |β
|β 2
Z | = | ð 2 | + | ð 0 | ,
2 2 2
(2.127a)
and

W
T
AA
T =T T BB AB A B = 12 |0 |2 + 12 |2 |2 . (2.127b)
Equations (2.127a), (2.127b) and the inequalities (2.124) combine to give the
estimate

r2
T dμSr ≤ |β|2 + |β Z | dμSr ,
2
W T (2.128)
Sr 2 Sr T
cf. [11, Lemma 15]. From the form (2.120) of the Morawetz current Pa , the
definition of βAA and the properties of βa given in Lemma 2.5 we get

−(D1,1 P) = − β AA β̄ B B (T1,1 A)ABA B

+ β AA β̄A A 14 (D1,1 A) + ABB UBB − q (2.129)
AA B
+ AB A B U (T0,0 q) − 2 (T1,1 T0,0 q)
BB 1 ABA
.
With the explicit choices (2.123a) and (2.123b) for the Morawetz vector field
Aa and the scalar q, respectively, the above estimates now yield

1 M
−(D1,1 P)dμ
≥ |βAA |21,deg + |AB |22 dμ
, (2.130)

8 100r4
for any spherically symmetric spacetime region
of the Schwarzschild
spacetime.
We now make use of Gauss’s formula to evaluate the left-hand side of
(2.130). Theorem 2.5 and the estimates just proved then yield the follow-
ing energy bound and Morawetz estimate for the Maxwell field on the
Schwarzschild spacetime.
Theorem 2.6 ([11, Theorem 2]) Let 1 and 2 be spherically symmetric
spacelike hypersurfaces in the exterior region of the Schwarzschild spacetime
such that 2 lies in the future of 1 and 2 ∪ − 1 is the oriented boundary of
a spacetime region
.
If φAB is a solution of the Maxwell equations on the Schwarzschild exterior,
and AB and βAA are defined by equations (2.115d) to (2.115c), then
Eξ ( 2 ) = Eξ ( 1 ), (2.131)

2M 72
|βAA |21,deg + 4
|AB |22 dμ
≤ Eξ ( 1 ), (2.132)

25r 5
where Eξ ( i ) is the energy associated with ξ a , evaluated on i , and |βAA |1,deg
and |AB |2 are, respectively, the degenerate norm of βAA and the norm of AB
defined by
(r − 3M)2 M(r − 2M)
|βAA |21,deg = 3
|β 2
Y| +
X | + |β
2
|β
Z|
2
r r3
M(r − 3M)2 (r − 2M)
+ |β
T| ,
2
r5
(r − 2M)
|AB |22 = W T.
T
r
Acknowledgments
We are grateful to Steffen Aksteiner, Siyuan Ma, Marc Mars, and Claudio
Paganini for helpful remarks. Pieter Blue and Thomas Bäckdahl were sup-
ported by EPSRC grant EP/J011142/1. Lars Andersson thanks the Institut
Henri Poincaré, Paris, for hospitality and support during part of the work on
this chapter.
References
[1] B. P. Abbott et al. Observation of gravitational waves from a binary black hole
merger. Physical Review Letters, 116:061102, Feb. 2016.
[2] S. Aksteiner. Geometry and analysis in black hole spacetimes. PhD thesis,
Gottfried Wilhelm Leibniz Universität Hannover, 2014. http://d-nb.info/
1057896721/34.
[3] S. Aksteiner and L. Andersson. Linearized gravity and gauge conditions.
Classical and Quantum Gravity, 28(6):065001, Mar. 2011. arXiv.org:1009.5647.
[4] S. Aksteiner and L. Andersson. Charges for linearized gravity. Classical and
Quantum Gravity, 30(15):155016, Aug. 2013. arXiv.org:1301.2674.
[5] S. C. Anco and J. Pohjanpelto. Conserved currents of massless fields of spin
s ≥ 12 . Proceedings of the Royal Society of London. Series A, Mathematical and
Physical Sciences, 459(2033):1215–1239, 2003.
[6] S. C. Anco and J. Pohjanpelto. Symmetries and currents of massless neutrino
fields, electromagnetic and graviton fields. In Symmetry in physics, vol. 34
of CRM Proceedings and Lecture Notes, pp. 1–12. American Mathematical
Society, Providence, RI, 2004. math-ph/0306072.
[7] L. Andersson. The global existence problem in general relativity. In Piotr T.
Chrusciel and Helmut Friedrich, editors, The Einstein equations and the large
scale behavior of gravitational fields, pp. 71–120. Birkhäuser, Basel, 2004.
[8] L. Andersson, T. Bäckdahl, and P. Blue. A new tensorial conservation law for
Maxwell fields on the Kerr background. Dec. 2014. arXiv.org:1412.2960, to
appear in the Journal of Differential Geometry.
[9] L. Andersson, T. Bäckdahl, and P. Blue. Second order symmetry operators.
Classical and Quantum Gravity, 31(13):135015, Jul. 2014. arXiv.org:1402.6252.
[10] L. Andersson, T. Bäckdahl, and P. Blue. Spin geometry and conservation
laws in the Kerr spacetime. In L. Bieri and S.-T. Yau, editors, One hundred
years of general relativity, pp. 183–226. International Press, Boston, 2015.
arXiv.org:1504.02069.
[11] L. Andersson, T. Bäckdahl, and P. Blue. Decay of solutions to the Maxwell
equation on the Schwarzschild background. Classical and Quantum Gravity,
33(8):085010, 2016. arXiv.org:1501.04641.
[12] L. Andersson and P. Blue. Hidden symmetries and decay for the wave equation
on the Kerr spacetime. Annals of Mathematics (2), 182(3):787–853, 2015.
[13] L. Andersson and P. Blue. Uniform energy bound and asymptotics for the
Maxwell field on a slowly rotating Kerr black hole exterior. Journal of
Hyperbolic Differential Equations, 12(04):689–743, Oct. 2015.
[14] L. Andersson, P. Blue, and J.-P. Nicolas. A decay estimate for a wave equation
with trapping and a complex potential. International Mathematics Research
Notices, (3):548–561, 2013.
[15] L. Andersson, M. Mars, J. Metzger, and W. Simon. The time evolution of
marginally trapped surfaces. Classical and Quantum Gravity, 26(8):085018,
Apr. 2009.
[16] L. Andersson, M. Mars, and W. Simon. Local Existence of Dynamical and
Trapping Horizons. Physical Review Letters, 95(11):111102, Sep. 2005.
[17] L. Andersson and J. Metzger. The area of horizons and the trapped region.
Communications in Mathematical Physics, 290(3):941–972, 2009.
[18] L. Andersson and V. Moncrief. Einstein spaces as attractors for the Einstein
flow. Journal of Differential Geometry, 89(1):1–47, 2011.
[19] T. Bäckdahl and J. A. Valiente Kroon. Geometric invariant measuring the
deviation from Kerr Data. Physical Review Letters, 104(23):231102, Jun. 2010.
[20] T. Bäckdahl and J. A. Valiente Kroon. On the construction of a geometric
invariant measuring the deviation from Kerr data. Annals Henri Poincaré,
11(7):1225–1271, 2010.
[21] T. Bäckdahl and J. A. Valiente Kroon. The “non-Kerrness” of domains of outer
communication of black holes and exteriors of stars. Royal Society of London.
Series A, Mathematical and Physical Sciences, 467:1701–1718, Jun. 2011.
arXiv.org:1010.2421.
[22] T. Bäckdahl and J. A. Valiente Kroon. Constructing “non-Kerrness” on compact
domains. Journal of Mathematical Physics, 53(4):042503, Apr. 2012.
[23] R. A. Bartnik and P. T. Chruściel. Boundary value problems for Dirac-type
equations. Journal fur die reine und angewandte Mathematik, 579:13–73, 2005.
arXiv.org:math/0307278.
[24] R. Beig and P. T. Chruściel. Killing vectors in asymptotically flat space–times.
I. Asymptotically translational Killing vectors and the rigid positive energy
theorem. Journal of Mathematical Physics, 37:1939–1961, Apr. 1996.
[25] R. Beig and N. Ó Murchadha. The Poincaré group as the symmetry group of
canonical general relativity. Annals of Physics, 174(2):463–498, 1987.
[26] I. Bengtsson and J. M. M. Senovilla. Region with trapped surfaces in spherical
symmetry, its core, and their boundaries. Physical Review, 83(4):044012, Feb.
2011.
[27] G. Bergqvist. Positivity of General Superenergy Tensors. Communications in
Mathematical Physics, 207:467–479, 1999.
[28] A. N. Bernal and M. Sánchez. Further results on the smoothability of Cauchy
hypersurfaces and Cauchy time functions. Letters in Mathematical Physics,
77:183–197, Aug. 2006.
[29] D. Bini, C. Cherubini, R. T. Jantzen, and R. Ruffini. Teukolsky master
equation–de Rham wave equation for the gravitational and electromagnetic
fields in vacuum. Progress of Theoretical Physics, 107:967–992, May 2002.
arXiv.org:gr-qc/0203069.
[30] D. Bini, C. Cherubini, R. T. Jantzen, and R. Ruffini. De Rham wave equation
for tensor valued p-forms. International Journal of Modern Physics D,
12:1363–1384, 2003.
[31] H. L. Bray. Proof of the Riemannian Penrose inequality using the positive mass
theorem. Journal of Differential Geometry, 59(2):177–267, 2001.
[32] H. Buchdahl. On the compatibility of relativistic wave equations for particles
of higher spin in the presence of a gravitational field. Il Nuovo Cimento,
10(1):96–103, 1958.
[33] B. Carter. Killing tensor quantum numbers and conserved currents in curved
space. Physical Review, 16:3395–3414, Dec. 1977.
[34] S. Chandrasekhar. The mathematical theory of black holes, vol. 69 of

International Series of Monographs on Physics. The Clarendon Press, Oxford
University Press, New York, 1992. Revised reprint of the 1983 original, Oxford
Science Publications.
[35] Y. Choquet-Bruhat and R. Geroch. Global aspects of the Cauchy problem in
general relativity. Communications in Mathematical Physics, 14:329–335, Dec.
1969.
[36] D. Christodoulou. The Formation of Black Holes in General Relativity. European
Mathematical Society, Zurich, 2009.
[37] D. Christodoulou and S. Klainerman. The global nonlinear stability of
the Minkowski space, vol. 41 of Princeton Mathematical Series. Princeton
University Press, Princeton, NJ, 1993.
[38] D. Christodoulou and N. O’Murchadha. The boost problem in general relativity.
Communications in Mathematical Physics, 80:271–300, Jun. 1981.
[39] P. T. Chruściel, G. J. Galloway, and D. Solis. Topological censorship for
Kaluza–Klein space–times. Annales Henri Poincaré, 10:893–912, Jul. 2009.
[40] P. T. Chruściel, J. Jezierski, and J. Kijowski. Hamiltonian field theory in
the radiating regime, vol. 70 of Lecture Notes in Physics. Monographs.
Springer-Verlag, Berlin, 2002.
[41] C. J. S. Clarke. A condition for forming trapped surfaces. Classical Quantum
Gravity, 5(7):1029–1032, 1988.
[42] J. M. Cohen and L. S. Kegeles. Electromagnetic fields in curved spaces: A
constructive procedure. Physical Review, 10:1070–1084, Aug. 1974.
[43] C. D. Collinson and P. N. Smith. A comment on the symmetries of Kerr black
holes. Communications in Mathematical Physics, 56:277–279, Oct. 1977.
[44] R. M. Corless, G. H. Gonnet, D. E. G. Hare, D. J. Jeffrey, and D. E. Knuth. On the
LambertW function. Advances in Computational Mathematics, 5(1):329–359,
1996.
[45] M. Dafermos and I. Rodnianski. A proof of the uniform boundedness of
solutions to the wave equation on slowly rotating Kerr backgrounds. Inventiones
Mathematicae, 185(3):467–559, 2011.
[46] M. Dafermos, I. Rodnianski, and Y. Shlapentokh-Rothman. Decay for solutions
of the wave equation on Kerr exterior spacetimes III: The full subextremal case
|a| < M. Feb. 2014. arXiv.org:1402.7034.
[47] D. M. Eardley. Gravitational collapse of vacuum gravitational field
configurations. Journal of Mathematical Physics, 36(6):3004–3011, 1995.
[48] W. E. East, F. M. Ramazanoǧlu, and F. Pretorius. Black hole superradiance in
dynamical spacetime. Physical Review, 89(6):061503, Mar. 2014.
[49] S. B. Edgar, A. G.-P. Gómez-Lobo, and J. M. Martı́n-Garcı́a. Petrov D vacuum
spaces revisited: Identities and invariant classification. Classical and Quantum
Gravity, 26(10):105022, May 2009. arXiv.org:0812.1232.
[50] A. Einstein. Kosmologische Betrachtungen zur allgemeinen Relativitätstheorie.
Sitzungsberichte der Königlich Preußischen Akademie der Wissenschaften
(Berlin), Seite 142–152., 1917.
[51] J. J. Ferrando and J. A. Sáez. An intrinsic characterization of the Kerr metric.
Classical and Quantum Gravity, 26(7):075013, Apr. 2009. arXiv.org:0812.3310.
[52] F. Finster, N. Kamran, J. Smoller, and S.-T. Yau. A Rigorous Treatment

of Energy Extraction from a Rotating Black Hole. Communications in
Mathematical Physics, 287:829–847, May 2009.
[53] A. E. Fischer and V. Moncrief. The Einstein flow, the σ -constant and the
geometrization of 3-manifolds. Classical Quantum Gravity, 16(11):L79–L87,
1999.
[54] Y. Fourès-Bruhat. Théorème d’existence pour certains systèmes d’équations aux
dérivées partielles non linéaires. Acta Mathematica, 88(1): 141–225, 1952.
[55] H. Friedrich. Hyperbolic reductions for Einstein’s equations. Classical and
Quantum Gravity, 13:1451–1469, Jun. 1996.
[56] V. Frolov and A. Zelnikov. Introduction to Black Hole Physics. Oxford
University Press Oxford, 2011.
[57] J. R. Gair, C. Li, and I. Mandel. Observable properties of orbits in exact bumpy
spacetimes. Physical Review, 77(2):024035, Jan. 2008. arXiv.org:0708.0628.
[58] R. Geroch. Spinor structure of space–times in general relativity. II. Journal of
Mathematical Physics, 11(1):343–348, 1970.
[59] R. Geroch, A. Held, and R. Penrose. A space–time calculus based on pairs of
null directions. Journal of Mathematical Physics, 14:874–881, Jul. 1973.
[60] J. Haláček and T. Ledvinka. The analytic conformal compactification of the
Schwarzschild spacetime. Classical and Quantum Gravity, 31(1):015007, Jan.
2014.
[61] Q. Han and M. Khuri. The Conformal Flow of Metrics and the General Penrose
Inequality. Aug. 2014. arXiv.org:1409.0067.
[62] S. W. Hawking and G. F. R. Ellis. The large scale structure of space–time.
Cambridge University Press, London and New York, 1973. Cambridge
Monographs on Mathematical Physics, No. 1.
[63] L.-H. Huang. On the center of mass of isolated systems with general asymptotics.
Classical and Quantum Gravity, 26(1):015012, Jan. 2009.
[64] L. P. Hughston and P. Sommers. The symmetries of Kerr black holes.
Communications in Mathematical Physics, 33:129–133, Jun. 1973.
[65] G. Huisken and T. Ilmanen. The inverse mean curvature flow and the Riemannian
Penrose inequality. Journal of Differential Geometry, 59(3):353–437, 2001.
[66] V. Iyer and R. M. Wald. Some properties of the Noether charge and a proposal
for dynamical black hole entropy. Physical Review, 50:846–864, Jul. 1994.
[67] E. G. Kalnins, W. Miller, Jr., and G. C. Williams. Teukolsky-Starobinsky
identities for arbitrary spin. Journal of Mathematical Physics, 30:2925–2929,
Dec. 1989.
[68] R. P. Kerr. Gravitational field of a spinning mass as an example of algebraically
special metrics. Physical Review Letters, 11:237–238, Sep. 1963.
[69] W. Kinnersley. Type D Vacuum Metrics. Journal of Mathematical Physics,
10:1195–1203, July 1969.
[70] S. Klainerman, J. Luk, and I. Rodnianski. A fully anisotropic mechanism for
formation of trapped surfaces in vacuum. Feb. 2013. arXiv.org:1302.5951.
[71] S. Klainerman and I. Rodnianski. Rough solutions of the Einstein-vacuum
equations. Annals of Mathematics (2), 161(3):1143–1193, 2005.
[72] S. Klainerman, I. Rodnianski, and J. Szeftel. Overview of the proof of the
Bounded L2 Curvature Conjecture. Apr. 2012. arXiv.org:1204.1772.
[73] S. Klainerman, I. Rodnianski, and J. Szeftel. The bounded L2 curvature

conjecture. Inventiones Mathematicae, 202(1):91–216, 2015.
[74] J.-P. Lasota, E. Gourgoulhon, M. Abramowicz, A. Tchekhovskoy, and
R. Narayan. Extracting black-hole rotational energy: The generalized Penrose
process. Physical Review, 89(2):024041, Jan. 2014.
[75] A. László and I. Rácz. Superradiance or total reflection? Springer Proceedings
in Physics, 157:119–127, 2014. arXiv.org:1212.4847.
[76] H. Lindblad and I. Rodnianski. Global existence for the Einstein vacuum
equations in wave coordinates. Communications in Mathematical Physics,
256:43–110, May 2005. arXiv.org:math/0312479.
[77] C. O. Lousto and B. F. Whiting. Reconstruction of black hole metric perturba-
tions from the Weyl curvature. Physical Review, 66(2):024026, July 2002.
[78] J. Luk. Weak null singularities in general relativity. Nov. 2013.
arXiv.org:1311.4970.
[79] G. Lukes-Gerakopoulos, T. A. Apostolatos, and G. Contopoulos. Observable
signature of a background deviating from the Kerr metric. Physical Review,
81(12):124005, Jun. 2010. arXiv.org:1003.3120.
[80] M. Mars. Uniqueness properties of the Kerr metric. Classical and Quantum
Gravity, 17:3353–3373, Aug. 2000.
[81] M. Mars, T.-T. Paetz, J. M. M. Senovilla, and W. Simon. Characterization
of (asymptotically) Kerr–de Sitter-like spacetimes at null infinity. Mar. 2016.
arXiv.org:1603.05839.
[82] B. Michel. Geometric invariance of mass-like asymptotic invariants. Journal of
Mathematical Physics, 52(5):052504–052504, May 2011.
[83] J.-P. Michel, F. Radoux, and J. Šilhan. Second Order Symmetries of the
Conformal Laplacian. SIGMA, 10:016, Feb. 2014.
[84] C. W. Misner, K. S. Thorne, and J. A. Wheeler. Gravitation. W. H. Freeman and
Co., San Francisco, CA, 1973.
[85] C. S. Morawetz. Time decay for the nonlinear Klein–Gordon equations.
Proceedings of the Royal Society of London. Series A, Mathematical and
Physical Sciences, 306:291–296, 1968.
[86] P. Mösta, L. Andersson, J. Metzger, B. Szilágyi, and J. Winicour. The merger
of small and large black holes. Classical and Quantum Gravity, 32(23):235003,
Dec. 2015.
[87] E. T. Newman, E. Couch, K. Chinnapared, A. Exton, A. Prakash, and
R. Torrence. Metric of a rotating, charged mass. Journal of Mathematical
Physics, 6:918–919, Jun. 1965.
[88] E. T. Newman and A. I. Janis. Note on the Kerr spinning-particle metric. Journal
of Mathematical Physics, 6:915–917, Jun. 1965.
[89] I. D. Novikov and V. P. Frolov. Physics of black holes. (Fizika chernykh dyr,
Moscow, Izdatel’stvo Nauka, 1986) Dordrecht, Netherlands, Kluwer Academic
Publishers, 1989. Translation. Previously cited in issue 19, p. 3128, Accession
no. A87-44677., 1989.
[90] B. O’Neill. The geometry of Kerr black holes. A K Peters Ltd., Wellesley, MA,
1995.
[91] J. R. Oppenheimer and H. Snyder. On continued gravitational contraction.
Physical Review, 56:455–459, Sep. 1939.
[92] D. Parlongue. Geometric uniqueness for non-vacuum Einstein equations and

applications. Sep. 2011. arXiv.org:1109.0644.
[93] R. Penrose. Zero rest-mass fields including gravitation: Asymptotic behaviour.
Proceedings of the Royal Society of London. Series A, Mathematical and
Physical Sciences, 284:159–203, 1965.
[94] R. Penrose and W. Rindler. Spinors and space–time I & II. Cambridge Mono-
graphs on Mathematical Physics. Cambridge University Press, Cambridge, 1986.
[95] Z. Perjés and Á. Lukács. Canonical quantization and black hole perturbations. In
J. Lukierski and D. Sorokin, editors, Fundamental interactions and twistor-Like
methods, volume 767 of American Institute of Physics Conference Series,
pp. 306–315, Apr. 2005.
[96] E. Poisson. A relativist’s toolkit. Cambridge University Press, Cambridge, 2004.
The mathematics of black-hole mechanics.
[97] T. Regge and C. Teitelboim. Role of surface integrals in the Hamiltonian
formulation of general relativity. Annals of Physics, 88:286–318, Nov. 1974.
[98] H. Ringström. Cosmic censorship for Gowdy spacetimes. Living Reviews in
Relativity, 13(2), 2010.
[99] H. Ringström. Origins and development of the Cauchy problem in general
relativity. Classical and Quantum Gravity, 32(12):124003, Jun. 2015.
[100] M. P. Ryan. Teukolsky equation and Penrose wave equation. Physical Review,
10:1736–1740, Sep. 1974.
[101] J. Sbierski. On the existence of a maximal Cauchy development for the Einstein
equations: A dezornification. Annales Henri Poincaré, 17(2):301–329, 2016.
[102] R. Schoen and S. T. Yau. The existence of a black hole due to condensation of
matter. Communications in Mathematical Physics, 90(4):575–579, 1983.
[103] J. M. M. Senovilla. Super-energy tensors. Classical and Quantum Gravity,
17:2799–2841, Jul. 2000.
[104] C. D. Sogge. Lectures on non-linear wave equations. International Press,
Boston, MA, second edition, 2008.
[105] E. Teo. Spherical photon orbits around a Kerr black hole. General Relativity
Gravitation, 35(11):1909–1926, 2003.
[106] S. A. Teukolsky. Rotating black holes: separable wave equations for gravitational
and electromagnetic perturbations. Physical Review Letters, 29:1114–1118, Oct.
1972.
[107] S. A. Teukolsky. Perturbations of a rotating black hole. I. Fundamental
equations for gravitational, electromagnetic, and neutrino-field perturbations.
Astrophysical Journal, 185:635–648, Oct. 1973.
[108] S. A. Teukolsky. The Kerr metric. Classical and Quantum Gravity,
32(12):124006, Jun. 2015.
[109] R. M. Wald. General relativity. University of Chicago Press, Chicago, IL, 1984.
[110] R. M. Wald. Gravitational collapse and cosmic censorship. Oct. 1997.
arXiv.org:gr-qc/9710068.
[111] R. M. Wald and V. Iyer. Trapped surfaces in the Schwarzschild geometry and
cosmic censorship. Physical Review, 44:R3719–R3722, Dec. 1991.
[112] M. Walker and R. Penrose. On quadratic first integrals of the geodesic
equations for type {2,2} spacetimes. Communications in Mathematical Physics,
18:265–274, Dec. 1970.
[113] J. Winicour and L. Tamburino. Lorentz-covariant gravitational energy-

momentum linkages. Physical Review Letters, 15:601–605, Oct. 1965.
[114] W. W.-Y. Wong. A comment on the construction of the maximal globally
hyperbolic Cauchy development. Journal of Mathematical Physics,
54(11):113511–113511, Nov. 2013.
[115] S. T. Yau. Geometry of three manifolds and existence of black hole due
to boundary effect. Advances in Theoretical and Mathematical Physics,
5(4):755–767, 2001.
[116] R. L. Znajek. Black hole electrodynamics and the Carter tetrad. Monthly Notices
of the Royal Astronomical Society, 179:457–472, May 1977.
[117] P. Hintz and A. Vasy. The global non-linear stability of the Kerr–de Sitter family
of black holes. Preprint, arXiv:1606.04014, 2016.
Albert Einstein Institute, Am Mühlenberg 1, D-14476 Potsdam, Germany

Mathematical Sciences, Chalmers University of Technology and University of
Gothenburg, SE-412 96 Gothenburg, Sweden and The School of Mathematics,
University of Edinburgh, James Clerk Maxwell Building, Peter Guthrie Tait Road,
Edinburgh EH9 3FD, UK
The School of Mathematics and the Maxwell Institute, University of Edinburgh, James
Clerk Maxwell Building, Peter Guthrie Tait Road, Edinburgh EH9 3FD,UK
3
An Introduction to Conformal Geometry
and Tractor Calculus, with a view to
Applications in General Relativity
Sean N. Curry and A. Rod Gover
Abstract. The chapter consists of expanded notes for the course of eight
one-hour lectures given by the second author at the 2014 summer school
entitled Asymptotic Analysis in General Relativity held in Grenoble by
the Institut Fourier. The first four lectures deal with conformal geometry
and the conformal tractor calculus, taking as primary motivation the search
for conformally invariant tensors and differential operators. The final four
lectures apply the conformal tractor calculus to the study of conformally
compactified geometries, motivated by the conformal treatment of infinity
in general relativity.
3.1. Introduction
Definition 3.1 A conformal n-manifold (n ≥ 3) is the structure (M, c) where
• M is an n-manifold,
• c is a conformal equivalence class of signature (p, q) metrics,
def.
g = 2 g and C∞ (M) > 0.
g ∈ c ⇐⇒
that is g,
To any pseudo-Riemannian n-manifold (M, g) with n ≥ 3 there is an
associated conformal manifold (M, [g]) where [g] is the set of all metrics ĝ
which are smooth positive multiples of the metric g. In Riemannian signature
(p, q) = (n, 0) passing to the conformal manifold means geometrically that
we are forgetting the notion of lengths (of tangent vectors and of curves)
and retaining only the notion of angles (between tangent vectors and curves)
and of ratios of lengths (of tangent vectors at a fixed point) associated to
2010 Mathematics Subject Classification. Primary 53A30, 35Q75, 53B15, 53C25; Secondary
83C05, 35Q76, 53C29.
Key words and phrases. Einstein metrics, conformal differential geometry, conformal infinity.
ARG gratefully acknowledges support from the Royal Society of New Zealand via Marsden
Grant 13-UOA-018.
86
Conformal Geometry and Tractor Calculus 87
the metric g. In Lorentzian signature (n − 1, 1) passing to the conformal

manifold means forgetting the ‘spacetime interval’ (analogous to length in
Riemannian signature) and retaining only the light cone structure of the
Lorentzian manifold. On a conformal Lorentzian manifold one also has the
notion of angles between intersecting spacelike curves and the notion of
orthogonality of tangent vectors at a point, but the conformal structure itself
is determined by the light cone structure, justifying the use of the word ‘only’
in the previous sentence.
The significance of conformal geometry for general relativity largely stems
from the fact that the light cone structure determines the causal structure of
spacetime (and, under some mild assumptions, the causal structure determines
the light cone structure). On top of this we shall see in the lectures that the
Einstein field equations admit a very nice interpretation in terms of conformal
geometry.
In these notes we will develop the natural invariant calculus on conformal
manifolds, the (conformal) tractor calculus, and apply it to the study of
conformal invariants and conformally compactified geometries. The course is
divided into two parts consisting of four lectures each. The first four lectures
deal with conformal geometry and the conformal tractor calculus, taking as
primary motivation the problem of constructing conformally invariant tensors
and differential operators. The tools developed for this problem however allow
us to tackle much more than our original problem of invariants and invariant
operators. In the final four lectures they will be applied in particular to the study
of conformally compactified geometries, motivated by the conformal treatment
of infinity in general relativity. Along the way we establish the connection
between the conformal tractor calculus and Helmut Friedrich’s conformal field
equations. We also digress for one lecture, discussing conformal hypersurface
geometry, in order to facilitate the study of the relationship of the geometry
of conformal infinity to that of the interior. Finally we show how the tractor
calculus may be applied to treat aspects of the asymptotic analysis of boundary
problems on conformally compact manifolds. For completeness an appendix
has been added which covers further aspects of the conformal tractor calculus
as well as discussing briefly the canonical conformal Cartan bundle and
connection.
The broad philosophy behind our ensuing discussion is that conformal
geometry is important not only for understanding conformal manifolds, or
conformally invariant aspects of pseudo-Riemannian geometry (such as con-
formally invariant field equations), but that it is highly profitable to think of
a pseudo-Riemannian manifold as a kind of symmetry breaking (linked to
holonomy reduction) of a conformal manifold whenever there are any (even
88 S. Curry and R. Gover
remotely) conformal geometry related aspects of the problem being consid-

ered. Our discussion of the conformal tractor calculus will lead us naturally
to the notions of almost Einstein and almost pseudo-Riemannian geometries,
which include Einstein and pseudo-Riemannian manifolds (respectively) as
well as their respective conformal compactifications (should they admit one).
The general theory of Cartan holonomy reductions then enables us to put
constraints on the smooth structure and the geometry of conformal infinities of
Einstein manifolds, and the tractor calculus enables us to generalise partially
these results to pseudo-Riemannian manifolds.
We deal exclusively with conformal manifolds of at least three dimensions
in these notes. That is not to say that two-dimensional conformal manifolds
cannot be fitted into the framework which we describe. However, in order to
have a canonical tractor calculus on two-dimensional conformal manifolds the
conformal manifold needs to be equipped with an extra structure, weaker than a
Riemannian structure but stronger than a conformal structure, called a Möbius
structure. In higher dimensions a conformal structure determines a canonical
Möbius structure via the construction of the canonical conformal Cartan bundle
and connection (outlined in Section A.1). In two dimensions there is no
canonical Cartan bundle and connection associated with a conformal manifold,
so this (Möbius) structure must be imposed as an additional assumption if
we wish to work with the tractor calculus. We note that to any Riemannian
2-manifold, or to any (non-degenerate) two-dimensional submanifold of a
higher dimensional conformal manifold, there is associated a canonical Möbius
structure and corresponding tractor calculus.
Also left out in these notes is any discussion of conformal spin geometry.
In this case there is again a canonical tractor calculus, known as spin tractor
calculus or local twistor calculus, which is a refinement of the usual conformal
tractor calculus in the same way that spinor calculus is a refinement of the usual
tensor calculus on pseudo-Riemannian spin manifolds. The interested reader is
referred to [4, 51].
3.1.1. Notation and Conventions

We may use abstract indices, or no indices, or frame indices according to con-
venience. However, we will make particularly heavy use of the abstract index
notation. For instance if L is a linear endomorphism of a finite dimensional
vector space V then we may choose to write L using abstract indices as La b

(or Lb c , or La b , it makes no difference as the indices are just place holders
meant to indicate tensor type, and contractions). In this case we would write

a vector v ∈ V as va (or vb , or va , ...) and we would write the action of L on
v as La b vb (repeated indices denote tensor contraction so La b vb simply means

L(v)). Similarly if w ∈ V ∗ then using abstract indices we would write w as wa
(or wb , or wa , ...) and w(v) as wa va , whereas the outer product v ⊗ w ∈ End(V)
would be written as va wb (or vb wc , or wb va , ...). A covariant 2-tensor T ∈ ⊗2 V ∗
may be written using abstract indices as Tab , the symmetric part of T is then
denoted by T(ab) and the antisymmetric part by T[ab] , that is
1 1
T(ab) = (Tab + Tba ) and T[ab] = (Tab − Tba ) .
2 2
Note that swapping the indices a and b in Tab amounts to swapping the ‘slots’
of the covariant 2-tensor (so that b becomes the label for the first slot and
a for the second), which gives in general a different covariant 2-tensor from
Tab whose matrix with respect to any basis for V would be the transpose
of that of Tab . We can similarly define the symmetric or antisymmetric
part of any covariant tensor Tab···e , and these are denoted by T(ab···e) and
T[ab···e] respectively. We use the same bracket notation for the symmetric and
antisymmetric parts of contravariant tensors. Note that we do not have to
symmetrise or skew-symmetrise over all indices, for instance T a[bc] d denotes
1 abc
T d − T acb d .
2
The abstract index notation carries over in the obvious way to vector and
tensor fields on a manifold. The virtue of using abstract index notation on
manifolds is that it makes immediately apparent the type of tensorial object
one is dealing with and its symmetries without having to bring in extraneous
vector fields or 1-forms. We will commonly denote the tangent bundle of M
by E a , and the cotangent bundle of M by Ea . We then denote the bundle of
covariant 2-tensors by Eab , its subbundle of symmetric 2-tensors by E(ab) , and
so on. In order to avoid confusion, when working on the tangent bundle of a
manifold M we will always use lower case Latin abstract indices taken from
the beginning of the alphabet (a, b, etc.) whereas we will take our frame indices
from a later part of the alphabet (starting from i, j, etc.).
It is common when working with abstract index notation to use the same
notation E a for the tangent bundle and its space of smooth sections. Here
however we have used the notation (V) for the space of smooth sections of
a vector bundle V consistently throughout, with the one exception that for a
differential operator D taking sections of a vector bundle U to sections of V we
have written
D:U →V
in order to simplify the notation.
Consistent with our use of E a for TM we will often denote by E the trivial
R-bundle over our manifold M, so that (E) = C∞ (M). When using index
free notation we denote the space of vector fields on M by X(M), and we use
the shorthand k for k T ∗ M (when the underlying manifold M is understood).
Unless otherwise indicated [ · , · ] is the commutator bracket acting on pairs of
endomorphisms. Note that the Lie bracket arises in this way when we consider
vector fields as derivations of the algebra of smooth functions. In all of the
following all structures (manifolds, bundles, tensor fields, etc.) will be assumed
smooth, meaning C∞ .
3.1.1.1. Coupled Connections

We assume that the reader is familiar with the notion of a linear connection on
a vector bundle V → M and the special case of an affine connection (being a
connection on the tangent bundle of a manifold). Given a pair of vector bundles
V and V over the same manifold M and linear connections ∇ and ∇ defined
on V and V respectively, then there is a natural way to define a connection on
the tensor product bundle V ⊗ V → M. The coupled connection ∇ ⊗ on V ⊗ V
is given on a simple section v ⊗ v of V ⊗ V by the Leibniz formula
∇X⊗ (v ⊗ v ) = (∇X v) ⊗ v + v ⊗ (∇X v )
for any X ∈ X(M). Since (V ⊗ V ) is (locally) generated by simple sections,
this formula determines the connection ∇ ⊗ uniquely. In order to avoid clumsy
notation we will often simply write all of our connections as ∇ when it is clear
from the context which (possibly coupled) connection is being used.
3.1.1.2. Associated Bundles

We assume also that the reader is familiar with the notion of a vector bundle on
a smooth manifold M as well as that of an H-principal bundle over M, where
H is a Lie group. If π : G → M is a (right) H-principal bundle and V is
a (finite dimensional) representation of H then the associated vector bundle
G ×H V → M is the vector bundle with total space defined by
G ×H V = G × V/ ∼
where ∼ is the equivalence relation
(u, v) ∼ (u · h, h−1 · v), h∈H
on G × V; the projection of G ×H V to M is simply defined by taking [(u, v)]
to π(u). For example if F is the linear frame bundle of a smooth n-manifold
M then
TM = F ×GL(n) Rn and T ∗ M = F ×GL(n) (Rn )∗ .
Similarly if H is contained in a larger Lie group G then one can extend any
principal H-bundle G → M to a principal G-bundle G̃ → M with total space
G̃ = G ×H G = G × G/ ∼ where (u, g) ∼ (u · h, h−1 · g).
For example if (M, g) is a Riemannian n-manifold and O denotes its orthonor-

mal frame bundle then
F = O ×O(n) GL(n)
is the linear frame bundle of M.
3.2. Lecture 1: Riemannian Invariants

and Invariant Operators
Recall that if ∇ is an affine connection then its torsion is the tensor field
T ∇ ∈ (TM ⊗ 2 T ∗ M) defined by
T ∇ (u, v) = ∇u v − ∇v u − [u, v] for all u, v ∈ X(M).
It is interesting that this is a tensor; by its construction one might expect a

differential operator. Importantly torsion is an invariant of connections. On a
smooth manifold the map
∇ → T ∇
taking connections to their torsion depends only the smooth structure. By its
construction here it is clear that it is independent of any choice of coordinates.
For any connection ∇ on a vector bundle V its curvature is defined by
R∇ (u, v)W := [∇u , ∇v ]W − ∇[u,v] W for all u, v ∈ X(M), and W ∈ (V),
and R∇ ∈ (2 T ∗ M ⊗ End(V)). Again the map taking connections to their

curvatures
∇ → R∇
clearly depends only on the (smooth) vector bundle structure of V over the
smooth manifold M. The curvature is an invariant of linear connections. In
particular this applies to affine connections, that is when V = TM.
3.2.1. Ricci Calculus and Weyl’s Invariant Theory

The most familiar setting for these objects is pseudo-Riemannian geometry. In
this case we obtain a beautiful local calculus that is sometimes called the Ricci
calculus. Let us briefly recall how this works. Recall that a pseudo-Riemannian
manifold consists of an n-manifold M equipped with a metric g of signature
(p, q), that is a section g ∈ (S2 T ∗ M) such that pointwise g is non-degenerate
and of signature (p, q). Then g canonically determines a distinguished affine
connection called the Levi-Civita connection. This is the unique connection ∇
satisfying:
• ∇g = 0 (metric compatibility) and
• T∇ = 0 (torsion freeness).
Thus on a smooth manifold we have a canonical map from each metric to its
Levi-Civita connection
g → ∇ g
and, as above, a canonical map which takes each Levi-Civita connection to its
curvature ∇ g → R∇ , called the Riemannian curvature. Composing these we
g
get a canonical map that takes each metric to its curvature

g −→ Rg ,
and this map depends only on the smooth structure of M. So we say that Rg is
an invariant of the pseudo-Riemannian manifold (M, g). How can we construct
more such invariants? Or ‘all’ invariants, in some perhaps restricted sense?
The first, and perhaps most important, observation is that using the Levi-
Civita connection and Riemannian curvature one can proliferate Riemannian
invariants. To simplify the explanation let’s fix a pseudo-Riemannian manifold
(M, g) and use abstract index notation (when convenient). Then the metric is
written gab and we write Rab c d for the Riemannian curvature. So if u, v, w are
tangent vector fields then so is R(u, v)w and this is written
ua vb Rab c d wd .
From curvature we can form the Ricci and Scalar curvatures, respectively:
Ricab := Rca c b and Sc := gab Ricab ,
and these are invariants of (M, g). As also are:
∇b Rcdef , ∇a ∇b Rcdef ,
which are tensor valued invariants and
Ricab Ricab Sc = | Ric |2 Sc, Rabcd Rabcd = |R|2 , (∇a Rbcde )∇ a Rbcde = |∇R|2 ,
which are some scalar valued invariants. Here we have used the metric (and its
inverse) to raise and lower indices, and contractions are indicated by repeated
indices.
Since this is a practical and efficient way to construct invariants, it would

be useful to know: Do all local Riemannian invariants arise in this way? That
is from partial or complete contractions of expressions made using g, R and
its covariant derivatives ∇ · · · ∇R (and the metric volume form volg , if M is
oriented). We shall term invariants constructed this way Weyl invariants.
Before answering this, one first needs to be careful about what is meant by a
local invariant. For example, the following is a reasonable definition for scalar
invariants.
Definition 3.2 A scalar Riemannian invariant P is a function which assigns to

each pseudo-Riemannian n-manifold (M, g) a function P(g) such that:
(i) P(g) is natural, in the sense that for any diffeomorphism φ : M → M we
have P(φ ∗ g) = φ ∗ P(g).
(ii) P is given by a universal polynomial expression of the nature that, given
a local coordinate system (xi ) on (M, g), P(g) is given by a polynomial in
the variables gmn , ∂i1 gmn , · · · , ∂i1 ∂i2 · · · ∂ik gmn , (det g)−1 , for some positive
integer k.
Then with this definition, and a corresponding definition for tensor-valued
invariants, it is true that all local invariants arise as Weyl invariants, and
the result goes by the name of Weyl’s classical invariant theory, see for
example [1, 53]. Given this result, in the following when we mention pseudo-
Riemannian invariants we will mean Weyl invariants.
3.2.2. Invariant Operators, and Analysis

In a similar way we can use the Ricci calculus to construct invariant differ-
ential operators on pseudo-Riemannian manifolds. For example the (Bochner)
Laplacian is given by the formula
∇ a ∇a = : E −→ E,
in terms of the Levi-Civita connection ∇. There are also obvious ways to make
operators with curvature in coefficients, for example
Racbd ∇c ∇d : E −→ E(ab) .
With suitable restrictions imposed, in analogy with the case of invariants,

one can make the statement that all local natural invariant differential operators
arise in this way. It is beyond our current scope to make this precise;
suffice to say that when we discuss invariant differential operators on pseudo-
Riemannian manifolds we will again mean operators constructed in this way.
Remark 3.1 If a manifold has a spin structure, then essentially the above is
still true but there is a further ingredient involved, namely the Clifford product.
This allows the construction of important operators such as the Dirac operator.
The main point here is that the ‘Ricci calculus’ provides an effective and
geometrically transparent route to the construction of invariants and invariant
operators.
Invariants and invariant operators are the basic objects underlying the first
steps (and often significantly more than just the first steps) of treating problems
in general relativity and, more generally, in:
• The global analysis of manifolds;
• The study and application of geometric PDE;
• Riemannian spectral theory;
• Physics and mathematical physics.
Furthermore from a purely theoretical point of view, we cannot claim to
understand a geometry if we do not have a good theory of local invariants
and invariant operators.
3.3. Lecture 2: Conformal Transformations

and Conformal Covariance
A good theory of conformal geometry should provide some hope of treating
the following closely related problems.
Problem 1 Describe a practical way to generate/construct (possibly all) local
natural invariants of a conformal structure.
Problem 2 Describe a practical way to generate/construct (possibly all) natural
linear differential operators that are canonical and well-defined on (i.e. are
invariants of) a conformal structure.
We have not attempted to be precise in these statements, since here they are
mainly for the purpose of motivation. Let us first approach these naı̈vely.
3.3.1. Conformal Transformations

Recall that for any metric g we can associate its Levi-Civita connection ∇.
Let ei = ∂x∂ i = ∂i be a local coordinate frame and Ei its dual. Locally, any
connection is determined by how it acts on a frame field. For the Levi-Civita
connection the resulting connection coefficients
ijk := Ei (∇k ej ), where ∇k := ∇ek ,
are often called the Christoffel symbols, and are given by the Koszul formula:
1 il
ijk := g glj,k + glk,j − gjk,l
2
where gij = g(ei , ej ) and glj,k = ∂k glj .
Using this formula for the Christoffel symbols we can easily compute the
transformation formula for ∇ under a conformal transformation g →
g = 2 g.
−1
Let ϒa := ∇a , v ∈ (E ) and ωb ∈ (Eb ). Then we have:
a a
∇
a v = ∇a v + ϒa v − ϒ va + ϒ vc δa ,
g b b b b c b
(3.1)
∇
a ωb = ∇a ωb − ϒa ωb − ϒb ωa + ϒ ωc gab .
g c
(3.2)
For ωb ∈ (Eb ) we have from (3.2) that
ωb → ∇a ωb − ∇b ωa
is conformally invariant. But this is just the exterior derivative ω → dω, its
conformal invariance is better seen from the fact that it is defined on a smooth
manifold without further structure: for u, v ∈ X(M)
dω(u, v) = uω(v) − vω(u) − ω([u, v]).
Inspecting (3.2) it is evident that this is the only first-order conformally

invariant linear differential operator on T ∗ M that takes values in an irreducible
bundle.
The Levi-Civita connection acts also on other tensor bundles. We can use the
formulae (3.1) and (3.2) along with the product rule to compute the conformal
transformation of the result. For example for a simple covariant 2-tensor
ub ⊗ wc we have a (ub ⊗ wc ) = (∇
∇ a ub ) ⊗ wc + ub ⊗ (∇
a wc ).
Thus we can compute ∇ a (ub ⊗ wc ) by using (3.2) for each term on the right
hand side. But locally any covariant 2-tensor is a linear combination of simple
2-tensors and so we conclude that for a covariant 2-tensor Fbc
a Fbc = ∇a Fbc − 2ϒa Fbc − ϒb Fac − ϒc Fba + ϒ d Fdc gab + ϒ d Fbd gac . (3.3)
∇
By the obvious extension of this idea one quickly calculates the formula for the
conformal transformation for the Levi-Civita covariant derivative of an (r, s)-
tensor.
From the formula (3.3) we see that the completely skew part ∇[a Fbc] is
conformally invariant. In the case F is skew, in that Fbc = −Fcb , this recovers
that dF is conformally invariant. A more interesting observation arises with the

divergence ∇ b Fbc . We have
b Fbc =
∇ a Fbc ,
gab ∇
gab = −2 gab . Thus we obtain
and

b Fbc = −2 gab ∇a Fbc − 2ϒa Fbc − ϒb Fac − ϒc Fba + ϒ d Fdc gab + ϒ d Fbd gac
∇

= −2 ∇ b Fbc + (n − 3)ϒ d Fdc + ϒ d Fcd − ϒc Fb b .
In particular then, if F is skew then Fb b = 0 and we have simply

∇b Fbc = −2 ∇ b Fbc + (n − 4)ϒ d Fdc . (3.4)
So we see that something special happens in dimension 4. Combining with our
earlier observation we have the following result.
Proposition 3.1 In dimension 4 the differential operators
Div : 2 → 1 and Max : 1 → 1
given by
Fbc → ∇ b Fbc and uc → ∇ b ∇[b uc]
respectively, are conformally covariant, in that
b Fbc = −2 ∇ b Fbc
∇ and ∇ [b uc] = −2 ∇ b ∇[b uc] .
b ∇ (3.5)
The non-zero powers of (precisely −2 ) appearing in (3.5) mean that
these objects are only covariant rather than invariant. Conformal covariance is
still a strong symmetry property however, as we shall see. Before we discuss
that in more detail note that, for the equations, these factors of make little
difference:
∇ b Fbc = 0 ⇔ ∇ b Fbc = 0
and
∇ b ∇[b uc] = 0 ⇔ ∇ [b uc] = 0.
b ∇
In this sense these equations are conformally invariant.

Remark 3.2 In fact these equations are rather important. If we add the
condition that F is closed then on Lorentzian signature 4-manifolds the system
dF = 0 and Div(F) = 0
is the field strength formulation of the source-free Maxwell equations of
electromagnetism. The locally equivalent equations Div(du) = 0 give the
potential formulation of the (source-free) Maxwell equations. The conformal
invariance of these has been important in physics.
Returning to our search for conformally covariant operators and equations,

our preliminary investigation suggests that such things might be rather rare.
From (3.3) we see that the divergence of a 2-form is not conformally covariant
except in dimension 4. In fact, in contrast to what this might suggest, there
is a rich theory of conformally covariant operators. However there are some
subtleties involved. Before we come to this it will be useful to examine how
conformal rescaling affects the curvature.
3.3.2. Conformal Rescaling and Curvature

Using (3.1), (3.2) and the observations following these we can compute, for
example, the conformal transformation formulae for the Riemannian curvature
and its covariant derivatives and so forth. At the very lowest orders this
provides a tractable approach to finding conformal invariants.
Let us fix a metric g. With respect to metric traces, we can decompose the
curvature tensor of g into a trace-free part and a trace part in the following way:
Rabcd = Wabcd + 2gc[a Pb]d + 2gd[b Pa]c .

trace-free trace part
Here Pab , so defined, is called the Schouten tensor, while the tensor Wab c d is
called the Weyl tensor. In dimensions n ≥ 3 we have Ricab = (n−2)Pab +Jgab ,
where J := gab Pab . So the Schouten tensor Pab is a trace modification of the
Ricci tensor.
Exercise 1 Prove using (3.1) that under a conformal transformation g →
g=
2
g, as above, the Weyl and Schouten tensors transform as follows:

g c g c
W ab d =W ab d
and

g 1
P ab = Pab − ∇a ϒb + ϒa ϒb − gab ϒc ϒ c . (3.6)
2
Thus the Weyl curvature is a conformal invariant, while objects such as
|W|2 := Wabcd W abcd may be called conformal covariants because under
the conformal change they pick up a power of the conformal factor |W |2 =
−4
|W| (where for simplicity we are hatting the symbol for the object rather
2
than the metric). We will see shortly that such objects correspond to invariants.
Here we are defining conformal invariants to be Riemannian invariants
that have the additional property of being unchanged under conformal
transformation.
Exercise 2 Consider computing the conformal transformation of the deriva-

tives of the curvature: ∇R, ∇∇R, and so on then possibly using this and
an undetermined coefficient approach to finding conformal invariants or
conformal covariants. This rapidly gets intractable.
3.3.3. Conformally Invariant Linear Differential Operators

We may try the corresponding approach for constructing further conformally
invariant linear differential operators:
Dg : U → V (3.7)
such that Dg = Dg . Namely, first consider the possible Riemannian invariant
linear differential operators between the bundles concerned and satisfying
some order constraint. Next compute their transformation under a conformal
change. Finally seek a linear combination of these that forms a conformally
invariant operator. For our purposes this also defines what we mean by
conformally invariant linear differential operator. For many applications we
require that the domain and target bundles U and V are irreducible.
It turns out that irreducible tensor (or even spinor) bundles are not sufficient
to deal with conformal operators. Let us see a first glimpse of this by
recalling the construction of one of the most well known conformally invariant
differential operators.
3.3.3.1. The Conformal Wave Operator

For analysis on pseudo-Riemannian manifolds (M, g) the Laplacian is an
extremely important operator. The Laplacian on functions, which is also
called the Laplace–Beltrami operator, is given by
:= ∇ a ∇a : E → E,
where ∇ is the Levi-Civita connection for g.
Let us see how this behaves under conformal rescaling. For a function f ,
∇a f is simply the exterior derivative of f and so is conformally invariant. So to
compute the Laplacian for g = 2 g we need only use (3.2) with u := df :
a ∇a f
g f = ∇
= b ∇a f
gab ∇
−2 ab

= g ∇b ∇a f − ϒb ∇a f − ϒa ∇b f + gab ϒ c ∇c f .
So

ĝ f = −2 f + (n − 2)ϒ c ∇c f . (3.8)
By inspecting this formula we learn two things. Let us summarise with a

proposition.
Proposition 3.2 The Laplacian on functions is conformally covariant in
dimension 2, but not in other dimensions.
This is reminiscent of our observations surrounding the expression (3.4) and
the Maxwell system (cf. Proposition 3.1).
We need to modify the Laplacian to have any hope of obtaining an invariant
operator in other dimensions. A key idea will be to introduce a curvature
into the formula for a new Laplacian. However, inspecting the formulae for
curvature transformation in Section 3.3.2 it is easily seen that this manoeuvre
alone will not deal with the term ϒ c ∇c f in (3.8).
To eliminate the term (n − 2)ϒ c ∇c f we will make what seems like a strange
move (and we will explain later the mathematics behind this). We will allow
the domain ‘function’ to depend on the choice of metric in the following way.
We decree that the function f on (M, g) corresponds to
n
f := 1− 2 f on (M, g),
where g = 2 g as above. Now let us calculate ĝ f̂ . First we have:
n
n 1− n n
∇a (1− 2 f ) = 1 − 2 ϒa f + 1− 2 df
2 n
1− n2
= ∇a f + 1 − ϒa f .
2
We use this in the next step:

g b 1− 2 ∇a f + 1 − n ϒa f .
n
f = −2 gab ∇
2
−1− n2 ab n n
= g ∇b ∇a f + 1 − ϒa ∇b f + 1 − ϒb ∇a f
2 2
n 2 n
+ 1− ϒa ϒb f + 1 − f ∇b ϒa − ϒb ∇a f
2 2
n n
− 1− ϒb ϒa f − ϒa ∇b f − 1 − ϒa ϒb f
2 n 2
+gab ϒ c ∇c f + 1 − ϒ 2f
2
n n a n
= −1− 2 f + 1 − ∇ ϒa + ϒ 2 −1 f ,
2 2
where we have used (3.2) once again and have written ϒ 2 as a shorthand for
ϒ a ϒa .
Now in the last line of the formula for g f the terms involving ϒ do not
involve derivatives of f . Thus there is hope of matching this with a curvature
transformation. Indeed contracting (3.6) with ĝ−1 gives (with Jg := gab Pab )
g
n 2
Jg = −2 J − ∇ a ϒa + 1 − ϒ , (3.9)
2
and so
n g n
n
g + 1 − J f = −1− 2 + 1 − J f, (3.10)
2 2
and we have found a Laplacian operator with a symmetry under conformal
rescaling. Using the relation between J and the scalar curvature this is written
as in the definition here.
Definition 3.3 On a pseudo-Riemannian manifold (M, g) the operator
n−2
Yg : E → E defined by Y g := g − Scg
4(n − 1)
is called the conformal Laplacian or, in Lorentzian signature, the conformal
wave operator.
Remark 3.3 On Lorentzian signature manifolds it is often called the confor-
mal wave operator because the leading term agrees with the operator giving
the usual wave equation. It seems that it was in this setting that the operator
was first discovered and applied [54, 16]. On the other hand in the setting of
Riemannian signature Y is often called the Yamabe operator because of its role
in the Yamabe problem of scalar curvature prescription.
According to our calculations above this has the following remarkable symme-
try property with respect to conformal rescaling.
Proposition 3.3 The conformal Laplacian is conformally covariant in the
sense that
n n
Yg ◦ 1− 2 = −1− 2 ◦ Y g .
This property of the conformal Laplacian motivates a definition.
Definition 3.4 On pseudo-Riemannian manifolds a natural linear differential
operator Pg , on a function or tensor/spinor field, is said to be a conformally
covariant operator if for all positive functions
Pg ◦ w1 = w2 ◦ Pg ,
where g = 2 g, (w1 , w2 ) ∈ R × R, and where we view the powers of as
multiplication operators.
In this definition it is not meant that the domain and target bundles are
necessarily the same. The example in our next exercise will be important for
our later discussions.
Exercise 3 On pseudo-Riemannian manifolds (M n≥3 , g) show that
g
Aab : E −→ E(ab)0 given by
(3.11)
f −→ ∇(a ∇b)0 f + P(ab)0 f
is conformally covariant with (w1 , w2 ) = (1, 1). That is if

g = 2 g, for some
positive function , then
Ag (f ) = (Ag f ).
3.3.4. Conformal Geometry

Recall that we defined a conformal manifold as a manifold M equipped only
with an equivalence class of conformally related metrics (see Definition 3.1).
Conformally covariant operators, as in Definition 3.4, have good conformal
properties but (in general) fail to be invariant in the sense of (3.7). This is
not just an aesthetic shortcoming, it means that they are not well-defined on
conformal manifolds (M, c). To construct operators on (M, g) that do descend
to the corresponding conformal structure (M, c = [g]) we need the notion of
conformal densities.
3.3.4.1. Conformal Densities and the Conformal Metric

Let (M, c) be a conformal manifold of signature (p, q) (with p + q = n). For
a point x ∈ M, and two metrics g and ĝ from the conformal class, there is an
element s ∈ R+ such that ĝx = s2 gx (where the squaring of s is a convenient
convention). Thus, we may view the conformal class as being given by a
smooth ray subbundle Q ⊂ S2 T ∗ M, whose fibre at x is formed by the values of
gx for all metrics g in the conformal class. By construction, Q has fibre R+ and
the metrics in the conformal class are in bijective correspondence with smooth
sections of Q.
Denoting by π : Q → M the restriction to Q of the canonical projection
S2 T ∗ M → M, we can view this as a principal bundle with structure group R+ .
The usual convention is to rescale a metric g to ĝ = 2 g. This corresponds to
a principal action given by ρ s (gx ) = s2 gx for s ∈ R+ and gx ∈ Qx , the fibre of
Q over x ∈ M.
With this, we immediately have a family of basic real line bundles E[w] →
M for w ∈ R by defining E[w] to be the associated bundle to Q with respect to
the action of R+ on R given by s·t := s−w t. The usual correspondence between
sections of an associated bundle and equivariant functions on the total space
of a principal bundle then identifies the space (E[w]) of smooth sections of
E[w] with the space of all smooth functions f : Q → R such that f (ρ s (gx )) =
sw f (gx ) for all s ∈ R+ . We shall call E[w] the bundle of conformal densities
of weight w. Note that each bundle E[w] is trivial and inherits an orientation
from that on R. We write E+ [w] for the ray subbundle consisting of positive
elements.
If
g = 2 g ∈ c then the conformally related metrics
g and g each determine
sections of Q. We may pull back f via these sections and we obtain functions
on M related by
f g = w f g .
With w = 1 − n2 this explains the ‘strange move’ in Section 3.3.3.1 for the
domain function of the conformal wave operator, which is really an operator
on the bundle E[1 − n2 ].
Although the bundle E[w] as we defined it depends on the choice of the
conformal structure, it is naturally isomorphic to a density bundle (which is
independent of the conformal structure). Recall that the bundle of α–densities
is associated with the full linear frame bundle of M with respect to the
one-dimensional representation A → | det(A)|−α of the group GL(n, R). In
particular, 2-densities may be canonically identified with the oriented bundle
(n T ∗ M)2 , and one-densities are exactly the geometric objects on manifolds
that may be integrated (in a coordinate-independent way).
To obtain the link with conformal densities, as defined above, recall that any
metric g on M determines a nowhere vanishing 1-density, the volume density

vol(g). In a local frame, this density is given by | det(gij )|, which implies that
for a positive function we get vol(2 g) = n vol(g). So there is bijectively
a map from 1-densities to functions Q → R that are homogeneous of degree
−n given by the map
φ → φ(x)/ vol(g)(x),
and this gives an identification of the 1-density bundle with E[−n] and thus an
identification of E[w] with the bundle of (− wn )-densities on M.
So we may think of conformal density bundles as those bundles associated
with the frame bundle via one-dimensional representations, just as tensor
bundles are associated with higher rank representations. Given any vector
bundle B we will use the notation
B[w] := B ⊗ E[w],
and say the bundle is weighted of weight w. Note that E[w]⊗E[w ] = E[w+w ]
and that in the above we assume that B is not a density bundle itself and is
unweighted (weight zero).
Clearly, sections of such weighted bundles may be viewed as homogeneous
(along the fibres of Q) sections of the pull-back along π : Q → M. Now
the tautological inclusion of g̃ : Q → π ∗ S2 T ∗ M is evidently homogeneous of
degree 2, as for (s2 gx , x) ∈ Q, we have g̃(s2 gx , x) = (s2 gx , x) ∈ π ∗ S2 T ∗ M. So g̃
may be identified with a canonical section of g ∈ (S2 T ∗ M[2]) which provides
another description of the conformal class. We call g the conformal metric.
Another way to recover g is to observe that any metric g ∈ c is a section

of Q, and hence determines a section σg ∈ (E+ [1]) with the characterising
property that the corresponding homogeneous function σ̃g on Q takes the value
1 along the section g. Then
g = (σg )2 g (3.12)
and it is easily verified that this is independent of the choice of g ∈ c.

Conversely it is clear that any section σ ∈ (E+ [1]) determines a metric via
g := σ −2 g.
On a conformal manifold we call σ ∈ (E+ [1]), or equivalently the corre-

sponding g ∈ c, a choice of scale.
A nice application of g is that we can use it to raise, lower and contract tensor
indices on a conformal manifold, for example
gab : E a → Eb [2] by va → gab va ,
just as we use the metric in pseudo-Riemannian geometry. Also g gives the

isomorphism
2
⊗n g : n TM −→ E[2n]. (3.13)
3.3.4.2. Some Calculus with Conformal Densities

Observe that a choice of scale g ∈ c determines a connection on E[w] via the
formula

∇ g τ := σ w d(σ −w τ ) , τ ∈ (E[w]), (3.14)
where d is the exterior derivative and σ ∈ (E+ [1]) satisfies g = σ 2 g, as σ −w τ

is a function (i.e. is a section of E[0] = E). Coupling this to the Levi-Civita
connection for g (and denoting both ∇ g ) we have at once that ∇ g σ = 0 and
hence
∇ g g = 0. (3.15)
On the other hand the Levi-Civita connection directly determines a linear

connection on E[w] since the latter is associated with the frame bundle, as
mentioned above. But (3.15), with (3.13), shows that this agrees with (3.14).
That is, (3.14) is the Levi-Civita connection on E[w]. Thus we have:
Proposition 3.4 On a conformal manifold (M, c) the conformal metric g is
preserved by the Levi-Civita connection ∇ g for any g ∈ c.
In view of Proposition 3.4 it is reasonable to use the conformal metric
to raise and lower tensor indices even when working in a scale! We shall
henceforth do this unless we state otherwise. This enables us to give formulae

for natural conformally invariant operators acting between density bundles. For
example, now choosing g ∈ c and forming the Laplacian, and the trace of the
Schouten tensor,
:= gab ∇a ∇b and J := gab Pab ,
we have the following. The conformal Laplacian can be interpreted as a
differential operator
n n n
Y :E 1− → E −1 − , given by + 1 − J,
2 2 2
that is conformally invariant, meaning that it is well-defined on conformal
manifolds.
g
Similarly the operator Aab of (3.11) is equivalent to a conformally invariant
operator,
Aab : E[1] → E(ab)0 [1].
3.3.4.3. Conformal Transformations

The above constructions enable us to understand how conformally covariant
objects may be reinterpreted as objects that descend to invariant operators
on conformal manifolds. A similar result applies to curvature covariants.
However, we have not advanced the problem of constructing these.
It follows at once from the formula (3.14) that under conformal transforma-
tions g → g = 2 g the Levi-Civita connection on E[w] transforms by
∇
a τ = ∇a τ + wϒa τ ,
g g
(3.16)
since σg = −1 σg .
We can combine this with the transformation formulae
(3.1), (3.2) and (3.6) to compute the conformal transformations of weighted
tensors and Riemannian invariants. However this remains a hopeless approach
to finding conformal invariants.
3.4. Lecture 3: Prolongation and the Tractor Connection

If we are going to be successful at calculating conformal invariants we are
going to need a better way to calculate, one which builds in conformal
invariance from the start. In the next two lectures we develop such a calculus,
the conformal tractor calculus, which can be used to proliferate conformally
invariant tensor (or tractor) expressions.
For treating conformal geometry it would be clearly desirable to find an
analogue of the Ricci calculus available in the pseudo-Riemannian setting.
On the tangent bundle a conformal manifold (M, c) has a distinguished

equivalence class of connections but no distinguished connection from this
class. So at first the situation does not look promising. However we will see
that if we pass from the tangent bundle to a vector bundle with two more
dimensions (the standard tractor bundle), then there is indeed a distinguished
connection.
There are many ways to see how the tractor calculus arises on a conformal
manifold; we will give a very explicit construction which facilitates calcula-
tion, but first it will be very helpful to examine how the tractor calculus arises
on the flat model space of (Riemannian signature) conformal geometry, the
conformal sphere.
Remark 3.4 Recall that inverse stereographic projection maps Euclidean

space conformally into the sphere as a one point (conformal) compactification,
so that it makes sense to think of the sphere as the conformally ‘flat’ model
of Riemannian signature conformal geometry. The real reason however is that
the conformal sphere arises naturally as the geometry of a homogeneous space
of Lie groups and that conformal geometry can be thought of as the geometry
of curved analogues of this homogeneous geometry in the sense of Élie Cartan
(see, e.g. [12]).
3.4.1. The Model of Conformal Geometry: The conformal Sphere

We now look at the conformal sphere, which is an extremely important
example. We shall see that the sphere can be viewed as a homogeneous
space on which is naturally endowed a conformal structure, and that the
conformal tractor calculus arises naturally from this picture. This is the model
for Riemannian signature conformal geometry, but a minor variation of this
applies to other signatures.
First some notation. Consider an (n + 2)-dimensional real vector space
V∼ = Rn+2 . Considering the equivalence relation on V \ {0} given by
v ∼ v , if and only if v = rv for some r>0
we now write
P+ (V) := {[v] | v ∈ V \ {0}}
where [v] denotes the equivalence class of v. We view this as a smooth manifold
by the identification with one, equivalently any, round sphere in V.
Suppose now that V is equipped with a non-degenerate bilinear form H of
signature (n + 1, 1). The null cone N of zero-length vectors forms a quadratic
N+ π(x)
P+
x
P+ (N+ ) ∼
= Sn
Figure 3.1. The conformal sphere Sn as the ray projectivisation of the forward
null cone N+
variety in V. Choosing a time orientation, let us write N+ for the forward part
of N \ {0}. Under the R+ -ray projectivisation of V, meaning the natural map to
equivalence classes V → P+ (V), the forward cone N+ is mapped to a quadric
in P+ (V). This image is topologically a sphere Sn and we will write π for the
submersion N+ → Sn = P+ (N+ ) (see Figure 3.1).
Each point x ∈ N+ determines a positive definite inner product gx on Tπ (x) Sn
by gx (u, v) = Hx (u , v ) where u , v ∈ Tx N+ are lifts of u, v ∈ Tπ (x) Sn ,
meaning that π(u ) = u, π(v ) = v. For a given vector u ∈ Tπ (x) Sn two
lifts to x ∈ N+ differ by a vertical vector (i.e. a vector in the kernel of
dπ ). By differentiating the defining equation for the cone we see that any
vertical vector is normal to the cone with respect to H (null tangent vectors
to hypersurfaces are normal), and so it follows that gx is independent of the
choices of lifts. Clearly then, each section of π determines a metric on Sn
and by construction this is smooth if the section is. Evidently the metric
agrees with the pull-back of H via the section concerned. We may choose
coordinates X A , A = 0, · · · , n + 1, for V so that N is the zero locus of the
form −(X 0 )2 + (X 1 )2 + · · · + (X n+1 )2 , in which terms the usual round sphere
arises as the section X 0 = 1 of π.
Now, viewed as a metric on TV, H is homogeneous of degree 2 with respect
to the standard Euler (or position) vector field E on V, that is LE H = 2H,
where L denotes the Lie derivative. In particular this holds on the cone, which
we note is generated by E. Write g for the restriction of H to vector fields in
TN+ which are the lifts of vector fields on Sn . Note that u is the lift of a vector
field u on Sn , which means that for all x ∈ N+ , dπ(u (x)) = u(π(x)), and
so LE u = 0 (modulo vertical vector fields) on N+ . Thus for any pair u, v ∈
(TSn ), with lifts to vector fields u , v on N+ , g(u , v ) is a function on N+
homogeneous of degree 2, and which is independent of how the vector fields
were lifted. It follows that if s > 0 then gsx = s2 gx , for all x ∈ N+ . Evidently
N+ may be identified with the total space Q of a bundle of conformally related

metrics on P+ (N+ ). Thus g(u , v ) may be identified with a conformal density
of weight 2 on Sn . That is, this construction canonically determines a section of
E(ab) [2] that we shall also denote by g. This has the property that if σ is any sec-
tion of E+ [1] then σ −2 g is a metric gσ . Obviously different sections of E+ [1]
determine conformally related metrics and, by the last observation in the pre-
vious paragraph, there is a section σo of E+ [1] so that gσo is the round metric.
Thus we see that P+ (N+ ) is canonically equipped with the standard
conformal structure on the sphere, but with no preferred metric from this class.
Furthermore g, which arises here from H by restriction, is the conformal metric
on P+ (N+ ). In summary then we have the following.
Lemma 3.1 Let c be the conformal class of Sn = P+ (N+ ) determined canoni-
cally by H. This includes the round metric. The map N+ (π(x), gx ) ∈ Q gives
an identification of N+ with Q, the bundle of conformally related metrics on
(Sn , c).
Via this identification: functions homogeneous of degree w on N+ are
equivalent to functions homogeneous of degree w on Q and hence correspond
to conformal densities of weight w on (Sn , c); the conformal metric g on (Sn , c)
agrees with, and is determined by, the restriction of H to the lifts of vector
fields on P+ (N+ ).
The conformal sphere, as constructed here, is acted on transitively by
G = O+ (H) = ∼ O+ (n + 1, 1), where this is the time orientation preserving
subgroup of the orthogonal group preserving H, O(H) ∼ = O(n + 1, 1). Thus,
as a homogeneous space, P+ (N+ ) may be identified with G/P, where P is the
(parabolic) Lie subgroup of G preserving a nominated null ray in N+ .
3.4.1.1. Canonical Calculus on the Model

Here we sketch briefly one way to see the tractor connection on the model
P+ (N+ ).
Note that as a manifold V has some special structures that we have already
used. In particular an origin and, from the vector space structure of V, the Euler
vector field E which assigns to each point X ∈ V the vector X ∈ TX V, via the
canonical identification of TX V with V.
The vector space V has an affine structure and this induces a global
parallelism: the tangent space Tx V at any point x ∈ V may be canonically
identified with V. Thus, in particular, for any parameterised curve in V there
is a canonical notion of parallel transport along the given curve. This exactly
means that, viewing V as a manifold, it is equipped with a canonical affine
connection ∇ V . The affine structure gives more than this of course. It is
isomorphic to Rn+2 with its usual affine structure, and so the tangent bundle
to V is trivialised by everywhere parallel tangent fields. It follows that the
canonical connection ∇ V is flat and has trivial holonomy.
Next observe that H determines a signature (n + 1, 1) metric on V, where
the latter is viewed as an affine manifold. By the definition of its promotion
from bilinear form to metric, one sees at once that for any vector fields U, V
that are parallel on V the quantity H(U, V) is constant. This means that H is
itself parallel since for any vector field W we have
(∇W H)(U, V) = W · H(U, V) − H(∇W U, V) − H(U, ∇W V) = 0.
The second key observation is that a restriction of these structures descends
to the conformal n-sphere. We observed above that N+ is an R+ -ray bundle
over Sn . We may identify Sn with N+ / ∼ where the equivalence relation is
that x ∼ y if and only if x and y are points of the same fibre π −1 (x ) for some
x ∈ Sn . The restriction TV|N+ is a rank n + 2 vector bundle over N+ . Now we
may define an equivalence relation on TV|N+ that covers the relation on N+ .
Namely we decree Ux ∼ Vy if and only if x, y ∈ π −1 (x ) for some x ∈ Sn ,
and Ux and Vy are parallel. Considering parallel transport up the fibres of π , it
follows that TV|N+ / ∼ is isomorphic to the restriction TV|im(S) where S is any
section of π (that is S : Sn → N+ is a smooth map such that π ◦ S = idSn ). But
im(S) is identified with Sn via π and it follows that TV|N+ / ∼ may be viewed
as a vector bundle T on Sn . Furthermore it is clear from the definition of the
equivalence relation on TV|N+ that T is independent of S. The vector bundle
T on Sn is the (standard) tractor bundle of (Sn , c).
By restriction, H and ∇ V determine, respectively, a (signature (n + 1, 1))
metric and connection on the bundle TV|N+ that we shall denote with the same
notation. Since a vector field which is parallel along a curve γ in N+ may be
N+
a tractor
x at π(x)
Figure 3.2. An element of Tπ(x) corresponds to a homogeneous of degree zero

vector field along the ray generated by x
uniquely extended to a vector field which is also parallel along every fibre of π
through the curve γ , it is clear that ∇ V canonically determines a connection on
T that we shall denote ∇ T . Sections U, V ∈ (T ) are represented on N+ by
vector fields Ũ, Ṽ that are parallel in the direction of the fibres of π : N+ → Sn
(see Figure 3.2). On the other hand H is also parallel along each fibre of π and
so H(Ũ, Ṽ) is constant on each fibre. Thus H determines a signature (n + 1, 1)
metric h on T satisfying h(U, V) = H(Ũ, Ṽ). What is more, since ∇ V H = 0
on N+ , it follows that h is preserved by ∇ T , that is
∇ T h = 0.
Summarising the situation thus far we have the following.
Theorem 3.1 The model (Sn , c) is canonically equipped with the following: a
canonical rank n + 2 bundle T ; a signature (n + 1, 1) metric h on this; and a
connection ∇ T on T that preserves h.
Although we shall not go into details here it is straightforward to show the
following:
Proposition 3.5 The tractor bundle T of the model (Sn , c) has a composition
structure
n
T = E[1] + TS [−1] + E[−1]. (3.17)

+ E[−1] induces the con-
The restriction of h to the subbundle T 0 = TSn [−1]
formal metric g : TSn [−1] × TSn [−1] → E. Any null Y ∈ T [−1] satisfying
∼
=
h(X, Y) = 1 determines a splitting T −→ E[1] ⊕ TSn [−1] ⊕ E[−1] such that
the metric h is given by (σ , μ, ρ) → 2σρ + g(μ, μ) as a quadratic form.
It is easily seen how this composition structure arises geometrically. The
subbundle T 0 of T corresponds to the fact that TN+ is naturally identified
with a subbundle of TV|N+ . The vertical directions in there correspond to the

fact that E[−1] is a subbundle of T 0 , and the semidirect sum symbols +
record this structure.
3.4.1.2. The Abstract Approach to the Tractor Connection

An alternative way of seeing how the tractor connection arises on the flat model
is via the group picture. If G/H is a homogeneous space of Lie groups, then
the canonical projection G → G/H gives rise to a principal H-bundle over
G/H (with total space G). If V is a representation of H then one obtains a
homogeneous vector bundle V := G ×H V over G/H whose total space is
defined to be the quotient of G × V by the equivalence relation
(g, v) ∼ (gh, h−1 · v).
If V is in fact a representation of G, then the bundle V := G ×H V is trivial.

The isomorphism
G ×H V ∼ = (G/H) × V
is given by
[g, v] → (gH, g · v)
which is easily checked to be well defined. This trivialisation gives rise to a

flat connection on G ×H V. In the case where we have the conformal sphere
G/P ∼ = Sn and V is the defining representation of G (i.e. Rn+2 ) then G ×P V is
(naturally identified with) the tractor bundle T and the connection induced
by the trivialisation G ×P V ∼ = (G/P) × V is the tractor connection. The
∼
trivialisation T = (G/P) × V also immediately gives the existence of a bundle
metric preserved by the tractor connection induced by the bilinear form on V.
3.4.1.3. The Conformal Model in Other Signatures

With only a little more effort one can treat the case of general signature (p, q).
This again begins with a real vector space V ∼ = Rn+2 , but now we equip it with
a non-degenerate bilinear form H of signature (p + 1, q + 1), where p + q = n.
Writing again N for the quadratic variety of vectors which have zero-
length with respect to H, we see that the space of null rays in N \ {0} has
the topology of Sp × Sq . This is connected, unless p or q is zero in either,
in which case we get two copies of Sn . In any case an easy adaption of the
earlier discussion shows that P+ (N ) is equipped canonically with a conformal
structure of signature (p, q) and on this a tractor connection preserving now
a tractor metric of signature (p + 1, q + 1). (One can easily check that the
conformal class c of P+ (N ) contains a metric g which is of the form gSp − gSq
for some identification of P+ (N ) with Sp × Sq .)
As constructed here the conformal space P+ (N ) is acted on transitively by
G = O(H) ∼ = O(p + 1, q + 1), the orthogonal group preserving H. As a
homogeneous space, P+ (N ) may be identified with G/P, where P is the Lie
subgroup of G preserving a nominated null ray in N . We may think of this
group picture as a good model for general conformal manifolds. Of course
there are other possible choices of G/P (which may result in models which
are only locally equivalent to the ones here), as we have already seen in the
Riemannian case. See [37] for a discussion of the possible choices of (G, P)
and the connection with global aspects of the conformal tractor calculus.
Remark 3.5 Note that the model space S1 × Sn−1 of Lorentzian signature
conformal geometry is simply the quotient of the (conformal) Einstein universe
R × Sn−1 by integer times 2π translations. Thus the usual embeddings of
Minkowski and de Sitter space into the Einstein universe can be seen (by
passing to the quotient) as conformal embeddings into the flat model space. In
fact, S1 × Sn−1 can be seen as two copies of Minkowski space glued together
along a null boundary with two cone points, or as two copies of de Sitter space
glued together along a spacelike boundary with two connected components
which are (n − 1)-spheres. The significance of this will become clearer as we
continue to develop the tractor calculus and then move on to study conformally
compactified geometries.
3.4.2. Prolongation and the Tractor Connection

Here we construct the tractor bundle, connection and metric on a conformal
manifold of dimension at least three. We will see that the ‘conformal to
Einstein’ condition plays an important role. The tractor bundle and connection
are obtained by ‘prolonging’ the ‘almost Einstein equation’ (3.11).
First we state what is meant by Einstein here.
Definition 3.5 In dimensions n ≥ 3, a metric will be said to be Einstein if
Ricg = λg
for some function λ.
The Bianchi identities imply that, for any metric satisfying this equation, λ
is constant (as we assume M is connected). Throughout the following we shall
assume that n ≥ 3.
Recalling the model above, note that a parallel co-tractor I corresponds to a

homogeneous polynomial which, in standard coordinates X A on V −→ Rn+2 ,
is given by σ̃ = IA X A . Then σ̃ = 1 is a section of N+ that corresponds to
the intersection of N+ with the hyperplane IA X A = 1, in V. For at least some
of these distinguished (conic) sections the resulting metric on Sn , g = σ̃ −2 g
(on the open set where the corresponding density σ ∈ (E[1]) is nowhere
vanishing) is obviously Einstein. For example the round metric has already
been discussed. It is in fact true that all metrics obtained on open regions of Sn
in this way are Einstein, as we shall shortly see.
3.4.3. The Almost Einstein Equation

Let (M, c) be a conformalo manifold with dim(M) ≥ 3. It is clear that go ∈ c is
g
Einstein if and only if P(ab)0 = 0, and this is the link with the equation (3.11).
As a conformally invariant equation this takes the form
g g g
∇(a ∇b)0 σ + P(ab)0 σ = 0, (AE)
where σ ∈ E+ [1] encodes go = σ −2 g and we have used the superscripts

to emphasise that we have picked some metric g ∈ c in order to write the
equation.
Suppose that σ is a solution with the property that it is nowhere zero. Then,
without loss of generality, we may assume that σ is positive, that is σ ∈ E+ [1].
So σ is a scale, and go = σ −2 g is a well-defined metric. Since (3.11) is
conformal invariant, there is no loss if we calculate the equation (AE) in this
o
scale. But then ∇ g σ = 0. Thus we conclude that
P(ab)0 = 0.
go
Conversely suppose that P(ab)0 = 0 for some go ∈ c. Then go = σ −2 g for
some σ ∈ E+ [1]. Therefore σ solves (AE) by the reverse of the same argument.
Thus in summary we have the following, cf. [40].
Proposition 3.6 (M, c) is conformally Einstein (i.e. there is an Einstein metric
go in c) if and only if there exists σ ∈ E+ [1] that solves (AE). If σ ∈ E+ [1]
solves (AE) then go := σ −2 g is the corresponding Einstein metric.
There are some important points to make here.
Remark 3.6 Equation (AE) is equivalent to a system of (n+2)(n−1) 2 scalar

equations on one scalar variable. So it is overdetermined and we do not expect
it to have solutions in general.
3.4.4. The Connection Determined by a Conformal Structure

Proposition 3.6 has the artificial feature that it is a statement about nowhere
vanishing sections of E[1]. Let us rectify this.
Definition 3.6 Let (M, c) be a conformal manifold, of any signature, and σ ∈

E[1]. We say that (M, c, σ ) is an almost Einstein structure if σ ∈ E[1] solves
equation (AE).
We shall term (AE) the (conformal) almost Einstein equation.

Since the almost Einstein equation is conformally invariant it is natural
to seek integrability conditions that are also conformally invariant. There is
a systematic approach to this using a procedure known as prolongation that
goes as follows. We fix a metric g ∈ c to facilitate the calculations. With this
understood, for the most part in the following we will omit the decoration by
g of the natural objects it determines; for example we shall write ∇a rather
g
than ∇a .
As a first step observe that the equation (AE) is equivalent to the equation
∇a ∇b σ + Pab σ + gab ρ = 0 (3.18)
where we have introduced the new variable ρ ∈ E[−1] to absorb the trace
terms. The key idea is to attempt to construct an equivalent first-order closed
system. We introduce μa ∈ Ea [1], so our equation is replaced by the equivalent
system
∇a σ − μa = 0, and ∇a μb + Pab σ + gab ρ = 0 . (3.19)
This system is almost closed in the sense that the derivatives of σ and μb are
given algebraically in terms of the unknowns σ , μb and ρ. However to obtain a
similar result for ρ we must differentiate the system; by definition (differential)
prolongation is precisely concerned with this process of producing higher order
systems, and their consequences. Here we use notation from earlier.
The Levi-Civita covariant derivative of (3.18) gives
∇a ∇b ∇c σ + gbc ∇a ρ + (∇a Pbc )σ + Pbc ∇a σ = 0.
Contracting this using, respectively, gab and gbc yields
gab : ∇c σ + ∇c ρ + (∇ a Pac )σ + Pac ∇a σ = 0 (1)

gbc : ∇c σ + n∇c ρ + (∇c J)σ + J∇c σ = 0 (2).
Then the difference (2) − (1) is simply
(n − 1)∇c ρ + J∇c σ − Pac ∇a σ + Rcbbd ∇ d σ = 0
where we have used the contracted Bianchi identity ∇ a Pac = ∇c J and

computed the commutator [∇c , ] acting on σ . But
Rcbba ∇ a σ = −Rca ∇ a σ = (2 − n)Pc a ∇a σ − J∇c σ ,
and using this we obtain

∇c ρ − Pca μa = 0, (3.20)
after dividing by the overall factor (n − 1). So we have our closed system and,
what is more, this system yields a linear connection. We discuss this now.
On a conformal manifold (M, c) let us write [T ]g to mean the pair consisting
of a direct sum bundle and g ∈ c, as follows:

[T ]g := E[1] ⊕ Ea [1] ⊕ E[−1], g . (3.21)
Proposition 3.7 On a conformal manifold (M, c), fix any metric g ∈ c. There
is a linear connection ∇ T on the bundle
E[1]
⊕
[T ]g ∼
= Ea [1]
⊕
E[−1]
given by
⎛ ⎞ ⎛ ⎞
σ ∇a σ − μa
∇aT ⎝ μb ⎠ := ⎝ ∇a μb + gab ρ + Pab σ ⎠ . (3.22)
ρ ∇a ρ − Pab μb
Solutions of the almost Einstein equation (AE) are in one-to-one correspon-
dence with sections of the bundle [T ]g that are parallel for the connection ∇ T .
Proof It remains only to prove that ∇ T is a linear connection. But this is an
immediate consequence of its explicit formula which we see takes the form
∇ + where ∇ is the Levi-Civita connection on the direct sum bundle [T ]g =
E[1] ⊕ Ea [1] ⊕ E[−1] and is a section of End([T ]g ).
We shall call ∇ T , of (3.22), the (conformal) tractor connection. Interpreted

naı̈vely the statement in the proposition might appear to be not very strong: we
have already remarked that on a particular manifold it can be that the equation
(AE) has no non-trivial solutions. However this is a universal result, and so it in
fact gives an extremely useful tool for investigating the existence of solutions.
Note that the connection (3.22) is well defined on any pseudo-Riemannian
manifold. The point is that in using (3.22) for any application, we have
immediately available the powerful theory of linear connections (e.g. parallel
transport and curvature).
It is an immediate consequence of Proposition 3.7 that the almost Einstein
equation (AE) can have at most n + 2 linearly independent solutions. However
we shall see that far stronger results are available after we refine our under-
standing of the tractor connection (3.22). Furthermore this connection will be
seen to have a role that extends far beyond the almost Einstein equation.
3.4.5. Conformal Properties of the Tractor Connection

Although derived from a conformally invariant equation, the bundle and
connection described in (3.21) and (3.22) are expressed in a way depending
a priori on a choice of g ∈ c, so we wish to study their conformal properties.
Let us again fix some choice g ∈ c, before investigating the consequences of

changing this conformally.
Given the data of (σ , μb , ρ) ∈ [T ]g at x ∈ M, it follows from the general
properties of linear connections that we may always solve
∇ T (σ , μb , ρ) = 0, at x ∈ M. (3.23)
This imposes no restriction on either the conformal class c or the choice g ∈ c.

Examining the formula (3.22) for the tractor connection we see that if (3.23)
holds then, at the point x, we necessarily have
1
μb = ∇b σ , ρ = − (σ + Jσ ), (3.24)
n
where the second equation follows by taking a gab trace of the middle entry
on the right hand side of (3.22). Thus canonically associated to the tractor
connection there is the second-order differential E[1] → [T ]g given by
⎛ ⎞
nσ
[Dσ ]g = ⎝ n∇b σ ⎠, (3.25)
−(σ + Jσ )
where we have included the normalising factor n (=dim(M)) for later
convenience.
Recall that we know how the Levi-Civita connection, and hence also
and J here, transform conformally. Thus it follows that [Dσ ]g , or equivalently
(3.24), determines how the variables σ , μb and ρ of the prolonged system must
transform if they are to remain compatible with ∇ T under conformal changes.
If
g = 2 g, for some positive function , then a brief calculation reveals
1 g g σ ) = − 1 (g σ +Jg σ )−ϒ b ∇ σ − 1 σ ϒ b ϒ ,
∇b σ = ∇b σ +ϒb σ , and − (σ +J
g g
b b
n n 2
where as usual ϒ denotes d. Thus we decree
1

σ := σ , μb := μb + ϒb σ ,
:= ρ − ϒ b μb − σ ϒ b ϒb ,
ρ
2
or, writing ϒ 2 := ϒ a ϒa , this may be otherwise written using an obvious
matrix notation:
⎛ ⎞ ⎛ ⎞⎛ ⎞ ⎛ ⎞

σ 1 0 0 σ σ
[T ]g ⎝ μb ⎠ = ⎝ ϒb δbc 0 ⎠ ⎝ μc ⎠ ∼ ⎝ μb ⎠ ∈ [T ]g .

ρ −2ϒ
1 2
−ϒ 1 c ρ ρ
(3.26)
Note that at each point x ∈ M the transformation here is manifestly by a group
action. Thus in the obvious way this defines an equivalence relation among the
direct sum bundles [T ]g (of (3.21)) that covers the conformal equivalence of
metrics in c, and the quotient by this defines what we shall call the conformal
standard tractor bundle T on (M, c). More precisely we have the following
definition.
Definition 3.7 On a conformal manifold (M, c) the standard tractor bundle is

T := [T ]g / ∼,
g∈c
meaning the disjoint union of the [T ]g (parameterised by g ∈ c) modulo

equivalence relation given by (3.26). We shall also use the abstract index
notation EA for T .
We shall carry many conventions from tensor calculus over to bundles and
tractor fields. For example we shall write E(AB) [w] to mean S2 T ⊗ E[w], and
so forth.
There are some immediate consequences of Definition 3.7 that we should
observe. First, from this definition, the next statement follows tautologically.
Proposition 3.8 The formula (3.25) determines a conformally invariant
D : E[1] → T .
This operator is evidently intimately connected with the very definition of

the standard tractor bundle. Because of its fundamental role we make the
following definition.
Definition 3.8 Given a section of σ ∈ E[1] we shall call

1
I := Dσ
n
the scale tractor corresponding to σ .
Next observe also that from (3.26) it is clear that T is a filtered bundle; we
summarise this using a semi-direct sum notation

T = E[1] + Ea [1] + E[−1] (3.27)
meaning that T has a subbundle T 1 ⊂ T isomorphic to E[−1], Ea [1] is

isomorphic to a subbundle of the quotient T /T 1 bundle, and E[1] is the final
factor. We use X to denote the bundle surjection X : T → E[1] or in abstract
indices:
X A : EA → E[1]. (3.28)
Note that given any metric g ∈ c we may interpret X as the map [T ]g → E[1]
given by
⎛ ⎞
σ
⎝ μb ⎠ → σ .
ρ
For our current purposes the critical result at this point is that the tractor
connection ∇ T ‘intertwines’ with the transformation (3.26) in the following
sense.
Exercise 4 Let V = (σ , μb , ρ), a section of [T ]g , and
V = (
σ, ), a section
μb , ρ
of [T ]g , be related by (3.26), where
g = g. Show that then
2
⎛ ⎞ ⎛ ⎞⎛ ⎞
∇aσ − μa 1 0 0 ∇a σ − μa
⎜ ⎟
⎝∇a μb + gab ρ + Pab σ ⎠ = ⎝ ϒb δbc 0⎠ ⎝∇a μb + gac ρ + Pac σ ⎠.
∇a ρ −2ϒ −ϒ 1 ∇a ρ − Pac μc
1 2 c
− Pab μb
Here ϒ = d, as usual, and for example ∇a a

σ − μa means ∇ σ −
μa .
Given a tangent vector field va , the exercise shows that va ∇aT V transforms
conformally as a standard tractor field, that is by (3.26). Whence ∇ T descends
to a well defined connection on T . Let us summarise as follows.
Theorem 3.2 Let (M, c) be a conformal manifold of dimension at least 3. The
formula (3.22) determines a conformally invariant connection
∇ T : T → 1 ⊗ T .
For obvious reasons this will also be called the conformal tractor connection
(on the standard tractor bundle T ); the formula (3.22) is henceforth regarded
as the incarnation of this conformally invariant object on the realisation [T ]g
of T , as determined by the choice g ∈ c.
It is important to realise that the conformal tractor connection exists
canonically on any conformal manifold (of dimension at least 3). (One also
has the tractor connection on two-dimensional Möbius conformal manifolds.)
In particular its existence does not rely on solutions to the equation (AE).
Nevertheless by its construction in Section 3.4.4 (as a prolongation of the
equation (AE)), and using also equation (3.24), we have at once the following
important property.
Theorem 3.3 On a conformal manifold (M, c) we have the following. There
is a one-to-one correspondence between sections σ ∈ E[1], satisfying the
conformal equation
∇(a ∇b)0 σ + P(ab)0 σ = 0, (AE)
and parallel standard tractors I. The mapping from almost Einstein scales to
parallel tractors is given by σ → 1n DA σ while the inverse map is IA → X A IA .
So a parallel tractor is necessarily a scale tractor, as in Definition 3.8, but in
general the converse does not hold.
3.4.6. The Tractor Metric

It turns out that the tractor bundle has a beautiful and important structure that
is perhaps not expected from its origins via prolongation in Section 3.4.4.
Proposition 3.9 Let (M, c) be a conformal manifold of signature (p, q). The
formula
[VA ]g = (σ , μa , ρ) → 2σρ + gab μa μb =: h(V, V)
defines, by polarisation, a signature (p + 1, q + 1) metric on T .

Proof As a symmetric bilinear form field on the bundle [T ]g , h takes the form
⎛ ⎞⎛ ⎞
( σ μ ρ ) 0 0 1 σ
g
h(V , V) = ⎝ 0 g−1 0 ⎠ ⎝ μ ⎠ , (3.29)
1 0 0 ρ
g
where = should be read as ‘equals, calculating in the scale g’. So we see that
the signature is as claimed. By construction h(V, V ) has weight 0. It remains
to check the conformal invariance. Here we use the notation from (3.26):
1
2+
σρ μa
μa = 2σ (ρ − ϒc μc − ϒ 2 σ ) + (μa + ϒ a σ )(μa + ϒa σ )
2
= 2σρ + μa μa − 2σ ϒμ − σ 2 ϒ 2 + 2ϒμσ + ϒ 2 σ 2
= 2σρ + μa μa .
In the abstract index notation the tractor metric is hAB ∈ (E(AB) ), and its
inverse hBC .
The standard tractor bundle, as introduced in Sections 3.4.4 and 3.4.5, would
more naturally have been defined as the dual tractor bundle. But Proposition
(3.9) shows that we have not damaged our development; the tractor bundle is
canonically isomorphic to its dual and normally we do not distinguish these,
except by the raising and lowering of abstract indices using hAB .
From these considerations we see that there is no ambiguity in viewing X,
of (3.28), as a section of T ⊗ E[1] ∼ = T [1]. In this spirit we refer to X as the
canonical tractor. At this point it is useful to note that in view of this canonical
self duality T ∼
= T ∗ , and the formula (3.29), we have the following result.
Proposition 3.10 The canonical tractor X A is null, in that
hAB X A X B = 0.
Furthermore XA = hAB X B gives the canonical inclusion of E[−1] into EA :
XA : E[−1] → EA .
In terms of the decomposition of EA given by a choice of metric this is simply

⎛⎞
0
ρ → ⎝ 0 ⎠ .
ρ
As another immediate application of the metric we observe the following.

Given a choice of scale σ ∈ E+ [1], consider the corresponding scale tractor I,
and in particular it squared length h(I, I). This evidently has conformal weight
zero and so is a function on (M, c) determined only by the choice of scale σ .
Explicitly we have
g 2
h(I, I) = gab (∇a σ )(∇b σ ) − σ (J + )σ (3.30)
n
from Definition 3.8 with (3.25) and (3.29). Here we have calculated the
right hand side in terms of some metric g in the conformal class (hence the
g
notation =). But since σ is a scale we may, in particular, use g := σ −2 g. Then
∇ σ = 0 and we find the following result.
g
Proposition 3.11 On a conformal manifold (M, c), let σ ∈ E+ [1], and I = 1n Dσ

the corresponding scale tractor. Then
2
hAB IB IC = − Jσ ,
n
where Jσ := gab Pab , and gab = σ −2 gab .

Note we usually write J to mean the density gab Pab , as calculated in the scale g.
So here Jσ = σ 2 J. In a nutshell the conformal meaning of scalar curvature is
that it is the length squared of the scale tractor (up to a negative constant factor).
Next we come to the main reason the tractor metric is important, namely
because it is preserved by the tractor connection. With V as in Proposition 3.9
we have
g
∇a h(V, V) = 2[ρ∇a σ + σ ∇a ρ + gbc μb ∇a μc ]
= 2[ρ(∇a σ − μa ) + σ (∇a ρ − Pab μb ) + gbc μb (∇a μc + gac ρ + Pac σ )]
= 2h(V, ∇aT V).
We summarise this with the previous result.
Theorem 3.4 On a conformal manifold (M, c) of signature (p, q), the trac-
tor bundle carries a canonical conformally invariant metric h of signature
(p + 1, q + 1). This is preserved by the tractor connection.
We see at this point that the tractor calculus is beginning to look like an ana-
logue for conformal geometry of the Ricci calculus of (pseudo-)Riemannian
geometry: a metric on a manifold canonically determines a unique Levi-Civita
connection on the tangent bundle preserving the metric. The analogue here
is that a conformal structure of any signature (and dimension at least 3)
determines canonically the standard tractor bundle T equipped with the
connection ∇ T and a metric h preserved by ∇ T .
Note also that the tractor bundle, metric and connection seem to be
analogues of the corresponding structures found for the model in Theorem
3.1. Especially in view of the matching filtration structures; (3.27) should
be compared with that found on the model in Proposition 3.5 (noting that
TSn [−1] ∼= T ∗ Sn [1]). In fact the tractor connection here of Theorem 3.2, and
the related structures, generalise the corresponding objects on the model. This
follows by more general results in [7], or alternatively may be verified directly
by computing the formula for the connection of Theorem 3.1 in terms of the
Levi-Civita connection on the round sphere.
3.5. Lecture 4: The Tractor Curvature, Conformal

Invariants and Invariant Operators
Let us return briefly to our motivating problems: the construction of invariants
and invariant operators.
3.5.1. Tractor Curvature

Since the tractor connection ∇ T is well defined on a conformal manifold its
curvature κ on E A depends only on the conformal structure; by construction
it is an invariant of that. If we couple the tractor connection with any torsion
free connection (in particular with the Levi-Civita connection of any metric in
the conformal calss) then, according to our conventions from Lecture 1, the
curvature of the tractor connection is given by
(∇a ∇b − ∇b ∇a )U C = κ ab C D U D for all U C ∈ (E C ),
where ∇ denotes the coupled connection. Now using the fact that the tractor
connection preserves the inverse metric hCD we have
0 = (∇a ∇b − ∇b ∇a )hCD = κ ab C E hED + κ ab D E hCE .
In other words, raising an index with hCD
κ ab CD = −κ ab DC . (3.31)
It is straightforward to compute κ in a scale g. We have

⎛ ⎞ ⎛ ⎞⎛ ⎞
σ 0 0 0 σ
(∇a ∇b − ∇b ∇a ) ⎝ μc ⎠ = ⎝ Cab c Wab c d 0 ⎠ ⎝ μd ⎠ (3.32)
ρ 0 −Cabd 0 ρ
where, recall W is the Weyl curvature and C is the Cotton tensor,
Cabc := 2∇[a Pb]c .
In the expression for κ , the zero in the top right follows from the skew
symmetry (3.31), while the remaining zeros of the right column show that the
canonical tractor X D annihilates the curvature,
κ ab C D X D = 0.
This with the skew symmetry determines the top row of the curvature matrix.
It follows from the conformal transformation properties of the tractor splittings
that the central entry of the matrix is conformally invariant, and this is
consistent with the appearance there of the Weyl curvature Wab c d . Note that
in dimension 3 this necessarily vanishes, and so it follows that then the tractor
curvature is fully captured by and equivalent to the Cotton curvature Cabc .
Again this is consistent with the well known conformal invariance of that
quantity in dimension 3. Thus we have the following result.
Proposition 3.12 The normal conformal tractor connection ∇ T is flat if and
only if the conformal manifold is locally equivalent to the flat model.
So we shall say a conformal manifold (M, c) is conformally flat if κ = 0.
3.5.1.1. Application: Conformally Invariant Obstructions to Metrics

Being Conformal-to-Einstein
Note that as an immediate application we can use the tractor curvature to
manufacture easily obstructions to the existence of an Einstein metric in the
conformal class c.
From Proposition 3.6 and Theorem 3.3 a metric g is Einstein if and only if
the corresponding scale tractor IA = 1n DA σ is parallel, where g = σ −2 g. Thus
if g is Einstein then we have
κ ab C D I D = 0. (3.33)
Recall that κ ab C D X D = 0. If at any point p the kernel of κ ab C D : E D → Eab C is

exactly one dimensional then we say that tractor curvature has maximal rank.
We have:
Proposition 3.13 Let (M, c) be a conformal manifold. If at any point p ∈ M
the tractor curvature has maximal rank then there is no Einstein metric in the
conformal equivalence class c.
From this observation it is easy to manufacture conformal invariants that
must vanish on an Einstein manifold, see [29].
Proof If a metric g is Einstein then σ is a true scale and hence nowhere zero.
Thus IpD is not parallel to XpD , and the result follows from (3.33).
3.5.2. Toward Tractor Calculus

Although the tractor connection and its curvature are conformally invariant it
is still not evident how to manufacture easily conformal invariants. The tractor
curvature takes values in 2 (T ∗ M) ⊗ End(T ) and this is not a bundle on which
the tractor connection acts.
To deal with this and the related problem of constructing differential
invariants of tensors and densities we need additional tools.
The simplest of these is the Thomas-D operator. Recall that in Proposition
3.8 (cf. (3.25)) we constructed a differential operator D : E[1] → T that was,
by construction, tautologically conformally invariant. This generalises. Let us
write E [w] to denote any tractor bundle of weight w. Then:
Proposition 3.14 There is a conformally invariant differential operator
DA : E [w] → EA ⊗ E [w − 1],
defined in a scale g by
⎛ ⎞
(n + 2w − 2)wV
[DA V]g := ⎝ (n + 2w − 2)∇a V ⎠
V→
−(V + wJV)
where ∇ denotes the coupled tractor–Levi-Civita connection.
Proof Under a conformal transformation g →

g = 2 g, [DA f ]g transforms by
(3.26).
Note that this result is not as trivial as the written proof suggests since V (with
any indices suppressed) is a section of any tractor bundle, and the operator is
second order. In fact there are nice ways to construct the Thomas-D operator
from more elementary invariant operators [8, 25].
3.5.2.1. Application: Differential Invariants of Densities

and Weighted Tractors
A key point about the Thomas-D operator is that it can be iterated. For V ∈
(E [w]) we may form (suppressing all indices):
V → (V, DV, D ◦ DV, D ◦ D ◦ DV, · · · )
and, for generic weights w ∈ R, in a conformally invariant way this encodes
the jets of the section V entirely into weighted tractor bundles. Thus we can
proliferate invariants of V.
For example for f ∈ (E[w]) we can form
(DA f )DA f = −2w(n + 2w − 2)f (f + wJf ) + (n + 2w − 2)2 (∇ a f )∇a f .
By construction this is conformally invariant, for any weight w. So in fact it is
a family of invariants (of densities). Similarly we may form
(DA DB f )DA BB f = − 2(n + 2w − 4)(n + 2w − 2)w(w − 1)f 2 f
+ lower order terms,
and so forth.
3.5.2.2. Conformal Laplacian-Type Linear Operators

One might hope that the tractor-D operator is also effective for the construc-
tion of conformally invariant linear differential operators. In particular the
construction of Laplacian type operators is important. Certainly DA DA is by
construction conformally invariant but, on any weighted tractor bundle:
DA DA = 0.
This is verified by a straightforward calculation, but it should not be surprising
as by construction it would be invariant on, for example, densities of any
weight. On the standard conformal sphere it is well-known that there is no
non-trivial operator with this property.
From our earlier work we know the domain bundle of the conformal
Laplacian is E[1 − n2 ]. Observe that for V ∈ (E [1 − n2 ]) we have
(n + 2(1 − n2 ) − 2) = 0, and
⎛ ⎞
0
g
DA f = ⎝ 0 ⎠, that is DA V = −XA V
−( + 2−n
2 J)V
where is the (tractor-twisted) conformal Laplacian. In particular, the proof
of Proposition 3.14 was also a proof of this result:
g
Lemma 3.2 The operator = ( + 2−n 2 J) is a conformally invariant
n n
: E 1 − → E −1 − ,
2 2
where E is any tractor bundle.
This is already quite useful, as the next exercise shows.
Exercise 5 Show that if f ∈ E[2 − n2 ] then
DA f = −XA P4 f . (3.34)
Thus there is a conformally invariant differential operator

n n g
P4 : E 2 − → E −2 − where P4 = 2 + lower order terms.
2 2
In fact, this is the celebrated Paneitz operator discovered by Stephen Paneitz
in 1983, see [46] for a reproduction of his preprint from the time. See [25, 30]
for further discussion and generalisations.
An important point is that (3.34) does not hold if we replace f with a tractor
field of the same weight! Those who complete the exercise will observe the
first hint of this subtlety, in that during the calculation derivatives will need to
be commuted.
3.5.3. Splitting Tractors

Although the importance of the tractor connection stems from its conformal
invariance we need efficient ways to handle the decomposition of the tractor
bundle corresponding to a choice of scale.
Recall that a metric g ∈ c determines an isomorphism
E[1]
⎛ ⎞
⊕ σ

Ea [1] −→ E A , mapping ⎝ μa ⎠ = [U A ]g → U A ∈ E A . (3.35)
⊕ ρ
E[−1]
YA Z Ac XA
YA 0 0 1
ZAb 0 δb c 0
XA 1 0 0
Figure 3.3. Tractor inner product
The inclusion of E[−1] into the direct sum followed by this map is just X A :
E[−1] → E A , as observed earlier in Proposition (3.10). This is conformally
invariant. However let us now fix the notation
Z Aa : Ea [1] → E A , and Y A : E[1] → E A , (3.36)
for the other two bundle maps determined by (3.35). We call these (along
with X A ) the tractor projectors and view them as bundle sections Z Aa ∈
(E Aa )[−1], and Y A ∈ E A [−1]. So in summary [U A ]g = (σ , μa , ρ) is
equivalent to
U A = Y A σ + Z Aa μa + X A ρ. (3.37)
Using the formula (3.29) for the tractor metric it follows at once that X A YA = 1,
Z Aa ZAb = δba and all other quadratic combinations of the X, Y and Z are zero
as summarised in Figure 3.3. Thus we also have YA U A = ρ, XA U A = σ ,
ZAb U A = μb and the metric may be decomposed into a sum of projections,
hAB = ZA c ZBc + XA YB + YA XB .
The projectors Y and Z depend on the metric g ∈ c. If Ŷ A and Ẑ A b are the
corresponding quantities in terms of the metric ĝ = 2 g then altogether we
have
XA = XA,
Z Ab = Z Ab + ϒ b X A ,
Y A = Y A − ϒb Z Ab − 12 ϒb ϒ b X A (3.38)
as follows immediately from (3.37) and (3.26).
Remark 3.7 In the notation Z A a the tractor and tensor indices are both abstract.
If we move to a concrete frame field for TM, e1 , · · · , en and then write Z A i ,
i = 1, · · · , n, for the contraction Z A a eai , we come to the (weighted) tractor
frame field:
XA, ZA1, · · · , ZAn, Y A.
This is a frame for the tractor bundle adapted to the filtration (3.27) and metric,
as reflected in the conformal and inner product properties described in (3.38)
and Figure 3.3.
Of course tensor products of the tractor bundle are also decomposed by the
isomorphism (3.35) and this is described by the tensor products of the projec-
tors in an obvious way. To illustrate consider the case of the bundle of tractor
k-forms k T (which note is non-zero for k = 0, · · · , n + 2). The composition
∗
series T = E[1] + T M[−1] + E[−1] determines the composition series for
k T ,
k−2 [k − 2]
k−1
T ∼
k
= k−1 [k] +
⊕ +
[k − 2]. (3.39)
k
[k]
Given a choice of metric g from the conformal class there is a splitting of

this composition series corresponding to the splitting (3.35) of T , and this
is easily computed and dealt with using the ‘projectors’ X, Y and Z, see for
example [26].
3.5.4. The Connection

Just as connections are often described in terms of their action on a suitable
frame field it is useful to give the tractor connection in terms of its action on the
projectors X, Y and Z. The tractor covariant derivative of a field U A ∈ (E A ),
as in (3.37), is given by (3.22). Using the isomorphism (3.35) this is written
∇a U B = Y B (∇a σ − μa ) + Z Bb (∇a μb + gab ρ + Pab σ ) + X B (∇a ρ − Pab μb ).
On the other hand applying the connection directly to U A expanded as in (3.37)

we have
∇a U B = Y B ∇a σ + σ ∇a Y B + Z Bb ∇a μb + μb ∇a Z Bb + X B ∇a ρ + ρ∇a X B ,
where we have used the Leibniz rule for ∇a , viewed as the coupled tractor-
Levi-Civita connection. Comparing these we obtain:
∇a X B = Z B a , ∇a Z B b = −Pab X B − gab Y B , ∇a Y B = Pab Z Bb . (3.40)
This gives the transport equations for the projectors (and determines these
for the adapted frame as in Remark 3.7). From a practical point of view the
formulae (3.40) here enable the easy computation of the connection acting on
a tractor field of any valence.
3.5.4.1. Application: Computing and Conformal Laplacian Operators

The formulae are effective for reducing most tractor calculations to a routine
task. For example suppose we want to compute DA XA f = DA (XA f ) for
f ∈ (E [w]), that is a section of any weighted tractor bundle. We calculate in

some scale g. First note that
X A ( + (w + 1)J)XA f = X A XA f ,
= X A [, XA ] f ,
= X A ∇ b [∇b , XA ]f + X A [∇b , XA ]∇ b f
= X A (∇ b ZbA )f + 2X A ZbA ∇ b f
= X A (−JXA − nYA )f
= − nf ,
where we used X A XA = 0, X A ZbA = 0 and X A YA = 1. Thus we have
DA XA f = [(w + 1)(n + 2w)Y A + (n + 2w)Z Aa ∇a − X A ]XA f

= (w + 1)(n + 2w)f + (n + 2w)Z aA ZaA f + nf
= (w + 1)(n + 2w)f + (n + 2w)nf + nf ,
and collecting terms we come to
DA XA f = (n + 2w + 2)(n + w)f . (3.41)
A straightforward induction using (3.41) and (3.40) then enables us to

show that:
Theorem 3.5 On a conformal manifold (M n , c), for any tractor bundle E ,
and for each k ∈ Z≥1 if n is odd (or k ∈ Z≥1 and 2k < n if n is even) there is
a conformally invariant differential operator
n n
2k : E k − → E −k −
2 2
of the form k + lower order terms (up to a non-zero constant factor). These
are given by
2k := DA1 · · · DAk−1 DAk−1 · · · DA1 .
Remark 3.8 The operators in the theorem were first reported in [25], (and
with a different proof) as part of joint work of the second author with M. G.
Eastwood.
Acting on densities of weight (k−n/2) there are the Graham–Jenne–Mason–
Sparling (GJMS) operators of [36]. For k ≥ 3 these operators of Theorem
3.5 differ from the GJMS operators, as follows easily from the discussion
in [30].
3.5.5. Constructing Invariants

We can now put together the above tools and proliferate curvature invariants.
As a first step, following [27] we may form
3
WAB K L := DP X[P ZA a ZB]
b
κ ab K L .
n−2
This is conformally invariant by construction, as it is immediate from (3.38)
b is conformally invariant. It is exactly the object X3 which, for
that X[P ZA a ZB]
example, gives the conformally invariant injection
X3 : 2 [1] → 3 T .
It turns out that the W-tractor WABCD has the symmetries of an algebraic Weyl
tensor. In fact in a choice of conformal scale, WABCE is given by

(n − 4) ZA a ZB b ZC c ZE e Wabce − 2ZA a ZB b X[C ZE] e Cabe
(3.42)
−2X[A ZB] b ZC c ZE e Cceb + 4X[A ZB] b X[C ZE] e Beb ,
where Cabc is the Cotton tensor, and
Bab := ∇ c Ccba + Pdc Wdacb , (3.43)
see [31]. Note that from (3.42) it follows that, in dimension 4, Beb is
conformally invariant. This is the Bach tensor.
Since WABCD takes values in a weighted tractor bundle we may apply the
Thomas-D to this to capture jets of structure invariantly
WABCD → (W, DW, D ◦ DW, D ◦ D ◦ DW, · · · )
following the idea of Section 3.5.2.1. Contractions of these terms then yield
invariants. On odd dimensional manifolds this idea with some minor additional
input produces a generating set of scalar conformal invariants [27]. An
alternative (but closely related [7]) approach to the construction of conformal
invariants uses the Fefferman–Graham ambient metric [3, 17]. With either
approach, in even dimensions the situation is far more subtle, deeper ideas
are needed, and even for the construction of scalar conformal invariants open
problems remain (as mentioned in the sources referenced).
3.6. Lecture 5: Conformal Compactification

of Pseudo-Riemannian Manifolds
At this point we change directions from developing the theory of tractor
calculus for conformal geometries to applying this calculus to various prob-
lems inspired by asymptotic analysis in general relativity. We will see that
the conformal treatment of infinity in general relativity fits very nicely with
the conformal tractor calculus, and our basic motivation will be to produce
results which are useful in this setting. For the most part we will work fairly
generally however, and much of what is presented will be applicable in other
situations involving hypersurfaces or boundaries (such as Cauchy surfaces or
various kinds of horizons in general relativity, or to the study of Poincaré–
Einstein metrics in differential geometry). From this point on we will adopt
the convention that d = n + 1 (rather than n as before) is the dimension of our
manifold M, so that n will be the dimension of ∂M or of a hypersurface in M.
3.6.1. Asymptotic Flatness and Conformal Infinity

in General Relativity
It is natural in seeking to understand and describe the physics of general
relativity to want to study isolated systems. In particular we want to be able
to understand the mass and/or energy of the system (as well as other physical
quantities) and how the system radiates gravitational energy (or how the system
interacts with gravitational radiation coming in from infinity, i.e. from outside
the system). Clearly it is quite unnatural to try to isolate a physical system from
a spacetime by simply considering the inside of a timelike tube containing the
system (events inside the tube would depend causally on events outside, and
there would be no natural choice of such a tube anyway). From early on in
the history of general relativity then, physicists have sought to define isolated
systems in terms of spacetimes which are asymptotically flat, that is which
approach the geometry of Minkowski space in a suitable way as you approach
‘infinity’.
Definitions of asymptotic flatness typically involve the existence of special
coordinate systems in which the metric components and other physical fields
fall off sufficiently quickly as you approach infinity (infinity being defined by
the coordinate system). Once a definition of asymptotically flat spacetimes is
established one can talk about the asymptotic symmetry group of such space-
times (which is not in general the Poincare group) and the corresponding phys-
ical quantities such as mass and energy. One must be careful in defining the
asymptotic flatness condition not to require the fields to fall off too fast (thereby
excluding massive spacetimes or gravitationally radiating ones) or too slow (so
that one loses the asymptotic symmetries needed to define physical quantities).
A condition on asymptotically flat spacetimes which came to be seen as
important is the Sachs peeling-off property, which is the condition that along a
future directed null geodesic which goes out to infinity with affine parameter λ
the Weyl curvature W a bcd satisfies

W (4) W (3) W (2) W (1) 1
W= + 2 + 3 + 4 +O
λ λ λ λ λ5
where each tensor W (k) is of special algebraic type with the tangent vector
to the null geodesic as a k-fold repeated principal null direction. Typically
asymptotically flat spacetimes were required to be Ricci flat near infinity, so
the Weyl curvature was in fact the full curvature tensor in this region; the term
W (4) in the asymptotic expansion was interpreted as the gravitational radiation
reaching infinity.
The mass associated with this approach to isolated systems is called the
Bondi mass, which is not a conserved quantity but satisfies a mass loss formula
as energy is radiated away from the spacetime. The asymptotic symmetry
group (which is the same abstract group for any spacetime) is known as the
BMS (Bondi–Metzner–Sachs) group. There is an alternative way of defining
asymptotic flatness using the 3+1 formalism. In this case one talks about the
spacetime being asymptotically flat at spatial infinity, the corresponding mass
is the ADM (Arnowitt–Deser–Misner) mass, and the asymptotic symmetry
group is known as the spatial infinity (Spi) group. Questions of how these two
notions of asymptotic flatness are related to each other can be subtle and tricky.
For a good introductory survey of these issues see the chapter on asymptotic
flatness in [56].
After a great deal of work in these areas, a new approach was suggested by
Roger Penrose in the 1960s [47, 48, 50]. Penrose required of an asymptotically
flat spacetime that the conformal structure of spacetime extend to a pair of null
hypersurfaces called future and past null infinity. This mirrored the conformal
compactification of Minkowski space obtained by adding a lightcone at
infinity (see Figure 3.4). It was quickly shown that (i) Penrose’s notion of
conformal infinity satisfied the appropriate uniqueness property [22], (ii) this
form of asymptotic flatness implied the peeling property [47, 50], and (iii) the
corresponding group of asymptotic symmetries was the usual BMS group [23].
It is not totally surprising that this idea worked out so well: the causal (or light-
cone) structure of spacetime, which is encoded by the conformal structure, had
played an important role in the analysis of gravitational radiation up to that
point. The conformal invariance of the zero rest mass equations for arbitrary
spin particles and their peeling-off properties also fit nicely with Penrose’s
proposal, providing further motivation. What is really nice about this approach
however is that it is both natural and coordinate free.
Let us now give the formal definition(s), following [19].
Definition 3.9 A smooth (time and space-orientable) spacetime (M+ , g+ )
is called asymptotically simple if there exists another smooth Lorentzian
manifold (M, g) such that
R × Sn−1 i+
I+
i0
Mn
I−
i−
Figure 3.4. The standard conformal embedding of n-dimensional Minkowski

space Mn into the Einstein cylinder
(i) M+ is an open submanifold of M with smooth boundary ∂M+ = I ;

(ii) There exists a smooth scalar field on M, such that g = 2 g+ on M+ ,
and so that = 0, d = 0 on I ;
(iii) Every null geodesic in M+ acquires a future and a past endpoint on I .
An asymptotically simple spacetime is called asymptotically flat if in addition
Ricg+ = 0 in a neighbourhood of I .
The Lorentz manifold (M, g) is commonly referred to as the unphysical

spacetime, and g is called the unphysical metric. One can easily see that I
must have two connected components (assuming M+ is connected) I + and
I − consisting of the future and past endpoints of null geodesics respectively.
It is also easy to see that I + and I − are both smooth null hypersurfaces.
Usually (in the four-dimensional setting) I + and I − will have topology
S2 × R, however this is not necessarily the case.
Remark 3.9 The third condition is in some cases too strong a requirement,
for instance in a Schwarzschild black hole spacetime there are null geodesics
which circle about the singularity forever. To include Schwarzschild and other
such spacetimes one must talk about weakly asymptotically flat spacetimes (see
[49]).
Remark 3.10 (For those wondering ‘whatever happened to the cosmological

constant?’) One can also talk about asymptotically de Sitter spacetimes in
which case the boundary hypersurface(s) will be spacelike and one asks for
the spacetime to be Einstein, rather than Ricci flat, in a neighbourhood of the
conformal infinity. Similarly one can talk about asymptotically anti-de Sitter
spacetimes, which have a timelike hypersurface as conformal infinity.
Just how many spacetimes satisfy this definition of asymptotic flatness, and
how do we get our hands on them? These questions lead us to the discussion
of Friedrich’s conformal field equations. We will see that the conformal field
equations arise very naturally from the tractor picture. However, first we
will discuss more generally the mathematics of conformal compactification
(inspired by the conformal treatment of infinity above) for pseudo-Riemannian
manifolds, and the geometric constraints placed on such ‘compactifications’.
Again we shall see that this fits nicely within a tractor point of view.
3.6.2. Conformal Compactification

A pseudo-Riemannian manifold (M+ , g+ ) is said to be conformally compact
if M+ can be identified with the interior of a smooth compact manifold with
boundary M and there is a defining function r for the boundary = ∂M+
so that
g+ = r−2 g on M+
where g is a smooth metric on M. In calling r a defining function for we
mean that r is a smooth real valued function on M such that is exactly the
zero locus Z(r) of r and furthermore that dr is non-zero at every point of .
In the case that dr| is nowhere null, such a g induces a metric g on ,
but the defining function r may be replaced by r = f · r where f is any
non-vanishing function. This changes g conformally and so canonically the
boundary has a conformal structure determined by g+ , but no metric. The
metric g+ is then complete and (, c) is sometimes called the conformal
infinity of M+ . Actually for our discussion here we are mainly interested in the
structure and geometry of the boundary and the asymptotics of M+ near this,
so it is not really important to us that M is compact. An important problem
is how to link the conformal geometry and conformal field theory of to the
corresponding pseudo-Riemannian objects on M+ .
We will see that certainly this kind of ‘compactification’ is not always
possible, but it is useful in a number of settings. Clearly conformal geometry
is involved. In a sense it arises in two ways (that are linked). Most obviously
the boundary has a conformal structure. Secondly the interior was conformally
rescaled to obtain the metric g that extends to the boundary. This suggests a
strong role for conformal geometry – and it is this that we want to discuss.
Toward our subsequent discussion let us first make a first step by linking
this notion of compactification to our conformal tools and notations. Let us fix
a choice of r and hence g above. If τ ∈ (E[1]) is any non-vanishing scale on
M then σ := rτ is also a section of E[1] but now with zero locus Z(σ ) = .
This satisfies that ∇ g σ is nowhere zero along and so we say that σ is a
defining density for . Clearly we can choose τ so that
g = τ −2 g
where g is the conformal metric on M. Then
g+ = σ −2 g.
Thus we may think of a conformally compact manifold, as defined above, as a
conformal manifold with boundary (M, c) equipped with a section σ ∈ (E[1])
that is a defining density for the boundary ∂M.
The key examples of conformally compactified manifolds fit nicely within
this framework. Indeed, consider the Poincare ball model of hyperbolic space
Hd . This conformal compactification can be realised by considering the extra
structure induced on the conformal sphere (Sd , c) by a choice of constant
spacelike tractor field I on Sd (i.e. a constant vector I in V). This choice gives
a symmetry reduction of the conformal group G = SO+ (H) to the subgroup
H∼ = SO+ (d, 1) which stabilises I. On each of the two open orbits of Sd under
the action of H there is induced a hyperbolic metric, whereas the closed orbit
(an n-sphere) receives only the conformal structure induced from (Sd , c). The
two open orbits correspond to the two regions where the scale σ of I is positive
and negative respectively, and the closed orbit is the zero locus of σ . If we
choose coordinates X A for V as before, then the 1-density σ corresponds to
the homogeneous degree one polynomial σ̃ = IA X A (dualising I using H);
that σ −2 g gives a hyperbolic metric on each of the open orbits can be seen by
noting that the hyperplanes IA X A = ±1 intersect the future light cone N+ in
hyperbolic sections, and the fact that Z(σ ) receives only a conformal structure
can be seen from the fact that the hyperplane IA X A = 0 intersects N+ in
a subcone. Thus we see that the scale corresponding to a constant spacelike
tractor I on (Sd , c) gives rise to a decomposition of the conformal d-sphere into
two copies of conformally compactified hyperbolic space glued along their
boundaries (see Figure 3.5).
One can repeat the construction above in the case of the Lorentzian signature
model space Sn × S1 in which case one obtains a decomposition of Sn × S1 into
two copies of conformally compactified AdSd /Z glued together along their
σ̃ =0 σ̃ =1 Sd−1
N+
P+
Hd
σ̃ = IA X A
Figure 3.5. The orbit decomposition of the conformal sphere corresponding to the
subgroup H of the conformal group G preserving a fixed spacelike vector I. An
open orbit may be thought of as Hd conformally embedded into Sd
conformal infinities (where the Z-action on anti-de Sitter space is given by

integer times 2π time translations); this conformal compactification of AdSd /Z
corresponds to the usual embedding of AdSd into (half of) the Einstein universe
after we take the quotient by Z. If one instead takes a constant null tractor I on
Sn × S1 then the corresponding scale σ gives a decomposition of Sn × S1 into
two copies of conformally compactified Minkowski space glued along their
null boundaries (plus two isolated points where Z(σ ) has a double cone-like
singularity).
3.6.3. Geometry of Scale

There is an interesting perspective which emerges from the above correspon-
dence between constant (parallel) tractors and conformal compactifications.
This may be motivated by Proposition 3.7 and Theorem 3.3 which linked
almost Einstein scales to parallel tractors. Together they imply:
Proposition 3.15 An Einstein manifold (M, g) is the same as a conformal
manifold (M, c) equipped with a parallel standard tractor I such that σ :=
X A IA is nowhere zero: Given c and I the metric is recovered by
g = σ −2 g. (3.44)
Conversely a metric g determines c := [g], and σ ∈ E+ [1] by (3.44) again,

now used as an expression for this variable. Then I = 1n Dσ .
So this suggests that if we wish to draw on conformal geometry then using

the package (M, c, I) may give perspectives not easily seen via the (M, g)
framework. Some questions arise:
(1) First the restriction that σ := X A IA is nowhere zero seems rather unnatural
from this point of view. So what happens if we drop that? Then, if I is a
parallel standard tractor, recall we say that (M, c, I) is an almost Einstein
structure. What, for example, does the zero locus Z(σ ) of σ = X A IA look
like in this case?
(2) Is there a sensible way to drop the Einstein condition and use this approach
on general pseudo-Riemannian manifolds?
3.6.3.1. The Zero Locus of Almost Einstein and Almost

Scalar Constant (ASC) Manifolds
In fact there is now known a way to answer the first question via a very general
theory. Using ‘the package (M, c, I)’ amounts to recovering the underlying
pseudo-Riemannian structure (and its generalisations as below) as a type of
structure group reduction of a conformal Cartan geometry. If I is parallel then
this a holonomy reduction. On an extension of the conformal Cartan bundle to
a principal bundle with fibre group G = O(p + 1, q + 1), the parallel tractor
I gives a bundle reduction to a principal bundle with fibre H where this is:
(i) O(p, q + 1), if I is spacelike; (ii) O(p + 1, q), if I is timelike; and (iii) a
pseudo-Euclidean group Rd O(p, q), if I is null.
By the general theory of Cartan holonomy reductions [11] any such reduc-
tion yields a canonical stratification of the underlying manifold into a disjoint
union of curved orbits. These are parametrised by H\G/P, as are the orbits of
H on G/P. Each curved orbit is an initial submanifold carrying a canonically
induced Cartan geometry of the same type as that of the corresponding orbit
in the model. Furthermore this curved orbit decomposition must look locally
like the decomposition of the model G/P into H-orbits. This means that in an
open neighbourhood U of any point on M there is a diffeomorphism from U to
an open set U◦ in the model that maps each curved orbit (intersected with U)
diffeomorphically to the corresponding H-orbit (intersected with U◦ ) of G/P.
In our current setting P is the stabiliser in G of a null ray in Rp+1,q+1 and the
model G/P is isomorphic to conformal Sd × {1, −1}, if c is Riemannian, and
conformal Sp × Sq otherwise. The curved orbits arise here because as we move
around the manifold the algebraic relationship between the parallel object I and
the canonical tractor X changes. (For general holonomy reductions of Cartan
geometries the situation is a simple generalisation of this.) In particular, in
Riemannian signature (or if I 2 = 0), it is easily verified using these tools that
the curved orbits (and the H-orbits on the model) are distinguished by the strict
sign of σ = X A IA , see [11, Section 3.5]. By examining these sets on the model
we conclude.
Theorem 3.6 The curved orbit decomposition of an almost Einstein manifold
(M, c, I) is according to the strict sign of σ = IA X A . The zero locus satisfies:
• If I 2 = 0 (i.e. go Einstein and not Ricci flat) then Z(σ ) is either empty or is
a smooth embedded hypersurface.
• If I 2 = 0 (i.e. go Ricci flat) then Z(σ ) is, after possibly excluding isolated
points from Z(σ ), either empty or a smooth embedded hypersurface.
Here go means the metric σ −2 g on the open orbits (where σ is nowhere zero).
Remark 3.11 Much more can be said using the tools mentioned. For example
in the case of Riemannian signature it is easily shown that:
• If I 2 < 0 (i.e. go Einstein with positive scalar curvature) then Z(σ ) is empty.
• If I 2 = 0 (i.e. go Ricci flat) then Z(σ ) is either empty or consists of isolated
points.
• If I 2 > 0 (i.e. go Einstein with negative scalar curvature) then Z(σ ) is either
empty or is a smooth embedded separating hypersurface.
This holds because this is how things are locally on the flat model [24, 11];
if I 2 < 0 then on the model we have a round metric induced on the whole
conformal sphere, the one open orbit is the whole space; if I 2 = 0 then
on the model we are looking at the one point conformal compactification of
Euclidean space given by inverse stereographic projection, and there are two
orbits, one open and one an isolated point, so in the curved case Z(σ ) is
either empty or consists of isolated points; if I 2 > 0 then on the model we
are looking at two copies of conformally compactified hyperbolic space glued
along their boundaries as discussed earlier, in this case the closed orbit is a
separating hypersurface, so in the curved case Z(σ ) is either empty or is a
smooth embedded separating hypersurface (see Figure 3.6). In the Lorentzian
setting one can obtain a similar improvement of Theorem 3.6 by considering
the model cases.
In fact similar results hold in greater generality (related to question (2)

above) and this is easily seen using the earlier tractor calculus and elementary
considerations. We learn from the Einstein case above that an important role is
played by the scale tractor
1
IA = DA σ .
d
M−
M0 M+
σ =0
σ <0 σ >0
Figure 3.6. The curved orbit decomposition of an almost Einstein manifold with
Z (σ ) an embedded separating hypersurface
In the Einstein case this is parallel and hence non-zero everywhere. Let us drop
the condition that I is parallel and for convenience say that a structure
(M d , c, σ ) where σ ∈ (E[1])
is almost pseudo-Riemannian if the scale tractor IA := d1 DA σ is nowhere zero.

Note then that σ is non-zero on an open dense set, since DA σ encodes part
of the 2-jet of σ . So on an almost pseudo-Riemannian manifold there is the
pseudo-Riemannian metric go = σ −2 g on the same open dense set. In the
following the notation I will always refer to a scale tractor, so I = d1 Dσ , for
some σ ∈ (E[1]). Then we often mention I instead of σ and refer to (M, c, I)
as an almost pseudo-Riemannian manifold.
Now recall from (3.30) that
g 2
I 2 = gab (∇a σ )(∇b σ ) − σ (J + )σ (3.45)
d
where g is any metric from c and ∇ its Levi-Civita connection. This is
well-defined everywhere on an almost pseudo-Riemannian manifold, while
according to Proposition 3.11, where σ is non-zero, it computes
o
2 o Scg
I 2 = − Jg = − where go = σ −2 g.
d d(d − 1)
Thus I 2 gives a generalisation of the scalar curvature (up to a constant factor
−1/d(d − 1)); it is canonical and smoothly extends the scalar curvature to
include the zero set of σ . We shall use the term ASC manifold to mean an
almost pseudo-Riemannian manifold with I 2 = constant. Since the tractor
connection preserves the tractor metric, an almost Einstein manifold is a

special case, just as Einstein manifolds have constant scalar curvature.
Much of the previous theorem still holds in the almost pseudo-Riemannian
setting when I 2 = 0:
Theorem 3.7 Let (M, c, I) be an almost pseudo-Riemannian manifold with
I 2 = 0. Then Z(σ ), if not empty, is a smooth embedded separating hypersur-
face. This has a spacelike (resp. timelike) normal if g has negative scalar (resp.
positive) scalar curvature.
If c has Riemannian signature and I 2 < 0 then Z(σ ) is empty.
Proof This is an immediate consequence of (3.45). Along the zero locus Z(σ )
of σ (assuming it is non-empty) we have
I 2 = gab (∇a σ )(∇b σ ),
in particular ∇σ is nowhere zero on Z(σ ), and so σ is a defining density. Thus

Z(σ ) is a smoothly embedded hypersurface by the implicit function theorem.
Evidently Z(σ ) separates M according to the sign of σ . Also ∇σ is a (weight
1) conormal field along Z(σ ) so the claimed signs for the normal also follow
from the display.
Finally if c (and hence g) has Riemannian signature, then the display shows
that at any point of Z(σ ) the constant I 2 must be positive. This is a contradiction
if I 2 < 0 and so Z(σ ) = ∅.
Remark 3.12 Note that if M is compact then Theorem 3.7 gives a decompo-
sition of M into conformally compact manifolds glued along their conformal
infinities. Note also that if M is allowed to have a boundary then we only mean
that Z(σ ) is separating if it is not a boundary component of M.
What we can conclude from all this is that almost pseudo-Riemannian

manifolds (M, c, I) naturally give rise to nicely conformally compactified
metrics (at least in the case where I 2 = 0). If the scalar curvature of the
manifold (M+ , g+ ) admitting a conformal compactification (M, c) is bounded
away from zero then the conformal compactification must arise from an almost
pseudo-Riemannian manifold (M, c, I) in this way because I 2 is then bounded
away from zero on M+ and hence I extends to be nowhere zero on M (all
that we really need is that I 2 = 0 on ∂M+ ). Thus we can apply the results
above to the conformal compactification of (M+ , g+ ). This tells us for instance
that Riemannian manifolds with negative scalar curvature bounded away from
zero must have a nice smooth boundary as conformal infinity if they can be
conformally compactified. In the next section we develop this kind of idea.
3.6.4. Constraints on Possible Conformal Compactifications

Here we show that elementary geometric considerations restrict the topological
(and geometric) possibilities for a conformal infinity.
Theorem 3.8 Let (M+ , g+ ) be a geodesically complete pseudo-Riemannian
manifold and i : M+ → M an embedding of M+ as an open submanifold in a
closed conformal manifold (M, c) and σ a smooth section of E[1] on M so that
on the image i(M+ ), which we identify with M+ , we have that
g+ = σ −2 g
where g is the conformal metric on M. If the scalar curvature of g+ is bounded

away from zero then either M+ = M or the boundary points of M+ in M
form a smooth embedded hypersurface in M. This has a timelike normal field
if Scg+ > 0 and a spacelike normal field if Scg+ < 0.
Proof We can form the scale tractor
1
IA = DA σ .
d
Then I 2 is a smooth function on M that is bounded away from zero on the open
set M+ ⊂ M. Thus I 2 = 0 on the topological closure M + of M+ in M. Now
since I 2 = 0 on M + we have at any point p ∈ Z(σ )∩M + that ∇σ (p) = 0. Thus
the zero locus of σ (intersected with M + ) is a regular embedded submanifold
of M + . By construction this lies in the set ∂M+ of boundary points to M+ .
On the other hand any boundary point is in the zero locus, as otherwise for
any p ∈ ∂M+ such that σ (p) = 0 there is an open neighbourhood of p in M +
such that σ is nowhere zero and so g extends as a metric to this neighbourhood.
So then there is a geodesic from a point in M+ that reaches p in finite time,
which contradicts that M+ is geodesically complete.
In particular if g+ is Einstein with non-zero cosmological constant then I is

parallel, and hence I 2 is constant and non-zero. So the above theorem applies
in this case. In fact in the Einstein case we can also get results which in some
aspects are stronger. For example:
Theorem 3.9 Suppose that M+ is an open dense submanifold in a compact
connected conformal manifold (M d , c), possibly with boundary, and that one
of the following two possibilities holds: either M is a manifold with boundary
∂M and M \ M+ = ∂M, or M is closed and M \ M+ is contained in a smoothly
embedded submanifold of M of codimension at least 2. Suppose also that g+
is a geodesically complete Einstein, but not Ricci flat, pseudo-Riemannian
metric on M+ such that on M+ the conformal structure [g+ ] coincides with

the restriction of c. Then, either
• M+ is closed and M+ = M; or
• we are in the first setting with M \ M+ the smooth n-dimensional boundary
for M.
Proof M+ is canonically equipped with a parallel standard tractor I such that
I 2 = c = 0, where c is constant.
Now [g+ ] on M+ is the restriction of a smooth conformal structure c on M.
Thus the conformal tractor connection on M+ is the restriction of the smooth
tractor connection on M. Working locally it is straightforward to use parallel
transport along a congruence of curves to give a smooth extension of I to a
sufficiently small open neighbourhood of any point in M \ M+ , and since M+
is dense in M the extension is parallel and unique. (See [15] for this extension
result in general for the case where M \ M+ lies in a submanifold of dimension
at least 2.) It follows that I extends as a parallel field to all of M.
It follows that σ := IA X A also extends smoothly to all of M. This puts us
back in the setting of Theorem 3.8, so that M \ M+ is either empty or is an
embedded hypersurface. If M \ M+ is empty then M is closed and M = N.
If M \ M+ is an embedded hypersurface then, by our assumptions, we must
be in the case where M is a manifold with boundary and M \ M+ = ∂M
(since an embedded hypersurface cannot be contained in a submanifold of
codimension two).
Remark 3.13 In Theorem 3.8 we had to assume that there was a smooth
globally defined conformal 1 density σ which glued the metric on M+
smoothly to the conformal structure of M. In Theorem 3.9 we are able to drop
this condition because we are requiring that the metric g+ on M+ is Einstein
with conformal structure agreeing with that on M, allowing us to recover the
smooth global 1 density σ by parallel extension of the scale tractor for g+ to
M \ M+ . However, it is not hard to see that one must place some restriction on
the submanifold M \ M+ in order for a parallel extension to exist; this is where
the condition that M \M+ must lie in a submanifold of codimension two comes
in for the case where M is closed.
3.6.5. Friedrich’s Conformal Field Equations and Tractors

In this section we briefly discuss Friedrich’s conformal field equations and their
applications. We will see that the equations can be very easily arrived at using
tractor calculus, and that the tractor point of view is also nicely compatible
with the way the equations are used in applications.
3.6.5.1. The Equations Derived

Suppose that (M+ , g+ ) is an asymptotically flat spacetime with corresponding
unphysical spacetime (M, g) and conformal factor . (We may assume that
M = M+ ∪ I .) Then since (M+ , g+ ) is Ricci flat near the conformal infinity
I , the conformal scale tractor I defined on M+ corresponding to the metric
g+ is parallel (and null) near I and thus has a natural extension (via parallel
transport or simply by taking the limit) to all of M. Let us assume for simplicity
that (M+ , g+ ) is globally Ricci flat, then the naturally extended tractor I is
globally parallel on the non-physical spacetime (M, g). If we write out the
equation ∇a I B = 0 on M in slots using the decomposition of the standard
tractor bundle of (M, [g]) induced by g then we get the system of equations
∇a σ = μa
∇a μb = −σ Pab − ρgab
∇a ρ = Pab μb
g
where I = (σ , μa , ρ). If we also trivialise the conformal density bundles using
g then σ becomes , g becomes g, and we recognise the above equations as
the first three equations in what are commonly known as Friedrich’s conformal
field equations. A fourth member of the conformal field equations can be
obtained by writing IA I A = 0 out as
2σρ + μa μa = 0.
Next we observe that ∇a I B = 0 clearly implies κ ab C D I D = 0, and the

projecting part of this equation gives the conformal C-space equation
∇b Pac − ∇a Pbc = Wabcd σ −1 μd
after multiplication by σ −1 . (The ‘bottom slot’ of κ ab C D I D = 0 taken w.r.t.

g is (∇b Pac − ∇a Pbc )μc = 0 which also follows from contracting the above
displayed equation with μc .) The contracted Bianchi identity states that
∇ d Wabcd = ∇b Pac − ∇a Pbc
and from this and the previous display it follows that
∇ d (σ −1 Wabcd ) = 0
(where we have used that ∇a σ = μa ). It can be shown [19] that the field
σ −1 Wab c d is regular at I (recall the Sachs peeling property), and we will
write this field as Kab c d . If we substitute Kabcd = gce Kab e d for σ −1 Wabcd in the
C-space equation and in the last equation displayed above we get the remaining
two equations from Friedrich’s conformal system
∇b Pac − ∇a Pbc = Kabcd μd

and
∇ a Kabcd = 0,
which hold not only on M+ but on all of M. This pair of equations encode
not only the fact that (M+ , g+ ) is Cotton flat (both equations being in some
sense conformal C-space equations), but also by their difference they encode
the contracted Bianchi identity (and hence the full Bianchi identity if M+ is
four-dimensional [21]).
The equations from Friedrich’s conformal system are thus seen to be
elementary consequences of the system of four tractor equations
∇a I B = 0
I A IA = 0
κ ab C D I D = 0
∇[a κ bc]DE = 0.
Remark 3.14 There are in fact more equations that could be considered
as a part of Friedrich’s conformal system, however these simply define the
connection and the curvature terms used in the equations we have given. These
equations would be the expressions ∇g = 0 and T ∇ = 0 satisfied by the Levi-
Civita connection of g, and the decomposition of the Riemannian curvature
tensor of ∇ as
Rabcd = σ Kabcd + 2Pa[c gd]b − 2Pb[c gd]a
which serves to define the curvature tensors Kabcd and Pab (along with the
appropriate symmetry conditions on Kabcd and Pab as well as the condition
that gac Kabcd = 0).
Note that we can easily allow for a non-zero cosmological constant in our
tractor system by taking I A IA to be a constant rather than simply zero.
3.6.5.2. The Purpose of the Equations

What are the conformal field equations for? The answer is that one seeks
solutions to them in the same way that one seeks solutions to Einstein’s
field equations. A global solution to the conformal field equations gives
a conformally compactified solution of Einstein’s field equations. (In fact
solutions of the conformal field equations may extend to regions ‘beyond
infinity’ where becomes negative.) We should note one can very easily
allow for a non-zero cosmological constant and even for non-zero matter
fields (especially ones with nice conformal behaviour) in the conformal field
equations. The conformal field equations are then a tool for obtaining and
investigating isolated systems in general relativity.
Like Einstein’s field equations, the conformal field equations have an initial
value formulation where initial data is specified on a Cauchy hypersurface. But
when working with the conformal field equations it is also natural to prescribe
data on the conformal infinity; when initial data is prescribed on (part of) I −
as well as on an ingoing null hypersurface which meets I − transversally we
have the characteristic initial value problem. When initial data is specified on
a spacelike hypersurface which meets conformal infinity transversally we have
the hyperboloidal initial value problem. (The name comes from the fact that
spacelike hyperboloids in Minkowski space are the prime examples of such
initial data hypersurfaces.) If the hyperboloidal initial data hypersurface meets
I − (rather than I + ) then one should also prescribe data on the part of I − to
the future of the hypersurface, giving rise to an initial-boundary value problem.
What is being sought in the study of these various geometric PDE problems?
Firstly information about when spacetimes will admit a conformal infinity
and what kind of smoothness it might have. Secondly information about
gravitational radiation produced by various gravitational systems as well as
the way that such systems interact with gravitational radiation (scattering
properties). Thirdly, one is obviously interested in the end in having a general
understanding of the solutions of the conformal field equations (though this is
a very hard problem). These problems have been studied a good deal, both
analytically and numerically, however there are many questions left to be
answered. For helpful overviews of this work consult [19, 21].
3.6.5.3. Regularity
An important feature of the conformal field equations is that they are regular
at infinity. From the tractor calculus point of view this is obvious since the
conformal structure extends to the conformal infinity so the tractor system
displayed above cannot break down there, but one can also easily see this
from the usual form of the equations. The significance of this is highlighted
when one considers what field equations one might naively expect to use in this
setting: the most obvious guess would be to rewrite Ricg+ = 0 in terms of the
Ricci tensor of the unphysical metric g and its Levi-Civita connection yielding

d−2 1 c d−1 c
Ricab + ∇a ∇b − gab ∇ ∇c − (∇ )∇c = 0
2
which degenerates as → 0.
It is worth noting the significance of the regularity of the conformal
field equations in the area of numerical relativity. The appeal of ‘conformal
compactification’ to numerical relativists should be obvious: it means that one

is dealing with finite domains.
3.6.5.4. Different Reductions and Different Forms of the Equations

The conformal field equations are a system of geometric partial differential
equations. In order to study them analytically or numerically they need to be
reduced to a classical system of PDE; this involves introducing coordinates
and adding conditions on various fields to pin down the natural gauge freedom
in the equations (gauge fixing). There is a significant amount of freedom
in how one reduces the system. If this process is done carefully with the
conformal field equations one can obtain (in four dimensions) a symmetric
hyperbolic system of evolution equations together with an elliptic system of
constraint equations (reflecting the fact that the conformal field equations are
overdetermined). The constraint equations can be taken as conditions on initial
data for the Cauchy problem for the conformal field equations, whereas the
evolution equations can be taken as prescribing how such data will evolve off
the Cauchy hypersurface. For further discussion of the reduction process and a
demonstration of how the conformal field equations can be reduced see [19].
One may employ different reductions of the conformal field equations in
different settings and for different purposes. Indeed, the form of Friedrich’s
conformal field equations which we presented above is by no means the only
form of the equations that is used, and the significance of the other forms is
that they allow for a still broader range of different reductions. As an example
of this, note that one could instead have employed the splitting of the tractor
bundle induced by a Weyl connection [8] to obtain a system of conformal field
equations in a similar way to what we did above; this would result in a different
form of the equations which have different gauge freedoms and from which we
can obtain different reduced systems; the equations which one would obtain
this way are what Friedrich calls the general conformal field equations whereas
the equations we presented above are referred to as the metric conformal field
equations [21]. Friedrich also frequently casts the equations in spinor form,
and we have now seen that the conformal field equations take a very simple
tractor form
∇a I B = 0
I A IA = 0
κ ab C D I D = 0
∇[a κ bc]DE = 0
(where we also need to impose the appropriate conditions on ∇ and κ as

variables). It may indeed prove profitable to see the various reductions of
the conformal field equations as being reductions of this tractor system.
The gauge freedom(s) in Friedrich’s field equations can be seen as coming
from the splitting of the tractor equations into tensor equations (along with
freedom to choose coordinates). Friedrich’s use of conformal geodesics [21] in
constructing coordinates as part of the reduction process also fits quite nicely
with the tractor picture (see, e.g., [42, 44]).
Remark 3.15 In [20] Fraundiener and Sparling observed that in the four-
dimensional (spin) case Friedrich’s conformal field equations could be recast
in terms of local twistors involving the so-called ‘infinity twistor’ I αβ . In
this case the bundle of real bitwistors is canonically isomorphic to the
standard tractor bundle, and the ‘infinity twistor’ I αβ corresponds (up to a
constant factor) to the scale tractor I A , so that the local twistor system in
[20] is closely related to the tractor system above. Indeed the connection
is so close (the modern approach to conformal tractors having developed
out of study of the local twistor calculus [2]) that we may consider the
local twistor formulation to be the origin of the above tractor system. The
conformal field equations have also been presented in terms of tractors
by Christian Lübbe [42]. Lübbe and Tod have applied the tractor calculus
to the study of conformal gauge singularities in general relativity (see,
e.g., [44]).
3.7. Lecture 6: Conformal Hypersurfaces

In order to progress in our study of conformal compactification we first need
to spend some time considering the geometry of embedded hypersurfaces
in conformal manifolds (or of boundaries to conformal manifolds if you
like). In particular we will need to examine how the standard tractor bundle
(and connection, etc.) of the hypersurface with its induced conformal
structure is related to the standard tractor bundle (and connection, etc.) of
the ambient space. Although our main motivation is the study of conformal
compactification for this lecture we will consider conformal hypersurfaces
more generally since there are many other important kinds of hypersurface
which turn up in general relativity. We will however restrict ourselves to the
case of non-degenerate hypersurfaces; in particular we do not give a treatment
of null hypersurfaces here.
3.7.1. Conformal Hypersurfaces

By a hypersurface in a manifold M we mean a smoothly embedded codi-
mension 1 submanifold of M. We recall some facts concerning hypersurfaces
in a conformal manifold (M d , c), d ≥ 3. In fact we wish to include the case
that might be a boundary component. In the case of a pseudo-Riemannian (or
conformal pseudo-Riemannian) manifold with boundary then, without further
comment, we will assume that the conformal structure extends smoothly to the
boundary.
Here we shall restrict ourselves to hypersurfaces with the property
that any conormal field along is nowhere null (i.e. to non-degenerate
hypersurfaces). In this case the restriction of any metric g ∈ c gives a
metric ḡ on . Different metrics, among the metrics so obtained, are related
conformally, and so the conformal class c determines a conformal structure
c on . To distinguish from the ambient objects, we shall overline the
corresponding objects intrinsic to this conformal structure. For example g
denotes the conformal metric on .
Then by working locally we may assume that there is a section na ∈
(Ea [1]) on M such that, along , na is a conormal satisfying |n|2g :=
gab na nb = ±1. This means that if g = σ −2 g is any metric in c then na =
g
σ −1 na is an extension to M of a unit conormal to in (M, g). So na | is the

conformally invariant version of a unit conormal in pseudo-Riemannian geom-
etry; na must have conformal weight 1 since g−1 has conformal weight −2.
We choose to work with the weighted (extended) conormal field na even in
the presence of a metric g from the conformal class. Thus we end up with a
weight 1 second fundamental form Lab by restricting, along , ∇a nb to T ×
T ⊂ (TM × TM)| , where ∇ = ∇ g . (Here we are viewing T ∗ as the
subbundle of T ∗ M| orthogonal to na .) Explicitly Lab is given by
Lab := ∇a nb ∓ na nc ∇c nb along ,
since |n|2g is constant along . From this formula, it is easily verified that Lab is
independent of how na is extended off . It is timely to note that Lab harbours
a hypersurface conformal invariant: using the formulae (3.2) and (3.16) we
compute that under a conformal rescaling, g → g = e2ω g, Lab transforms
according to

g g
Lab = Lab + gab ϒc nc ,
where as usual ϒ is the exterior derivative of ω (which is equal to the log

exterior derivative −1 d of = eω ) and we use this notation below without
further mention. Thus we see easily the following well-known result:
Proposition 3.16 The trace-free part of the second fundamental form

◦ 1 cd
Lab = Lab − Hgab , where H := g Lcd
d−1
is conformally invariant.
The averaged trace of L (= Lg ), denoted H above, is the mean curvature
of . Evidently this a conformal −1-density and under a conformal rescaling,
g = e2ω g, H g transforms to Hg = H g + na ϒa .
g →
Thus we obtain a conformally invariant section N of T |
⎛ ⎞
0
g
NA = ⎝ na ⎠ ,
−H g
and from (3.29) h(N, N) = ±1 along ; the conformal invariance follows
because the right hand side of the display transforms according to the ‘tractor
characterising’ transformation (3.26). Obviously N is independent of any
choices in the extension of na off . This is the normal tractor of [2] and
may be viewed as a tractor bundle analogue of the unit conormal field from the
theory of pseudo-Riemannian hypersurfaces.
3.7.2. Umbilicity
A point p in a hypersurface is said to be an umbilic point if, at that point, the
◦
trace-free part L of the second fundamental form is zero. Evidently this is a
conformally invariant condition. A hypersurface is totally umbilic if this holds
at all points. As an easy first application of the normal tractor we recall that it
leads to a nice characterisation of the umbilicity condition.
Differentiating N tangentially along using ∇ T , we obtain the following
result.
Lemma 3.3 ⎛ ⎞
0
gcb ⎜ ◦ ⎟
LaB := ∇ a NB = ⎝ Lab ⎠ (3.46)
◦
− d−2 ∇ b Lab
1
where ∇ is the pull-back to of the ambient tractor connection.

Proof Using the formula (3.22) for the tractor connection, we have
⎛ ⎞
−nc
g
∇c NB = ⎝ ∇c nb − gcb H ⎠ .
−∇c H − Pcb nb
Thus applying the orthogonal TM| → T projector ca := (δac ∓ nc na )

we obtain immediately the top two terms on the right hand side of (3.46).
The remaining term follows after using the hypersurface Codazzi equation for
pseudo-Riemannian geometry

∇ a Lbc − ∇ b Lac = aa bb Ra b cd nd ,
where ∇ denotes the Levi-Civita connection for the metric g induced by g (for
full details see [55, 32]).
Thus we recover the following result.
Proposition 3.17 [2] Along a conformal hypersurface , the normal tractor
N is parallel, with respect to ∇ T , if and only if the hypersurface is totally
umbilic.
3.7.3. Conformal Calculus for Hypersurfaces

The local calculus for hypersurfaces in Riemannian geometry is to a large
extent straightforward because there is a particularly simple formula, known
as the Gauss formula, which relates the ambient Levi-Civita connection to the
Levi-Civita connection of the induced metric. It is natural to ask if we have the
same with our conformal tractor calculus.
Given an ambient metric g we write ∇ to denote the pull-back of the ambient
Levi-Civita connection along the embedding of the hypersurface , that is the
ambient connection differentiating sections of TM| in directions tangent to
the hypersurface. Then the Gauss formula may be expressed as
∇ a vb = ∇ a vb ∓ nb Lac vc
for any tangent vector field v to (thought of as a section va of E a | satisfying
va na = 0). We now turn towards deriving a tractor analogue of this.
Note that by dint of its conformal structure the hypersurface (, c) has its
own intrinsic tractor calculus, and in particular a rank d + 1 standard tractor
bundle T . Before we could hope to address the question above we need to
relate T to the ambient tractor bundle T .
First observe that, along , T has a natural rank (d + 1)-subbundle, namely
N ⊥ the orthogonal complement to NB . As noted in [5, 38], there is a canonical
(conformally invariant) isomorphism

N ⊥ −→ T . (3.47)
Calculating in a scale g on M, the tractor bundle T , and hence also N ⊥ ,
decomposes into a triple. Then the mapping of the isomorphism is
⎛ ⎞ ⎛ ⎞
σ σ
[N ⊥ ]g ⎝ μb ⎠ → ⎝ μb ∓ Hnb σ ⎠ ∈ [T ]g (3.48)
ρ ρ ± 12 H 2 σ
where, as usual, H denotes the mean curvature of in the scale g and g is the
pull-back of g to . Since (σ , μb , ρ) is a section of [N ⊥ ]g we have na μa = Hσ .
Using this one easily verifies that the mapping is conformally invariant: if we
transform to g = e2ω g, ω ∈ E, then (σ , μb , ρ) transforms according to (3.26).
Using that H = H +na ϒa one calculates that the image of (σ , μb , ρ) (under the
map displayed) transforms by the intrinsic version of (3.26), that is by (3.26)
except where ϒa is replaced by ϒ a = ϒa ∓ na nb ϒb (which on agrees with
dω, the tangential derivative of ω). This signals that the explicit map displayed
in (3.48) descends to a conformally invariant map (3.47). We henceforth use
this to identify N ⊥ with T , and write Proj : T | → T for the orthogonal
projection afforded by N (or using abstract indices AB = δBA ∓ N A NB ).
It follows easily from (3.48) that the tractor metric h on T agrees with the
restriction of the ambient tractor metric h to N ⊥ . In summary we have:
Theorem 3.10 Let (M d , c) be a conformal manifold of dimension d ≥ 4 and
a regular hypersurface in M. Then, with T deonting the intrinsic tractor bundle
of the induced conformal structure c , there is a canonical isomorphism
T → N⊥.
Furthermore the tractor metric of c coincides with the pull-back of the

ambient tractor metric, under this map.
Henceforth we shall simply identify T and N ⊥ .
Remark 3.16 Note that if is minimal in the scale g, that is if H g = 0, then

the isomorphism (3.48) is simply
⎛ ⎞ ⎛ ⎞
σ σ
[N ⊥ ]g ⎝ μb ⎠ → ⎝ μb ⎠ ∈ [T ]g . (3.49)
ρ ρ
Moreover, it is easy to see that one can always find such a minimal scale g
for . Let g = σ −2 g be any metric in c and let ω := ∓sσ H g , where s is a
normalised defining function for (at least in a neighbourhood of ) and H g
has been extended off arbitrarily. Then if ĝ = e2ω g we have that, along ,
H ĝ = H g + na ϒa = H g + na ∇a ω = 0
since by assumption ∇a s = σ −1 na along and s| = 0.
Now we have two connections on T that we may compare, namely

T
the intrinsic tractor connection ∇ , meaning the normal tractor connection
determined by conformal structure (, c), and the projected ambient tractor
v
connection ∇. On U ∈ (T ) the latter is defined by
v
∇a U B := BC (ca ∇c U C ) along ,
where we view U ∈ (T ) as a section of N ⊥ , and make an arbitrary smooth

extension of this to a section of T in a neighbourhood (in M) of . It is then
v
easily verified that ∇ is a connection on T . By construction it is conformally
invariant. Thus the difference between this and the intrinsic tractor connection
is some canonical conformally invariant section of T ∗ ⊗ End(T ).
The difference between the projected ambient and the intrinsic tractor
connections can be expressed using the tractor contorsion Sa B C defined by the
equation
v
∇a V B = ∇ a V B ∓ Sa B C V C (3.50)
B
where V B ∈ (E ) is an intrinsic tractor. The intrinsic tractor contorsion can
be computed explicitly in any scale g to take the form
g
Sa B C = X̄ B Z̄C c Fac − Z̄ B b X̄C Fa b (3.51)
or in other words
SaBC = XBC c Fac , (3.52)
where evidently Fac must be some conformal invariant of hypersurfaces (here

X̄ B and Z̄ B b are standard tractor projectors for and XBC c = 2X̄[B Z̄C] c , which
is conformally invariant). In fact the details of the computation (see [52, 55])
reveal this is the Fialkow tensor (cf. [18])
◦
◦
|L|2
Fab = n−21
Wacbd nc nd + Lab
2
− 2(n−1) gab , (3.53)
◦ ◦ ◦
2 := L c L and n = d − 1. Altogether we have,
where Lab a cb
∇ a V B = B C ∇ a V C ± NB NC ∇ a V C
v
= ∇a V B ∓ NB LaC V C (3.54)
= ∇ a V − Sa B B
CV
C
∓ N LaC V
B C
B
for an intrinsic tractor V B ∈ T . The tractor Gauss formula is therefore
∇ a V B = ∇ a V B ∓ Sa B C V C ∓ NB LaC V C (3.55)
B
for any V B ∈ T . Here we see that the object
LaB := ∇ a NB
is a tractor analogue of the second fundamental form; we shall therefore call it
the tractor shape form.
These results provide the first steps in a calculus for conformal hypersurfaces
that is somewhat analogous to the local invariant calculus for Riemannian
hypersurfaces. In particular it can be used to proliferate hypersurface confor-
mal invariants and conformally invariant operators [55, 32]. We will apply this
calculus to the study of conformal infinities in the final two lectures.
3.8. Lecture 7: Geometry of Conformal Infinity

3.8.1. Geometry of Conformal Infinity and its Embedding
We return now to the study of conformally compact geometries. We will con-
sider in particular those which near the conformal infinity are asymptotically
of constant non-zero scalar curvature. By imposing a constant dilation we may
assume that I 2 approaches ±1.
We begin by observing that the normal tractor is linked, in an essential way,
to the ambient geometry off the hypersurface := Z(σ ).
Proposition 3.18 Let (M d , c, I) be an almost pseudo-Riemannian structure
with scale singularity set = ∅ and I 2 = ±1 + σ 2 f for some smooth (weight
−2) density f . Then by Theorem 3.7 is a smoothly embedded hypersurface
and, with N denoting the normal tractor for , we have N = I| .
Proof For simplicity let us first assume I 2 = ±1 (so f = 0 and the structure is
ASC). As usual let us write σ := h(X, I). By definition
⎛ ⎞
σ
1 g
IA = DA σ = ⎝ ∇a σ ⎠,
d
− d (σ + Jσ )
1
where g ∈ c and ∇ denotes its Levi-Civita connection. Let us write na := ∇a σ .

Along we have σ = 0, therefore
⎛ ⎞
0
g
I| = ⎝ na ⎠,
− d1 σ
along . Clearly then |n|2g = ±1, along , since I 2 = ±1. So na | is a
conformal weight 1 conormal field for .
Next we calculate the mean curvature H = H g in terms of σ . Recall

(d − 1)H = ∇ a na ∓ na nb ∇b na , on . We calculate the right hand side in a
neighbourhood of . Since na = ∇a σ , we have ∇ a na = σ . On the other
hand

1 b 1 b 2 2 2
n n ∇b na = n ∇b (n na ) = n ∇b ±1 + σ σ + Jσ ,
a b a
2 2 d d
where we used that | d1 Dσ |2 = ±1 and so na na = ±1 + d2 σ σ + d2 Jσ 2 . Now

along we have ±1 = na na = na ∇a σ , and so there this simplifies to
1
na nb ∇b na = ± σ .
d
Putting these results together, we have
1 1
(d − 1)H = (d − 1)σ ⇒ H= σ along .
d d
Thus
⎛ ⎞
0
g
I| = ⎝ na ⎠ ,
−H
as claimed. Now note that if we repeat the calculation with I 2 = ±1 + σ 2 f

then the result still holds, as in the calculation this relation was differentiated
just once.
Corollary 3.1 Let (M d , c, I) be an almost pseudo-Riemannian structure with
scale singularity set = ∅, and that is asymptotically Einstein in the sense
that I 2 | = ±1, and ∇a IB = σ faB for some smooth (weight −1) tractor valued
1-form faB . Then is a totally umbilic hypersurface.
Proof The assumptions on the scale tractor I imply that I 2 = ±1 + σ 2 f for
some smooth function f (since we assume c and σ smooth). Thus it follows
from Proposition 3.18 that, along the singularity hypersurface, I agrees with
the normal tractor N. Thus N is parallel along and so, from Proposition
3.17, is totally umbilic.
Note that a hypersurface is totally umbilic if and only if it is conformally

totally geodesic: if is totally umbilic then locally it is straightforward to
g
find a metric g ∈ c so that H g = 0 whence Lab = 0 (this was demonstrated
in Remark 3.16). In this scale any geodesic on the submanifold , with its
induced metric g, is also a geodesic of the ambient (M, g). So the condition of
being totally umbilic is a strong matching of the conformal structures.
In fact for a conformally compact metric that is asymptotically Einstein, as

in Corollary 3.1, there is an even stronger compatibility (involving a higher
order of contact) between the geometry of c and c. First a preliminary result.
Proposition 3.19 Let (M d≥4 , c, I) be an almost pseudo-Riemannian structure
with scale singularity set = ∅, and that is asymptotically Einstein in the
sense that I 2 | = ±1, and ∇a IB = σ 2 faB for some smooth (weight −2) tractor
valued 1-form field faB . Then the Weyl curvature Wab c d satisfies
Wab c d nd = 0, along ,
where nd is the normal field.
Proof Since, along , IB is parallel to the given order we have that the tractor
curvature satisfies
κ ab C D I D = κ ab C D N D = 0 along ,
and from the formulae for κ and N the result is immediate.
Theorem 3.11 Let (M d≥4 , c, I) be an almost pseudo-Riemannian structure
with scale singularity set = ∅, and that is asymptotically Einstein in the
sense that I 2 | = ±1, and ∇a IB = σ 2 faB for some smooth (weight −2) tractor
valued 1-form field faB . Then the tractor connection of (M, c) preserves the
intrinsic tractor bundle of , where the latter is viewed as a subbundle of
the ambient tractors: T ⊂ T . Furthermore the restriction of the parallel
transport of ∇ T coincides with the intrinsic tractor parallel transport of ∇ T .
Proof From Theorem 3.10 the tractor bundle T of (, c) may be identified
with the subbundle N ⊥ in T | , consisting of standard tractors orthogonal to
the normal tractor N. Since N = I| is parallel along this subbundle is
preserved by the ambient tractor connection and the projected ambient tractor
v
connection ∇ coincides with restriction to T of the pull-back of the ambient
tractor connection:
v
∇a = ba ∇bT on T ⊂T.
Thus the result follows from the tractor Gauss formula if the Fialkow tensor
◦
◦
|L|2
Fab = n−21
Wacbd nc nd + Lab
2
− 2(n−1) gab ,
◦
of (3.53) vanishes. Now from Corollary 3.1 is totally umbilic, and so Lab =
0, while Proposition 3.19 states that Wacbd nd = 0 along . Thus Fab = 0.
Remark 3.17 In the case of an almost Einstein manifold Theorem 3.11
can also be seen to follow directly from the general theory of [11], see
Theorem 3.5.
3.9. Lecture 8: Boundary Calculus and Asymptotic Analysis

In the previous lectures we have seen that for conformally compact manifolds
with a non-degenerate conformal infinity there are useful tools for studying the
link between the conformal structure on the conformal infinity and its relation
to the ambient pseudo-Riemannian structure. These showed for example that
conditions like asymptotically ASC and then asymptotically Einstein lead
to a higher order of contact between the ambient and boundary conformal
structures.
Here we show that the same tools, which are canonical to the geometry of
almost-pseudo Riemannian manifolds, lead first to an interesting canonical
calculus along the boundary and then to canonical boundary problems that
can be partly solved by these tools. Further details of the boundary calculus
for conformally compactified manifolds presented in this lecture, and on
applications, can be found in [28, 33, 34, 35].
3.9.1. The Canonical Degenerate Laplacian

On an almost pseudo-Riemannian manifold (M, c, I) there is a canonical
degenerate Laplacian type differential operator, namely
I·D := I A DA .
This acts on any weighted tractor bundle, preserving its tensor type but
lowering the weight:
I·D : E [w] → E [w − 1].
In the following it will be useful to define the weight operator w. On sections
of a conformal density bundle this is just the linear operator that returns the
weight. So if τ ∈ (E[w0 ]) then
w τ = w0 τ .
Then this is extended in an obvious way to weighted tensor or tractor bundles.
So if B is some vector bundle of conformal weight zero then w acts as the zero
operator on its sections and then if β ∈ (B[w0 ]) we have
w β = w0 β.
Now expanding I·D in terms of some background metric g ∈ c, we have
⎛ ⎞
w(d + 2w − 2)
g
I·D = − d1 (σ + Jσ ) ∇ a σ σ ⎝ ∇a (d + 2w − 2) ⎠ .
−( + Jw)
As an operator on any density or weighted tractor bundle of weight w each

occurrence of w evaluates to w. So then
g
w 2w
I D = −σ +(d +2w−2) (∇ a σ )∇a − (σ ) − (d +w−1)σ J (3.56)
d d
on E [w]. Now if we calculate in the metric g+ = σ −2 g, away from the zero
locus of σ , and trivialise the densities accordingly, then σ is represented by 1
in the trivialisation, and we have

g+ 2w(d + w − 1) g+
I·D = − g+
+ J .
d
In particular if g+ satisfies Jg+ = ∓ d2 ( i.e. Scg+ = ∓d(d − 1) or equivalently
I 2 = ±1) then, relabeling d + w − 1 =: s and d − 1 =: n, we have
g+
I·D = − g+ ± s(n − s) . (3.57)
On the other hand, looking again to (3.56), we see that I·D degenerates along
the conformal infinity = Z(σ ) (assumed to be non-empty), and there the
operator is first order. In particular if the structure is asymptotically ASC in the
sense that I 2 = ±1 + σ f , for some smooth (weight −1) density f , then along
I·D = (d + 2w − 2)δn ,
on E [w], where δn is the conformal Robin operator,
g
δn = na ∇ag − wH g ,
of [13, 5] (twisted with the tractor connection); here na is a length ±1 normal
and H g the mean curvature, as measured in the metric g.
3.9.2. A Boundary Calculus for the Degenerate Laplacian

Let (M, c) be a conformal structure of dimension d ≥ 3 and of any signature.
Given σ a section of E[1], write IA for the corresponding scale tractor as usual.
That is IA = d1 DA σ . Then σ = X A IA .
3.9.2.1. The sl(2)-algebra

Suppose that f ∈ E [w], where E
denotes any tractor bundle. Select g ∈ c for
g
the purpose of calculation, and write IA = (σ , νa , ρ) to simplify the notation.
Then using νa = ∇a σ , we have

I·D σ f = (d + 2w) (w + 1)σρf + σ νa ∇ a f + f νa ν a

−σ σ f + 2νa ∇ a f + f σ + (w + 1)Jσ f ,
while

−σ I·D f = −σ (d + 2w − 2) wρf + νa ∇ a f + σ 2 (f + wJf ) .
So, by virtue of the fact that ρ = − d1 (σ + Jσ ), we have
[I·D, σ ] f = (d + 2w)(2σρ + νa ν a )f .
g
Now I A IA = I 2 = 2σρ + νa ν a , whence the last display simplifies to
[I·D, σ ] f = (d + 2w)I 2 f .
Denoting by w the weight operator on tractors, we have the following.
Lemma 3.4 Acting on any section of a weighted tractor bundle we have
[I·D, σ ] = I 2 (d + 2w),
where w is the weight operator.
Remark 3.18 A similar computation to above shows that, more generally,

I · D σ α f − σ α I · Df = σ α−1 α I 2 (d + 2w + α − 1)f ,
for any constant α.
The operator I·D lowers conformal weight by 1. On the other hand, as an
operator (by tensor product) σ raises conformal weight by 1. We can record
this by the commutator relations
[w, I·D] = −I·D and [w, σ ] = σ ,
so with the lemma we see that the operators σ , I·D and w, acting on weighted
scalar or tractor fields, generate an sl(2) Lie algebra, provided I 2 is nowhere
vanishing. It is convenient to fix a normalisation of the generators; we record
this and our observations as follows.
Proposition 3.20 Suppose that (M, c, σ ) is such that I 2 is nowhere vanishing.
Setting x := σ , y := − I12 I·D and h := d + 2w we obtain the commutation
relations
[h, x] = 2x, [h, y] = −2y, [x, y] = h,
of standard sl(2)-algebra generators.
In the case of I 2 = 0 the result is an Inönü–Wigner contraction of the sl(2)-
algebra:
[h, x] = 2x, [h, y] = −2y, [x, y] = 0,
where h and x are as before, but now y = −I·D.
Subsequently g will be used to denote this (sl(2)) Lie algebra of operators.

From Proposition 3.20 (and in concordance with Remark 3.18) follow some
useful identities in the universal enveloping algebra U (g).
Corollary 3.2
[xk , y] = xk−1 k(d + 2w + k − 1) = xk−1 k(h + k − 1)

and (3.58)
[x, yk ] = yk−1 k(d + 2w − k + 1) = yk−1 k(h − k + 1) .
3.9.3. Tangential Operators and Holographic Formulæ

Suppose that σ ∈ (E[1]) is such that IA = d1 DA σ satisfies that I A IA = I 2 is
nowhere zero. As explained in Section 3.6.3, the zero locus Z(σ ) of σ is then
either empty or forms a smooth hypersurface.
Conversely if is any smooth oriented hypersurface then, at least in a
neighbourhood of , there is a smooth defining function s. Now take σ ∈
(E[1]) to be the unique density which gives s in the trivialisation of E[1]
determined by some g ∈ c. It follows then that = Z(σ ) and ∇ g σ is non-zero
at all points of . If ∇ g σ is nowhere null along then is non-degenerate and
I 2 is nowhere vanishing in a neighbourhood of , and we are in the situation
of the previous paragraph. We call such a σ a defining density for (recall our
definition from Lecture 6), and to simplify the discussion we shall take M to
agree with this neighbourhood of . Until further notice σ will mean such a
section of E[1] with = Z(σ ) non-empty and non-degenerate. Note that
has a conformal structure c induced in the obvious way from (M, c) and is a
conformal infinity for the metric g+ := σ −2 g on M \ .
3.9.3.1. Tangential Operators

Suppose that σ is a defining density for a hypersurface . Let P : E [w1 ] →
E [w2 ] be some linear differential operator defined in a neighbourhood of .
We shall say that P acts tangentially (along ) if P ◦ σ = σ ◦ P, where

P : E [w1 − 1] → E [w2 − 1] is some other linear operator on the same
neighbourhood. The point is that for a tangential operator P we have
P( f + σ h) = Pf + σ
Ph.
Thus along the operator P is insensitive to how f is extended off . It is

easily seen that if P is a tangential differential operator then there is a formula
for P, along , involving only derivatives tangential to , and the converse
also holds.
Using Corollary 3.2 we see at once that certain powers of I·D act tangentially
on appropriately weighted tractor bundles. We state this precisely. Suppose
that is a (non-degenerate) hypersurface in a conformal manifold (M n+1 , c),
and σ a defining density for . Then recall = Z(σ ) and I 2 is nowhere
1
zero in a neighbourhood of , where IA := n+1 DA σ is the scale tractor. The
following holds.
Theorem 3.12 Let E be any tractor bundle and k ∈ Z≥1 . Then, for each
k ∈ Z≥1 , along
k
k−n −k − n 1
Pk : E →E given by Pk := − 2 I·D (3.59)
2 2 I
is a tangential differential operator, and so determines a canonical differential
operator Pk : E [ k−n −k−n
2 ]| → E [ 2 ]| .
Proof The Pk are differential by construction. Thus the result is immediate

k
from Corollary 3.2 with
P = − I12 I·D .
3.9.4. The Extension Problems and their Asymptotics

Henceforth we consider an almost pseudo-Riemannian structure (M, c, σ ) with
σ a defining density for a hypersurface and I 2 nowhere zero. We consider
the problem of solving, off asymptotically,
I·Df = 0 ,
for f ∈ (E [w
0 ]) and some given weight w0 . For simplicity we henceforth
calculate on the side of where σ is non-negative, so effectively this amounts
to working locally along the boundary of a conformally compact manifold.
3.9.4.1. Solutions of the First Kind

Here we treat an obvious Dirichlet-like problem where we view f | as the
initial data. Suppose that f0 is an arbitrary smooth extension of f | to a section
of E [w0 ] over M. We seek to solve the following problem:
Problem 3.1 Given f | , and an arbitrary extension f0 of this to E [w0 ] over M,
find fi ∈ E [w0 − i] (over M), i = 1, 2, · · · , so that
f () := f0 + σ f1 + σ 2 f2 + · · · + O(σ +1 )
solves I·Df = O(σ ), off , for ∈ N ∪ ∞ as high as possible.
Remark 3.19 For i ≥ 1 we do not assume that the fi are necessarily non-
vanishing along . Also, note that the stipulation ‘given f | , and an arbitrary
extension f0 of this’ can be rephrased as ‘given f0 in the space of sections with

a fixed restriction to (denoted f | )’.
We write h0 = d + 2w0 so that hf0 = h0 f0 , for example. The existence or

not of a solution at generic weights is governed by the following result.
Lemma 3.5 Let f () be a solution of Problem 3.1 to order ∈ Z≥0 . Then
provided = h0 − 2 = n + 2w0 − 1 there is an extension
f (+1) = f () + σ +1 f+1 ,
unique modulo σ +2 , which solves
I·Df (+1) = 0 modulo O(σ +1 ).
If = h0 − 2 then the extension is obstructed by P+1 f0 | , where P+1 =

(− I12 I·Df )+1 is the tangential operator on densities of weight w0 given by
Theorem 3.12.
Proof Note that I·Df = 0 is equivalent to − I12 I·Df = 0 and so we can recast
this as a formal problem using the Lie algebra g = x, y, h from Proposition
3.20. Using the notation from there
yf (+1) = yf () − x ( + 1)(h + )f+1 + O(x+1 ).

Now hf+1 = h0 − 2( + 1) f+1 , thus
yf (+1) = yf () − x ( + 1)(h0 − − 2)f+1 + O(x+1 ). (3.60)
By assumption yf () = O(x ), thus if = h0 − 2 we can solve yf (+1) =

O(x+1 ) and this uniquely determines f+1 | .
On the other hand if = h0 − 2 then (3.60) shows that, modulo O(x+1 ),

yf () = y f () + x+1 f+1 ,
regardless of f+1 . It follows that the map f0 → x− yf () is tangential and
x− yf () | is the obstruction to solving yf (+1) = O(x+1 ). By a simple
induction this is seen to be a non-zero multiple of y+1 f0 | .
Thus by induction we conclude the following.

Proposition 3.21 For h0 ∈ / Z≥2 Problem (3.1) can be solved to order = ∞.
For h0 ∈ Z≥2 the solution is obstructed by [Ph0 −1 f ]| ; if, for a particular f ,
[Ph0 −1 f ]| = 0 then there is a family of solutions to order = ∞ parametrised
by sections fh0 −1 ∈ E [−d − w0 + 1]| .
Appendix: Conformal Killing Vector Fields

and Adjoint Tractors
Here we discuss conformal Killing vector fields and the prolongation of the
conformal Killing equation which is given by a connection on a conformal
tractor bundle called the adjoint tractor bundle. Although the adjoint bundle is
not a new tractor bundle per se (being the bundle of skew endomorphisms
of E A , naturally isomorphic to E[AB] ), it is a very important tractor bundle
and is worthy of our further attention; we discuss briefly the connection of
the conformal Cartan bundle with the adjoint tractor bundle and its calculus.
We finish by applying the adjoint tractor calculus and the prolongation of
the conformal Killing equation to Lie derivatives of tractors with respect to
conformal Killing vector fields.
A.1. The Conformal Cartan Bundle and the Adjoint

Tractor Bundle
In the lectures we have been discussing tractor calculus on the so-called
standard tractor bundle on a conformal manifold (M, c), which can be seen
as the associated bundle G ×P Rp+1,q+1 to the P-principal conformal Cartan
bundle G → M. We have not discussed the construction of the conformal
Cartan bundle in the lectures because it is easier to construct the standard
tractor bundle directly from the underlying conformal structure. Indeed, having
understood the conformal standard tractor bundle one could simply define the
conformal Cartan bundle as an adapted frame bundle for the standard tractor
bundle T ; from this point of view the conformal Cartan connection simply
encodes the tractor parallel transport of the frames, cf. (3.40). Although for
calculational purposes the standard tractor approach presented in the lectures is
optimal, we have seen (e.g. in Section 3.6.3.1) that the Cartan geometric point
of view can afford additional insights. This is also the case when it comes to
discussing the adjoint tractor bundle.
If P ⊂ G = O(p, q) is the stabiliser of a null ray in Rp+1,q+1 then P
contains a subgroup which may be identified with CO(p, q). The conformal
Cartan bundle of (M, c) is then easily defined as the extension of the conformal
orthogonal frame bundle G0 → M to a principal P-bundle G → M
corresponding to the inclusion CO(p, q) ⊂ P. The total space of this bundle
is then G0 ×CO(p,q) P.
The conformal Cartan connection ω is a 1-form on G taking values in the
Lie algebra of G, which can be written as a (vector space) direct sum
∗
Rd ⊕ co(p, q) ⊕ Rd
where d = p + q. The Rd component of ω is simply the trivial extension of

the soldering form of G0 (which is tautologically defined on any reduction of
the frame bundle of a manifold) to G; this component corresponds to the terms
−μa and +gab ρ in the formula (3.22) for the standard tractor connection. The
next component arises from noting that sections of the bundle G → G0 are
in one-to-one correspondence with affine connections on M which preserve
the conformal metric g (Weyl connections); the co(p, q) component of ω is
the 1-form γ on G whose pull-back to G0 by any section is the connection 1-
form for the corresponding Weyl connection. In order to obtain the formula for
the tractor connection from that of the Cartan connection one must first pull
everything down from G to G0 using a Weyl connection, allowing one to break
tractors (in associated bundles to G) up into weighted tensors (in associated
bundles to G0 ). If one does this using the Levi-Civita connection ∇ of a metric
g then one can see that the component γ of ω gives rise to the three terms
involving ∇ in (3.22). The final component of ω gives rise to the two terms
involving the Schouten tensor in (3.22) and is determined by the other two
components of ω and an algebraic condition on the curvature
dω + ω ∧ ω
of ω (which can be identified with the tractor curvature). For more details see
the first chapter of [12].
The conformal Cartan connection has the following three properties:
• ωu : Tu G → G × g is a linear isomorphism for each u ∈ G;
• ω is P-equivariant, that is (rp )∗ ω = Ad(p−1 ) ◦ ω for all p ∈ P (where rp
denotes the right action of p on G);
• ω returns the generators of fundamental vector fields, that is

d
ω u · exp(tX) = X
dt 0
for all u ∈ G and X ∈ p = Lie(P).
These properties more generally define what is called a Cartan connection of
type (G, P) on a P-principal bundle G. In the model case the Lie group G can
be seen as a P-principal bundle over the model space G/P and the Maurer-
Cartan 1-form ωMC , which evaluates left invariant vector fields at the identity,
is a Cartan connection which has vanishing curvature by the Maurer–Cartan
structure equations
dωMC + ωMC ∧ ωMC = 0.
The adjoint tractor bundle A → M of a conformal manifold (M, c) is
the associated bundle to the conformal Cartan bundle G corresponding to
the adjoint representation of G (restricted to P), that is A = G ×P g. Since

g = so(p+1, q+1) can be identified with the skew-symmetric endomorphisms
of Rp+1,q+1 the adjoint tractor bundle can similarly be identified with the
bundle of skew-symmetric endomorphisms of the standard tractor bundle T
(with respect to the tractor metric). Clearly then we may identify A with E[AB]
by lowering a tractor index. An adjoint tractor LA B can be written in terms of
the direct sum decomposition of T (and hence A) as
⎛ ⎞
−ν −lb 0
⎝ −ρ a μa b la ⎠ (3.61)
0 ρb ν
where μab = μ[ab] and the matrix acts on standard tractors from the left, as
the tractor curvature does in (3.32) (the tractor curvature is in fact an adjoint
tractor valued 2-form). It is easy to see that the two appearances of la make up
the ‘top slot’ of LA B , so that there is an invariant projection from A to TM
that takes LA B to la .
By writing an adjoint tractor LA B in terms of the splitting tractors (X A , ZaA ,
Y A ) corresponding to a choice of metric g one can easily obtain the formula
for the tractor connection acting on LA B using (3.40). If LA B is given in matrix
form (w.r.t. g) by (3.61) we then have
⎛ ⎞
∗ ∗ 0
g
∇a LB C = ⎝ ∗ gbb (∇a μb c − 2Pa[b lc] + 2ga[b ρc] ) ∇a lb + μa b + νδab ⎠
0 ∇a ρc − Pba μbc + νPac
∇a ν − Pba lb − ρa
(3.62)
where the entries marked with a ∗ are determined by skew-symmetry.
A.2. Prolonging the Conformal Killing Equation

Note that if L is a parallel adjoint tractor, and l = (L), then from the above
display we must have
∇a lb + μab + νgab = 0
where μab is skew. This implies that ∇(a lb)o = 0, in other words la is a
conformal Killing vector field. A conformal Killing vector field ka is a non-
zero solution of the conformally invariant equation
∇(a kb)o = 0,
where kb = gbc kc . Geometrically this equation says that the local flow of k
preserves any metric g ∈ c up to a conformal factor, equivalently, the Lie
derivative of any metric g ∈ c with respect to k is proportional to g. It is natural
to ask whether there is a one-to-one correspondence between conformal Killing

vector fields and (non-zero) parallel adjoint tractor fields – the answer to this
question turns out to be no, except for on the flat model (where one can use
this correspondence to easily write the (d + 2)(d + 1)/2-dimensional space of
Killing vector fields explicitly).
One can however construct a different conformally invariant connection on
the adjoint tractor bundle which does prolong the conformal Killing equation,
that is for which parallel sections are in one-to-one correspondence with
solutions. One can obtain this system directly by (fixing g and ∇ = ∇ g and
then) writing the equation ∇(a kb)o = 0 as
∇a kb = μab + νgab ,
introducing the new variables μab ∈ (E[ab] [2]) and ν ∈ (E). Beyond this
the key step is to introduce the fourth variable
ρa = ∇a ν + Pab kb ∈ (Ea )
(rather than ρa = ∇a ν), cf. (3.62). The remainder of the prolongation process
simply involves taking covariant derivatives of the above two displays and then
skew-symmetrising over certain pairs of indices in them in order to bring out
curvature terms (as well as using Bianchi identities to simplify expressions);
from these ‘differential consequences’ of the above two displays one may
derive expressions for ∇a μbc and ∇a ρb which are linear in the other three
variables (respectively). (The way to accomplish this process efficiently is to
suppose you have a flat Levi-Civita connection first and go through the process
to obtain both expressions, then go back through the same steps and take into
account the non-vanishing curvature for the general case.) The result is the
following system of differential equations
∇a kb = νgab + μab
∇a μbc = −2Pa[b kc] − 2ga[b ρc] + Wdabc kd
∇a ν = ρa − Pab kb
∇a ρb = −Pca μbc − Pab ν − Ccab kc
where Cabc = 2∇[a Pbc] is the Cotton tensor.
From the above system we can see that if we define the connection ∇˜ on
A by
˜ = ∇L + i(L) κ for all L ∈ (A)
∇L (3.63)
˜
then there is a one-to-one correspondence between (non-zero) ∇-parallel
sections of A and conformal Killing vector fields on (M, c). To check this we
simply calculate (in a scale g) that
⎛ ⎞ ⎛ ⎞
∗ ∗ 0 ∗ ∗ 0

˜
∇a ⎝ ∗ μb c l ⎠ = ⎝ ∗ gbb (∇a μb c − 2Pa[b lc] + 2ga[b ρc] )
b ∇a lb + μa b + νδab ⎠
0 ρc ν 0 ∇a ρc − Pba μbc + νPac ∇a ν − Pba lb − ρa
⎛ ⎞
∗ ∗ 0
+ ⎝ ∗ Wda b c ld 0 ⎠
0 −Cdac ld 0
and observe that by setting the right hand side equal to zero (and substituting
ka = −la ) we recover our prolonged system for the conformal Killing
equation. Note that it is possible to take a much more abstract and theoretical
approach to obtaining this prolonged system and the corresponding connection
∇˜ on A (see, e.g. [6, 12]). From the general theory (or direct observation) we
also have the invariant linear differential operator L : TM → A which takes a
vector field la on M to the adjoint tractor LA B given in a scale g by (3.61) with
μab = −∇[a lb] , ν = − d1 ∇a la and ρa = ∇a ν − Pab lb . Clearly ◦ L = idTM
and consequently L is referred to as a differential splitting operator.
A.3. The Fundamental Derivative and Lie

Derivatives of Tractors
Let V be a representation of P and let V = G ×P V. If a function ṽ : G → V
satisfies
ṽ(u · p) = p−1 · ṽ(u) for all u ∈ G, p ∈ P
then we say that ṽ is P-equivariant. Observe that if ṽ : G → V is a smooth
P-equivariant map then the map v : M → V which takes x to [(u, v(u))] for
some u ∈ Gx defines a smooth section of V → M (being independent of the
choice of u ∈ Gx for each x). It is easy to see that sections of V → M are
in one-to-one correspondence with such functions. (We have used this already
in discussing conformal densities where R+ replaces P and Q replaces G, see
Section 3.3.4.) Now observe that the conformal Cartan connection ω allows us
to view smooth P-equivariant functions L̃ : G → g as smooth vector fields VL
on G (since ωu : Tu G → g is an isomorphism for each u ∈ G and ω is smooth).
From the properties of the Cartan connection it is not hard to see that any such
vector field VL must be P-invariant, that is,
(rp )∗ VL = VL for all p∈P
where rp : G → G denotes the right action of p ∈ P. Moreover, the
Cartan connection gives a one-to-one correspondence between P-equivariant
functions L̃ : G → g and P-invariant vector fields on G. Combining this with
the previous observation we see that one may naturally identify (A) with the
space X(G)P of P-invariant vector fields on G.
The observations of the preceding paragraph allow us to define a new

canonical differential operator acting on sections of any vector bundle V
associated with the conformal Cartan bundle G. Fix L ∈ (A) and let v ∈
(V) = (G ×P V), then VL ṽ = dṽ(VL ) is again a P-equivariant function from
G to V and thus defines a section DL v of V. Thus for each section L ∈ (A)
we have a first order differential operator DL : V → V. It is easy to see that
Df L = fDL for any f ∈ C∞ (M) so that we really have a (first order) differential
operator taking sections of V to sections of A∗ ⊗ V. The operator
D : V → A∗ ⊗ V
defined in this way is referred to as the fundamental derivative (or fundamental

D operator). This operator was introduced by Čap and Gover in [9].
Now if k is a conformal Killing vector field on (M, c) and v is a section of
V = G ×P V then we may talk about the Lie derivative of v with respect to k;
since V → M is a natural bundle in the category of conformal manifolds (with
diffeomorphisms as maps) one may pull the section v of V back by the (local)
flow of k and define the Lie derivative by

d
Lk v = (Fltk )∗ (v).
dt t=0
Notice that if L is a section of the adjoint tractor bundle then VL ṽ (which
gives the P-equivariant function corresponding to DL v) may be written as the
Lie derivative LVL ṽ, it should not come as a total surprise then that the Lie
derivative and the fundamental derivative are connected. In fact, one has that
Lk v = DL(k) v
for any conformal Killing vector field k on (M, c) and any section v of a natural
bundle V = G ×P V (this is proven in [14]).
Using the results of [9] (and carefully comparing sign conventions) we
have that
g
DL τ = ∇l τ + wντ (3.64)
for sections τ of the density bundle E[w] (which can be thought of as an

associated bundle to G0 and hence to G), where L is given in the scale g by
(3.61). We also have that
T
DL V = ∇(L) V − L(V)
for all L ∈ (A) and V ∈ (T ). The operator DL is easily seen to satisfy the
Leibniz property and hence for any tractor field T A···B C···D we have
T
DL T A···B C···D = ∇(L) T A···B C···D − LA A T A ···B C···D − · · · − LB B T A···B C···D

+ LC C T A···B C ···D + · · · + LD D T A···B C···D .
From all of this we can finally write down an explicit formula for the Lie
derivative of a standard tractor field V A in terms of ‘slots’: if k is a conformal
g
Killing vector field on (M, c) and V A = (σ , μa , ρ) then
L k V A = k b ∇b V B + K A B V B
⎛ ⎞ ⎛ ⎞⎛ ⎞
∇b σ − μb −ν kb 0 σ
g
= kb ⎝ ∇b μa + ρδba + σ Pab ⎠ + ⎝ −ρ a μa b −ka ⎠ ⎝ μb ⎠
∇b ρ − Pab μa 0 ρb ν ρ
⎛ ⎞ ⎛ ⎞
kb ∇b σ − νσ Lk σ
g
= ⎝ k ∇b μ + μa b μb − σ ∇ a ν ⎠ = ⎝ Lk μa − σ ∇ a ν ⎠
b a
kb ∇b ρ + νρ + μa ∇a ν Lk ρ + μa ∇a ν
where K = L(−k) and we have used that ρa = ∇a ν+Pab kb and that Lk = D−K
on densities so that by (3.64) we have Lk σ = ∇k σ − νσ , Lk ρ = ∇k ρ + νρ, and
Lk μa = (kb ∇b μa − μb ∇b ka ) + νμa
= kb ∇b μa − μb (μb a + νδba ) + νμa
= kb ∇b μa + μa b μb
since μa has conformal weight −1.
Note that one can also, of course, calculate the expression for Lk on densities
and on standard tractor fields directly from the definition (which was done in
[14]) by looking at what you get when you pull back densities and standard
tractors by the local flow of k. As a check of our above formula for Lk V A we
observe that in the case where k is a Killing vector field for g then the flow
of k preserves g and hence also preserves the splitting tractors X A , ZaA and Y A
so that Lk X A = 0, Lk ZaA = 0 and Lk Y A = 0; thus by the Leibniz property
and linearity one immediately has that if V A = σ X A + μa ZaA + ρY A then
Lk V A = (Lk σ )X A + (Lk μa )ZaA + (Lk ρ)Y A which is consistent with our above
g
formula for Lk V A since Lk g = 0 forces ν = d1 ∇a ka to be zero.
Remark 3.20 On vector fields the fundamental derivative acts according to
DL vb = la ∇a vb + (μa b + νδab )va
where L is given in the scale g by (3.61) and ∇ = ∇ g . Thus if k is a conformal
Killing vector field then applying Lk = DL(k) on vector fields simply returns
the usual formula for the Lie derivative in terms of a torsion free (in this case
Levi-Civita) connection:
Lk vb = ka ∇a vb − (∇a kb )va .
Similarly, using the Leibniz property of the fundamental derivative, we

have that
DL T b···c d···e
= la ∇a T b···c d···e − (μa b + νδab )T a···c d···e − · · · − (μa c + νδac )T b···a d···e
+ (μd a + νδda )T b···c a···e + · · · + (μe a + νδea )T b···c d···a ,
and again applying Lk = DL(k) for a conformal killing vector field k
simply yields the standard formula for the Lie derivative. These observations
further demonstrate the consistency of our claims. What’s more, we may
now calculate the fundamental derivative of any weighted tensor-tractor field
T b···c d···e B···C D···E . For instance, by writing gab as σ 2 gab and using the Leibniz
property one may easily show that
DL gab = 0
for all L ∈ (A).
A.3.1. Static and Stationary Spacetimes

As an application of the above, we observe that if g ∈ c is an Einstein metric
with corresponding scale tractor I and k is a Killing vector field for g with
K = L(−k) then K(I) = 0. This follows from the fact Lk I = 0 since the flow
of k preserves g (and hence also I), and from the above formula for the Lie
derivative of I,
Lk I A = ∇k I A + KA B I B = KA B I B ,
since I is parallel. Moreover, if k is hypersurface orthogonal (i.e. its orthogonal
distribution is integrable) then it is easy to see that K[AB KC]D X D = 0 so that
KAB is simple and can be written as 2v[A KB] where KB = X A KAB and vA K A =
0; on top of this, from KAB I B = 0 we obtain KA I A = 0 and hence also vA I A =
0. These observations form the starting point for the development of conformal
tractor calculus adapted to static and stationary spacetimes. In particular, we
note that in the case where (M, g, k) is a static spacetime then KA = uNA where
NA is the normal tractor to each of the spacelike hypersurfaces in the foliation
given by k⊥ ⊂ TM and u is the so called static potential (if we trivialise the
density bundles using g); thus KA I A = 0 and KA vA = 0 imply that both I A and
vA lie in the intrinsic tractor bundle of the foliating spacelike hypersurfaces and
one can carry out dimensional reduction using tractors. One can in fact still
carry out dimensional reduction in the stationary case by identifying tractors
which are Lie dragged by k and orthogonal to KB = X A KAB with conformal
tractors on the manifold of integral curves of k (see [14] for further details in
both cases).
References
[1] M. Atiyah, R. Bott, and V.K. Patodi, On the heat equation and the index theorem,
Inventiones Mathematicae, 19 (1973), 279–330.
[2] T.N. Bailey, M.G. Eastwood, and A.R. Gover, Thomas’s structure bundle for
conformal, projective and related structures, Rocky Mountain J. Math., 24 (1994),
1191–1217.
[3] T.N. Bailey, M.G. Eastwood, and C.R. Graham, Invariant theory for conformal
and CR geometry, Annals of Mathematics, 139 (1994), 491–552.
[4] T. Branson, Conformal structure and spin geometry, in Dirac Operators: Yesterday
and Today, (J.-P. Bourguignon, T. Branson, A. Chamseddine, O. Hijazi and
R. Stanton, Eds.), International Press, (2005), 163–191.
[5] T. Branson and A.R. Gover, Conformally invariant non-local operators, Pacific
Journal of Mathematics, 201 (2001), 19–60.
[6] A. Čap, Infinitesimal automorphisms and deformations of parabolic geometries,
J. Eur. Math. Soc., 10 (2008), 415–437.
[7] A. Čap and A.R. Gover, Standard tractors and the conformal ambient metric
construction, Annals Global Anal.Geom., 24 (2003), 231–259.
[8] A. Čap and A.R. Gover, Tractor bundles for irreducible parabolic geometries, in
S.M.F. Colloques, Seminaires & Congres, 4 (2000), 129–154.
[9] A. Čap and A.R. Gover, Tractor calculi for parabolic geometries, Trans. Amer.
Math. Soc., 354 (2002), 1511–1548.
[10] A. Čap, A.R. Gover, and M. Hammerl, Projective BGG equations, algebraic sets,
and compactifications of Einstein geometries, J. London Math. Soc., 86 (2012),
433–454.
[11] A. Čap, A.R. Gover, and M. Hammerl, Holonomy reductions of Cartan geometries
and curved orbit decompositions, Duke Math. J., 163 (2014), 5, 1035–1070.
[12] A. Čap, J. Slovák, Parabolic Geometries I: Background and General Theory,
Mathematical Surveys and Monographs. American Mathematical Society, Provi-
dence, RI, 2009.
[13] P. Cherrier, Problèmes de Neumann non linéaires sur les variètès riemanniennes,
J. Funct. Anal., 57 (1984), 154–206.
[14] S. Curry, Conformal Tractor Calculus for Stationary and Static Spacetimes, MSc
thesis, University of Auckland, 2012.
[15] A.J. Di Scala and G. Manno, On the extendability of parallel sections of linear
connections, Ann. Mat. Pura Appl. (4), 195 (2016), 1237–1253.
[16] P.A.M. Dirac, The electron wave equation in de-Sitter space, Ann. Math., 36
(1935), 657–669.
[17] C. Fefferman and C.R. Graham, The Ambient Metric, Annals of Mathematics
Studies, 178. Princeton University Press, 2012. Available online: http://arxiv.org/
pdf/0710.0919v2.pdf
[18] A. Fialkow, Conformal geometry of a subspace, Trans. Amer. Math. Soc., 56
(1944), 309–433.
[19] J. Frauendiener, Conformal infinity, Living Rev. Relativity, 7 (2004), 1. [Online
Aricle]: cited 21/04/2014, http://www.livingreviews.org/lrr-2004-1
[20] J. Frauendiener and G.A.J. Sparling, Local twistors and the conformal field
equations, J. Math. Phys., 41 (2000), 437–443.
[21] H. Friedrich, Conformal Einstein evolution, in Proceedings of the Tubin-

gen Workshop on the Conformal Structure of Space-times, (H. Friedrich and
J. Frauendiener, Eds.), Springer Lecture Notes in Physics, 604 (2002), 1–50.
[22] R. Geroch, Local characterization of singularities in general relativity, J. Math.
Phys., 9 (1968), 450–465.
[23] R. Geroch, Asymptotic structure of space-time, in Asymptotic Structure of Space-
Time, (F.P. Esposito and L. Witten, Eds.), Springer, (1977), 1–105.
[24] A.R. Gover, Almost Einstein and Poincare–Einstein manifolds in Riemannian
signature, J. Geometry and Physics, 60 (2010), 182–204.
[25] A.R. Gover, Aspects of parabolic invariant theory, Supplemento ai Rendiconti del
Circolo Matematico di Palermo, Serie II, 59 (1999), 25–47.
[26] A.R. Gover, Conformal de Rham Hodge theory and operators generalising the
Q-curvature, Supplemento ai Rendiconti del Circolo Matematico di Palermo,
Serie II, 75 (2005), 109–137.
[27] A.R. Gover, Invariant theory and calculus for conformal geometries, Advances in
Mathematics, 163 (2001), 206–257.
[28] A.R. Gover, E. Latini, and A. Waldron, Poincare–Einstein holography for forms
via conformal geometry in the bulk, Mem. Amer. Math. Soc., 235 (2014), 1–101.
[29] A.R. Gover and P. Nurowski, Obstructions to conformally Einstein metrics in n
dimensions, J. Geom. Phys., 56 (2006), 450–484.
[30] A.R. Gover and L.J. Peterson, Conformally invariant powers of the Laplacian,
Q-curvature, and tractor calculus, Communications in Mathematical Physics, 235
(2003), 339–378.
[31] A.R. Gover and L.J. Peterson, The ambient obstruction tensor and the conformal
deformation complex, Pacific Journal of Mathematics, 226 (2006), 309–351.
[32] A.R. Gover and Y. Vyatkin, forthcoming.
[33] A.R. Gover and A. Waldron, Boundary calculus for conformally compact mani-
folds, Indiana Univ. Math. J., 63 (2014), 119–163.
[34] A.R. Gover and A. Waldron, Generalising the Willmore equation: sub-
manifold conformal invariants from a boundary Yamabe problem, preprint,
arXiv:1407.6742v1 [hep-th], (2014).
[35] A.R. Gover and A. Waldron, Conformal hypersurface geometry via a boundary
Loewner-Nirenberg-Yamabe problem, arXiv:1506.02723 [math.DG] (2015).
[36] C.R. Graham, R. Jenne, L.J. Mason, and G.A.J. Sparling, Conformally invariant
powers of the Laplacian. I. Existence, J. London Math. Soc., 46 (1992), 557–565.
[37] C.R. Graham and T. Willse, Subtleties concerning conformal tractor bundles,
Cent. Eur. J. Math., 10 (2012), 1721–1732.
[38] D.H. Grant, A Conformally Invariant Third Order Neumann-Type Operator for
Hypersurfaces, MSc thesis, University of Auckland, 2003.
[39] C. Kozameh, E.T. Newman, K.P. Tod, Conformal Einstein spaces, Gen. Relativity
Gravitation, 17 (1985), 343–352.
[40] C. LeBrun, Ambitwistors and Einstein’s equations, Class. Quant. Grav., 2 (1985),
555–563.
[41] F. Leitner, Conformal Killing forms with normalisation condition, Rend. Circ.
Mat. Palermo Suppl. No. 75 (2005), 279–292.
[42] C. Lübbe, A conformal extension theorem based on null conformal geodesics,
J. Math. Phys., 50, 112502, (2009).
[43] C. Lübbe, The conformal field equations and the tractor formalism in general
relativity, conference presentation, BritGrav 10, Dublin, 7th April 2010. Slides
available online: http://www.dcu.ie/∼nolanb/Luebbe.pdf
[44] C. Lübbe and K.P. Tod, An extension theorem for conformal gauge singularities,
J. Math. Phys., 50, 112501, (2009).
[45] R.S. Palais, Seminar on the Atiyah-Singer index theorem. With contributions by
M. F. Atiyah, A. Borel, E. E. Floyd, R. T. Seeley, W. Shih and R. Solovay. Annals
of Mathematics Studies, No. 57, Princeton University Press, 1965.
[46] S.M. Paneitz, A quartic conformally covariant differential operator for arbitrary
pseudo-Riemannian manifolds (Summary), SIGMA, 4 (2008), Paper 036, 3 pp.
[47] R. Penrose, Asymptotic properties of fields and space-times, Phys. Rev. Lett., 10
(1963), 66–68.
[48] R. Penrose, The light cone at infinity, in Relativistic Theories of Gravitation,
(L. Infeld, Ed.), Pergamon Press, Oxford, UK, (1964), 369–373.
[49] R. Penrose, Structure of space-time, in Battelle Rencontres, (DeWitt, C.M., and
Wheeler, J.A., Eds.), W.A. Benjamin, Inc., New York, NY, USA, (1968), 121–235.
[50] R. Penrose, Zero rest-mass fields including gravitation: asymptotic behaviour,
Proc. R. Soc. Lond., A284 (1965), 159–203.
[51] R. Penrose, M. MacCallum, Twistor theory: an approach to the quantisation of
fields and space-time, Physics Reports (Section C of Physics Letters), 6 (1972),
241–316.
[52] R. Stafford, Tractor calculus and invariants for conformal submanifolds, MSc
thesis, University of Auckland, 2006.
[53] P. Stredder, Natural differential operators on Riemannian manifolds and represen-
tations of the orthogonal and special orthogonal groups, J. Differential Geom., 10
(1975), 647–660.
[54] O. Veblen, A conformal wave equation, Proc. Nat. Acad. Sci. USA, 21 (1935),
484–487.
[55] Y. Vyatkin, Manufacturing Conformal Invariants of Hypersurfaces, PhD thesis,
University of Auckland, 2013.
[56] R.M. Wald, General Relativity, University of Chicago Press, 1984.
SNC & ARG: Department of Mathematics, The University of Auckland, Private Bag
92019, Auckland 1142, New Zealand
4
An Introduction to Quantum Field Theory
on Curved Spacetimes
Christian Gérard
4.1. Introduction
The purpose of these notes is to provide an introduction to some recent aspects
of quantum field theory on curved spacetimes, emphasizing its relations with
partial differential equations and microlocal analysis.
4.1.1. Quantum Field Theory

Quantum field theory arose from the need to unify quantum mechanics with
special relativity. However, trying to treat the two basic relativistic field
equations, the Klein–Gordon equation
∂t2 φ(t, x) − x φ(t, x) + m2 φ(t, x) = 0,
and the Dirac equation
γ 0 ∂t ψ(t, x) + γ i ∂xi ψ(t, x) − mψ(t, x) = 0,
(where the γ i are the Dirac matrices) in a way parallel to the non-relativistic
Schroedinger equation
i
∂t ψ(t, x) − x ψ(t, x) + iV(x)ψ(t, x) = 0
2m
leads to difficulties (see e.g. [3]). For the Klein–Gordon equation, there exists
a conserved scalar product

φ1 |φ2 = i ∂t φ 1 (t, x)φ2 (t, x) − φ 1 (t, x)φ2 (t, x)dx
R3
which is however not positive definite, hence cannot lead to a probabilistic
interpretation. However, we do have
φ|i∂t φ ≥ 0, (positivity of the energy).
171
172 C. Gérard
For the Dirac equation the situation is the opposite: the conserved scalar
product

ψ1 |ψ2 = ψ 1 (t, x) · ψ2 (t, x)dx
R3
is positive, but
ψ|i∂t ψ is indefinite.
The reason behind these difficulties is that, although all these equations are
partial differential equations, their nature is very different: the Klein–Gordon
and Dirac equations are classical equations, while the Schroedinger equation
is a quantum equation, obtained by quantizing the classical Newton equation
ẍ(t) = −∇x V(x(t)), x ∈ Rn
or equivalently the Hamilton equations

ẋ(t) = ∂ξ h(x(t), ξ(t)),
ξ̇(t) = −∂x h(x(t), ξ(t))
for the classical Hamiltonian
1 2
ξ + V(x).
h(x, ξ ) =
2
We denote by X = (x, ξ ) the points in T ∗ Rn and introduce the coordinate
functions
q : X → x, p : X → ξ .
If (t) : T ∗ Rn → T ∗ Rn is the flow of Hh and q(t) := q ◦ (t), p(t) := p ◦ (t)
then
∂t q(t) = p(t),
∂t p(t) = −∇V(q(t)),
where {·, ·} is the Poisson bracket. Note that
{pj (t), qk (t)} = δjk , {pj (t), pk (t)} = {qj (t), qk (t)} = 0.
To quantize the Liouville equation means to find a Hilbert space H and func-
tions R t → p(t), q(t) with values in self-adjoint operators on H such that
[pj (t), iqk (t)] = δjk 1l, [pj (t), ipk (t)] = [qj (t), iqk (t)] = 0, ∂t q(t) = p(t),
∂t p(t) = −∇V(q(t)).
The last two equations are called Heisenberg equations. The solution is as
follows.
An Introduction to Quantum Field Theory on Curved Spacetimes 173
(1) Find operators p, q satisfying

[pj , iqk ] = δjk 1l, [pj , ipk ] = [qj , iqk ] = 0.
(2) Construct the self-adjoint operator on H
1 2
H= p + V(q).
2
(3) Then
q(t) := eitH qe−itH , p(t) := eitH pe−itH
solve Heisenberg equations.
The Stone–von Neumann theorem says that there is no choice in Step (1):
modulo some technical conditions and multiplicity one has only one choice,
up to unitary equivalence:
H = L2 (Rn ), q = x, p = i−1 ∇x .
Then H = − 12 +V(x) is the Schroedinger operator. Step (2) is then a standard
problem in the theory of self-adjoint operators, and Step (3) is straightforward.
The Klein–Gordon equation is also a Hamiltonian equation, though with an
infinite dimensional phase space, which can be taken for example as C0∞ (Rd )⊕
C0∞ (Rd ). The classical Hamiltonian is then

1
h(ϕ, π ) := π 2 (x) + |∇x ϕ(x)|2 + m2 ϕ 2 (x)dx,
2 Rd
for the linear case, or

1
h(ϕ, π ) := π 2 (x) + |∇x ϕ(x)|2 + m2 ϕ 2 (x) + ϕ n (x)dx,
2 Rd
for some nonlinear version. Here the symbols ϕ(x), π(x) are (coordinate)
functions, parametrized by a point x ∈ Rd , on the space of smooth solutions
of the Klein–Gordon equation, with compactly supported Cauchy data. If φ is
such a solution then
ϕ(x)(φ) := φ(0, x), π(x)(φ) := ∂t φ(0, x).
It is well-known that these are symplectic coordinates, that is
{ϕ(x), ϕ(x )} = {π(x), π(x )} = 0, {π(x), φ(x )} = δ(x, x ), ∀ x, x ∈ Rd .
One would like to follow the same path and consider families of operators on
a Hilbert space H, ϕ(x), π(x), x ∈ Rd such that
[π(x), iϕ(x )] = δ(x − x )1l, [ϕ(x), iϕ(x )] = [π(x), iπ(x )] = 0, ∀ x, x ∈ Rd .
174 C. Gérard
The fundamental difference with non-relativistic quantum mechanics is that,

since the phase space is infinite dimensional, the Stone–von Neumann theorem
cannot be applied anymore: there exists an infinite number of inequivalent
representations of commutation relations.
In other words, when one tries to quantize a classical field equation, the
Hilbert space has to be constructed together with the quantum Hamiltonian:
one cannot work on our familiar Hilbert space and then use tools from operator
theory to construct the quantum Hamiltonian.
This is the reason why the rigorous construction of quantum field theory
models is so difficult, except for non-interacting theories. For interacting
theories this has been achieved only in two and three spacetime dimensions;
see the construction of the P(ϕ)2 and ϕ34 models, which were the landmark
successes of constructive field theory. In four spacetime dimensions one has to
rely instead on perturbative methods.
Another lesson learned from quantum field theory (and also from quantum
statistical mechanics) is that Hilbert spaces do not play such a central role
anymore. Instead one focuses on algebras and states.
Let us finish this discussion by recalling a well-known anecdote. At the
Solvay conference in 1927, Dirac told Bohr that he was trying to find a
relativistic quantum theory of the electron (i.e. the Dirac equation). Bohr
replied that this problem had already been solved by Klein, who had found
the Klein–Gordon equation. We know now that these two equations are of a
different nature, the first describing fermionic fields, the second bosonic ones,
and that they can be interpreted as quantum equations only via quantum field
theory.
4.1.2. Quantum Field Theory on Curved Spacetimes

Given the difficulties with the construction of interacting field theories on
Minkowski spacetime, one may wonder why one should consider quantum
field theories on curved spacetimes, which have no reason to be simpler.
One reason comes from attempts to quantize gravitation, where one starts by
linearizing Einstein equations around a curved background metric g. Another
argument is that there are several interesting quantum effects appearing in the
presence of strong gravitational fields. The most famous one is the Hawking
effect, which predicts that a black hole can emit quantum particles. There are
several new challenges one has to face when moving from flat Minkowski
spacetime to an arbitrary curved spacetime.
On the computational side, one cannot rely anymore on the Fourier
transform and related analyticity arguments, which are natural and useful
on Minkowski space, since the Klein–Gordon equation has then constant

coefficients.
From a more conceptual angle, a curved spacetime does not have the large
group of isometries (the Poincar group) of the Minkowski space. It follows that
on a curved spacetime there seems to be no natural notion of a vacuum state,
which is defined on Minkowski space as the unique state which is invariant
under spacetime translations, and which has an additional positive energy
condition.
In the 1980s, physicists managed to define a class of states, the so-called
“Hadamard states,” which were characterized by properties of their 2-point
functions, which had to have a specific asymptotic expansion near the diagonal,
connected with the well-known Hadamard parametrix construction for the
Klein–Gordon equation on a curved spacetime. Later in 1995, in a seminal
paper, Radzikowski reformulated the old Hadamard condition in terms of the
wave front set of the 2-point function. The wave front set of a distribution,
introduced in 1970 by Hörmander, is one of the important notions of microlocal
analysis, a theory which was precisely developed to extend Fourier analysis, in
the study of general partial differential equations.
The introduction of tools from microlocal analysis had a great influence
on the field, leading for example to the proof of renormalizability of scalar
interacting field theories by Brunetti and Fredenhagen.
The goal of these notes is to provide an introduction to the modern notion of
Hadamard states for a mathematically oriented audience.
4.2. A Quick Introduction to Quantum Mechanics

This section provides a very quick introduction to the mathematical formalism
of quantum mechanics, which is (or is expected to be) still relevant to quantum
field theory.
4.2.1. Hilbert Space Approach

In ordinary quantum mechanics, the description of a physical system starts
with a Hilbert space H, whose scalar product is denoted by (u|v). The states of
the system are described by unit vectors ψ ∈ H with ψ = 1.
The various physical quantities which can be measured (like position,
momentum, energy, spin) are represented by self-adjoint operators on H,
that is (forgetting about important issues with unbounded operators) linear
operators A on H, assumed to be bounded for simplicity, such that A = A∗
(where A∗ is the adjoint of A), and which are called observables.
176 C. Gérard
If ψ ∈ H , ψ = 1 is a state vector, then the map
A → ωψ (A) = (ψ|Aψ)
computes the expectation value of A in the state ψ which represents the

average value of actual measurements of the physical quantity represented
by A.
Rather quickly people were also led to consider mixed states, where the
state of the system is only incompletely known. For example if ψi , i ∈ N is an

orthonormal family and 0 ≤ ρi ≤ 1 are real numbers with ∞ i=0 ρi = 1, then
we can consider the trace-class operator
∞

ρ= ρi |ψi )(ψi |, Trρ = 1,
i=0
called a density matrix, and the map
A → ωρ (A) := Tr(ρA)
which is called a mixed state. Vector states are also called pure states.
4.2.2. Algebraic Approach

The framework above is sufficient to cover all of non-relativistic quantum
mechanics, that is in practice quantum systems consisting of a finite number of
non-relativistic particles. However, when one considers systems with an infinite
number of particles, as in statistical mechanics or quantum field theory, where
the notion of particles is dubious, an algebraic framework is more relevant.
It starts with the following observation about the space B(H) of bounded
operators on H.
If we equip the space with the operator norm, it is a Banach space, and a
Banach algebra, that is an algebra with the property such that AB ≤ AB.
It is also an involutive Banach algebra, that is the adjoint operation A → A∗
has the properties such that
(AB)∗ = B∗ A∗ , A∗ = A.
Finally one can easily check that
A∗ A = A2 , A ∈ B(H).
This last property has very important consequences; for example one can
deduce from it the functional calculus and spectral theorem for self-adjoint
operators.
An abstract algebra A equipped with a norm and an involution with these

properties, which is moreover complete, is called a C∗ algebra. The typical
example of a C∗ algebra is of course the algebra B(H) of bounded operators
on a Hilbert space.
If H is a Hilbert space, a ∗−homomorphism
A A → π(A) ∈ B(H)
is called a representation of A in H. An injective representation is called

faithful.
The need for such a change of point of view comes from the fact that a
physical system, like a gas of electrons, can exist in many different physical
realizations, for example at different temperatures. In other words it does not
come equipped with a canonical Hilbert space.
Observables, like for example the electron density, have a meaning irrelevant
of the realizations, and are described by self-adjoint elements in some C∗
algebra A. However, the various Hilbert spaces and the representations of the
observables on them are very different from one temperature to another.
One can also describe the possible physical realizations of a system with the
language of states. A state ω on A is a linear map
ω : A → C
such that
ω(A∗ A) ≥ 0, A ∈ A.
Assuming that A has a unit (which can always be assumed by adjoining one),
one also requires that
ω(1l) = 1.
The set of states on a C∗ algebra is a convex set, its extremal points being called
pure states. If A ⊂ B(H) and ψ is a unit vector, or if ρ is a density matrix, then
ωψ (A) := (ψ|Aψ), ωρ (A) := Tr(ρA)
are states on A. If A = B(H), then ωψ is a pure state. It is important to be

aware of the fact that if A is only a C∗ subalgebra of B(H), then ωψ may not
be a pure state on A.
4.2.3. The Gelfand–Naimark–Segal Construction

After being told that one should use C∗ algebras and states, one may wonder
where the Hilbert spaces have gone. Given a C∗ algebra A and a state ω on
178 C. Gérard
it, it is quite easy to construct a canonical Hilbert space and a representation

of A on it, as proved by Gelfand, Naimark, and Segal. There exists a triple
(Hω , πω , ω ) where Hω is a Hilbert space, πω : A → B(Hω ) is a faithful
representation, and ω ∈ Hω is a unit vector such that
ω(A) = (ω |πω (A)ω ), A ∈ A.
4.3. Notation
In this section we collect some notation that will be used in these notes. If X
is a real or complex vector space we denote by X # its dual. Bilinear forms on
X are identified with elements of L(X , X # ), which leads to the notation x1 ·bx2
for b ∈ L(X , X # ), x1 , x2 ∈ X . The space of symmetric (resp. anti-symmetric)
bilinear forms on X is denoted by Ls (X , X # ) (resp. La (X , X # )).
If σ ∈ Ls (X , X # ) is non-degenerate, we denote by O(X , σ ) the linear
(pseudo-)orthogonal group on X . Similarly if σ ∈ La (X , X # ) is non-
degenerate, that is (X , σ ) is a symplectic space, we denote by Sp(X , σ ) the
linear symplectic group on X .
If Y is a complex vector space, we denote by YR its realification, that is Y
considered as a real vector space. We denote by Y a conjugate vector space
to Y, that is a complex vector space Y with an anti-linear isomorphism Y
y → y ∈ Y. The canonical conjugate vector space to Y is simply the real
vector space YR equipped with the complex structure −i, if i is the complex
structure of Y. In this case the map y → y is chosen as the identity. If a ∈
L(Y1 , Y2 ), we denote by a ∈ L(Y 1 , Y 2 ) the linear map defined by
ay1 := ay1 , y1 ∈ Y 1 . (4.1)
We denote by Y ∗ the anti-dual of Y, that is the space of anti-linear forms

on Y. Clearly Y ∗ can be identified with Y # ∼ Y .
#
Sesquilinear forms on Y are identified with elements of L(Y, Y ∗ ), and we

use the notation (y1 |by2 ) or y1 ·by2 for b ∈ L(Y, Y ∗ ), y1 , y2 ∈ Y.
The space of hermitian (resp. anti-hermitian) sesquilinear forms on Y is
denoted by Ls (Y, Y ∗ ) (resp. La (Y, Y ∗ )).
If q ∈ Lh (Y, Y ∗ ) is non-degenerate, that is (Y, q) is a pseudo-unitary space,
we denote by U(Y, q) the linear pseudo-unitary group on Y.
If b is a bilinear form on the real vector space X , its canonical sesquilinear
extension to CX is by definition the sesquilinear form bC on CX given by
(w1 |bC w2 ) := x1 ·bx2 + y1 ·by2 + ix1 ·by2 − iy1 ·bx2 , wi = xi + iyi
for xi , yi ∈ X , i = 1, 2. This extension maps (anti-)symmetric forms on X onto

(anti-)hermitian forms on CX .
Conversely if Y is a complex vector space and YR is its realification, that
is Y considered as a real vector space, then for b ∈ Lh/a (Y, Y ∗ ) the form Reb
belongs to Ls/a (YR , YR# ).
4.4. CCR and CAR Algebras

4.4.1. Introduction
It is useful to discuss the canonical commutation relations (CCR) and canonical
anticommutation relations (CAR) without making reference to a Fock space.
There are some mathematical subtleties with CCR algebras, coming from the
fact that the field operators are “unbounded.” These subtleties can mostly be
ignored for our purposes.
4.4.2. Algebras Generated by Symbols and Relations

In physics many algebras are defined by specifying a set of generators and the
relations they satisfy. This is completely sufficient to perform computations,
but mathematicians may feel uncomfortable with such an approach. However,
it is easy (and actually rather useless) to give a rigorous definition.
Assume that A is a set. We denote by cc (A, K) the vector space of functions
A → K with finite support (usually K = C). If for A ∈ A, we denote the
indicator function 1l{A} simply by A, we see that any element of cc (A, K) can

be written as A∈B λA A, B ⊂ A finite, λA ∈ K.
Then cc (A, K) can be seen as the vector space of finite linear combinations
of elements of A. We set
al
A(A, 1l) := ⊗ cc (A, K),
called the universal unital algebra over K with generators A. Usually one
writes A1 · · · An instead of A1 ⊗ · · · ⊗ An for Ai ∈ A.
Let us denote by A another copy of A. We denote by a the element a ∈ A.
We then set ∗a := a, ∗a := a and extend ∗ to A(A A, 1l) by setting
(b1 b2 · · · bn )∗ = b∗n · · · b∗2 b∗1 , bi ∈ A A, 1l = 1l∗ .
The algebra A(A A, 1l) equipped with the involution ∗ is called the universal
unital ∗ −algebra over K with generators A.
180 C. Gérard
Now let R ⊂ A(A, 1l) (the set of “relations”). We denote by I(R) the ideal
of A(A, K) generated by R. Then the quotient
A(A, 1l)/I(R)
is called the unital algebra with generators A and relations R = 0, R ∈ R.

Similarly if R ⊂ A(A∪A, 1l) is ∗-invariant, then A(AA, 1l)/I(R) is called
the unital ∗-algebra with generators A A and relations R = 0, R ∈ R.
4.4.3. Polynomial CCR Algebra

We fix a (real) presymplectic space (X , σ ), that is σ ∈ La (X , X # ) is not
supposed to be injective.
Definition 4.1 The polynomial CCR ∗-algebra over X , denoted by

CCRpol (X , σ ), is defined to be the unital complex ∗-algebra generated by
elements φ(x), x ∈ X , with relations
φ(λx) = λφ(x), λ ∈ R, φ(x1 + x2 ) = φ(x1 ) + φ(x2 ),

φ ∗ (x) = φ(x), φ(x1 )φ(x2 ) − φ(x2 )φ(x1 ) = ix1 ·σ x2 1l.
4.4.4. Weyl CCR Algebra

One problem with CCRpol (X , σ ) is that (unless σ = 0) its elements cannot
be faithfully represented as bounded operators on a Hilbert space. To cure this
problem one has to work with Weyl operators, which lead to the Weyl CCR
∗−algebra.
Definition 4.2 The algebraic Weyl CCR algebra over X denoted by

CCRWeyl (X , σ ) is the ∗-algebra generated by the elements W(x), x ∈ X ,
with relations
W(0) = 1l, W(x)∗ = W(−x),

i
W(x1 )W(x2 ) = e− 2 x1 ·σ x2 W(x1 + x2 ), x, x1 , x2 ∈ X .
It is possible to equip CCRWeyl (X , σ ) with a unique C∗ −norm. Its comple-

tion for this norm is called the Weyl CCR algebra over X and still denoted by
CCRWeyl (X , σ ). We will mostly work with CCRpol (X , σ ), which will simply
be denoted by CCR(X , σ ). Of course the formal relation between the two
approaches is
W(x) = eiφ(x) , x ∈ X ,
which does not make sense a priori, but from which mathematically correct
statements can be deduced.
4.4.5. Charged Symplectic Spaces

Definition 4.3 A complex vector space Y equipped with a non-degenerate anti-
hermitian sesquilinear form σ is called a charged symplectic space. We set
q := iσ ∈ Lh (Y, Y ∗ ),
which is called the charge.
4.4.6. Kähler Spaces

Let (Y, σ ) be a charged symplectic space. Its complex structure will be denoted
by j ∈ L(YR ) (to distinguish it from the complex number i ∈ C). Note that
(YR , Reσ ) is a real symplectic space with j ∈ Sp(YR , Reσ ) and j2 = −1l.
We have
y1 qy2 = y1 · Reσ jy2 + iy1 · Reσ y2 , y1 , y2 ∈ Y.
The converse construction is as follows. A real (pre-)symplectic space (X , σ )

with a map j ∈ L(X ) such that
j2 = −1l, j ∈ Sp(X , σ )
is called a pseudo-Kähler space. If in addition ν := σ j is positive definite, it is

called a Kähler space. We now set
Y = (X , j),
which is a complex vector space, whose elements are logically denoted by y.

If (X , σ , j) is a pseudo-Kähler space we can set
y1 qy2 := y1 · σ jy2 + iy1 · σ y2 , y1 , y2 ∈ Y
and check that q is sesquilinear hermitian on Y equipped with the complex

structure j. One may consider the CCR algebra CCRpol (YR , Reσ ), with
selfadjoint generators φ(y) and relations
[φ(y1 ), φ(y2 )] = iy1 · Reσ y2 1l.
But one can instead generate CCRpol (YR , Reσ ) by the charged fields
1 1
ψ(y) := √ (φ(y) + iφ(jy)), ψ ∗ (y) := √ (φ(y) − iφ(jy)), y ∈ Y.
2 2
182 C. Gérard
The map Y y → ψ ∗ (y) (resp. Y y → ψ(y)) is C−linear (resp. C−anti-

linear). The commutation relations take the form
[ψ(y1 ), ψ(y2 )] = [ψ ∗ (y1 ), ψ ∗ (y2 )] = 0,

[ψ(y1 ), ψ ∗ (y2 )] = y1 · qy2 1l, y1 , y2 ∈ Y.
Note the similarity with the CCR expressed in terms of creation/annihilation

operators, the difference being the fact that q is not necessarily positive. In this
context, it is natural to denote CCR(YR , Reσ ) by CCR(Y, σ ).
4.4.7. CAR Algebra

Here we fix a Euclidean space (X , ν) (possibly infinite dimensional).
Definition 4.4 The algebraic CAR algebra over X , denoted CARalg (X , ν),
is the complex unital ∗-algebra generated by elements φ(x), x ∈ X , with
relations
φ(λx) = λφ(x), λ ∈ R, φ(x1 + x2 ) = φ(x1 ) + φ(x2 ),

∗
φ (x) = φ(x), φ(x1 )φ(x2 ) + φ(x2 )φ(x1 ) = 2x1 ·νx2 1l.
Again CARalg (X , ν) has a unique C∗ −norm, and its completion is denoted

by CAR(X , ν).
4.4.8. Kähler Spaces

Now let (Y, ν) be a hermitian space, denoting again its complex structure by j.
Then (YR , Reν) is a Euclidean space, with j ∈ U(YR , Reν). We have
y1 · νy2 = y1 · Reνy2 − iy1 · Rejy2 , y1 , y2 ∈ Y.
Denoting by φ(y) the self-adjoint fields which generate CAR(YR , Reν), we

can introduce the charged fields
ψ(y) := φ(y) + iφ(jy), ψ ∗ (y) := φ(y) − iφ(jy), y ∈ Y.
Again the map Y y → ψ ∗ (y) (resp. Y y → ψ(y)) is C−linear (resp.

C−anti-linear). The anti-commutation relations take the form
[ψ(y1 ), ψ(y2 )]+ = [ψ ∗ (y1 ), ψ ∗ (y2 )]+ = 0,

[ψ(y1 ), ψ ∗ (y2 )]+ = 2y1 · νy2 1l, y1 , y2 ∈ Y.
4.5. States on CCR/CAR Algebras

4.5.1. Introduction
Let A be a unital ∗−algebra. A state on A is a linear map ω : A → K such that
ω(A∗ A) ≥ 0, ∀A ∈ A, ω(1l) = 1.
Elements of the form A∗ A are called positive.

Let X be either a presymplectic or Euclidean space and let ω be a state on
CCR(X , σ ) or CAR(X , ν). One can associate to ω a bilinear form on X called
the 2-point function:
X × X (x1 , x2 ) → ω(φ(x1 )φ(x2 )).
Of course to specify completely the state ω one also needs to know the n-point
functions
X n (x1 , . . . xn ) → ω(φ(x1 ) . . . φ(xn )).
A particularly useful class of states are the quasi-free states, which are defined
by the fact that all n-point functions are determined by the 2-point function.
4.5.2. Bosonic Quasi-Free States

Let (X , σ ) be a presymplectic space and ω a state on CCRWeyl (X , σ ). The
function
X x → ω(W(x)) =: G(x)
is called the characteristic function of the state ω, and is a non-commutative

version of the Fourier transform of a probability measure.
There is also a non-commutative version of Bochner’s theorem (the theorem
which characterizes these Fourier transforms).
Proposition 4.1 A map G : X → C is the characteristic function of a state on
CCRWeyl (X , σ ) iff for any n ∈ N, xi ∈ X the n × n matrix
i

G(xj − xi )e 2 xi ·σ xj
1≤i,j≤n
is positive.
Proof ⇒: for x1 , . . . , xn ∈ X , λ1 , . . . , λn ∈ C set

n
A := λj W(xj ) ∈ CCRWeyl (X , σ ).
j=1
184 C. Gérard
Such A are dense in CCRWeyl (X , σ ). One computes A∗ A using the CCR and
obtains
n
i
A∗ A = λj λk W(xj − xk )e 2 xj ·σ xk ,
j,k=1
from which ⇒ follows.

⇐: one uses exactly the same argument, defining ω using G, and the above
formula shows that ω is positive.
Definition 4.5 (1) A state ω on CCRWeyl (X , σ ) is a quasi-free state if there

exists η ∈ Ls (X , X # ) (a symmetric form on X ) such that
1
ω W(x) = e− 2 x·ηx , x ∈ X . (4.2)
(2) The form η is called the covariance of the quasi-free state ω.
Quasi-free states should be considered as non-commutative versions of

Gaussian measures. To explain this remark, consider the Gaussian measure
on Rd with covariance η:
1 1 −1 y
dμη := (2π )d/2 det η− 2 e− 2 y·η dy.
We have

1
eix·y dμη (y) = e− 2 x·ηx .
Note also that if xi ∈ Rd , then

2n+1

xi · ydμη (y) = 0,
1

2n
n
xi · ydμη (y) = xσ (2j−1) · ηxσ (2j) ,
1 σ ∈Pair2n j=1
which should be compared with Definition 4.6.

Proposition 4.2 Let η ∈ Ls (X , X # ). Then the following are equivalent.
1
(1) X x → e− 2 x·ηx is a characteristic function and hence there exists a
quasi-free state satisfying (4.2).
(2) ηC + 2i σC ≥ 0 on CX , where ηC , σC ∈ L CX , (CX )∗ are the canonical
sesquilinear extensions of η, σ .
1 1
(3) |x1 ·σ x2 | ≤ 2(x1 ·ηx1 ) 2 (x2 ·ηx2 ) 2 , x1 , x2 ∈ X .
Proof The proof of (1) ⇒ (2) is easy, by considering complex fields φ(w) =
φ(x1 ) + iφ(x2 ), w = x1 + ix2 and noting that the positivity of ω implies that
ω(φ ∗ (w)φ(w)) ≥ 0, for any w ∈ CX .
The proof of (2) ⇒ (1) is more involved: let us fix x1 , . . . , xn ∈ X and set
i
bjk = xj · ηxk + xj · σ xk .
2
Then, for λ1 , . . . , λn ∈ C,
i n
λj bjk λk = w·ηC w + w·ωC w, w = λj xj ∈ CX .
2
1≤j,k≤n j=1
By (2), the matrix [bjk ] is positive. One has then to use an easy lemma, stating
that the pointwise product of two positive matrices is positive. From this it
1 1
follows also that [ebjk ] is positive, and hence the matrix [e− 2 xj ·ηxj ebjk e− 2 xk ·ηxk ]
is positive. Hence:

n
1 i
e− 2 (xk −xj )·η(xk −xj ) e 2 xj ·ωxk λj λk
j,k=1
n
1 1
= e− 2 xj ·ηxj ebjk e− 2 xj ·ηxj λj λk
j,k=1
n
i
= G(xj − xk )e 2 xj ·σ xk ≥ 0,
j,k=1
1
for G(x) = e− 2 x·ηx . By Proposition 4.1, this means that G is a characteristic
function. The proof of (2) ⇔ (3) is an exercise in linear algebra.
It is easy to deduce from ω the corresponding state acting on CCRpol (X , σ ),

by setting
ω(φ(x1 ) · · · φ(xn ))
d d
:= ··· ω(W(t1 x1 + · · · + tn xn ))|t1 =···tn =0 .
dt1 dtn
In particular
i
ω(φ(x1 )φ(x2 )) = x1 · ηx2 + x1 · σ x2 .
2
The corresponding definition of a quasi-free state on CCRpol (X , σ ) is as
follows.
186 C. Gérard
Definition 4.6 A state ω on CCRpol (X , σ ) is quasi-free if

ω φ(x1 ) · · · φ(x2m−1 ) = 0,

m

ω φ(x1 ) · · · φ(x2m ) = ω φ(xσ (2j−1) )φ(xσ (2j) .
σ ∈Pair2m j=1
We recall that Pair2m is the set of pairings, that is the set of partitions of
{1, . . . , 2m} into pairs. Any pairing can be written as
{i1 , j1 }, · · · , {im , jm }
for ik < jk and ik < ik+1 , hence can be uniquely identified with a permutation
σ ∈ S2m such that σ (2k − 1) = ik , σ (2k) = jk .
4.5.3. Gauge-Invariant Quasi-Free States

Let us now assume that (X , σ , j) is a pseudo-Kähler space, that is that there
exists an anti-involution j ∈ Sp(X , σ ). Note that we have
ejθ = cos θ + j sin θ , θ ∈ R,
and that the map
[0, 2π] θ → ejθ ∈ Sp(X , σ )
is a 1-parameter group called the group of (global) gauge transformations.
A quasi-free state ω on CCRWeyl (X , σ ) is called gauge invariant if
ω(W(x)) = ω(W(ejθ x)), x ∈ X , θ ∈ R.
We can of course let ω act on CCRpol (X , σ ). It is much more convenient then
to use the charged fields ψ (∗) (x) as generators of CCRpol (X , σ ). We have then:
Proposition 4.3 A state ω on CCRpol (X , σ ) is gauge invariant quasi-free iff

ω n1 ψ ∗ (yi )1 ψ(xi ) = 0, if n = p
p

n
ω n1 ψ ∗ (yi )n1 ψ(xi ) = ω(ψ ∗ (yi )ψ(xσ (i) )).
σ ∈Sn i=1
It follows that a gauge invariant quasi-free state ω is uniquely determined by

the sesquilinear form
ω(ψ(y1 )ψ ∗ (y2 )) =: y1 · λ+ y2 .
Clearly λ+ ∈ Lh (Y, Y ∗ ). Let
ω(ψ ∗ (y2 )ψ(y1 )) =: y1 · λ− y2 ,
with λ− ∈ Lh (Y, Y ∗ ). From the commutation relations we have of course
λ+ − λ− = q,
so λ− is determined by λ+ , but nevertheless it is convenient to work with the

pair (λ± ) and to call λ± the complex covariances of ω.
The link between the real and complex covariances is as follows.
Lemma 4.1 We have

1 1
η = Re λ± ∓ q , η̂ = λ± ∓ q,
2 2
where η̂ ∈ Lh (Y, Y ∗ ) is given by
y1 η̂y2 := y1 ·ηy2 − iy1 ·ηjy2 .
Note that the fact that η̂ is sesquilinear follows from the gauge invariance
of ω. From the above lemma one easily gets the following characterization of
complex covariances.
Proposition 4.4 Let λ± ∈ Lh (Y, Y ∗ ). Then the following are equivalent.
(1) λ± are the covariances of a gauge-invariant quasi-free state on
CCRpol (Y, q),
(2) λ± ≥ 0 and λ+ − λ− = q.
Proof Since ω is gauge invariant we have
j ∈ O(YR , η) ∩ Sp(YR , Reσ ) = O(YR , η) ∩ O(YR , Req).
From this fact and Lemma 4.1 we deduce that η ≥ 0 ⇔ λ+ ≥ 12 q, and that the
second condition in Proposition 4.2 (with σ replaced by Reσ ) is equivalent to
±q ≤ 2λ+ − q ⇔ λ± ≥ 0.
This completes the proof of the proposition.
4.5.4. Complexifying Bosonic Quasi-Free States

If (X , σ ) is real symplectic, we can form (CX , σC ) which is charged symplec-
tic. As real symplectic space it equals (X , σ )⊕ (X , σ ). If ω is a quasi-free state
on (X , σ ) with real covariance η, we form a state ωC on ((CX )R , ReσC ), with
covariance ReηC , which is by definition gauge invariant. Hence, possibly after
complexification, we can always reduce ourselves to gauge invariant quasi-free
states.
188 C. Gérard
4.5.5. Pure Quasi-Free States

In this subsection we discuss pure quasi-free states, which turn out to be
precisely vacuum states. To start the discussion, let us recall that from
Proposition 4.2 (3) one has
1 1
|x1 ·σ x2 | ≤ 2(x1 ·ηx1 ) 2 (x2 ·ηx2 ) 2 , x1 , x2 ∈ X . (4.3)
We can complete the real vector space X w.r.t. the (semi-definite) symmetric
form η, after taking the quotient by the vectors of zero norm as usual. Note that
from (4.3) σ passes to quotient and to completion.
Denoting once again (X /Kerη)cpl by X , we end up with the following
situation: (X , η) is a real Hilbert space and σ is a bounded, anti-symmetric
form on X . However, σ may very well not be non-degenerate anymore, that
is (X , σ ) may just be presymplectic. One can show that if σ is degenerate,
then the state on CCR(X , σ ) with covariance η is not pure. In the sequel we
hence assume that σ is non-degenerate on X . One can then prove the following
theorem. The proof uses some more advanced tools, like the Araki-Woods
representation and its properties.
Theorem 4.1 The state ω of covariance η is pure iff the pair (2η, σ ) is Kähler,
that is there exists j ∈ Sp(X , σ ) such that j2 = 1l and 2η = σ j.
The link with Fock spaces and vacuum states is now as follows. If one equips
X with the complex structure j and the scalar product
(x1 |x2 ) := x1 · 2ηx2 + ix1 · ηjx2 ,
then Z := (X , (·|·)) is a complex Hilbert space. One can build the bosonic
Fock space s (Z) and the Fock representation X x → eiφ(x) ∈ U(s (Z)).
This representation is precisely the GNS representation of the state ω, with
GNS vector ω equal to the Fock vacuum.
For reference let us state the version of Theorem 4.1 using charged fields.
Theorem 4.2 Let λ± ∈ Lh (Y, Y ∗ ). Then the following are equivalent.
(1) λ± are the covariances of a pure gauge-invariant quasi-free state on
CCRpol (Y, σ ).
(2) There exists an involution κ ∈ U(Y, q) such that
1
qκ ≥ 0, λ± = q(κ ± 1l).
2
(3) λ± ≥ ± 12 q, λ± q−1 λ± = ±λ± , λ+ − λ− = q.
Theorem 4.2 can be easily deduced from Theorem 4.1.
4.5.6. Fermionic Quasi-Free States

We consider now a Euclidean space (X , ν). Without loss of generality we can
assume that X is complete, that is (X , ν) is a real Hilbert space. We consider
the CAR algebra CAR(X , ν), with the self-adjoint fermionic fields φ(x) as
generators.
Definition 4.7 (1) A state ω on CAR(X , ν) is called quasi-free if

ω φ(x1 ) · · · φ(x2m−1 ) = 0,

m

ω φ(x1 ) · · · φ(x2m ) = sgn(σ ) ω φ(xσ (2j−1) )φ(xσ (2j) ) ,
σ ∈Pair2m j=1
for all x1 , x2 , · · · ∈ X , m ∈ N.
(2) The anti-symmetric form β ∈ La (X , X # ) defined by
x1 · βx2 := i−1 ω([φ(x1 ), φ(x2 )])
is called the covariance of the quasi-free state ω.
From the CAR it follows that
i
ω φ(x1 )φ(x2 ) = x1 ·νx2 + x1 ·βx2 , x1 , x2 ∈ X . (4.4)
2
Proposition 4.5 Let β ∈ La (X , X # ). Then the following are equivalent.
(1) β is the covariance of a fermionic quasi-free state ω;
(2) νC + 2i βC ≥ 0 on CX ;
1 1
(3) |x1 ·βx2 | ≤ 2(x1 ·νx1 ) 2 (x2 ·νx2 ) 2 , x1 , x2 ∈ X .
Proof As in the bosonic case (1) ⇒ (2) and (2) ⇔ (3) are easy to prove.
The proof of (2) ⇒ (1) is more difficult, since it relies on the Jordan–Wigner
representation of CAR(X , ν) for X finite dimensional.
4.5.7. Gauge-Invariant Quasi-Free States

We now assume that (X , ν) is equipped with a Kähler anti-involution j. This
implies that ejθ ∈ O(X , ν) hence that there exists the group of global gauge
transformations τθ , θ ∈ [0, 2π] defined by
τθ φ(x) := φ(ejθ x), x ∈ X .
As in the bosonic case, a state ω on CAR(X , ν) is called gauge invariant if
ω ◦ τθ = ω, ∀θ ∈ [0, 2π]. Again it is more convenient to use the charged fields
ψ(y) := φ(y) + iφ(jy), ψ ∗ (y) := φ(y) − iφ(jy),
190 C. Gérard
as generators of CAR(Y). We recall that one sets q := 2νC ∈ Lh (Y, Y ∗ ), which

is moreover positive definite.
∗
Proposition 4.6 A state ω on CARC (Y) is gauge-invariant quasi-free iff:

ω n1 ψ ∗ (yi )1 ψ(xi ) = 0, if n = p
p

n
ω n1 ψ ∗ (yi )n1 ψ(xi ) = sgn(σ ) ω(ψ ∗ (yi )ψ(xσ (i) )).
σ ∈Sn i=1
Again we can introduce the two complex covariances

ω(ψ(y1 )ψ ∗ (y2 )) =: y1 · λ+ y2 , ω(ψ ∗ (y2 )ψ(y1 )) =: y1 · λ− y2 .
Proposition 4.7 Let λ± ∈ Lh (Y, Y ∗ ). Then the following are equivalent.
(1) λ± are the covariances of a gauge-invariant quasi-free state on
CAR(Y, q);
(2) λ± ≥ 0 and λ+ + λ− = q.
4.5.8. Pure Quasi-Free States

We discuss pure quasi-free states in the fermionic case. We consider the general
case, that is we do not assume the states to be gauge invariant.
Theorem 4.3 Let β ∈ La (X , X # ). Then β is the covariance of a pure quasi-
free state on CAR(X , ν) iff (ν, 12 β) is Kähler, that is there exists j ∈ O(X , ν)
such that j2 = −1l and ν = 12 βj.
We leave the formulation of the gauge-invariant version of this theorem as
an exercise.
4.6. Lorentzian Manifolds

4.6.1. Causality
Let M be a smooth manifold of dimension n = d + 1. We assume that M is
equipped with a Lorentzian metric g, that is a smooth map
M x → gμν (x) ∈ Ls (Tx M, Tx∗ M),
with signature (−1, d). The inverse g−1 (x) ∈ Ls (Tx∗ M, Tx M) is traditionally
denoted by gμν (x). We also set |g|(x) := | det gμν (x)|.
Using the metric g we can define time-like, causal, etc. tangent vectors
and vector fields on M as on the Minkowski spacetime. The set {v ∈ Tx M :
vg(x)v = 0} is called the lightcone at x.
M is then called time-orientable if there exists a global continuous time-like

vector field v. Once a time-orientation is chosen, one can define future/past
directed time-like vector fields.
One can similarly define time-like, causal etc. piecewise C1 curves by
requiring the said property to hold for all its tangent vectors.
Definition 4.8 Let x ∈ M. The causal, resp. time-like future, resp. past of x
is the set of all y ∈ M that can be reached from x by a causal, resp. time-like
future, resp. past-directed curve, and is denoted J ± (x), resp. I ± (x). For U ⊂ M,
its causal, resp. time-like future, resp. past is defined as

J ± (U) = J ± (x), I ± (U) = I ± (x).
x∈U x∈U
We define also the causal, resp. time-like shadow:
J(U) = J + (U) ∪ J − (U), I(U ) = I + (U ) ∪ I − (U ).
The classification of tangent vectors in Tx M can be naturally extended to

linear subspaces of Tx M.
Definition 4.9 A linear subspace E of Tx M is space-like if it contains only

space-like vectors, time-like if it contains both space-like and time-like vectors,
and null (or light-like) if it is tangent to the lightcone at x.
If E ⊂ Tx M we denote by E⊥ ⊂ Tx M its orthogonal for g(x).

Lemma 4.2 A subspace E ⊂ Tx M is space-like, resp. time-like, resp. null iff
E⊥ is time-like, resp. space-like, resp. null.
We refer to [9, Lemma 3.1.1] for the proof.
4.6.2. Globally Hyperbolic Manifolds

Definition 4.10 A Cauchy hypersurface is a hypersurface ⊂ M such that
each inextensible time-like curve intersects at exactly one point.
The following deep result is originally due to Geroch, with a stronger

condition instead of (1b) in this form due to Bernal-Sanchez.
Theorem 4.4 Let M be a connected Lorentzian manifold. The following are
equivalent.
(1) The following two conditions hold:
(1a) for any x, y ∈ M, J + (x) ∩ J − (y) is compact,
(1b) (causality condition) there are no closed causal curves.
192 C. Gérard
(2) There exists a Cauchy hypersurface.

(3) M is isometric to R × with metric −βdt2 + gt , where β is a smooth
positive function, gt is a Riemannian metric on depending smoothly on
t ∈ R, and each {t} × is a smooth space-like Cauchy hypersurface in M.
Definition 4.11 A connected Lorentzian manifold satisfying the equivalent
conditions of the above theorem is called globally hyperbolic.
The adjective “hyperbolic” comes from the connection with hyperbolic
partial differential equations: roughly speaking globally hyperbolic spacetimes
are those on which the Cauchy problem for the Klein–Gordon equation can be
formulated and uniquely solved.
Definition 4.12 A function f : M → C is called space-compact if there exists
K M such that supp f ⊂ J(K). It is called future/past space-compact if there
exists K M such that supp f ⊂ J ± (K). The spaces of such smooth functions
∞ (M), resp. C∞ (M).
are denoted by Csc ±sc
4.7. Klein–Gordon Fields on Lorentzian Manifolds

4.7.1. The Klein–Gordon Operator
Let (M, g) be a Lorentzian manifold. The Klein–Gordon operator on M is
1 1
P(x, ∂x ) = −|g|− 2 ∂μ |g| 2 gμν (x)∂ν + r(x),
acting on functions u : M → R. Here r ∈ C∞ (M, R) represent a (variable)
mass. Using the metric connection, one can write
P(x, ∂x ) = −∇a ∇ a + r(x).
One can generalize the Klein–Gordon operator to sections of vector bundles,
the most important example being the bundle of 1-forms on M, which appears
in the quantization of Maxwell’s equation. One would like to interpret some
space of smooth solutions of the Klein–Gordon equation
P(x, ∂x )u = 0 (KG)
as a symplectic space. Following the discussion in Section 4.5 we will consider
complex solutions.
4.7.2. Conserved Current

Definition 4.13 We set for u1 , u2 ∈ C∞ (M):
J a (u1 , u2 ) := u1 .∇ a u2 − ∇ a u1 u2 ∈ C∞ (M).
J a (u1 , u2 ) is a vector field called a current in the physics literature. Using

the rules to compute with connections we obtain
∇a J a (u1 , u2 ) = ∇a u1 ∇ a u2 + u1 ∇a ∇ a u2 − ∇a ∇ a u1 u2 − ∇ a u1 ∇a u2
= u1 ∇a ∇ a u2 − ∇a ∇ a u1 u2 = Pu1 u2 − u1 Pu2 .
It follows that if Pui = 0 the vector field J a (u1 , u2 ) is divergence free.
Moreover from the Gauss formula we obtain:
Lemma 4.3 (Green’s formula) Let U ⊂ M be an open set with ∂U non
characteristic. Then

u1 Pu2 − Pu1 u2 dμg = − u1 ∇a u2 − ∇a u1 u2 na dσg .
U ∂U
4.7.3. Advanced and Retarded Fundamental Solutions

The treatment of the Klein–Gordon operator is quite similar to its Riemannian
analog, that is the Laplace operator, and starts with the construction of
fundamental solutions, that is inverses. An important difference is the
hyperbolic nature of the Klein–Gordon operator: to obtain a unique solution
of the equation
Pu = v,
say for v ∈ C0∞ (M),
one has to impose extra support conditions on u, since
there exists plenty of solutions of Pu = 0. The condition that (M, g) is globally
hyperbolic is the natural condition so as to be able to construct fundamental
solutions. In the sequel we will assume that (M, g) is globally hyperbolic.
The main result is the following theorem, due to Leray.
Theorem 4.5 For any v ∈ C0∞ (M), there exist unique functions u± ∈ C±sc
∞ (M)
that solve
P(x, ∂x )u± = v.
Moreover,

u± (x) = (E± v)(x) := E± (x, y)v(y)dμg (y),
where E± ∈ D (M × M, L(V)) satisfy

P ◦ E± = E± ◦ P = 1l, supp E± ⊂ (x, y) : x ∈ J ± (y) .
Note that P is self-adjoint for the scalar product

(u|v) = uvdμg ,
M
194 C. Gérard
which by uniqueness implies that (E± )∗ = E∓ . This also implies that by

duality E± can be applied to distributions of compact support.
Definition 4.14 E+ , resp. E− , is called the retarded, resp. advanced Green’s

function.
E := E+ − E−
is called the Pauli–Jordan function.
Note that E = −E∗ , that is E is anti-hermitian.
4.7.4. Cauchy Problem

Once the existence of E± is established, it is easy to solve the Cauchy problem
(one can also proceed the other way around). Let us denote by Solsc (KG) the
space of smooth, space compact solutions of (KG). Let be a smooth Cauchy
hypersurface. We denote by ρ : Csc∞ (M) → C∞ () ⊕ C∞ () the map
0 0
ρu := (ρ0 u, ρ1 u) = (u| , nμ ∂μ u)| ).
Theorem 4.6 Let f ∈ C0∞ () ⊗ C2 . Then there exists a unique u ∈ Solsc (KG)
such that ρu = f . This satisfies suppu ⊂ J(suppf0 ∪ suppf1 ) and is given by

u(x) = − n ∇y E(x, y)f0 (y)dσ (y) +
μ μ
E(x, y)f1 (y)dσ (y). (4.5)

We will set u =: Uf , for u given by (4.5).

Remark 4.1 Let us denote by ρi∗ : D () → D (M) the adjoints of ρi . Let us
also set

0 1l
q = ∈ L(C0∞ () ⊕ C0∞ ()).
−1l 0
Then (4.5) can be rewritten as
1l = Eρ ∗ q ρ, on Solsc (KG), (4.6)
or equivalently as
U = E ◦ ρ ∗ ◦ q , on C0∞ () ⊕ C0∞ (). (4.7)
Proof We will just prove that the solution u of the Cauchy problem is given
by the above formula. We fix f ∈ C0∞ (M) and apply Green’s formula to
u0 = E∓ f , u1 = u, and U = J ± (). We obtain

f udμg = E− f ∇a u − ∇a E− f u na dσg ,
J + ()

f udμg = − E+ f ∇a u − ∇a E+ f u na dσg .
J − ()
Summing these two identities we obtain

f udμg = −Ef ∇a u + ∇a Ef u dσg .
M
To complete the proof of the formula, it suffices to introduce the distribution

kernel E(x, y) to express Ef as an integral and to use E = −E∗ . Details are left
to the reader.
4.7.5. Symplectic Structure of the Space of Solutions

By Theorem 4.6 we know that
ρ : Solsc (KG) → C0∞ () ⊕ C0∞ ()
is bijective, with inverse U. For u, v ∈ Solsc (KG) with ρu =: f , ρv =: g we set

u1 · σ u2 := f 1 g0 − f 0 g1 dσg = J a (u, v)na dσg

(4.8)
f σ g := f 1 g0 − f 0 g1 dσg .

It is obvious from Theorem 4.6 that (Solsc (KG), σ ) is a (complex) symplectic

space. From the Gauss formula we know that σ is independent of the choice
of the Cauchy surface .
4.7.6. Spacetime Fields

Theorem 4.7 (1) Consider E : C0∞ (M) → C∞ (M). Then
RanE = Solsc (KG), KerE = PC0∞ (M).
(2) One has

Ef 1 · σ Ef2 = ( f1 |Ef2 ), fi ∈ C0∞ (M).
Part (1) of the theorem can be nicely rephrased by saying that the sequence
P E P
0 −→ C0∞ (M) −→ C0∞ (M) −→ Solsc (KG) −→ 0
is exact.
196 C. Gérard
Proof (1) Since PE± = 1l, we have PE = 0, and taking adjoints also EP = 0.
This shows that EC0∞ (M) ⊂ Solsc (KG) and PC0∞ (M) ⊂ KerE. It remains to
prove the converse inclusions.
(1a) Solsc (KG) ⊂ EC0∞ (M): let u ∈ Solsc (KG). Since u is space-compact,
we can find cutoff functions χ ± ∈ C±sc ∞ (M) such that χ + + χ − = 1 on suppu.
We have some compact sets K ⊂ K1 ⊂ K2 such that suppu ⊂ J(K), suppχ ± ⊂

J ± (K2 ), and χ ± = 1 on J ± (K1 ). It follows that supp∇χ ± ⊂ J ± (K2 )\J ± (K1 ),
hence supp∇χ ± ∩ suppu ⊂ J ∓ (K2 ) ∩ J ± (K). This set is compact by the global
hyperbolicity of M.
We set now u± = χ ± u, f = Pu+ = −Pu− . By the above discussion
f ∈ C0∞ (M), hence u± = ±E± f and u = u+ + u− = Ef .
(1b) KerE = PC0∞ (M): let u ∈ C0∞ (M) such that Eu = 0 and let f = E± u.
Then suppf ⊂ J + (suppu) ∩ J − (suppu), hence again by global hyperbolicity,
f ∈ C0∞ (M).
(2) From (4.6) we obtain
E = −Eρ ∗ q ρE = (ρ ◦ E)∗ ◦ q ◦ (ρ ◦ E).
This implies (2).
We now summarize the discussion as follows:

Theorem 4.8 (1) The following spaces are symplectic spaces:
(C0∞ (M)/PC0∞ (M), E), (Solsc (KG), σ ), (C0∞ () ⊕ C0∞ (), σ ).
(2) The following maps are symplectomorphisms:

E ρ
(C0∞ (M))/PC0∞ (M), E) −→(Solsc (KG), σ ) −→(C0∞ () ⊕ C0∞ (), σ ).
4.7.7. Quasi-Free States for the Free Klein–Gordon Field

We can now consider quasi-free states on any of the symplectic spaces in
Theorem 4.8. The most natural one is
(C0∞ (M))/PC0∞ (M), E)
which leads to spacetime fields. The associated CCR algebra will be denoted
simply by CCR(C0∞ (M), E), ignoring the need to pass to quotient to get a true
symplectic space.
Strictly speaking we would write symbols like
φ([f ]), [f ] ∈ C0∞ (M)/PC0∞ (M).
We write this as

φ( f )“ = ” φ(x)f (x)dμg “ = ”φ, f
M
if Pφ(x) = 0, that is the quantum field φ satisfies the KG equation.
Causality: denoting by φ( f ) the (self-adjoint) fields associated with
[f ] ∈ C0∞ (M))/PC0∞ (M), we have
[φ( f ), φ(g)] = Re( f |Eg) = 0,
if suppf , suppg are causally disjoint. This follows from the fact that
suppE(x, y) ⊂ {(x, y) ∈ M × M : x ∈ J(y)}.
Let us now consider a gauge-invariant quasi-free state ω, defined by the com-

plex covariances (λ± ). Recall that these complex covariances λ± are sesquilin-
ear forms on C0∞ (M))/PC0∞ (M). One may assume that they are obtained from
sesquilinear forms ± on C0∞ (M) which pass to quotient, that is such that
P ◦ ± = ± ◦ P = 0. (4.9)
It is also natural to assume that ± are continuous sesquilinear forms,

hence by the Schwartz kernel theorem we have distributional kernels
± (x, y) ∈ D (M × M). Then (4.9) becomes, using P = P∗ :
P(x, ∂x )± (x, y) = P(y, ∂y )± (x, y) = 0. (4.10)
4.8. Free Dirac Fields on Lorentzian Manifolds

4.8.1. The Dirac Operator
Let (M, g) be a Lorentzian manifold. To define correctly a Dirac operator we
need a spin structure. We keep the discussion simple and use a framework
in [7].
Let V be a complex, finite dimensional vector space. We assume that there
exists a map M x → γ a (x) ∈ L(V) such that
[γ a (x), γ b (x)]+ = 2gab (x).
Of course to make this definition clean one should use the language of bundles.
One can think of γ a as x dependent Dirac matrices.
Definition 4.15 Let M x → m(x) ∈ L(V). The operator on C∞ (M; V)
D := γ a ∂xa + m(x)
198 C. Gérard
is called a Dirac operator on M. The Dirac equation is
Dζ = 0. (D)
We need some more structure to be able to quantize the Dirac equation. We

assume that there exists a smooth map
M x → λ(x) ∈ Lh (V, V ∗ ) (4.11)
such that
γ a (x) is self-adjoint for λ(x),
1 (4.12)
m(x) − ∇a γ a (x) is anti-self-adjoint for λ(x).
2
We equip C0∞ (M; V) of the sesquilinear form

(ζ1 |ζ2 ) = ζ 1 (x) · λ(x)ζ2 (x)dμg ,
M
and we obtain that if (4.12) holds then D∗ = −D.
4.8.2. Conserved Current

We set for ζ1 , ζ2 ∈ C∞ (M; V)
J a (ζ1 , ζ2 ) := ζ 1 (x) · λ(x)γ a (x)ζ2 (x).
One proves that if ζi are solutions of the Dirac equation, then
∇a J a (ζ1 , ζ2 ) = 0.
We also obtain Green’s formula:

Lemma 4.4 Let U ⊂ M be an open set with ∂U non-characteristic. Then

ζ 1 · λDζ2 − Dζ 1 · λζ2 dμg = − ζ 1 · λγa ζ2 na dσg ,
U ∂U
for γa = gab γ b.
4.8.3. Advanced and Retarded Fundamental Solutions

We assume now that (M, g) is globally hyperbolic. It is easy to construct
fundamental solutions for the Dirac operator: in fact DD is a Klein–Gordon
operator, acting on vector valued functions, but with a scalar principal part
equal to −∇a ∇ a . The existence of fundamental solutions E± for D2 yields the
fundamental solutions S± = DE± . This is summarized in the next theorem.
Theorem 4.9 For any f ∈ C0∞ (M; V), there exist unique functions ζ ± ∈
∞ (M; V) that solve
C±sc
Dζ ± = f .
Moreover,

ζ ± (x) = (S± f )(x) := S± (x, y)fvy)dμg (y),
where S± ∈ D (M × M, L(V)) satisfy

D ◦ S± = S± ◦ D = 1l, suppS± ⊂ (x, y) : x ∈ J ± (y) .
If (4.12) holds then
λ(x)S± (x, y) = −S∓ (x, y)∗ λ(y),
that is S±∗ = −S∓ , if we equip C0∞ (M; V) with the (non-positive) scalar
product obtained from λ.
Definition 4.16 S± are called retarded/advanced Green’s functions.
S := S+ − S−
is called the Pauli–Jordan function.
Note that S = S∗ , that is S, is hermitian.
4.8.4. Cauchy Problem

We state without proof the existence and uniqueness result for the Cauchy
problem. We denote by ρ : Csc ∞ (M; V) → C∞ (; V) the trace on and by
0
Solsc (D) the space of smooth space-compact solutions of the Dirac equation.
Theorem 4.10 Let be a Cauchy surface. Then for any f ∈ C0∞ (; V) there
exists a unique ζ ∈ Solsc (D) such that ρζ = f . This satisfies suppζ ⊂ J(suppf )
and is given by

ζ (x) = − S(x, y)γ a (y)na (y)f (y)dμg (y).

4.8.5. Hermitian Structure on the Space of Solutions

We would like to equip Solsc (D) with a (positive) hermitian structure. To do
this we need an additional positivity condition. We assume that there exists a
global, time-like future directed vector field v such that
λ(x)γ a (x)va (x) > 0, ∀x ∈ M. (4.13)
200 C. Gérard
It can be shown that if this is true for one such vector field, it is automatically
true for all such vector fields, in particular for the normal vector to a given
Cauchy surface. For ζ1 , ζ2 ∈ Solsc (D) with ρζi = fi we set

ζ 1 · νζ2 := J a (ζ1 , ζ2 )na dσg = f 1 · λ(y)γ a (y)f2 (y)na (y)dσg (y),

(4.14)
f 1 · ν f 2 = f 1 · λ(y)γ (y)f2 (y)na (y)dσg (y).
a

From Theorem 4.10 we see that (Solsc (D), ν) is a hermitian space, and from
(4.13) ν is positive definite. Moreover from the Gauss formula ν is independent
of the choice of a Cauchy surface.
4.8.6. Spacetime Fields

Theorem 4.11 (1) Consider S : D(M; V) → C∞ (M; V). Then RanS =
Solsc (D) and KerS = DC0∞ (M; V).
(2) One has
Sf 1 · νSf2 = ( f1 |Sf2 ), fi ∈ C0∞ (M; V).
Proof (1) is left to the reader.
(2) Theorem 4.10 can be rewritten as
1l = Sρ ∗ γ a∗ na ρ, on Solsc (D),
hence by (1) as
S = Sρ ∗ γ a∗ na ρS,
which implies (2) since S∗ = S.
Theorem 4.12 (1) The following spaces are pre-Hilbert spaces:

(C0∞ (M; V)/DC0∞ (M; V), S), (Solsc (D), ν), (C0∞ (; V), ν ).
(2) The following maps are unitary:
S ρ
(C0∞ (M; V)/DC0∞ (M; V), S) −→(Solsc (D), ν) −→(C0∞ (; V), ν ).
4.8.7. Quasi-Free States for the Free Dirac Field

As for the Klein–Gordon case we choose the pre-Hilbert space
(C0∞ (M; V)/DC0∞ (M; V), S).
The associated CAR algebra is denoted by CAR(C0∞ (M; V), S).
The causality is a bit different: we obtain
[φ( f ), φ(g)]+ = Re( f |Sg) = 0,
if suppf , suppg are causally disjoint, that is fields supported in causally disjoint
regions anti-commute. This puzzle is solved by considering only even elements
of CAR(C0∞ (M; V)/DC0∞ (M; V), S) as true physical observables.
Let us consider a gauge-invariant quasi-free state ω, defined by the complex
covariances (λ± ) which are sesquilinear forms on C0∞ (M; V)/DC0∞ (M; V)).
Again one assumes that they are obtained from sesquilinear forms ± on
C0∞ (M; V) which pass to quotient, that is such that
D ◦ ± = ± ◦ D = 0. (4.15)
Introducing as before the distributional kernels ± (x, y) ∈ D (M × M) ⊗

V ⊗ V ∗ . Then (4.15) becomes
D(x, ∂x )± (x, y) = D(y, ∂y )± (x, y) = 0. (4.16)
4.9. Microlocal Analysis of Klein–Gordon Quasi-Free States

4.9.1. The Need for Renormalization
The stress-energy tensor for a classical Klein–Gordon field is given by
1
Tμν (x) = ∇μ φ(x)∇ν φ(x) − gμν (x) gab (x)∇a φ(x)∇b φ(x) − m2 φ 2 (x) .
2
For a quantized Klein–Gordon field one would like to be able to define Tμν (x)
as an operator valued distribution. This means the following.
We choose a state ω (say a quasi-free state) and fix f1 , . . . , fn , g1 , . . . , gp ∈
C0∞ (M). Then
n

p
x → ω φ( fi )Tμν (x) φ(gj )
1 1
should be a distribution on M. This is never the case, even on Minkowski

space: for example ω(φ 2 (x)) should be the trace on x = x of ω(φ(x)φ(x )) =
ω2 (x, x ), which makes no sense. We need to subtract the singular part of
ω2 (x, x ) near the diagonal, that is to perform a renormalization.
Another requirement is that the renormalization procedure should be
“covariant”: it should depend only on the metric in an arbitrarily small neigh-
borhood of x. This implies that it will be covariant under isometric embeddings.
202 C. Gérard
The procedure is as follows. One first performs the “point-splitting,” that is

consider
1
T μν (x, y) = ∇μ φ(x)∇ν φ(y)− gμν (x) gab (x)∇a φ(x)∇b φ(y) − m2 φ(x)φ(y) .
2
One then removes the singular part by setting
: φ(x)φ(y) : = φ(x)φ(y) − cHad (x, y)1l,
where cHad (x, y) is a well chosen distributional kernel. Then one has to check
that the distributions
n

p
ω φ(xi )Tμν (x, y) : φ(yj )
1 1
have well-defined restrictions to x = y.
4.9.2. Old Form of Hadamard States

Choose a Cauchy surface . A causal normal neighborhood N of in M is an
open neighborhood of such that is a Cauchy surface of (N, g) and for each
x, x ∈ N such that x ∈ J + (x ) there exists a convex normal open set containing
J + (x ) ∩ J − (x). We fix a time function T : M → R , that is a smooth function
which increases towards the future, and an open neighborhood O ⊂ M × M
of the set of pairs of causally related points (x, x ) such that J ± (x) ∩ J ∓ (x ) are
contained in a convex, normal open neighborhood.
One fixes also O an open neighborhood in N × N of the set of pairs of
causally related points such that O ⊂ O.
The squared geodesic distance σ (x, x ) is smooth on O and well defined
on the subset of N × N consisting of causally related points. The van Vleck–
Morette determinant is
1 1
(x, x ) := − det(−∇α ∇β σ (x, x ))|g|− 2 (x)|g|− 2 (x ).
One can then construct a sequence of functions v(n) ∈ C∞ (O × O) of the form

n
v(n) (x, x ) = σ (x, x )i vi (x, x ),
1
where the vi ∈ C∞ (O × O) are uniquely determined by some transport

equations (the so-called Hadamard recursion relations).
One then defines a sequence of distributions c(n)

Had ∈ D (O × O) for n ≥ 1 by
1
(2π )2 (x, x ) 2
c(n)
Had (x, x ) = lim→0+
σ (x, x ) + 2i(T(x) − T(x )) + 2

+ lim→0+ v(n) (x, x ) ln σ (x, x ) + 2i(T(x) − T(x )) .
Let us now give the old definition of Hadamard states, restricting ourselves to
real Klein–Gordon fields.
Definition 4.17 A quasi-free state ω on CCR(C0∞ (M), E) is a Hadamard state
if for any m ∈ N there exists n ∈ N such that ω2 −c(n)
Had is of class C in O ×O.
m
4.9.3. The Wavefront Set on a Manifold

Let M be a manifold and T ∗ M be its cotangent bundle. The zero section T0∗ M
will be denoted by Z.
We recall the spaces: D(M) (smooth compactly supported functions), D (M)
(distributions), E(M) (smooth functions with well-known topology), E (M)
(distributions with compact support).
4.9.4. Operations on Conic Sets

A set ⊂ T ∗ M\Z is conic if (x, ξ ) ∈ ⇒ (x, tξ ) ∈ for all t > 0. Let
i ⊂ T ∗ M\Z, i = 1, 2 be conic sets. We set
− := {(x, −ξ ) : (x, ξ ) ∈ },
1 ⊕ 2 := {(x, ξ1 + ξ2 ) : (x, ξi ) ∈ i }.
Let Mi , i = 1, 2 be two manifolds and ⊂ T ∗ M1 × M2 \Z be a conic set.
The elements of T ∗ M1 × M2 \Z will be denoted by (x1 , ξ1 , x2 , ξ2 ) which allows
us to consider as a relation between T ∗ M2 and T ∗ M1 , still denoted by .
Clearly maps conic sets into conic sets. We set:
:= {(x1 , ξ1 , x2 , −ξ2 ) : (x1 , ξ1 , x2 , ξ2 ) ∈ },
M1 := {(x1 , ξ1 ) : ∃x2 such that (x1 , ξ1 , x2 , 0) ∈ } = (Z2 ),
M2 := {(x2 , ξ2 ) : ∃x1 such that (x1 , 0, x2 , ξ2 ) ∈ } = −1 (Z1 ).
4.9.5. Operations on Distributions

(1) Complex conjugation: if u ∈ D (M) then WFu = −WF(u).
(2) Tensor product: if ui ∈ D (Mi ), i = 1, 2 then
204 C. Gérard
WF(u1 ⊗ u2 ) ⊂ (WF(u1 ) × WF(u2 )) ∪ (suppu1 × {0})

× WF(u2 ) ∪ WF(u1 ) × (suppu2 × {0})
⊂ (WF(u1 ) × WF(u2 )) ∪ Z1 × WF(u2 ) ∪ WF(u1 ) × Z2 .
(3) Restriction to a submanifold: let S ⊂ M be a submanifold. The conormal

bundle TS∗ M is defined as:
TS∗ M := {(x, ξ ) ∈ T ∗ M\Z : x ∈ S, ξ · v = 0 ∀v ∈ Tx S}.
If u ∈ D (M) the restriction u|S of u to S is well defined if
WFu ∩ TS∗ M = ∅.
One has
WFu|S ⊂ {(x, ξ|Tx S ) : x ∈ S, (x, ξ ) ∈ WF(u)}.
(4) Product: if ui ∈ D (M), i = 1, 2 then u1 u2 is well defined if WF(u1 ) ⊕

WF(u2 ) ∩ Z = ∅ and then
WF(u1 u2 ) ⊂ WF(u1 ) ∪ WF(u2 ) ∪ WF(u1 ) ⊕ WF(u2 ).
(5) Kernels: let K : D(M2 ) → D (M1 ) be linear continuous and K(x1 , x2 ) ∈

D (M1 × M2 ) its distributional kernel.
(K) = ∅ and
Then Ku is well defined for u ∈ E (M2 ) if WF(u) ∩ WFM 2
then

WF(Ku) ⊂ M1WF(K) ∪ WF (K) ◦ WF(u).
(6) Composition: let K1 ∈ D (M1 × M2 ), K2 ∈ D (M2 × M3 ), where K2 is

properly supported, that is the projection suppK2 → M2 is proper. Then
K1 ◦ K2 is well defined if
WF (K1 )M2 ∩ M2WF(K2 ) = ∅,
and then
WF (K1 ◦K2 ) ⊂ WF (K1 )◦WF (K2 )∪(M1WF(K1 )×Z3 ) ∪ (Z1 ×WF(K2 )M3 ).
4.9.6. Parametrices for the Klein–Gordon Operator

Once we choose an orientation, we can define for x ∈ M the open future/past
light cones Vx± ⊂ Tx M. We denote by Vx±
∗ ⊂ T ∗ M the dual cones
x
∗
Vx± = {ξ ∈ Tx∗ M : ξ · v > 0, ∀v ∈ Vx± , v = 0}.
For simplicity we write

∗ ∗
ξ 0 (resp. ξ 0) if ξ ∈ Vx+ ( resp. Vx− ),
∗ cl ∗ cl
ξ 0 (resp. ξ 0) if ξ ∈ (Vx+ ) ( resp. (Vx− ) ).
The principal symbol of the Klein–Gordon operator P is p(x, ξ ) = ξμ gμν (x)ξν

and one sets
N := p−1 ({0}) ∩ T ∗ M\Z,
called the characteristic manifold of P. Note that N splits into its two connected
components (positive/negative energy shells):
N = N + ∪ N − , N ± = N ∩ {±ξ 0}.
We denote by Hp the Hamilton field of p. We denote by X = (x, ξ ) the points

in T ∗ M\Z. The bicharacteristic (Hamilton curve for p) passing through X is
denoted by B(X). For X, Y ∈ N we write X ∼ Y if Y ∈ B(X). Clearly this is an
equivalence relation.
For X ∼ Y, we write X # Y (resp. X ≺ Y) if X comes strictly after (before)
Y w.r.t. the natural parameter on the bicharacteristic curve through X and Y.
We recall Hörmander’s propagation of singularities theorem:
Theorem 4.13 Let u ∈ D (M) such that Pu ∈ C∞ (M). Then
WF(u) ⊂ N, X ∈ WF(u) ⇒ B(X) ⊂ WF(u).
The bicharacteristic relation of P is the set
C := {(X, Y) ∈ N × N : X ∼ Y}.
We set
N := {(X, X)} ∩ N × N
the diagonal in N × N.
Parametrices (i.e. inverses modulo smoothing operators) of operators of real
principal type (of which Klein–Gordon operators are an example) were studied
by Duistermaat and Hörmander in the famous paper [8]. They introduced the
notion of distinguished parametrices, that is parametrices which are uniquely
determined (modulo smoothing terms of course) by the wavefront set of their
kernels. Distinguished parametrices are in one-to-one correspondence with
orientations of C, defined below.
Definition 4.18 An orientation of C is a partition of C\N as C1 ∪ C2 where

Ci are open sets in C\N and inverse relations (i.e. Exch(C1 ) = C2 ).
206 C. Gérard
Note that Ci = ∅, Ci = C\N (because they are inverse relations) and Ci are
open and closed in C\N . Therefore Ci are a union of connected components
of C\N .
Theorem 4.14 (D-H) Let C\N = C1 ∪ C2 be an orientation of C. Then there
exists parametrices Ei , i = 1, 2 of P such that
WF (Ei ) ⊂ ∗ ∪ Ci ,
where ∗ is the diagonal in T ∗ M\Z × T ∗ M\Z. Any left or right parametrix

with the same property is equal to E1 or E2 modulo C∞ .
Orientations of C are themselves in one-to-one correspondence with the
partitions of N 1 ∪ N 2 = N into open and closed sets, that is into connected
components of N. For the Klein–Gordon operator N has two connected
components:
∗
N± := {X ∈ N : ξ ∈ Vx± },
which are invariant under the bicharacteristic flow, hence two orientations,
hence four distinguished parametrices. The two connected components are
N+ , N− . The two orientations are
C\N = C+ ∪ C− for
C+ := {(X, Y) ∈ C : X # Y}, C− := {(X, Y) ∈ C : X ≺ Y}
and
C\N = C+ ∪ C− for
C+ := {(X, Y) ∈ C : x ∈ J + (y)}, C− := {(X, Y) ∈ C : x ∈ J − (y)}.
The associated parametrices are well-known in physics, we have already
encountered two of them, namely E± .
Feynman: denoted EF :
WF(EF ) = ∗ ∪ C+ .
Anti-Feynman: denoted EF :
WF(EF ) = ∗ ∪ C− .
Retarded: denoted E+ :
WF(E+ ) = ∗ ∪ C+ .
Advanced: denoted E− :
WF(E− ) = ∗ ∪ C− .
The parametrices E± are more fundamental since they are used to define the
symplectic form E. The parametrices EF , EF appear in connection with the
vacuum state on Minkowski space and with Hadamard states on general curved
spacetimes.
4.9.7. Examples
Assume that P = ∂t2 + 2 (whatever 2 is, e.g. a real number). Then:
sin t
E+ (t) = θ (t) ,

sin t
E− (t) = −θ (−t) ,

1 it
EF (t) = e θ (t) + e−it θ (−t) ,
2i
1 −it
EF (t) = − e θ (t) + eit θ (−t) ,
2i
for θ (t) = a Heaviside function.

We prove some properties of the distinguished parametrices.
Lemma 4.5 We have
(1) WF (E+ − E− ) = C,
(2) WF (E+ − EF ) = C ∩ N− × N− ,
(3) WF (E− − EF ) = C ∩ N+ × N+ .
Proof (1) Since E+ and E− have disjoint wavefront sets above {x1 = x2 }, we
see that above {x1 = x2 }
WF (E+ − E− ) = WF (E+ ) ∪ WF (E− ) = C\N .
Since P1 (E+ − E− ) ∈ C∞ by propagation of singularities we obtain that

N ⊂ WF (E+ − E− ). This proves (1).
(2) Above {(x1 , x2 ) : x1 ∈ J − (x2 )} we have
WF (E+ − EF ) = WF (EF ) = {(X1 , X2 ) : x1 ∈ J − (x2 ), ξ1 0}.
Using again P1 (E+ −EF ) ∈ C∞ we obtain that WF (E+ −EF ) = C∩N− ×N− .
The proof of (3) is similar.
208 C. Gérard
4.9.8. The Theorem of Radzikowski

Definition 4.19 Let ± : D(M) → E(M) be linear continuous. The pair ±
satisfies the Hadamard condition if
(Had) WF(± ) = {(X1 , X2 ) ∈ N ± × N ± : X1 ∼ X2 }.
The pair ± satisfies the generalized Hadamard condition if there exists conic
sets ± with (X1 , X2 ) ∈ ± ⇒ ±ξ1 0, ±ξ2 0 such that
(genHad) WF(± ) ⊂ ± .
We introduce the following conditions on a pair ± :
(KG) P ◦ ± , ± ◦ P smoothing,
(CCR) + − − − iE smoothing.
The following theorem is the theorem of Radzikowski (extended to the
complex case).
Theorem 4.15 The following three conditions are equivalent:
(1) ± satisfy (Had), (KG), and (CCR),
(2) ± satisfy (genHad), (KG), and (CCR),
(3) one has
± = i(EF − E∓ ) modulo C∞ (M × M).
Proof (1) ⇒ (2): obvious.
(2) ⇒ (3): set S± = i(EF − E∓ ). By Lemma 4.5 we have WF(S± ) ⊂
C ∩(N± ×N± ). By (KG) and Theorem 4.13 we obtain that WF(± ) ⊂ N ×N.
Using then (genHad) we obtain that WF(± ) ⊂ ± ∩ N × N ⊂ N+ × N+ .
This implies that
WF(± − S± ) ⊂ N± × N± ,
which implies in particular that
WF(+ − S+ ) ∩ WF(− − S− ) = ∅. (4.17)
By (CCR) we have also
(+ − S+ ) − (− − S− ) = iE − (S+ − S− ) = iE − iE mod C∞ (M × M).
By (4.17) this implies that both ± − S± are smooth.
(3) ⇒ (1): (KG) and (CCR) are obvious, (Had) follows from Lemma 4.5.
There is a remaining painful step, which I will only briefly explain. I will
consider real fields for simplicity. The conclusion of the above theorem for the
real covariance ω2 (x, x ) is that ω2 = i(EF − E+ ) mod C∞ . Then one has to

perform some painful computations to prove that i(EF − E+ ) = cHad mod C∞ .
4.9.9. Return to the Stress-Energy Tensor

Let us forget the derivatives and juste consider : φ 2 (x) :. We put just one field
on each side (extension to several fields is straightforward).
We compute the expectation value of
φ(x1 )φ(x)φ(x )φ(x2 ) − φ(x1 )φ(x2 )cHad (x, x ),
in the state ω. By the quasi-free property, we obtain a sum of two types of
terms:
ω2 (x1 , x)ω2 (x , x2 ), ω2 (x, x2 )ω2 (x1 , x )
and
ω2 (x1 , x2 ) × (ω2 (x, x ) − cHad (x, x )).
The second term has a restriction to the diagonal x = x , since ω2 − cHad is
smooth.
For the first term we consider the wavefront set of the distribution. By the
tensor product rule we obtain
WF(ω2 ) × WF(ω2 ) ∪ (suppω2 × {0}) × WF(ω2 ) ∪ WF(ω2 ) × (suppω2 × {0}).
We compute the conormal to S = {x = x }:
TS∗ M = {(x1 , ξ1 , x, ξ , y, η, x2 , ξ2 ) : ξ1 = ξ2 = 0, x = y, ξ = −η}.
We have to show the intersection is empty, which is obvious since (X, Y) ∈
WF(ω2 ) implies ξ , η = 0.
4.10. Construction of Hadamard States

It is not a priori obvious that Hadamard states exist on an arbitrary globally
hyperbolic spacetime. In this section we explain some constructions of
Hadamard states.
4.10.1. Ultrastatic Spacetimes

Consider an ultrastatic spacetime (R×, g), g = −dt2 +h. Let a = −h +m2
1
on and = a 2 . One can then define the vacuum state ωvac , whose two-point
function is given by the kernel
210 C. Gérard
1 ±it
± (t) =
vac e ,
2
that is writing u ∈ C∞ (R × ) as R t → u(t) ∈ C∞ ():

1 ±i(t−s)
± u(t) =
vac
e u(s)ds.
R 2
Sahlmann and Verch [18] have shown that ωvac is a Hadamard state. Another
proof can be given by using the arguments of Section 4.10.3.
Similarly one can define the thermal state at temperature β −1 , β > 0 with
the kernel
1
(e±it + e∓it e−β ).
β
± (t) :=
2(1 − e−β )
This is again a Hadamard state since e−β is a smoothing operator.
The conclusion is that vacuum or thermal states on ultrastatic (or static)
spacetimes are Hadamard states.
4.10.2. The FNW Deformation Argument

Let (M, g) be a globally hyperbolic spacetime. The deformation argument of
Fulling, Narcowich, and Wald [11] is based on two facts.
The first fact is the so-called time slice property: if U ⊂ M is a neighborhood
of a Cauchy surface , then for any u ∈ C0∞ (M) there exists v ∈ C0∞ (U) such
that u − v ∈ PC0∞ (M). In other words
C0∞ (M)/PC0∞ (M) = C0∞ (U)/PC0∞ (M).
This implies that a pair of distributions ± ∈ C0∞ (U × U) satisfying:
P ◦ ± = ± ◦ P = 0,
+ − − = −iE on U × U,
± ≥ 0 on C0∞ (U)
generate a quasi-free state on M.
The second fact is Hörmander’s propagation of singularities theorem
(Theorem 4.13): If ± satisfy (Had) (or (genHad)) over U × U, and
P ◦ ± = ± ◦ P = 0, then ± satisfy (Had) or (genHad) globally.
The conclusion of these two facts is that if g1 , g2 are two Lorentzian metrics
such that (M, gi ) is globally hyperbolic, if they have a common Cauchy
surface , and if they coincide in a neighborhood of , then a Hadamard
state for the Klein–Gordon field on (M, g1 ) generates a Hadamard state for the
Klein–Gordon field on (M, g2 ).
One argues then as follows. Let us fix a Cauchy surface for (M, g) and
identify (M, g) with (R×, −c2 (t, x)dt2 +hij (t, x)dxi dxj ). We set t = {t}×.
We fix a real function r ∈ C∞ (M) and consider P = −∇ a ∇a + r(x) – the
associated Klein–Gordon operator. One chooses an ultra-static metric
gus = −dt2 + hjk,us (x)dxj dxk , rus (x) = m2 > 0,
and an interpolating metric gint = −c2int (t, x)dt2 + hjk,int (x)dxj dxk , and real
function rint ∈ C∞ (M) such that (gint , rint ) = (g, r) near T , (gint , rint ) =
(gus , m2 ) −T .
The vacuum state ωvac for (M, gus ) is Hadamard for Pus , hence generates a
Hadamard state for Pint , which itself generates a Hadamard state ω for P.
Using the Cauchy evolution operator from −T to T , one sees that ω is
pure, since ωvac is pure.
4.10.3. Construction of Hadamard States by

Pseudodifferential Calculus
Let us now briefly describe another construction given in a joint work with
Michal Wrochna [12]. It relies on the choice of a Cauchy surface and on the
global pseudodifferential calculus on .
We identify M with R × , with the split metric g = −c2 (t, x)dt2 +
hij (t, x)dxi dxj , and we set s := {s} × ⊂ M. For simplicity we will assume
that is either equal to Rd or to a compact manifolds, but much more general
situations can be treated as well.
We denote by m () the space of (uniform) pseudodifferential operators of
order m on , corresponding to quantization of symbols in S1,0 m (T ∗ ), and by
ph () the subspace of pseudodifferential operators with poly-homogeneous

symbols. We set

∞ () = m (), −∞ () = m ().
m∈R m∈R
We denote by H s () the Sobolev space of order s on and

H ∞ () = H s (), H −∞ () = H s ().
s∈R s∈R
We use the third version of the phase space, namely (C0∞ ()⊗C2 , q), where
we recall that q = iσ . We embed C0∞ () ⊗ C2 into D () ⊗ C2 using the
natural scalar product on C0∞ () ⊗ C2 , that is we identify sesquilinear forms
with operators.
212 C. Gérard
Quasi-free states are now defined by a pair of covariances λ± on

C0∞ () ⊗ C2 , and the relationship with the spacetime covariances ± is
± = (ρ ◦ E)∗ ◦ λ± ◦ (ρ ◦ E).
It is natural to restrict attention to states with covariances λ± ∈ ∞ () ⊗

M(C2 ). It can be shown that if a state has pseudodifferential covariances on
s for some s it has pseudodifferential covariances on s for any s.
The most important part is to characterize the Hadamard condition in terms
of λ± , which relies on the construction of a parametrix for the Cauchy problem
on , see Proposition 4.8 below.
4.10.3.1. The Model Klein–Gordon Equation

By a change of coordinates and conjugation with a convenient weight, the
equation reduces to the general form
∂t2 u + a(t, x, ∂x )u = 0,
where a(t, x, ∂x ) is a second order, elliptic self-adjoint operator on L2 ().

If = Rd , a natural hypothesis (which if needed can be rephrased in terms
of the original metric g) is that

a(t, x, ∂x ) = − ajk (t, x)∂xj ∂xk + (t, x)∂xj + r(t, x),
ij aj
where
C−1 ξ 2 ≤ ajk (t, x)ξj ξk ≤ Cξ 2 ,
|∂tm ∂xα ajk |, |∂tm ∂xα aj ], ∂tm ∂xα r bounded locally uniformly in t,
an assumption of course related to the uniform pseudo-differential calculus
on . Let us set for s ∈ R ρs u := (us , i−1 ∂t us ) and Us the solution of the
Cauchy problem on s , that is
(∂t2 + a(t))Us = 0, ρs ◦ Us = 1l.
4.10.3.2. Parametrices for the Cauchy Problem

Let R t → b(t) be a map with values in linear operators on L2 () with
Dom b(t) = H 1 () such that b(t)−b∗ (t) is bounded, and R t → b(t)
t is norm
continuous with values in B(H 1 (), L2 ()). We denote by Texp(i s b(σ )dσ )
the strongly continuous group with generator b(t). A routine computation
shows that
t
(∂t + a(t))Texp i
2
b(σ )dσ = 0
s
if and only if b(t) solves the following Riccati equation:
i∂t b(t) − b2 (t) + a(t) = 0. (4.18)
This equation can be solved modulo C∞ (R, −∞ ()) by symbolic calculus,

which amounts to solving transport equations.
Theorem 4.16 There exists b(t) ∈ C∞ (R, 1 ()) which is unique modulo
C∞ (R, −∞ ()) such that
(i) b(t) = (t) + C∞ (R, 0 ()),

1 1
(ii) (b(t) + b∗ (t))−1 = (t)− 2 (1l + r−1 (t))(t)− 2 , r−1 (t) ∈ C∞ (R, −1 ()),
(iii) (b(t) + b∗ (t))−1 ≥ c(t)(t)−1 , for some c(t) > 0,
(iv) i∂t b − b2 + a = r−∞ ∈ C∞ (R, −∞ ()).
Note that if b+(t) := b(t) solves (4.18), so does b− ()t := −b∗ (t). Setting
u± (t, s) = Texpi s b± (σ )dσ , we obtain the following result.
t
Theorem 4.17 Set
rs0± ··= ∓(b+ (s) − b− (s))−1 b∓ (s) ∈ 0 (),

rs1± ··= ±(b+ (s) − b− (s))−1 ∈ −1 (),
rs± f ··= r0± f 0 + r1± f 1 , f = ( f 0 , f 1 ) ∈ H ∞ () ⊗ C2 .
Then
Us = u+ (·, s)rs+ + u− (·, s)rs− + C∞ (R, −∞ ()).
4.10.3.3. Pure Hadamard States

The following proposition summarizes conditions implying that a pair of
operators λ±
s are the Cauchy surface covariances (at time s) of a (pure)
Hadamard state.
Proposition 4.8 Let λ± ∞ ∞
s : H () ⊗ C → H () ⊗ C be continuous. Then
2 2
λ±
s are the Cauchy surface covariances of a Hadamard state ω if:
(i) λ±∗ ± ±
s = λs , λ ≥ 0,
(ii) λ+ −
s − λs = q,
(iii) f ∈ H −∞ () ⊗ C2 ∩ Kerλ∓ ⇒ WF(Us f ) ⊂ N ± .
If additionally c±
s := ±iq
−1 ◦ λ± are projections, then ω is pure.
s
214 C. Gérard
Note that if f ∈ H −∞ ()⊗C2 and rs∓ f = 0, then WFUs f ⊂ N ± . This allows

us easily to construct a pure Hadamard state associated with the asymptotic
solution b(t) of (4.18).
In fact if we set
∗
∗ − 12 b (s) 1l
Ts (b) = (b(s) + b (s)) ,
b(s) 1l
we note that

0 1l ∗ 1l 0
q= = Ts (b) Ts (b).
1l 0 0 1
It follows that

± ∗ ± + 1l 0 − 0 0
λ (s) := Ts (b) π Ts (b), π = , π = ,
0 0 0 1
are the Cauchy surface covariances of a pure Hadamard state.
Remark 4.2 One can show that if the Cauchy surface covariances of a state ω
are pseudodifferential at some time s, the same is true at any other time s .
Moreover one can show that any pure Hadamard state with pseudodifferen-
tial Cauchy surface covariances is of the form given above, for some t → b(t)
as in Theorem 4.16.
The above construction of Hadamard states by pseudodifferential calculus

can be generalized to more delicate situations, like linearized Yang–Mills
fields, where the deformation argument cannot be applied anymore, see [13].
4.10.4. Construction of Hadamard States by the

Characteristic Cauchy Problem
It is possible to construct Hadamard states by replacing the (space-like) Cauchy
surface by some null hypersurface C; typically C is chosen to be the forward
lightcone from some point p ∈ M. Then the interior M0 of C, when equipped
with g, is a globally hyperbolic spacetime in its own right. This approach was
introduced by Valter Moretti [16] and generalized afterwards to various similar
situations [4, 5, 6, 17] in order to construct a distinguished Hadamard state on
asymptotically flat spacetimes, for the conformal wave equation.
In [14] we construct a large family of pure Hadamard states on the cone C,
using again pseudodifferential calculus.
4.10.4.1. The Geometric Framework

Let (M, g) be a globally hyperbolic spacetime, p ∈ M a distinguished point
of M. Let
C ··= ∂J + (p)\{p}
be the future lightcone from p and
M0 = I + (p) its interior.
One can show that (M0 , g) is also a globally hyperbolic spacetime. Moreover
one can show that
J + (K; M0 ) = J + (K; M), J − (K; M0 ) = J − (K; M) ∩ M0 , ∀ K ⊂ M0 . (4.19)
Some global conditions are needed to avoid singularities of C. One assumes
that there exists f ∈ C∞ (M) such that:
(1) C ⊂ f −1 ({0}), ∇a f = 0 on C, ∇a f (p) = 0, ∇a ∇b f (p) = −2gab (p),
(2) the vector field ∇ a f is complete on C.
Note that since C is a null hypersurface the vector field ∇ a f is tangent to C.
From this hypothesis it is easy to construct coordinates ( f , s, θ ) near C, with
f , s ∈ R, θ ∈ Sd−1 such that C ⊂ {f = 0} and
gC = −2dfds + h(s, θ )dθ 2 , (4.20)
where h(s, θ )dθ 2 is a Riemannian metric on Sd−1 .
Such choice of coordinates allows one to identify C with C̃ ··= R × Sd−1 .
A natural space of smooth functions on C̃ is then provided by H ∞ (C̃) – the
intersection of Sobolev spaces of all orders, defined using the round metric
m(θ )dθ 2 on Sd−1 .
4.10.4.2. Bulk to Boundary Correspondence

Let us consider the restriction P0 of the Klein–Gordon operator P to C∞ (M0 ),
and E0± its advanced/retarded Green’s functions. From (4.19), one obtains
E0± = E±M0 ×M0 , hence:
E0 = EM0 ×M0 .
This implies that any solution φ0 ∈ Solsc (P0 ) uniquely extends to φ ∈
Solsc (P). In particular its trace φ0C is well defined.
It is convenient to introduce the coordinates (s, θ ) on C, and to set:
ρ : Solsc (P0 ) → C∞ (R × Sd−1 )
φ → β −1 (s, θ )φC (s, θ ),
216 C. Gérard
for
1 1
β(s, θ ) ··= |m| 4 (θ )|h|− 4 (s, θ ),
where h(s, θ )dθ 2 is defined in (4.20) and m(θ )dθ 2 is the round metric on Sd−1 .
We equip H ∞ (C̃) with the symplectic form:

1
g1 σC g2 ··= (∂s g1 g2 − g1 ∂s g2 )|m| 2 (θ )dsdθ , g1 , g2 ∈ H(C̃). (4.21)
R×Sd−1
Introducing the charge q ··= iσC we have:
g1 qg2 = 2(g1 |Ds g2 )L2 (C̃) , g1 , g2 ∈ H(C̃), (4.22)
where Ds = i−1 ∂s is self-adjoint on L2 (C̃) on its natural domain. Clearly

(H ∞ (C̃), σC ) is a complex symplectic space.
One can show the following result (the second statement follows from
Stokes theorem), which summarizes the bulk to boundary correspondence:
Proposition 4.9 (1) ρ maps Solsc (P0 ) into H(C̃);
(2) ρ : (Solsc (P0 ), σ ) → (H(C̃), σC ) is a monomorphism, that is:
ρφ 1 σC ρφ2 = φ 1 σ φ2 , ∀ φ1 , φ2 ∈ Solsc (P0 ).
4.10.4.3. Hadamard States on the Cone

From Proposition 4.9, we see that any quasi-free state ωC on CCR(H ∞ (C̃), σC )
generates a quasi-free state ω0 on CCR(C0∞ (M0 )/PC0∞ (M0 ), E0 ). The task is
now to give conditions on ωC implying that ω0 is Hadamard.
Let us denote by x = (r, s, y) the coordinates ( f , s, θ ) near C and by
ξ = (ρ, σ η) the dual coordinates. The complex covariances of ωC are denoted
by λ± ∈ D (C̃ × C̃). The associated covariances ± of ω0 are then:
± ··= (ρ ◦ E0 )∗ ◦ λ± ◦ (ρ ◦ E0 ).
One can show the following result.

Theorem 4.18 Assume that λ± : H ∞ (C̃) → H ∞ (C̃) and C̃WF(λ± ) =
WF(λ± ) = ∅. Then if:
C̃
(i) WF(λ± ) ∩ {(Y1 , Y2 ) : ±σ1 < 0 or ± σ2 < 0} = ∅,

(ii) WF(λ± ) ∩ {(Y1 , Y2 ) : ±σ1 > 0 and ± σ2 > 0} ⊂ ,
± satisfy (Had).
A state on CCR(H ∞ (C̃), σC ) satisfying thc assumptions of Theorem 4.18
will be called a Hadamard state on the cone.
4.10.4.4. Hadamard States on the Cone and Pseudodifferential Calculus

Recall that q defined in (4.22) equals 2Ds , whose resolvent (2Ds − z)−1 is not
an elliptic pseudodifferential operator on C̃, for the usual calculus. However,
it belongs to a larger bi-homogeneous class p1 ,p2 (C̃), which can be loosely
defined as p1 (R) ⊗ p2 (Sd−1 ).
It is then possible to construct Hadamard states on the cone by mimicking
the arguments in Section 4.10.3. Note that no construction of a parametrix
for the characteristic Cauchy problem on C̃ is necessary, since the Hadamard
condition on the cone is rather explicit.
4.10.4.5. Purity Inside the Cone

A natural issue with this construction is wether a pure state on the cone gener-
ates a pure state inside. This is not a priori obvious, since the map ρ is not sur-
jective. Nevertheless it is shown in [14] that purity is preserved using the results
of Hörmander [15] on the characteristic Cauchy problem. This is the only place
where the solvability of the characteristic Cauchy problem is important.
References
[1] C. Bär, K. Fredenhagen, editors. Quantum Field Theory on Curved Spacetimes,
Springer Lecture Notes in Physics, Springer, 2009.
[2] C. Bär, N. Ginoux, F. Pfäffle: Wave Equations on Lorentzian Manifolds and
Quantization, ESI Lectures in Mathematics and Physics, Springer, 2007.
[3] Bjorken, J.D., Drell, S.D. Relativistic Quantum Mechanics, McGraw-Hill, 1964.
[4] C. Dappiaggi, V. Moretti and N. Pinamonti, Distinguished Quantum States in a
Class of Cosmological Spacetimes and their Hadamard Property, J. Math. Phys.
50, 2009, 062304.
[5] C. Dappiaggi, V. Moretti and N. Pinamonti, Rigorous Construction and Hadamard
Property of the Unruh State in Schwarzschild Spacetime, Adv. Theor. Math.
Phys. 15, 2011, 355.
[6] C. Dappiaggi, D. Siemssen, Hadamard States for the Vector Potential on
Asymptotically Flat Spacetimes, Rev. Math. Phys. 25, 2013, 1350002.
[7] J. Derezinski, C. Gérard. Mathematics of Quantization and Quantum Fields,
Cambridge Monographs on Mathematical Physics, Cambridge University Press,
2013.
[8] Duistermaat, J.J., Hörmander, L. Fourier Integral Operators. II. Acta Math. 128,
183–269, 1972.
[9] Friedlander, F.G. The Wave Equation on a Curved Space-time Cambridge
University Press, 1975.
[10] S. Fulling. Aspects of Quantum Field Theory in Curved Space-Time, London
Mathematical Society Student Texts, Cambridge University Press, 1989.
218 C. Gérard
[11] Fulling, S.A., Narcowich, F.J., Wald, R.M. Singularity Structure of the Two-point
Function in Quantum Field Theory in Curved Space-time, II, Annals of Physics,
136, 243–272, 1981.
[12] C. Gérard, M. Wrochna, Construction of Hadamard States by Pseudo-differential
Calculus, Comm. Math. Phys. 325, 713–755, 2014.
[13] C. Gérard, M. Wrochna, Hadamard States for the Linearized Yang-Mills Equation
on Curved Spacetime, arXiv:1403.7153, 2014.
[14] C. Gérard, M. Wrochna, Construction of Hadamard States by Characteristic
Cauchy Problem, arXiv:1409.6691, 2014.
[15] L. Hörmander, A Remark on the Characteristic Cauchy Problem, J. Funct. Anal.
93, 270–277, 1990.
[16] V. Moretti, Uniqueness Theorem for BMS-invariant States of Scalar QFT on
the Null Boundary of Asymptotically Flat Spacetimes and Bulk-boundary
Observable Algebra Correspondence, Comm. Math. Phys. 268, 727–756, 2006.
[17] V. Moretti, Quantum Out-states Holographically Induced by Asymptotic
Flatness: Invariance under Space-time Symmetries, Energy Positivity and
Hadamard Property, Comm. Math. Phys. 279, 31–75, 2008.
[18] Sahlmann, H., Verch, R. Passivity and Microlocal Spectrum Condition, Comm.
Math. Phys. 214, 705–731, 2000.
Département de Mathématiques, Université de Paris XI, 91405 Orsay Cedex France

5
A Minicourse on Microlocal Analysis
for Wave Propagation
András Vasy
5.1. Introduction
This minicourse describes a microlocal framework for the linear analysis that
has been useful for the global understanding of wave propagation phenomena.
While it is useful to have some background in microlocal analysis since
relatively sophisticated frameworks and notions are discussed and it may be
easier for the reader to start with simpler cases, I will in fact cover the subject
from scratch.
Section 5.2 covers the differential operator aspects of totally characteristic,
or b-, operators, in particular pointing out the differences from the local theory,
or from the global theory on compact manifolds without boundary.
The basics of microlocal analysis in Section 5.3 are covered roughly
following Melrose’s lecture notes [31] (which introduced me to the subject!),
but in a generalized version. These generalizations concern both introducing
the scattering algebra and allowing variable order operators. The former
was first described systematically from a geometric perspective in [30] from
the beginning; see [41, 37] for earlier descriptions in Rn ; here it serves to
have a pseudodifferential algebra with Fredholm properties on a non-compact
space right from the outset to orient the reader. The latter was discussed by
The author was supported in part by National Science Foundation grants DMS-1068742 and
DMS-1361432. He is very grateful to the students at Stanford University in the Math 256B
course he taught in Winter 2014, for which these notes were originally written, for their patience
and comments. He is also grateful to the Université de Grenoble for its hospitality during the
summer school in 2014 on mathematical aspects of general relativity, as well as to the
participants; these notes again provided the backbone of the material covered. In addition, he is
grateful to the hospitality of the Université de Nice in summer 2015 where he gave a lecture
series based on the more advanced parts of these notes and where these notes were put in their
final form. Thanks are also due to Peter Hintz for a careful reading of the final version. But most
of all the author is grateful to Richard Melrose, from whom he learnt microlocal analysis, and
whose notes are followed, in a generalized form, in Section 5.3.
219
220 A. Vasy
Duistermaat and Unterberger [43, 7], though not quite in the setting we are
interested in. I would remark that, apart from solving elliptic partial differential
equations (PDEs), the elliptic scattering algebra results shown in this section
also allow one to deal with the local geodesic X-ray transform [42], and
were the key ingredient in showing boundary rigidity in a fixed conformal
class [40].
Section 5.4 gives a thorough description of propagation phenomena, from
generalizations of Hörmander’s propagation of singularities theorem [28] to
the scattering algebra (where it is due to Melrose in [30]) variable order
settings, and including complex absorption and radial points. These are then
used to discuss the limiting absorption principle in scattering theory as well
as the Klein–Gordon equation on Minkowski-like spaces. The approach of
this minicourse, given the limited space, is that the proofs are given in the
basic microlocal setting; extension to the b-operators (presented in the recent
literature in detail) is then straightforward, and I do not give the detailed proofs
in that setting here. Note that I also use large parameter, or semiclassical,
estimates, which are discussed briefly in Section 5.5. While their incorporation
in the earlier section would have been straightforward, they result in further
notational overheads, so it seemed better to sketch only the basic properties
and leave the details to the reader, who may also refer to Zworski’s recent
treatment of the topic [58].
Section 5.5 shows, following [48, 47], how the tools developed can be
applied to understand the analytic continuation of the resolvent of the Lapla-
cian on conformally compact, asymptotically hyperbolic, spaces. While I do
not discuss it here, another application of these microlocal tools is the work
of Dyatlov and Zworski [11] to dynamical zeta functions for Anosov flows;
I refer to [49] for a concise overview of some recent applications and to detailed
references.
Finally Section 5.6 introduces the b-pseudodifferential operator algebra of
Melrose [34] by local reduction to cylindrical models in Euclidean space.
This actually does not give quite as much of the full “small calculus” as the
analogous localization for the standard algebra on compact manifolds without
boundary and for the scattering algebra, though it is sufficient for us. This
completes the tools necessary for the linear analysis of waves on Lorentzian
scattering spaces, which are generalizations of asymptotic Minkowski spaces
[48, 2, 26, 12], as well as Kerr–de Sitter space, which was described in [48, 26],
apart from a treatment of the trapped set in the latter case (for which we
provide references at the end of the section). Indeed, it was the Kerr–de Sitter
project [48] that started the work on the non-elliptic microlocal framework
presented here.
A Minicourse on Microlocal Analysis for Wave Propagation 221
In fact, these techniques extend to non-linear problems. The extension to

semilinear wave equations was done in [26], to quasilinear problems without
trapping in [24], and to quasilinear problems with trapping in [22]. The
quasilinear works introduce b-pseudodifferential operators with coefficients
with Sobolev regularity (rather than smoothness) in the spirit of Beals and
Reed [3] and, to deal with trapping, tame linear estimates; it turns out that these
extensions of the “smooth” microlocal framework described in these notes is
not hard. Given these linear estimates, the non-linear techniques are rather
straightforward by now; an easy version of the Nash–Moser iteration due to
Saint Raymond [38] is used in [22] to deal with the losses of derivatives due to
the trapping. These quasilinear results were then used in the recent proof of the
stability of slowly rotating Kerr–de Sitter black holes by Hintz and Vasy [23].
I hope that these detailed lecture notes will provide a more accessible path,
available in a single place, to the microlocal material that has proved so useful
in our understanding of so many problems.
5.2. The Overview

The ultimate goal of this minicourse is to understand the analysis of not
necessarily elliptic totally characteristic, or b-(pseudo)differential, operators.
These are operators on manifolds M with boundaries or corners, though they
can also arise on complete manifolds without boundary after one appropriately
compactifies or bordifies them (i.e. makes them into a manifold with boundary
or corners, not necessarily compact), as we discuss below. We will study both
elliptic and hyperbolic operators, with the Laplacian or the d’Alembertian of
a b-metric (i.e. a metric corresponding to this structure), and provide main
examples. An important aspect of b-analysis is, however, that the principal
symbol, on which ellipticity etc. is based, does not completely capture the
problem: there is also an analytic family of operators at the boundary ∂M,
depending on the dual of the normal variable to the boundary, the poles of
whose inverse determine both Fredholm properties and asymptotic expansions.
To be more concrete, consider the vector space of vector fields Vb (M)
tangent to ∂M, M an n-dimensional manifold with boundary. Recall that
V ∈ Vb (M) is equivalent to the statement that for all f ∈ C ∞ (M) with f |∂M = 0
one has Vf |∂M = 0; this in turn is equivalent to the same statement for
a single f with a non-degenerate differential at ∂M (such a non-degenerate
function, when non-negative, is called a boundary definining function). This
space Vb (M) is a left C ∞ (M) module and a Lie algebra: if f is a function that
vanishes at ∂M, V, W ∈ Vb (M), then VWf , WVf vanish at ∂M, thus so does
222 A. Vasy
[V, W] f = (VW − WV)f .
Further, in a local coordinate chart mapping an open set O in M to an open set

U in [0, ∞)x × Rn−1 y , with ∂M mapped to x = 0 (thus x is a local boundary
defining function), any smooth vector field has the form

n−1
V = b 0 ∂x + bj ∂yj , bj ∈ C ∞ (M),
j=1
so the tangency of V to ∂M means Vx|x=0 = 0, that is b0 |x=0 = 0, so b0 = xa0 ,

a0 ∈ C ∞ (M). Correspondingly,

V = a0 (x∂x ) + aj ∂yj , aj ∈ C ∞ (M), (5.1)
j
locally; conversely, any vector field with such local coordinate expressions is
in Vb (M).
Our main interest is in differential operators generated by such V ∈ Vb (M).
Namely we let Diffm b (M) consist of finite sums of up to m-fold products of
elements of Vb (M). With the usual Fourier analysis convention that D = 1i ∂,
typical examples, in local coordinates, are operators such as

(xDx )2 + D2yj
j
which is elliptic, corresponding to a Riemannian space with a cylindrical end,

as well as, after appropriate conjugation and division by a factor, to conic
spaces, and
2
xDx − yj Dyj − D2yj ,
j j
which is hyperbolic, and corresponds to a neighborhood of the static patch in

a de Sitter space.
Typically the spaces of interest are not presented as manifolds with bound-
ary; we need to bordify them for this purpose. Thus, an exact cylindrical end
metric is traditionally written as
g = dr2 + h(y, dy), r → +∞,
with Laplacian
g = D2r + h .
Letting x = e−r (see Figure 5.1) means r → +∞ corresponds to x → 0, and

dr = − dxx , Dr = −xDx shows that such a metric and Laplacian indeed has
Figure 5.1. Left: a cylindrical end, with r → +∞. Right: the bordification, with
x = 0 corresponding to r = ∞
the stated form if one bordifies the end by gluing x = 0 to it, that is replacing
(r0 , +∞)r × Y by [0, e−r0 )x × Y, with the identification
(r, y) → (e−r , y)
of
(r0 , +∞)r × Y → (0, e−r0 )x × Y;
combining this with the compact core of the original manifold one obtains the
manifold with boundary M. (More precisely, if the original manifold is M̃, the
manifold M is the disjoint union of M̃ and [0, r0−1 ) × Y with the identification
of the cylindrical end in M̃ with (0, r0−1 ) × Y as above, with a base for the
topology of M given by the images under the equivalence relation of open sets
in either M̃ or [0, r0−1 ) × Y, and with coordinate charts given by those in M̃
as well as [0, r0−1 ) times coordinate charts in Y.) Note that a smooth function
f ∈ C ∞ (M) has a Taylor series expansion at ∂M in terms of powers of x; this
amounts to powers of e−r in the original coordinates.
While we postpone detailed discussion, it turns out that wave, or more
general Klein–Gordon, equations on asymptotic de Sitter spaces, or even
Kerr–de Sitter spaces, describing rotating black holes in a background with
a positive cosmological constant (reflecting our current understanding of the
universe) also give rise to b-problems that the tools we develop can be
used to analyze. A byproduct of this discussion is a new perspective to
analyze asymptotically hyperbolic spaces, and indeed it sheds light on the role
so-called even metrics play in these. Before continuing, we first comment on a
different class of spaces, which do not seem to be b-spaces at the outset, where
b-analysis in fact turns out to be useful.
On a manifold with boundary M, there is a conformally related class of
vector fields, called scattering vector fields Vsc (M). These arise by letting ρ be
a boundary defining function of M (i.e. ρ ≥ 0, ρ vanishes exactly on ∂M, and
224 A. Vasy
dρ is non-degenerate at ∂M; x was a local boundary defining function above),

and letting
Vsc (M) = ρVb (M).
In local coordinates, such vector fields thus have the form

V = a0 (x2 ∂x ) + aj (x∂yj ), aj ∈ C ∞ (M),
j
locally; conversely, any vector field with such local coordinate expressions is in
Vsc (M). Again, we can define scattering differential operators by taking finite
sums of finite products of these. A typical scattering differential operator is
(x2 Dx )2 + (xDy )2 − λ,
where λ ∈ C. Up to an inessential first order term, this is the spectral family of

the Laplacian of a conic metric on the “large end” of a cone, such as Euclidean
space near infinity. To see this, recall that an exact conic metric has the form
dr2 + r2 h(y, dy),
and thus, up to first order terms, its Laplacian has the form D2r + r−2 h . If we
let x = r−1 , dr = − dxx2
, and Dr = −x2 Dx , this takes the form described above
if we bordify by adding x = 0 to (0, r0−1 )x × Y. Note that the identification
(r0 , ∞) × Y → (0, r0−1 ) × Y
is now via
(r, y) → (r−1 , y),
so if the bordified space is denoted by M, elements of C ∞ (M) have an

expansion in powers of x, that is of r−1 , unlike e−r above. Thus, one has to
pay attention to the bordification/compactification one uses. Now, if λ = 0, we
can factor this operator as
x2 L, L = Dx x2 Dx + y ∈ Diff2b (M),
and thus b-analysis is applicable, but if λ = 0 this is no longer the case. Thus,
for λ = 0, one needs to use the scattering framework to analyze the problem.
(If λ = 0, the scattering framework can be used as a starting point, but the
operator is degenerate then in an appropriate sense, which forces one to work
with the b-framework at least implicitly.)
In the case of Euclidean space, a straightforward calculation shows that
compactification of Rn by reciprocal polar coordinates, that is identifying
Rn \ {0} with (0, ∞)r × Sn−1 , and compactifying to M = Rn = Bn (a closed
Figure 5.2. Left: two families of parallel lines in R2 . Right: the compactification
R2 and the image of the two families of parallel lines. Note that parallel families
end at the same two points at ∂R2 , while conic sectors, such as on the upper left
between the two thick lines, map to wedge shaped domains
ball) using the map x = r−1 , see Figure 5.2, means not only that the spectral
family of the Laplacian becomes an element of Diff2sc (M), but more generally
translation invariant differential operators become elements of Diffm sc (M) and
indeed become a basis, over C ∞ (M), of Diffm sc (M). In particular, g , where
g is the Minkowski metric, satisfies g ∈ Diff2sc (M). Further, much like the
Euclidean Laplacian, due to its homogeneity under dilations (which is what
the b-structure relates to), g = x2 L with L ∈ Diff2b (M). This means that
the Minkowski wave operator, and indeed more general operators of similar
form, called d’Alembertians of Lorentzian scattering metrics, are amenable to
b-analysis. Note, however, that the Klein–Gordon operator g −λ ∈ Diff2sc (M)
does not factor in this way for λ = 0, and it is thus analyzable in the scattering
framework.
The general approach I present to the analysis of differential operators P is
via Fredholm estimates, such as
uX ≤ C(PuY + uX̃ )
and
vY ∗ ≤ C(P∗ vX ∗ + vZ ),
where the inclusion of the spaces X into X̃ and Y ∗ into Z is compact. Since
our spaces correspond to geometrically complete non-compact manifolds, or
compact manifolds but with differentiability encoded via complete vector
fields such as Vb (M), such a compact inclusion has two ingredients: gain
in differentiability and gain in decay. (The simplest example is weighted
Sobolev spaces H s,r (Rn ), with H s (Rn ) the standard Sobolev space, which is a
weighted L2 -space on the Fourier transform side: H s (Rn ) = F −1 .−s L2 (Rn ),
and H s,r (Rn ) = .−r H s (Rn ), with z = (1 + |z|2 )1/2 ; see (5.47). Then the

inclusion H s,r (Rn ) → H s ,r (Rn ) being compact requires both s > s and r > r .
226 A. Vasy
s,r
These spaces in fact turn out to be the scattering Sobolev spaces Hsc (Rn ) and
s ,r
Hsc (Rn ).) The gain in differentiability, much as in standard local (pseudo)-
differential analysis, is based on properties of principal symbols, which capture
the operators modulo lower (differential) order terms. However, the gain in
decay is more subtle.
To illustrate this, notice a structural difference between the Lie algebras
Vb (M) and Vsc (M):
[Vb (M), Vb (M)] ⊂ Vb (M), [Vsc (M), Vsc (M)] ⊂ ρVsc (M),
that is while Vb (M) is merely a Lie algebra, Vsc (M) is commutative to the
leading (zeroth) order at ∂M. In order to analyze differential operators in
Diffb (M) or Diffsc (M), we develop pseudodifferential algebras b (M) and
sc (M). The leading order commutativity of Vsc (M) means that sc (M) has
another symbol at ∂M which admits similar constructions and estimates as
the standard principal symbol; the two symbols are indeed related via the
Fourier transform on Rn as we discuss in the next sections covering standard
and scattering pseudodifferential operators. On the other hand, to b (M)
corresponds an analytic family of operators on ∂M, called the indicial or
normal family. For instance, on cylindrical ends, one conjugates the operator
by the Mellin transform in x, that is the Fourier transform in log x = −r:

dx
(Mu)(σ , y) = x−iσ u(x, y) ,
x
with

1
(M−1 v)(x, y) = xiσ v(σ , y) dσ ,
2π Im σ =−α
where the choice of α corresponds to the weighted space we are working with.
Here we are interested in functions supported near x = 0 (i.e. r near +∞), so
the behavior of u as x → +∞ is irrelevant, and the Mellin transform gives a
result which is holomorphic in an upper half-plane: the more decay one has
at x = 0, the larger the half-plane is. The Mellin transform changes xDx to
multiplication by the b-dual variable σ , so the indicial family of
P = (xDx )2 + Y − λ
is
N̂(P)(σ ) = σ 2 + Y − λ.

This family has a meromorphic inverse, with poles at σ = σj , σj = ± λ − λj ,
as λj runs over the eigenvalues of Y (here we are assuming Y is compact).
Note that we are inverting operators (i.e. working with a non-commutative
algebra), but on a manifold without boundary, that is on a simpler space. The

Fredholm properties of P then will depend on the weighted b-Sobolev space
we use, with the choice of weight corresponding to the choice of the imaginary
part of σ : one needs to choose a weight xα such that no poles N̂(P)(σ )−1 lie on
the line Im σ = −α. (There are only finitely many poles in a strip | Im σ | < C,
so there are many good choices, but there are also a few bad choices.) Here the
b-Sobolev spaces of non-negative integer differentiability order s and of real
weight α are defined by
u ∈ Hbs,α (M) iff x−α u ∈ Hbs,0 (M) = Hbs (M),

u ∈ Hbs (M) iff ∀L ∈ Diffsb (M), Lu ∈ Lb2 (M),
where Lb2 (M) is the L2 -space with respect to any non-degenerate b-density (see
x dy1 . . . dyn−1 |. Further, even if we have a Fredholm choice
below), such as | dx
of space, the invertibility properties, even the Fredholm index, depend on this
choice. This contrasts much with say formally self-adjoint elliptic problems
on compact manifolds, for which any pair of Sobolev spaces, with orders
differing by the order of the operator, give rise to a Fredholm pair of the same
invertibility properties. One indication of this is that the kernel (nullspace) of
elliptic operators on such manifolds is in C ∞ (M), that is in every Sobolev
space, while in the b-setting, it is not in every weighted space, though the
differentiability order does not matter. In fact, if Pu ∈ C˙∞ (M), say, that is
C ∞ and vanishes with all derivatives at ∂M, then typically u has an asymptotic
expansion at ∂M of the form

xiσj uj , uj ∈ C ∞ (M),
j: Im σj <−α
kj
(with additional logarithmic terms, such as
=0 (log x)
xiσj uj,
in the general
s,α
case), where α is such that u ∈ Hb (M) for some s, and there are no σj
with Im σj = −α. The σj are also called resonances; they thus determine the
asymptotic behavior/expansion of solutions of P.
While we have discussed linear problems, with a bit of work the analysis
can be extended to nonlinear problems as well. In the non-elliptic context
(our main interest here) small data problems are quite well behaved: for
semilinear problems one can even use Picard iteration typically (much as
for ordinary differential equations (ODEs)) to obtain global solutions and
asymptotic expansions; for quasilinear problems a bit more care is needed.
We now return to Vb (M); many of the following considerations apply, with
simple modifications, to Vsc (M). Much like the set of all vector fields V(M)
is the set of all smooth sections of TM, and Vb (M) is the set of all smooth
228 A. Vasy
sections of a vector bundle b TM, called the b-tangent bundle. Indeed, a local
basis of b TM is given by x∂x , ∂y1 , . . . , ∂yn−1 , that is in terms of coordinates
(x, y) as above, (x, y1 , . . . , yn−1 , a0 , a1 , . . . , an−1 ) are local coordinates on b TM,
cf. (5.1). In M ◦ , b TM can be naturally identified with TM (corresponding to
tangency to ∂M being vacuous there); globally Vb (M) ⊂ V(M) gives a fiber-
preserving map
ι : b TM → TM;
with (x, y1 , . . . , yn−1 , b0 , b1 , . . . , bn−1 ) coordinates on TM as in the above

parameterization of vector fields, then the map is, in local coordinates,
ι(x, y1 , . . . , yn−1 , a0 , a1 , . . . , an−1 ) = (x, y1 , . . . , yn−1 , xa0 , a1 , . . . , an−1 ),
corresponding to
b0 = xa0 , bj = aj , j = 1, . . . , n − 1.
Note that ι is not injective at p ∈ ∂M; its kernel b Np ∂M is the span of x∂x , the
space of normal b-vector fields to the boundary, while its range is Tp ∂M, the
tangent space to ∂M (as a subspace of Tp M). Note that, unlike for Tp M, the
natural subspace of b Tp M is not vector fields tangent to ∂M, but vector fields
b-normal to ∂M.
The dual bundle, b T ∗ M, called the b-cotangent bundle, of b TM, then has a
local basis dx
x , dy1 , . . . , dyn−1 , that is smooth sections locally have the form
dx
σ + ηj dyj ,
x
with σ , η ∈ C ∞ (M). The adjoint of ι is the map π : T ∗ M → b T ∗ M, which
is sometimes (especially for boundary value problems) called the compressed
cotangent bundle map, and b T ∗ M the compressed cotangent bundle. In coordi-
nates, with (ξ , ζ1 , . . . , ζn−1 ) being dual coordinates to (x, y1 , . . . , yn−1 ),
π(x, y1 , . . . , yn−1 , ξ , ζ1 , . . . , ζn−1 ) = (x, y1 , . . . , yn−1 , xξ , ζ1 , . . . , ζn−1 ),
corresponding to
dx
ξ dx + ζj dyj = (xξ ) + ζj dyj .
x
j
The kernel of π on Tp∗ M is now Np∗ ∂M, the conormal bundle fiber of ∂M,
and the range is the span of the dyj , which is the cotangent bundle of ∂M
within b Tp∗ M. Note that the natural subspace of b Tp∗ M is the cotangent, not the
conormal, bundle of the boundary.
Intuitively, π projects out the normal momentum (momentum being dual

to position), but preserves the tangential momentum, and thus can be used to
encode the law of reflection for boundary problems. In fact, if we had more
time in this minicourse, in its final part we would study exactly this problem
(even on manifolds with corners): how do waves on, say, M = Y × Rt , (Y, h)
Riemannian with boundary, g = dt2 − h, behave at ∂M? Notice that g ∈
Diff2 (M), but is not in Diff2b (M). Nonetheless, the tools we use to analyze g
are in the b-category. The reason for this is that “standard” pseudo-differential
operators do not quite make sense on manifolds with boundary, and even in the
cases when they do, they typically do not preserve boundary conditions. Thus,
we use a two-algebra approach: the operator to analyze is in Diff(M), but the
toolkit we use is in b (M). I refer to [52] for details and further references.
In fact, I would point out that b-pseudodifferential operators originated in
Melrose’s work on propagation problems on manifolds with boundary [33].
We now turn to tensorial considerations. A smooth non-degenerate symmet-
ric bilinear form on b TM, that is a non-degenerate section of Sym2 b T ∗ M =
b T ∗ M b T ∗ M is, for each p ∈ M, a (symmetric, bilinear) map g : b T M ×
p p
b T M → R such that if V ∈ b T M and g (V, W) = 0 for all W ∈ b T M,
p p p p
then V = 0. Equivalently, any g ∈ Sym2 b T ∗ M gives a map b Tp M → b Tp∗ M,
and then non-degeneracy is equivalent to the injectivity of this map. In view
of the equality of dimensions of the domain and target spaces, this in turn is
equivalent to its invertibility. We call such a non-degenerate symmetric bilinear
form a metric. The signature of such a metric is (k, n − k) if the maximal
subspace on which it is positive definite is k; note that the signature is constant.
Taking the negative of g reverses the signature, so for example (n, 0) and (0, n)
are essentially equivalent signatures, corresponding to Riemannian metrics.
Lorentzian metrics have signature (1, n − 1) or (n − 1, 1); we mostly use the
former convention here (so time has a positive sign, space negative signs). One
can also takes metrics of other signatures (if the dimension n is at least four);
these can arise naturally.
In view of the invertibility, one obtains a dual bilinear form on b T ∗ M, as well
as on tensorially related bundles, such as the form bundles or the symmetric
bundles. One also has a metric density, which is a non-degenerate section
of the density bundle b M. The latter has basis | dx x dy1 . . . dyn−1 | in local
coordinates; the metric density is

dx
|dg| = | det g| dy1 . . . dyn−1 ,

x
where one takes the determinant of the matrix of coefficients of g with respect
to the b-frame x∂x , ∂y1 , . . . , ∂yn−1 . In particular, this gives rise to a positive
230 A. Vasy
definite inner product on C˙∞ (M), the space of C ∞ functions vanishing to

infinite order at ∂M:

u, v = uv |dg|,
as well as a non-positive definite inner product on, say, differential forms

u, v = Gp (u, v) |dg|,
where G is the dual metric on differential forms. While this inner product is
not positive definite, this is only due to the fiber inner product Gp (on finite
dimensional spaces), rather than to |dg|, so in particular formal adjoints of
differential operators are well-defined differential operators. Moreover, the
exterior derivative d maps C˙∞ (M) to C˙∞ (M, b T ∗ M), and more generally
C˙∞ (M, b k M) to C˙∞ (M, b k+1 M), and thus given a metric g we can define
the Laplacian or d’Alembertian
= d∗ d + dd∗ ,
with adjoint ∗ taken relative to g.
5.3. The Basics of Microlocal Analysis

In this section we discuss basic properties of pseudodifferential and scattering
pseudodifferential operators, introduced in this generality by Melrose [30],
formerly discussed by Parenti and Shubin on Rn [41, 37], where it can be also
considered an example of Hörmander’s Weyl calculus [27]. These operators
generalize differential operators of the form

A= aα Dα , with aα ∈ C ∞ (Rn ), (5.2)
|α|≤m
as we show below in (5.31). Indeed, the conditions on the coefficients aα are

relaxed to be “symbolic,” so that for instance a0 (z) = φ(z)|z|−ρ , φ ≡ 0 near
the origin, ≡ 1 near infinity is allowed. Thus, in particular operators such as
+ V, where V is the Coulomb potential, without its singularity at the origin,
fit into this framework. (The singularity at the origin would make the problem
into an elliptic b-problem, such as those discussed in Section 5.6, near 0, but
we do not discuss this here.)
More generally, we can consider Riemannian metrics g with gij ∈ C ∞ (Rn )

such that for all z ∈ Rn , ij gij (z)ζi ζj = 0 implies ζ = 0, that is g is positive
definite on the compact manifold Rn . Then, with V as above and with σ ∈ C,

g + V − σ is of the form (5.2) with m = 2.
The extension of this class to scattering pseudodifferential operators allows
one to construct approximate inverses (parametrices), showing Fredholm
properties, for operators that are elliptic in this class. Ellipticity here also
encodes behavior at spatial infinity, so for instance + V − σ , where V may be
Coulomb type with ρ > 0, is elliptic for σ ∈ C \ [0, ∞), but is not elliptic for
σ ∈ [0, ∞). It also allows one to develop tools to study non-elliptic operators.
For instance, the limiting absorption principle, that is the existence of the limits
R(σ ± i0) = lim ( + V − (σ ± i))−1

→0+
for V real valued and σ > 0 fits very nicely into this framework.
5.3.1. The Outline

Since there are technicalities along the way, we begin with an outline of this
section. First, for m,
,
∈ R, δ, δ ∈ [0, 1/2), we define two kinds of function
spaces,
m,
m,
∞
(R ; R ) ⊂ S∞,δ (R ; R ) ⊂ C (R ),
n n n n 2n
Sδ,δ
as well as analogues on R3n :

m,
1 ,
2
Sδ,δ
m,
1 ,
2
(Rn ; Rn ; Rn ) ⊂ S∞,δ (Rn ; Rn ; Rn ) ⊂ C ∞ (R3n ).
The elements of these spaces are called symbols; the important point is the
behavior of these symbols at infinity. Here the spaces become larger with
increasing m,
, and
j , and δ = 0 = δ gives the standard classes also
denoted by
m,
m,
S0,0 (Rn ; Rn ) = Sm,

(Rn ; Rn ), S∞,0 (Rn ; Rn ) = S∞
m,
(Rn ; Rn ),
and similarly for the R3n versions. The cases δ = 0 = δ are by far the most
important ones. We have projections πL , πR : R3n → R2n , with πL dropping
the second factor of R3n and πR dropping the first factor:
πL (z, z , ζ ) = (z, ζ ), πR (z, z , ζ ) = (z , ζ );
the subscripts L and R refer to z, resp. z , being the left, resp. right, “base” or
“position” variable. (The variable ζ will be the “dual” or “momentum” vari-
able.) Then πL∗ , πR∗ pull back elements of the R2n spaces to the corresponding
R3n spaces (with
1 =
,
2 = 0, resp,
2 =
,
1 = 0). With S denoting
232 A. Vasy
Schwartz functions on Rn , S denoting tempered distributions on Rn , and L

denoting continuous linear operators, we define an oscillatory integral map:
m,
1 ,
2
I : S∞,δ (Rn ; Rn ; Rn ) → L(S, S),
and also show by duality that
m,
1 ,
2
I : S∞,δ (Rn ; Rn ; Rn ) → L(S , S ),
and that the range of I is closed under Fréchet space or L2 -based adjoints. The
compositions
qL = I ◦ πL∗ , qR = I ◦ πR∗ ,
are called the left and right quantization maps. Now, it turns out that I is
m,
1 ,
2 m,
1 ,
2
redundant, and its range on Sδ,δ (Rn ; Rn ; Rn ), resp. S∞,δ (Rn ; Rn ; Rn ),
m,
m,
(R ; R ), resp. S∞,δ (R ; R ) with

=
1 +
2 ; the
is that of qL on Sδ,δ n n n n
analogous statement also holds with qL replaced by qR . This is called left, resp.
right, reduction; see Proposition 5.1. One defines pseudodifferential operators,
m,
m,
δ,δ , resp. ∞,δ , to be the range of qL (or equivalently qR ) on the spaces

m,
m,
(R ; R ), resp. S∞,δ (R ; R ), and writes

Sδ,δ n n n n
m,
m,
m,
= 0,0 m,
, ∞ = ∞,0 .
Once this reducibility is shown it is straightforward to see (using the general I,
m,
m ,

which is why it is introduced) that A ∈ δ,δ , B ∈ δ,δ implies AB ∈

m+m ,
+
∞,∞ m,
δ,δ , that is that δ,δ = ∪m,

δ,δ is an order-filtered algebra, with the
∞,∞
analogous statements holding for ∞,δ as well. One also shows that compo-
sition is commutative to the leading order, that is
m,
m ,
m+m −1+2δ,
+
−1+2δ
A ∈ δ,δ , B ∈ δ,δ ⇒ [A, B] = AB − BA ∈ δ,δ ;
the analogous statement here is
m,
m ,
m+m −1+2δ,
+

A ∈ ∞,δ , B ∈ ∞,δ ⇒ [A, B] = AB − BA ∈ ∞,δ ,
that is the gain is only in the first order. This is conveniently encoded by the
principal symbol maps
m,
m,
m−1+2δ,
−1+2δ
m,
m,
m−1+2δ,
σm,
: δ,δ → Sδ,δ /Sδ,δ , σ∞,m,
: ∞,δ → S∞,δ /S∞,δ ,
which are multiplicative (homomorphisms of filtered algebras); the leading
order commutativity of pseudodifferential operators correspond to the com-
mutativity of function spaces under multiplication. Here δ, δ are suppressed
in the principal symbol notation. An immediate consequence is the ellip-
m,
tic parametrix construction: for operators A ∈ δ,δ with invertible principal
symbol, which are called elliptic, one can construct an approximate inverse
−m,−
B ∈ δ,δ such that AB − Id, BA − Id : S → S are continuous, that is

m,
completely regularizing. In the case of A ∈ ∞,δ , we only have AB − Id,

∞
BA − Id : S → C (R ), that is are smoothing, but do not give decay
n
at infinity. Since completely regularizing operators are compact from any

weighted Sobolev space to any other weighted Sobolev space, and since we
show that (recalling the weighted Sobolev spaces from Section 5.2)
m,
A ∈ ∞,δ ⇒ A ∈ L(H r,s , H r−m,s−

)
m,
m,
for all r, s ∈ R (so analogous statements hold for δ,δ ⊂ ∞,δ ), we deduce
m,
that elliptic A ∈ δ,δ are Fredholm on any weighted Sobolev space, with the
nullspace of both A and A∗ lying in S(Rn ), and being independent of the choice
m,0
of the weighted Sobolev space. In particular, if A ∈ δ,δ , m > 0, elliptic,
is symmetric with respect to the L2 inner product, then one immediately

concludes that A ± i Id is invertible as a map H m,0 → L2 , and thus A is self-
adjoint with domain H m,0 .
Another important direction we explore is microlocalization, by introducing
the notion of the operator wavefront set, WF (A), or WF ∞ (A), which measures
where in phase space A is “trivial.” Thus, while σm,
, σ∞,m,
capture the leading
order behavior of operators, that is their behavior modulo one order lower
operators, WF (A) and WF ∞,
(A) give the locations where A is not residual,
−∞,
that is in −∞,−∞ , resp. ∞ , so for instance the emptiness of WF (A)

implies A ∈ −∞,−∞ . One should think of these as an analogue of the singular
support of distributions, which measures where a distribution is not C ∞ , except
that its location will not be in the base space Rn , but rather at infinity in phase
space, Rn × Rn . To make this concrete, it is useful to compactify Rn × Rn to
Rn × Rn , see Figure 5.3; then for A ∈ m,
, WF (A) ⊂ ∂(Rn × Rn ) while
for A ∈ ∞ m,
, WF ∞,
(A) ⊂ Rn × ∂Rn . Then one can perform a microlocal
version of the elliptic parametrix construction, that is one that is localized, in
the sense of WF , near points at which the operator A is elliptic; this is a first
step towards understanding non-elliptic operators.
It turns out that it is convenient to generalize the class of operators
considered here to allow their orders m and
to vary, namely m = m is a
function on ∂Rn × Rn and
= l a function on Rn × ∂Rn , so at different
points microlocally one has an operator of different order. This is the reason
we consider δ, δ > 0 here; we naturally end up with the classes Sδ,δ m,
m,
and S∞,δ

where δ, δ can be taken to be arbitrarily small but positive. (There is also the
possibility of taking logarithmic weight losses below, but we do not discuss it
here.)
234 A. Vasy
Figure 5.3. The product compactified phase space, Rn × Rn . The whole boundary
∂(Rn × Rn ) = (Rn × ∂Rn ) ∪ (∂Rn × Rn ) carries WF (A), while only Rn × ∂Rn
carries WF ∞,
(A)
5.3.2. The Definition of Pseudodifferential Operators

and Oscillatory Integrals
We now go through the details. Thus, starting with Rn , we consider operators
of the form

−n
Au(z) = (I(a)u)(z) = (2π ) eiζ ·(z−z ) a(z, z , ζ ) u(z ) dz , u ∈ S(Rn ),
Rn ×Rn
(5.3)
where a is a product-type symbol of class Sδ,δ m,
1 ,
2
, m,
,
∈ R, δ, δ ∈
1 2

[0, 1/2), that is differentiation in z, resp. z , resp. ζ , provides extra decay in
the respective variables:
m,
1 ,
2
a ∈ Sδ,δ (Rnz ; Rnz ; Rnζ )
⇐⇒ a ∈ C ∞ (Rnz × Rnz × Rnζ ),

|Dαz Dz Dζ a| ≤ Cαβγ z
1 −|α| z
2 −|β| ζ m−|γ | (z + z )δ |(α,β,γ )| ζ δ|(α,β,γ )|
β γ
with
|(α, β, γ )| = |α| + |β| + |γ |
and
· = (1 + | · |2 )1/2 .
One writes

a m,
,
= supz−
1 +|α| z −
2 +|β| (z + z )−δ |(α,β,γ )|
Sδ,δ 1 2 ,N
|α|+|β|+|γ |≤N
× ζ −m+|γ |−δ|(α,β,γ )| |Dαz Dz Dζ a|;

β γ
as N runs over N, these give a family of seminorms on Sm,

1 ,
2 , giving it a
Fréchet topology.
Note that the orders on S are reversed compared to the order of the factors,
that is z, z , ζ ; this is done in part to conform with the usual notation. Moreover,

(z + z )δ |(α,β,γ )| can be replaced by (z, z )δ |(α,β,γ )| . Also, z and z play an
equivalent role since, as mentioned before, and as we show below, one can even
eliminate, say, the z dependence. In fact, it turns out that the behavior of a is
essentially irrelevant in the region where zz is not bounded between M −1 and
M, M > 1 is any fixed number, in that if one cuts a off to be supported outside
−∞,−∞
such a set, one obtains an element of δ,δ , see (5.25), but since this is due
to the oscillatory nature of the integral in ζ , this is not obvious at this point.
However, we have already pointed out that fixing some χ ∈ Cc∞ (R), χ ≡ 1 on
m,
1 ,
2
[ 12 , 2], supported in [ 14 , 4], for a ∈ Sδ,δ we have the decomposition as

z z
a = a1 + a2 , a1 = χ a, a2 = 1 − χ a, (5.4)
z z
m,
1 ,
2
with aj depending continuously on a in the Sδ,δ topology; (5.25) shows that
the contribution of a2 is essentially irrelevant in the sense stated above.
In fact, in the beginning it is better to start with a larger (at least if δ = 0)
class of symbols, without extra decay in the z, z variables upon differentiation:
for δ ∈ [0, 1/2),
m,
1 ,
2
a ∈ S∞,δ (Rnz ; Rnz ; Rnζ ) ⇐⇒ a ∈ C ∞ (Rnz × Rnz × Rnζ ),
|Dαz Dz Dζ a| ≤ Cαβγ z
1 z
2 ζ m−|γ |+δ|(α,β,γ )| .
β γ
One writes

supz−
1 z −
2 ζ −m+|γ |−δ|(α,β,γ )| |Dαz Dz Dζ a|.
β γ
a m,
,
2 =
S∞,δ1 ,N
|α|+|β|+|γ |≤N
For
1 =
2 = 0, this is Hörmander’s uniform symbol class of type 1 − δ, δ
(i.e. ρ, δ with ρ = 1 − δ). Note that
m,
1 ,
2 m,
1 ,
2
Sδ,0 ⊂ S∞,δ ,
and the inclusion map

m,
1 ,
2 m,
1 ,
2
ι : Sδ,0 → S∞,δ
is continuous, with
a m,
,
2 ≤ a m,
,
2
S∞,δ1 ,N Sδ,0 1 ,N
for all N.
236 A. Vasy
Note that
j ≤
j , m ≤ m implies
m,
1 ,
2 m ,
,

Sδ,δ ⊂ Sδ,δ 1 2 ,
and similarly with S∞ . Further, if δ ≤ δ̃, δ ≤ δ̃ then

m,
1 ,
2
Sδ,δ ⊂ Sm,
1 ,
2 .
δ̃,δ̃
One writes
−∞,
1 ,
2 m,
1 ,
2 −∞,
1 ,−∞ m,
1 ,
2
Sδ,δ = ∩m∈R Sδ,δ , Sδ,δ = ∩m∈R,
2 ∈R Sδ,δ ,
and similarly again with S∞ . Notice that for all δ, δ ∈ [0, 1/2),
−∞,−∞,−∞
Sδ,δ = S(R3n )
−∞,0,0
while S∞,δ consists of C ∞ functions on R2n
z,z which are bounded with
all derivatives, and take values in S(Rn ). Thus, these residual spaces are
independent of δ, δ . One also writes
∞,∞,∞ m,
1 ,
2
Sδ,δ = ∪m,
1 ,
2 ∈R Sδ,δ .
∞,∞,∞
Further, note that Sδ,δ forms a commutative filtered *-algebra in the sense
m,
1 ,
2
that in addition to Sδ,δ being a vector space for each m,
1 ,
2 , closed under
complex conjugation, the (function-theoretic, i.e. pointwise) product (which is
commutative) satisfies
m,
1 ,
2 m ,
,
2 m+m ,
1 +
1 ,
2 +
2
a ∈ Sδ,δ , b ∈ Sδ,δ 1 ⇒ ab ∈ Sδ,δ ,
∞,∞,∞
as follows from Leibniz’s rule. Similarly S∞,δ forms a commutative filtered
*-algebra as well. Notice also that for δ = 0,
m,
1 ,
2 β γ m−|γ |+δ|(α,β,γ )|,
1 −|α|,
2 −|β|
a ∈ Sδ,0 ⇒ Dαz Dz Dζ a ∈Sδ,0 , (5.5)
while for general δ , the a1 piece, as defined in (5.4), satisfies
m,
1 ,
2 β γ m−|γ |+δ|(α,β,γ )|,
1 −|α|+δ |(α,β,γ )|,
2 −|β|
a1 ∈ Sδ,δ ⇒ Dαz Dz Dζ a1 ∈Sδ,δ ,
(5.6)
where by the support property of a1 , δ |(α, β, γ )| could also be shifted to the
last order (and recall that a2 will be shown to be essentially irrelevant). The
∞,∞,∞
analogue of (5.5) also holds for S∞,δ , in which case
1 and
2 are unaffected
by derivatives.
It is also useful to note the following lemma:
−∞,
1 ,
2
Lemma 5.1 For m > m, the residual spaces S∞,δ m̃,
1 ,
2
= ∩m̃∈R S∞,δ , resp.
−∞,
1 ,
2 m̃,
1 ,
2 m,
1 ,
2 m,
1 ,
2
Sδ,δ = ∩m̃∈R Sδ,δ , are dense in S∞,δ , resp. Sδ,δ , in the topology

m ,
1 ,
2
m ,
1 ,
2
of S∞,δ , resp. Sδ,δ .
Proof Let χ ∈ Cc∞ (Rn ) be such that 0 ≤ χ ≤ 1, χ (ζ ) = 1 for |ζ | ≤ 1,

χ (ζ ) = 0 for |ζ | ≥ 2, and let aj (z, z , ζ ) = χ (ζ /j)a(z, z , ζ ), where a ∈ S∞,δ
m,
1 ,
2
.
Then

Dαz D β Cμν j−|μ| (Dζ (χ − 1))(ζ /j)(Dαz D β
γ μ
z Dζ (aj − a) =
ν
z Dζ a)(z, z , ζ ),
μ+ν=γ
with Cμν combinatorial constants. The μ = 0 term is supported in |ζ | ≥ j, the

μ = 0 terms are supported in j ≤ |ζ | ≤ 2j. Correspondingly, for μ = 0, the
summand is bounded by
C0γ ζ m−|γ |+δ|(α,β,γ )| z

1 z
2 , (5.7)
while for μ = 0, j ∼ |ζ | on the support, so the summand is bounded by a

constant multiple of
ζ m−|μ|−|ν|+δ|(α,β,ν)| z
1 z
2 . (5.8)
Multiplying by

ζ −m +|γ |−δ|(α,β,γ )| z−
1 z−
2 ,
in either case we obtain a quantity bounded by a constant multiple of

ζ −(m −m) . Since the difference is supported in |ζ | ≥ j, and since m > m,
this goes to 0 as j → ∞, proving the claim.
m,
1 ,
2
The proof for a ∈ Sδ,δ is similar, with (5.7) replaced by

C0γ ζ m−|γ |+δ|(α,β,γ )| z
1 z
2 (z, z )δ |(α,β,γ )| , (5.9)
and (5.8) replaced by

ζ m−|μ|−|ν|+δ|(α,β,ν)| z
1 z
2 (z, z )δ |(α,β,ν)| , (5.10)
so multiplication by

ζ −m +|γ |−δ|(α,β,γ )| z−
1 z−
2 (z, z )−δ |(α,β,γ )| ,
gives the desired result.
As examples, recall that if a is a polynomial of order

1 ,
2 and m in the
m,
1 ,
2
three variables, then certainly a ∈ Sm,
1 ,
2 = S0,0 . More interestingly, if
3
a ∈ C ∞ (Rn × Rn × Rn ) = C ∞ (Rn ) then a ∈ S0,0,0 = S0,0
0,0,0
, so
a ∈ z
1 z
2 ζ m C ∞ ((Rn )3 ) ⇒ a ∈ Sm,
1 ,
2 = S0,0
m,
1 ,
2
.
Such a are called classical symbols; one writes

m,
1 ,
2
Scl = z
1 z
2 ζ m C ∞ ((Rn )3 ).
238 A. Vasy
3
m,
1 ,
2
Thus, Sδ,δ is a C ∞ (Rn )-module. A particular example is a = |z|−ρ φ(z),
where φ ≡ 0 near 0, φ ≡ 1 near ∞, then a ∈ S−ρ,0,0 , such an a can be thought
of as a potential which may decay only slowly at infinity; ρ = 1 would give
the Coulomb potential without its singularity at the origin.
On the flipside, we can rewrite the estimates for Sm,
1 ,
2 :

|α | ≤ |α|, |β | ≤ |β|, |γ | ≤ |γ | ⇒ |zα Dαz (z )β Dz ζ γ Dζ a|
β γ
≤ Cαβγ z
1 z
2 ζ m .
Since zi ∂zj and ∂zj generate all C ∞ vector fields over C ∞ (Rn ) which are tangent
to ∂Rn , whose set is denoted by Vb (Rn ), we can rewrite this equivalently as
follows: let Vj,k ∈ Vb (Rn ), j = 1, 2, 3, Nj ∈ N (possibly 0) and 1 ≤ k ≤ Nj
acting in the jth factor, then
Nj

3
−
1 −
2 −m
z z ζ Vj,k a ∈ L∞ .
j=1 k=1
3
This could be further rephrased, in terms of vector fields on Rn , tangent to all
boundary faces: if Vj are such, 1 ≤ j ≤ N (possibly N = 0), then
z−
1 z −
2 ζ −m V1 . . . VN a ∈ L∞ .
Since one can use any vector fields tangent to the various boundary faces, in
any product decomposition [0, 1)r−1 × Sn−1 near the boundary of each factor
Rn , one automatically has smoothness in the various angular variables; in the
radial variables one has iterated regularity with respect to r∂r . We contrast
m,
1 ,
2
this conormal or symbolic regularity with the classical regularity a ∈ Scl ,
which means
V1 . . . VN z−
1 z −
2 ζ −m a ∈ L∞
3
for all vector fields on Rn , without the tangency requirement. In particular, in
terms of a product decomposition [0, 1)r−1 × Sn−1 near the boundary of each
factor Rn , one has smoothness in the various angular variables and in the radial
variables, that is one has iterated regularity with respect to ∂r .
We are also interested in the generalization of this setting in which the orders
m,
1 ,
2 are allowed to vary. Concretely, to set this up, suppose that m, lj ∈
S0,0,0 are real valued symbols. We write
m,l1 ,l2
a ∈ Sδ,δ (Rnz ; Rnz ; Rnζ )
⇐⇒ a ∈ C ∞ (Rnz × Rnz × Rnζ ),

|Dαz Dz Dζ a| ≤ Cαβγ zl1 −|α| z l2 −|β| ζ m−|γ | (z+z )δ |(α,β,γ )| ζ δ|(α,β,γ )| .
β γ
Notice that replacing m by m where m − m ∈ S−,0,0 for some > 0 does

not change the class since ζ m−m = e(m−m ) logζ , and (m − m ) logζ
is a bounded function in this case. Since we are interested only in m, lj ∈
C ∞ (Rn × Rn × Rn ), we regard m as a function on Rn × Rn × ∂Rn , and take
an arbitrary (smooth) extension to Rn × Rn × Rn ; we proceed similarly with
the
j . Thus, with
m = sup m,
j = sup lj ,
where the sup may be taken over the appropriate boundary of the compactifi-
cation only, we have
m,l1 ,l2 m,
1 ,
2
a ∈ Sδ,δ ⇒ a ∈ Sδ,δ .
One can also define

m,l1 ,l2
a ∈ S∞,δ (Rnz ; Rnz ; Rnζ ) ⇐⇒ a ∈ C ∞ (Rnz × Rnz × Rnζ ),
|Dαz Dz Dζ a| ≤ Cαβγ zl1 z l2 ζ m−|γ |+δ|(α,β,γ )| ,
β γ
so with m,
j as above
m,l1 ,l2 m,
1 ,
2
a ∈ S∞,δ ⇒ a ∈ S∞,δ .
However, these variable order spaces provide more precise information than
simply taking m = sup m, etc., much like the Sm,
1 ,
2 spaces provide more
m,
1 ,
2
precise information that S∞ . Further, we note that we have introduced the

subscript δ and δ (limiting the gains under differentiation) since the function
b = ζ m = em logζ is in Sδ,0 m,0,0
for all δ > 0, but not for δ = 0. Indeed,
differentiating in, say, zj , gives
Dzj b = (Dzj m)(logζ )ζ m ,
so there is a logarithmic loss (unless m is constant). On the other hand, we

formally state the regularity result as a lemma:

Lemma 5.2 Let b(z, z , ζ ) = ζ m(z,z ,ζ ) . Then b ∈ Sδ,0
m,0,0
for all δ > 0.
Proof Observe that f = m logζ ∈ S,0,0 for all > 0 since this holds for
logζ , and as m ∈ S0,0,0 . Further, if f ∈ S0 ,1 ,2 with 0 ≤ 0 , 1 , 2 < 1 then
e−f Dαz Dz Dζ e f ∈ S−|γ |+0 |(α,β,γ )|,−|α|+1 |(α,β,γ )|,−|β|+2 |(α,β,γ )| ,

β γ
as follows by induction on |α| + |β| + |γ |. Indeed, it holds when α, β, γ all

vanish. Further,
e−f Dzj (Dαz Dz Dζ ef ) = Dzj (e−f Dαz Dz Dζ ef ) + (Dzj f )(e−f Dαz Dz Dζ ef ),

β γ β γ β γ
240 A. Vasy
and e−f Dαz Dz Dζ ef ∈ S−|γ |+0 |(α,β,γ )|,−|α|+1 |(α,β,γ )|,−|β|+2 |(α,β,γ )| by the
β γ
inductive hypothesis, and then the first term on the right hand side improves
the second order by 1 keeping all others unchanged, while Dzj f ∈ S0 ,1 −1,2 ,
so the second term on the right hand side adds 0 , 1 − 1, 2 to the orders,
while |α| is increased by 1 in both cases. The argument is symmetric for all
other derivatives, giving the conclusion. Applying this with 1 = 2 = 0,
0 = , > 0 arbitrary, we deduce that for all δ > 0 (namely, we take = δ),
m,0,0
ζ m ∈ Sδ,0 indeed.
We still have, analogously to the constant order setting, that

m,l1 ,l2 m ,l ,l 2 m+m ,l1 +l 1 ,l2 +l 2
a ∈ Sδ,δ , b ∈ Sδ,δ 1 ⇒ ab ∈ Sδ,δ ,
and for δ = 0
m,l1 ,l2 β γ m−|γ |+δ|(α,β,γ )|,l1 −|α|,l2 −|β|
a ∈ Sδ,0 ⇒ Dαz Dz Dζ a ∈Sδ,0 , (5.11)
while for general δ , the a1 piece, as defined in (5.4), satisfies

m,l1 ,l2 β γ m−|γ |+δ|(α,β,γ )|,l1 −|α|+δ |(α,β,γ )|,l2 −|β|
a1 ∈ Sδ,δ ⇒ Dαz Dz Dζ a1 ∈Sδ,δ ,
(5.12)
where by the support property of a1 , δ |(α, β, γ )| could also be shifted to
∞,∞,∞
the last order). The analogue of (5.11) also holds for S∞,δ , in which case
l1 and l2 are unaffected by derivatives.
Having discussed symbols in some detail, we now turn to operators, starting
with the constant order S∞,δ -type setting. Note that unless m < −n, the integral
m,
1 ,
2
(5.3) with a ∈ S∞,δ is not absolutely convergent; if m < −n, it is, with the
result Au ∈ C(Rn ), and for M >
2 + n,
sup |z−
1 Au(z)| ≤ Ca m,
,
2 uS ,0,M ,
S∞,δ1 ,0
where C is a universal constant (independent of a and u) and

uS ,k,M = sup |zβ Dαz u|
|α|≤k |β|≤M
are the Schwartz seminorms. However, if m < −n, one can also integrate by

parts as usual in z , noting that (1 + z )eiζ ·(z−z ) = ζ 2 eiζ ·(z−z ) , so

−n
Au(z) = (2π ) ζ −2N (1 + z )N eiζ ·(z−z ) a(z, z , ζ ) u(z ) dζ dz
R ×R
n n
−n
= (2π ) eiζ ·(z−z ) ζ −2N (1 + z )N (a(z, z , ζ ) u(z )) dζ dz .
Rn ×Rn
(5.13)
Expanding (1 + z )N (a(z, z , ζ ) u(z )), one deduces that
|(1 + z )N (a(z, z , ζ ) u(z ))| ≤ z

1 z
2 −M ζ m+2Nδ a m,
,
2 uS ,2N,M ,
S∞,δ1 ,2N
(5.14)
so for just m + 2Nδ < −n + 2N, that is
2(1 − δ)N > m + n,
the right hand side of (5.13) is integrable, and defining Au ∈ C(Rn ) to be the
result,
sup |z−
1 Au(z)| ≤ Ca m,
,
2 uS ,2N,M . (5.15)
S∞,δ1 ,2N
m,
1 ,
2 −∞,
1 ,
2
This gives an extension of A = I(a) to S∞,δ . Since S∞,δ is dense in
m,
1 ,

m,
1 ,
2
S∞,δ in the topology of S∞,δ 2 for m > m, and since for m < −n, the
expressions (5.13) for various N are all equal, the continuity property (5.15)
shows that A is independent of the choice of N provided m < −n + 2(1 − δ)N
(since one can then take m ∈ (m, −n + 2(1 − δ)N), and use the m -continuity
and density statements).
Now at least Au ∈ C(Rn ), with a suitable bound, is defined, but in fact it

is in S(Rn ). To see this, first note that Dαz eiζ ·(z−z ) = ζ α , so for N sufficiently
large, so that m + |α| < −n + 2(1 − δ)N, differentiating under the integral sign
and using Leibniz’s rule,

(Dαz Au)(z) = Cγ λ (2π )−n Dγz (eiζ ·(z−z ) )ζ −2N
γ +λ≤α Rn ×Rn
(1 + z )N (Dλz a(z, z , ζ ) u(z )) dζ dz

(5.16)

= Cγ λ (2π )−n eiζ ·(z−z ) ζ γ ζ −2N
γ +λ≤α Rn ×Rn
(1 + z )N (Dλz a(z, z , ζ ) u(z )) dζ dz ,
with Cγ λ combinatorial constants, so by (5.14) with a replaced by Dλz a, with

M > n +
2 still,
sup |z−
1 (Dαz Au)(z)| ≤ Ca m,
,
2 uS ,2N,M .
S∞,δ1 ,2N+|α|

Further, zj eiζ ·(z−z ) = z j eiζ ·(z−z ) + Dζj eiζ ·(z−z ) , so

zβ eiζ ·(z−z ) = (z + Dζ )β eiζ ·(z−z ) = Cμν (z )μ Dνζ eiζ ·(z−z ) ,
μ+ν≤β
242 A. Vasy
so integration by parts in ζ gives

(zβ Dαz Au)(z) = Cγ λ Cμν (2π )−n eiζ ·(z−z )
γ +λ≤α μ+ν≤β Rn ×Rn

Dνζ ζ γ ζ −2N (z )μ (1 + z )N (Dλz a(z, z , ζ ) u(z )) dζ dz

−n
= Cγ λ Cμν Cν ν (2π )
eiζ ·(z−z )
γ +λ≤α μ+ν≤β ν +ν ≤ν Rn ×Rn

Dνζ (ζ γ ζ −2N )(z )μ (1 + z )N (Dνζ Dλz a(z, z , ζ ) u(z ))) dζ dz .
(5.17)
Thus with
M > n +
2 + |β| and m + |γ | − |ν | − 2N + (2N + |ν | + |λ|)δ < −n,
the latter of which is implied by
m + |α| + |β|δ < −n + 2(1 − δ)N,
we have
sup |z−
1 zβ Dαz Au(z)| ≤ Ca m,
,
2 uS ,2N,M ,
S∞,δ1 ,2N+|α|
with C independent of a, u. Now for

1 ≤ 0, z−
1 can simply be dropped,

while for
1 > 0 the z−
1 factor can be absorbed into a sum zβ terms with
|β | ≤ M where M ≥
1 , so we obtain that for
M ≥ max(0,
1 ), M > n +
2 + |β| + M , m + |α| + |β|δ < −n + 2(1 − δ)N
we have
sup |zβ Dαz Au(z)| ≤ Ca m,
,
2 uS ,2N,M ,
S∞,δ1 ,2N+|α|
so Au ∈ S(Rn ), and the map A : S → S is continuous, and in fact the stronger

continuity property, namely that
m,
1 ,
2
S∞,δ × S (a, u) → I(a)u ∈ S
is continuous, holds. Thus, we have the first claim of the following lemma, as
well as the second in case δ = 0:
Lemma 5.3 The maps
m,
1 ,
2
S∞,δ × S (a, u) → I(a)u ∈ S,
m,
1 ,
2
Sδ,δ × S (a, u) → I(a)u ∈ S,
are continuous.
Proof To deal with general (not necessarily vanishing) δ ∈ [0, 1/2), proceed
by using χ ∈ Cc∞ (R), χ ≡ 1 on [ 12 , 2], supported in [ 14 , 4]. Then we can write
m,
1 ,
2
a ∈ Sδ,δ as

z z
a = a1 + a2 , a1 = χ a, a2 = 1 − χ a,
z z
m,
1 ,
2
with aj depending continuously on a in the Sδ,δ topology. Now, since
z ∼ z ∼ (z, z )
on supp a1 , and since differentiation is local, a1 satisfies estimates

|Dαz Dz Dζ a1 | ≤ Cαβγ z
1 +
2 −|α|−|β|+δ |(α,β,γ )| ζ m−|γ |+δ|(α,β,γ )| .
β γ
Denoting the corresponding seminorms by . m,

+
2 temporarily, note that
S̃δ,δ 1 ,N
m,
1 +
2
a1 in S̃δ,δ depends continuously on a. The right hand side of (5.14)
becomes

z
1 +
2 +2Nδ z −M ζ m+2Nδ a1 m,
,
2 uS ,2N,M ,
S̃δ,δ 1 ,2N
so for M > n and m + 2Nδ < −n + 2N the right hand side of (5.13) is
integrable, and (5.15) becomes

sup |z−
1 −
2 −2Nδ A1 u(z)| ≤ Ca1 m,
,
2 uS ,2N,M . (5.18)
S̃δ,δ 1 ,2N
In fact, using z ∼ z on supp a1 , taking M > n +

1 +
2 + 2Nδ + |β|,
m + 2Nδ < −n + 2N (i.e. first choose N sufficiently large, then M sufficiently
large), this even gives
sup |zβ A1 u(z)| ≤ Ca m,

,
2 uS ,2N,M .
S̃δ,δ 1 ,2N
To deal with derivatives, use (5.16) and note that the integrand is bounded by
a constant multiple of

sup z
1 +
2 +2Nδ +|λ|δ z −M ζ m+2Nδ−2N+|γ |+|λ|δ
|γ |+|λ|=|α|
a1 m,
1 ,
2 uS ,2N,M ,
S̃δ,δ ,2N+|α|
which in turn is bounded by

sup z
1 +
2 +2Nδ +|α|δ z −M ζ m+2Nδ−2N+|α| a1 m,
,
2 uS ,2N,M ,
S̃δ,δ 1 ,2N+|α|
|γ |+|λ|=|α|
244 A. Vasy
so in view of the support of a1 first choosing N such that m+2Nδ −2N +|α| <
−n and then M such that M > n +
1 +
2 + 2Nδ + |α|δ + |β|, the estimate
sup |zβ Dαz A1 u(z)| ≤ Ca1 m,

,
2 uS ,2N,M
S̃δ,δ 1 ,2N
follows, with C independent of a1 , u. This shows that a1 satisfies the conclusion

of the lemma.
Now, to deal with a2 , integrate by parts in ζ , starting with (5.13) for A2 =
I(a2 ) in place of A = I(a), using

ei(z−z )·ζ = z − z −2 (1 + ζ )ei(z−z )·ζ ,
so first for m < −n

A2 u(z) = (2π )−n eiζ ·(z−z ) z − z −2K
Rn ×Rn

(1 + ζ )K ζ −2N (1 + z )N (a2 (z, z , ζ ) u(z )) dζ dz

C̃μν (2π )−n eiζ ·(z−z ) z − z −2K (Dζ ζ −2N )
μ
=
|μ|+|ν|≤2K Rn ×Rn
(1 + z )N (Dνζ a2 (z, z , ζ ) u(z )) dζ dz ,

(5.19)
where C̃μν are combinatorial constants. On the support of a2 ,
z − z ≥ C (z + z )
for some C > 0, and now the integrand on the right hand side is bounded by a
constant multiple of

z
1 z
2 −M (z, z )−2K+(2N+2K)δ ζ −2N+m+(2N+2K)δ
a2 m,
,
2 uS ,2N,M .
Sδ,δ 1 ,2N+2K
For a given β, we can now even take M = 0, and take N, K so that
2Nδ − (1 − δ )2K < −n − |β| −

1 −
2
and
−(1 − δ)2N + 2Kδ + m < −n;
to see that such a choice exists, take K = N, in which case sufficiently large N
works as 1 − 2δ, 1 − 2δ > 0. We then deduce
sup |zβ A2 u(z)| ≤ Ca m,

,
2 uS ,2N,M .
Sδ,δ 1 ,2N+2K
To deal with derivatives, we again use a calculation similar to (5.16) to obtain

that

Dαz A2 u(z) = Cγ κλ C̃μν (2π )−n
γ +κ+λ≤α |μ|+|ν|≤2K

(5.20)
ζ γ eiζ ·(z−z ) (Dκz z − z −2K )(Dζ ζ −2N )
μ
Rn ×Rn
(1 + z )N (Dνζ Dλz a2 (z, z , ζ ) u(z )) dζ dz .
Since
Dκz z − z −2K ≤ Cz − z −2K
(indeed, one even has a bound Cz − z −2K−|κ| ), so now the integrand on the
right hand side is bounded by a constant multiple of

z
1 z
2 −M (z, z )−2K+(2N+2K+|α|)δ ζ −2N+m+(2N+2K+|α|)δ
a2 m,
,
2 uS ,2N,M ,
Sδ,δ 1 ,2N+2K+|α|
which gives
sup |zβ Dαz A2 u(z)| ≤ Ca2 m,

,
2 uS ,2N,M
Sδ,δ 1 ,2N+2K+|α|
when M = 0, and take N, K so that
2Nδ − (1 − δ )2K + δ |α| < −n − |β| −

1 −
2
and
−(1 − δ)2N + 2Kδ + m + |α|δ < −n,
which can be arranged exactly as in the α = 0 case above. This completes the
proof of the lemma.
Note that for such an A with m < −n to start, u ∈ S, φ ∈ S,

i(−ζ )·(z −z)
Au(z)φ(z) dz = u(z ) e a(z, z , ζ )φ(z) dz dζ dz

iζ ·(z −z)
= u(z ) e a(z, z , −ζ )φ(z) dz dζ dz

= u(z )(I(b)φ)(z ) dz ,
where b(z, z , ζ ) = a(z , z, −ζ ), so b ∈ S∞,δ

m,
2 ,
1
. Let j to be the transposition
map j(z, z , ζ ) = (z , z, ζ ), ρ the reflection map ρ(z, z , ζ ) = (z, z , −ζ ), so

246 A. Vasy
ρ ∗ : S∞,δ
m,
1 ,
2 m,
1
2 ∗
→ S∞,δ m,
1 ,
2
, j : S∞,δ m,
2 ,
1
→ S∞,δ are continuous for all m,
1 ,
2 .
We then have at first for m < −n,

(I(a)u)φ = u(I(ρ ∗ j∗ a)φ),
m,
1 ,
2
so both sides being continuous trilinear maps S∞,δ × S × S → C for all
−∞,
1 ,
2 m,
1 ,
2 m ,
1 ,
2
m,
1 ,
2 , by the density of S∞,δ in S∞,δ in the S∞,δ topology for
m > m, the identity extends to all m. Thus, the Fréchet space adjoint, I(a)† :
S → S , defined by
(I(a)† φ)(u) = φ(I(a)u), φ ∈ S , u ∈ S,
satisfies
I(a)† φ = I(ρ ∗ j∗ a)φ, φ ∈ S,
that is by the weak-* density of S in S , I(a)† is the unique continuous
extension of I(ρ ∗ j∗ a) from S to S ; one simply writes I(ρ ∗ j∗ a) = I(a)†
even as maps S → S . Since ρ ∗ j∗ ρ ∗ j∗ a = a, we deduce that for any a,
I(a) = I(ρ ∗ j∗ a)† : S → S is continuous.
Here we have used the bilinear distributional pairing; if one uses the
sesquilinear L2 -pairing, one has

Au(z)φ(z) dz = u(z ) eiζ ·(z −z) a(z, z , ζ )φ(z) dz dζ dz

= u(z ) (I(b̃)φ)(z ) dz ,
b̃(z, z , ζ ) = a(z , z, ζ ), so using ∗ to denote the corresponding (Hilbert-space-

type) adjoint
(I(a))∗ = I(cj∗ a), (5.21)
where c is the complex conjugation map.
m,
1 ,
2
Note that if a ∈ Sδ,δ then cj∗ a ∈ Sδ,δ
m,
2 ,
1
, thus the adjoint of operators
given by our scattering symbols is still in the same class, with
2 and
1
reversed.
While we have two indices
1 and
2 for growth in the spatial variables, this
is actually redundant,
1 +
2 is the relevant quantity, as we have already seen
m,
1 ,
2
signs of in the proof of Lemma 5.3 in the case of Sδ,δ : for the a1 term the
orders were interchangeable due to support properties, while the a2 term was
irrelevant.
Lemma 5.4 Given
∈ R, the range of the map a → I(a) is independent of
the choice of
1 and
2 as long as
1 +
2 =
.
Definition 5.1 We now define

m,
m,
,0
∞,δ (Rn ) = {I(a) : a ∈ S∞,δ }
and
m,
m,
,0
(R ) = {I(a) : a ∈ Sδ,δ };
n
δ,δ
m,
1 ,
2 m,
1 ,
2
we could have used S∞,δ , resp. Sδ,δ instead for any
1 ,
2 with

1 +
2 =
.
m,
1 ,
2
Proof To see this lemma for S∞,δ , we note as in the proof of Lemma 5.3
)·ζ
that (1 + ζ )e i(z−z )·ζ = z − z e 2 i(z−z , so at first for m < −n, as usual, for
m,
1 ,
2
a ∈ S∞,δ ,

−n
(I(a)u)(z) = (2π ) z − z −2N (1 +ζ )N (eiζ ·(z−z ) )a(z, z , ζ ) u(z ) dz ,
R ×R
n n
−n
= (2π ) eiζ ·(z−z ) (z − z −2N (1 +ζ )N a(z, z , ζ )) u(z ) dz
Rn ×Rn
= (I(b)u)(z),
(5.22)
where
b(z, z , ζ ) = z − z −2N (1 + ζ )N a(z, z , ζ ). (5.23)
Notice that
z2 = 1 + |z|2 ≤ 1 + (|z − z | + |z|)2 ≤ 1 + 2|z |2 + 2|z − z |2 ≤ 2z − z 2 z 2 ,
(5.24)
and the analogous inequality also holds with z and z interchanged, and
Dαz Dz z − z −2N ≤ Cαβ z − z −2N ,
β
m,
1 ,
2
so for any m,
1 ,
2 , a ∈ S∞,δ , with b defined by (5.23) satisfies b ∈
m,
1 +s,
2 −s
S∞,δ for −2N ≤ s ≤ 2N, and the map
m,
1 ,
2 m,
1 +s,
2 −s
S∞,δ a → b ∈ S∞,δ
is continuous, hence I(a) = I(b) holds for all m,
1 ,
2 (as it holds for m < −n).
m,
1 ,
2
Given any s, choosing sufficiently large N, shows that the range of I on S∞,δ
only depends on
1 +
2 .
m,
1 ,
2 m,
1 +s,
2 −s
Now, if a ∈ Sδ,δ then b defined by (5.23) is usually not in Sδ,δ , as

derivatives in z and z do not typically give extra decay when hitting z−z −2N .
However, for the decomposition a = a1 + a2 used in the proof of Lemma 5.3,
on the support of the a2 piece derivatives of z−z −2N have the required decay
(indeed, one has decay in (z, z ) jointly upon differentiation in either z or z ), so
248 A. Vasy
m,
1 −s,
2 −s
the corresponding b2 satisfies b2 ∈ Sδ,δ if s + s ≤ 2N(1 − δ ) (with
δ coming from the ζ derivatives), while for the a1 piece the weights
1 and
2
are directly equivalent as z ∼ z on supp a1 .
I use this opportunity to remark that for the a2 piece I(a2 ) of I(a) in fact one
has
m ,
−∞,−∞
I(a2 ) ∈ ∩m ,
∈R δ,δ = δ,δ . (5.25)
We have already seen above that the analogue of this holds with m = m fixed,
l ∈ R. In order to see that m can be taken arbitrarily as well, note that due to

the support of a2 , we can use ζ ei(z−z )·ζ = |z − z |2 ei(z−z )·ζ and integrate by
parts in ζ (noting that the diagonal singularity of |z − z |−2 is irrelevant due to
the support of a2 ) to see that

iζ ·(z−z )
(I(a2 )u)(z) = (2π )−n |z − z |−2N N ζ (e )a2 (z, z , ζ ) u(z ) dz ,
Rn ×Rn

−n
= (2π ) eiζ ·(z−z ) (|z − z |−2N N
ζ a2 (z, z , ζ )) u(z ) dz

Rn ×Rn
= (I(b2 )u)(z), (5.26)
where
m−(1−δ)2N,
1 −s,
2 −s
b2 (z, z , ζ ) = |z − z |−2N N
ζ a2 (z, z , ζ ) ∈ Sδ,δ (5.27)
if s + s ≤ 2N(1 − δ ). This shows (5.25). The analogue also holds on S∞,δ

m,
1 ,
2
,
−∞,−∞
namely in that case the similarly defined a2 gives rise to I(a2 ) ∈ ∞,δ .
5.3.3. Left and Right Reduction

m,
One very useful property of ∞,δ (Rn ) is that it is in fact exactly the range of I
acting on symbols of a special form, namely those independent of z . Thus, let
m,
a ∈ S∞,δ (Rnz ; Rnζ ) ⇐⇒ a ∈ C ∞ (Rnz × Rnζ ),

|Dαz Dζ a| ≤ Cαγ z
ζ m−|γ |+δ|(α,γ )| ;
γ
so with
πL : Rnz × Rnz × Rnζ → Rnz × Rnζ
the projection map dropping z , a ∈ S∞,δ

m,
(Rnz ; Rnζ ) if and only if
πL∗ a ∈ S∞,δ
m,
,0
(Rnz ; Rnz ; Rnζ ).
As usual, the seminorms

supz−
ζ −m+|γ |−δ|(α,γ )| |Dαz Dζ a|
γ
aSm,
,N =
∞,δ
|α|+|γ |≤N
give a Fréchet topology. With πR the projection dropping the z variables, one
m,
also has a ∈ S∞,δ (Rn ; Rn ) if and only if πR∗ a ∈ S∞,δ

m,0,
(Rnz ; Rnz ; Rnζ ).

Then:
m,
1 ,
2
Proposition 5.1 For any
=
1 +
2 and a ∈ S∞,δ (Rnz ; Rnz ; Rnζ ) there exists
m,
a unique aL ∈ S∞,δ (Rnz ; Rnζ ) such that I(a) = I(πL∗ aL ); one writes qL = I ◦πL∗ :
m,
m,
S∞,δ → ∞,δ . Here aL is called the left reduced symbol of I(a), and qL is the
left quantization map.
m,
1 ,
2
Similarly, for any
=
1 +
2 and a ∈ S∞,δ (Rnz ; Rnz ; Rnζ ) there exists a
m,
unique aR ∈ S∞,δ (Rnz ; Rnζ ) such that I(a) = I(πR∗ aR ); one writes qR = I ◦ πR∗ :
m,
m,
S∞,δ → ∞,δ . Here aR is called the right reduced symbol of I(a), and qR is
the right quantization map.
Moreover, the maps a → aL , a → aR are continuous.
Further, with ι : Rn ×Rn → Rn ×Rn ×Rn the inclusion map as the diagonal
in the first two factors, that is ι(z, ζ ) = (z, z, ζ ),
i|α|
aL ∼ ι∗ Dαz Dαζ a, (5.28)
α
α!
and
(−i)|α|
aR ∼ ι∗ Dαz Dαζ a,
α
α!
−∞,
with the summation asymptotic in ζ , that is is modulo S∞,δ ; see (5.36).

m,
1 ,
2 m,
If instead a ∈ Sδ,δ , then the conclusions hold with aL , aR ∈ Sδ,δ , with the
asymptotic summation being asymptotic both in z and in ζ , that is is modulo
S−∞,−∞ .
m,l1 ,l2
In the case of variable orders, stated for Sδ,δ only:
m,l1 ,l2 m,l
Corollary 5.1 If a ∈ Sδ,δ then aL , aR ∈ Sδ,δ , where
l(z, ζ ) = l1 (z, z, ζ ) + l2 (z, z, ζ ).
This corollary is an immediate consequence of the asymptotic expansion in

m−(1−2δ)|α|,l−(1−2δ )|α|
Proposition 5.1, for the αth term there is in Sδ,δ .
250 A. Vasy
m,
Notice that for a ∈ S∞,δ ,

qL (a)u(z) = (2π )−n eiζ ·z a(z, ζ ) (Fu)(ζ ) dζ (5.29)
Rn
for m < −n, but now, for u ∈ S, the right hand side extends continuously to
m,
S∞,δ for all m, so one could have directly defined qL (a) for all m. Similarly,

qR (a)u = F −1 (ζ → e−iz ·ζ a(z , ζ ) u(z ) dz ), (5.30)
Rn
where now the right hand side makes sense directly as a tempered distribution
for all m. However, relating qL and qR , as well as performing other important
calculations, would be rather hard without having defined the map I in general,
−∞,−∞
via a continuity/regularization argument! Note that for a ∈ S∞ , in either
case, one deduces that directly that qR (a)u and qL (a)u are in S.
m,
We remark that if a ∈ S∞ is a polynomial in ζ , that is a(z, ζ ) =

α
|α|≤m aα (z)ζ , then one can pull the factors aα (z) out of the integral (5.29),
and thus ζ α F = FDα and the Fourier inversion formula yields

qL (a)u(z) = aα (z)(Dα u)(z),
|α|≤m
that is with aα acting as multiplication operators,

qL (a) = aα Dα . (5.31)
|α|≤m
Similarly,

qR (a)u(z) = (Dα (aα u))(z),
|α|≤m
that is

qR (a) = Dα aα .
|α|≤m
So differential operators of order m on Rn with coefficients in S

(Rn ) lie in
m,
. In particular, differential operators with coefficients in C ∞ (Rn ) lie in
m,0 (Rn ).
We now prove Proposition 5.1. We only consider the left reduction, that is
the L subscript case, as the R case is completely analogous. First, we note that
m,
1 ,
2
the uniqueness is straightforward. Any operator A = I(a), a ∈ S∞,δ , has

a Schwartz kernel, KA ∈ S (as it is a continuous linear map S → S, thus
S → S ). When m < −n, the Schwartz kernel satisfies

KA (φ ⊗ u) = (Au)(z)φ(z) dz

= (2π )−n eiζ ·(z−z ) a(z, z , ζ ) u(z ) φ(z) dζ dz dz (5.32)

= (Fζ−1 a)(z, z , z − z )u(z )φ(z) dz dz,
where Fζ−1 is the inverse Fourier transform in the third variable, ζ . (F3−1 is a
logically better, but less self-explanatory, notation.) Thus, for such a, KA is the
polynomially bounded function (hence tempered distribution) given by
Fa (z, z ) = (Fζ−1 a)(z, z , z − z ) = (F3−1 a)(z, z , z − z ). (5.33)
m,
If a ∈ S∞,δ , then, with 2 denoting that the inverse Fourier transform is in the
second slot, we have
FπL∗ a (z, z ) = (F2−1 a)(z, z − z ) = (G∗ F2−1 a)(z, z )
where G : R2n → R2n is the invertible linear map G(z, z ) = (z, z − z ), thus
one can pull back tempered distributions by it. Thus,
KI(πL∗ a) = G∗ F2−1 a,
and correspondingly
a = F2 (G−1 )∗ KI(πL∗ a) ,
m,
first for m < −n, but then as both sides are continuous maps S∞,δ → S , this
m,
1 ,
2
identity holds in general. In particular, given ã ∈ S∞,δ there exists at most
m,
1 +
2
one a ∈ S∞,δ such that I(πL∗ a) = I(ã), for
a = F2 (G−1 )∗ KI(ã) . (5.34)
Now for existence. In principle (5.34) solves this problem, but then one
needs to show that the a it provides, that is aL in the notation of the proposition,
is not merely a tempered distribution, but is in an appropriate symbol class. So
we proceed differently.
For the following discussion it is useful to replace a by a1 ; recall that
−∞,−∞
I(a2 ) ∈ ∞,δ in this case, thus does not affect the argument below. Hence,
to minimize subscripts, we simply write a below, but we actually apply the
argument to a1 . With the notation of the proposition, one expands a in Taylor
series in z around the diagonal z = z, with the integral remainder term:
252 A. Vasy
(z − z)α
a(z, z , ζ ) = ((∂z )α a)(z, z, ζ ) + RN (z, z , ζ )
α!
|α|≤N−1
(z − z)α 1
RN (z, z , ζ ) = N (1 − t)N−1 ((∂z )α a)(z, (1 − t)z + tz , ζ ) dt.
α! 0
|α|=N
(5.35)

Now, for m < −n, as (z j − zj )eiζ ·(z−z ) = −Dζj eiζ ·(z−z ) ,

(I((z j − zj )a)u)(z) = (2π )−n (−Dζj )eiζ ·(z−z ) a(z, z , ζ ) u(z ) dz dζ

−n
= (2π ) eiζ ·(z−z ) (Dζj a)(z, z , ζ ) u(z ) dz dζ
= (I(Dζj a)u)(z),
so as
m,
1 ,
2
S∞,δ × S (a, u) → I((z j − zj )a)u ∈ S
and
m,
1 ,
2
S∞,δ × S (a, u) → I(Dζj a)u ∈ S
m,
1 ,
2
are both continuous bilinear maps, the density of S∞,δ in the topology of

m ,
1 ,
2
S∞,δ for m > m shows that
I((z − z)α a) = I(Dαζ a)

m,
1 ,
2
for all m and a ∈ S∞,δ .
Thus, for a as in (5.35),
1
I(a) = I((Dζ )α ι∗ ∂zα a) + I(R N ),
α!
|α|≤N−1

1 1
RN (z, z , ζ ) = N (1 − t)N−1 (Dαζ (∂z )α a)(z, (1 − t)z + tz , ζ ) dt.
α! 0
|α|=N
But keeping in mind the support properties of a (recall that it stands for the a1
piece!),
m−(1−2δ)N,
1 +
2 ,0
(Dζ )α ι∗ ∂zα a ∈ S∞,δ , R N ∈ S∞,δ
m−(1−2δ)|α|,
1 ,
2
,
with the map

m−(1−2δ)|α|,
1 +
2
m,
1 ,
2
S∞,δ a → (Dζ )α ι∗ ∂zα a ∈ S∞,δ
continuous, and similarly with R N . Since (Dζ )α ι∗ ∂zα a is independent of z , and

for this the original a and a1 give exactly the same expression, this proves the
m,
1 ,
2
following weaker version of Proposition 5.1: for all a ∈ S∞,δ and for all N
m,
1 +
2
there exists aN ∈ S∞,δ such that
m−(1−2δ)N,
1 +
2 ,0
I(a) − I(aN ) = I(R N ), R N ∈ S∞,δ .
m,
1 ,
2
Notice that if a ∈ Sδ,δ then writing a = a1 + a2 , we already know by (5.25)
m ,
,

that for any m ,
1 ,
2 we can write I(a2 ) = I(b2 ), b2 ∈ Sδ,δ 1 2 , while for a1
m,
1 ,
2
the analogous conclusions to the S∞,δ setting hold but with
m−(1−2δ)|α|,
1 +
2 −(1−2δ)|α|
(Dζ )α ι∗ ∂zα a1 ∈ Sδ,δ ,
m−(1−2δ)N,
1 +
2 −(1−2δ )N,0
R 1,N ∈ Sδ,δ .
An asymptotic summation argument allows one to improve this. This notion
m−(1−2δ)j,
means the following: suppose aj ∈ S∞,δ for j ∈ N. Then there exists

m,
a ∈ S∞,δ such that

N−1
m−(1−2δ)N,
a− aj ∈ S∞,δ . (5.36)
j=0
To see this, we take χ ∈ C ∞ (Rn ) with χ (ζ ) = 1 for |ζ | ≥ 2, χ (ζ ) = 0 for

|ζ | ≤ 1. For 0 < j < 1 to be determined, but with j → 0, consider
∞

a(z, ζ ) = χ (j ζ )aj (z, ζ );
j=0
the sum is finite for (z, ζ ) with |ζ | ≤ R, with only the finitely many terms with
j ≥ R−1 contributing. Thus, a is C ∞ ; the question is convergence in S∞,δ m,
, and
the property (5.36). But by Leibniz’s rule,
∞
|γ | α−γ
(Dαζ Dβz a)(z, ζ ) = Cαγ j (Dγ χ )(j ζ )(Dζ Dβz aj )(z, ζ ).
j=0 γ ≤α
m−(1−2δ)N,
To get convergence of the tail in S∞,δ , we need to estimate the sup

norm of
⎛ ⎛ ⎞⎞
∞

ζ −m+(1−2δ)N−δ(|α|+|β|)+|α| z−
⎝Dαζ Dβz ⎝ χ (j ζ )aj (z, ζ )⎠⎠
j=N
∞

(N−j)(1−2δ)+|γ |
Cαγ ζ−δ|γ | j ζ|γ |+(N−j)(1−2δ) j
(j−N)(1−2δ)
= (Dγχ)(j ζ )
j=N γ ≤α

ζ −m+(1−2δ)j+(1−δ)(|α|−|γ |)−δ|β| z−
(Dζ
α−γ
Dβz aj )(z, ζ ) ;
(5.37)
254 A. Vasy
where we have used the above expansion. For γ = 0, we use |ζ | ≥ j−1 on

supp χ (j .), so for j ≥ N (as δ ∈ [0, 1/2)),
(N−j)(1−2δ)
j ζ (N−j)(1−2δ) = (j2 + j2 |ζ |2 )(1−2δ)(N−j)/2 ≤ 1,
while for γ = 0 we use j−1 ≤ |ζ | ≤ 2j−1 on supp(Dγ χ )(j .), so
1 ≤ ζ j = (j2 + j2 |ζ |2 )1/2 ≤ 51/2
on supp(Dγ χ )(j .) for all γ = 0, and thus for j ≥ N,

(N−j)(1−2δ)+|γ |
ζ |γ |+(N−j)(1−2δ) j ≤ 5|γ |/2
there. Thus, adding up the terms with |α| + |β| = M as required by the
symbolic seminorms, there are constants CM > 0 (arising from finitely many
combinatorial constants, from suprema of finitely many derivatives of χ and
from finite powers of 51/2 ) such that the series is absolutely summable, and
hence convergent, if for all M
∞

CM j aj Sm−(1−2δ)j,
,M
∞,δ
j≥N+(1−2δ)−1
(j−N)(1−2δ)
converges; here j is from j ≤ j on the right hand side of (5.37),
taking advantage of j ≥ N + (1 − 2δ)−1 in our sum. Now, if aj Sm−(1−2δ)j,
,M ≤
∞,δ
Rj,M , where Rj,M are specified constants, then one can arrange the convergence
by for instance requiring that for j > M, the corresponding summand is ≤ 2−j ,
that is for j > M,
−1 −1
j ≤ 2−j CM Rj,M .
Note that for each j this is finitely many constraints (as only the values of M
with M < j matter), which can thus be satisfied. Correspondingly, the tail of
m−(1−2δ)N,
m,
the series converges for each N in S∞,δ , and thus a ∈ S∞,δ and also
(5.36) holds. This gives a continuous asymptotic summation map on arbitrary
bounded subsets of the product of the symbol spaces. (One can make the map
globally defined and continuous by letting j be the minimum of, say,
−1
2−j CM (1 + aj Sm−(1−2δ)j,
,M )−1 ,
∞,δ
over M = 0, 1, . . . , j − 1, but this is actually not important below.)

Now, let
1 m,
1 +
2
ã ∼ (Dζ )α ι∗ ∂zα a ∈ S∞,δ ;
α
α!
asymptotic summation can be done so that the map a → ã is continuous. Then

m−(1−2δ)N,
1 ,
2
ã − aN ∈ S∞,δ for all N, and thus
m−(1−2δ)N,
1 ,
2
I(a) − I(ã) ∈ ∩N I(S∞,δ ).
m,
1 ,
2
If a ∈ Sδ,δ then with
1 m,
1 +
2
ã ∼ (Dζ )α ι∗ ∂zα a ∈ Sδ,δ ,
α
α!
where we asymptotically sum both in the z and in the ζ variables (this can be
done at the same time, adding a factor of χ (j z)),
m−(1−2δ)N,
1 ,
2 −(1−2δ )N
I(a) − I(ã) ∈ ∩N I(Sδ,δ ).
The following lemma then finishes the proof of Proposition 5.1:
m,
1 ,
2 m−N,
1 ,
2
Lemma 5.5 Suppose b ∈ S∞,δ satisfies I(b) ∈ ∩N I(S∞,δ ), that is for
m−N,
1 ,
2
all N ∈ N there is bN ∈ S∞,δ such that I(b) = I(bN ). Then there exists
−∞,
1 +
2
c ∈ S∞,δ such that I(c) = I(b). Moreover, if there are continuous maps
jN : b → bN , then the map b → c is continuous.
m,
1 ,
2 m−N,
1 ,
2 −N
Suppose instead b ∈ Sδ,δ satisfies I(b) ∈ ∩N I(Sδ,δ ), that is for
m−N,
1 ,
2 −N
all N ∈ N there is bN ∈ Sδ,δ such that I(b) = I(bN ). Then there exists
c ∈ S −∞,−∞ such that I(c) = I(b). Moreover, if there are continuous maps
jN : b → bN , then the map b → c is continuous.
Proof The idea of the proof is to use (5.34), as in the present setting the
Schwartz kernel can be shown to be well-behaved, so (5.34) immediately gives
the appropriate symbolic properties of c. Thus, we note that for all N there is
m−N,
1 ,
2
bN ∈ S∞,δ such that I(b) = I(bN ), so taking N such that m − N < −n,
(5.32)–(5.33) give the Schwartz kernel (which is independent of N) as the
continuous polynomially bounded function
KI(bN ) (z, z ) = (Fζ−1 bN )(z, z , z − z );
taking m − N < −n − k, this is in fact Ck with polynomial bounds up to the kth

derivatives. Correspondingly, it satisfies, for |α| + |β| + δ|γ | ≤ k, and writing
Dαj for the αth derivative in the jth slot, Mjα for the multiplication by the αth
coordinate in the jth slot,
z−
1 z −
2 (z − z )γ Dαz Dz KI(bN ) (z, z )
β

= .−
1 −
2 γ α β −1
1 .2 M3 (D1 + D3 ) (D2 − D3 ) (F3 bN ) (z, z , z − z )

−1 −
1 −
2 γ
= F3 .1 .2 D3 (D1 + M3 )α (D2 − M3 )β bN (z, z , z − z ).
256 A. Vasy
As
.−
−
2 γ
1 .2 D3 (D1 + M3 ) (D2 − M3 ) bN
1 α β
is bounded in C∞ (Rn ×Rn ; L1 (Rnζ )) by a seminorm of bN as |α|+|β|+δ|γ | ≤

k, m − N < −n − k, where C∞ stands for bounded continuous functions,
F3−1 .−
−
2 γ
1 .2 D3 (D1 + M3 ) (D2 − M3 ) bN
1 α β
is bounded in C∞ (Rn × Rn × Rn ) by a seminorm of bN , hence the same holds

for the pull back by the map (z, z ) → (z, z , z − z ). Since N is arbitrary, we
can take arbitrary α, β, γ and deduce that
sup |z−
1 z −
2 (z − z )γ (Dαz Dz KI(b) )(z, z )| < ∞.
β
Using (5.24) and that γ is arbitrary, we deduce that

sup |z−
1 −
2 (z − z )γ Dαz Dz KI(b) | < ∞.
β
(5.38)
Since we want KI(c) = KI(b) , we need
(F2−1 c)(z, z − z ) = KI(b) (z, z ),
that is with w = z − z ,
(F2−1 c)(z, w) = KI(b) (z, z − w).
Now, a linear change of variables for KI(b) gives that
sup |z−
1 −
2 wγ (Dαz Dβw F2−1 c)(z, w)| < ∞,
so z−
1 −
2 Dαz F2−1 c is Schwartz in w, uniformly in z, and thus z−
1 −
2
−∞,
1 +
2
Dαz c is Schwartz in the second variable, ζ , uniformly in z, that is c ∈ S∞,δ .
This also shows that any seminorm of c depends only on the seminorms of bN
for some N, and does so continuously, and thus depends on b continuously.
m,
1 ,
2
The argument in the case of Sδ,δ is completely analogous, but now even
z−
1 z −
2 (z )μ (z − z )γ Dαz Dz KI(bN ) (z, z )
β

= .−
1 −
2 μ γ α β −1
1 .2 M2 M3 (D1 + D3 ) (D2 − D3 ) (F3 bN ) (z, z , z − z )

−1 −
1 −
2 μ γ
= F3 .1 .2 M2 D3 (D1 + M3 )α (D2 − M3 )β bN (z, z , z − z ),
with the result that
sup |z−
1 z −
2 (z )μ (z − z )γ (Dαz Dz KI(b) )(z, z )| < ∞.
β
Using (5.24) and that γ , μ are arbitrary, we deduce that

sup |(z )μ (z − z )γ Dαz Dz KI(b) | < ∞.
β
This gives KI(b) ∈ S(R2n ), and the argument is finished as before. This
completes the proof of Lemma 5.5.
As already mentioned, this completes the proof of Proposition 5.1.

−∞,
As a corollary of the lemma, we note that elements of ∞,δ have a C ∞

∞
Schwartz kernel, of the form C (Rz ; S(Rz )), and thus give continuous linear
n n
maps S → C ∞ (Rn ), that is are smoothing. Note that this does not mean
decay at infinity. On the other hand, elements of −∞,−∞ are completely
regularizing, as their Schwartz kernel is in S(R2n ), and thus they give maps
S → S. Note that maps S → S are actually compact on all polynomially
weighted Sobolev spaces H r,s .
m,
m,
m,
The isomorphism qL : S∞,δ → ∞,δ can be used to topologize ∞,δ . Since

−1 −1
qR ◦ qL , qL ◦ qR are continuous, this is the same topology as that induced
by qR .
5.3.4. The Principal Symbol

m−1+2δ,
1 +
2
Note that if a ∈ S∞,δ m,
1 ,
2
then ι∗ a − aL , ι∗ a − aR ∈ S∞,δ , while if
m,
1 ,
2 ∗ ∗ m−1+2δ,
1 +
2 −1+2δ
a ∈ Sδ,δ then ι a − aL , ι a − aR ∈ Sδ,δ . We thus make the
following definition:
m,
Definition 5.2 The principal symbol σ∞,m,

(qL (a)) in ∞,δ of qL (a), a ∈
m,
m,
m−1+2δ,
S∞,δ , is the equivalence class [a]∞ of a in S∞,δ /S∞,δ .

m,
m,
The joint principal symbol σm,

(qL (a)) in δ,δ of qL (a), a ∈ Sδ,δ , is the
m,
m−1+2δ,
−1+2δ
equivalence class [a] of a in Sδ,δ /Sδ,δ .
In case the orders are variable, the principal symbols
σ∞,m,l (qL (a)), resp. σm,l (qL (a)),

m,l m−1+2δ,l m,l m−1+2δ,l−1+2δ
are defined analogously in S∞,δ /S∞,δ , resp. Sδ,δ /Sδ,δ .
Thus, the principal symbol also satisfies
σ∞,m,
(qR (a)) = [a]∞ , σm,
(qR (a)) = [a],
with analogues for variable orders.

For a ∈ C ∞ (Rn × Rn ) ⊂ S0,0 , there is a natural identification of the
equivalence class, namely the restriction of a to ∂(Rn × Rn ) can be identified
with its equivalence class, namely changing a by any element of C ∞ (Rn × Rn )
which vanishes on the boundary, and thus being in S−1,−1 does not affect the
equivalence class, so the map a → [a] descends to a|∂(Rn ×Rn ) → [a], and
the result is injective. Note that Rn × Rn is a manifold with corners with two
258 A. Vasy
boundary hypersurfaces, ∂Rn × Rn and Rn × ∂Rn , so equivalently one can

restrict to each of these separately, and keep in mind that the restrictions must
agree at the corner, ∂Rn × ∂Rn ; see Figure 5.3. The restrictions to these two
hypersurfaces are denoted by
σfiber,0,0 (qL (a)) = a|Rn ×∂Rn
and
σbase,0,0 (qL (a)) = a|∂Rn ×Rn ,
with the subscript indicating whether we are considering the part of σ0,0 at
“fiber infinity,” that is as |ζ | → ∞, or “base infinity,” that is as |z| → ∞.
In the case of σ∞ , a common way of understanding it is in terms of the
R+ -action by dilations on the second factor of Rn × (Rn \ {0}):
R+ × Rn × (Rn \ {0}) (t, z, ζ ) → (z, tζ ) ∈ Rn × (Rn \ {0}).
The quotient of Rn \ {0} by the R+ action can be identified with the unit
sphere Sn−1 : every orbit of the R+ -action intersects the sphere at exactly
one point. A different identification of this quotient (which is actually more
relevant from the perspective of where our analysis actually takes place) is
the sphere at infinity, ∂Rn . Thus, homogeneous degree zero C ∞ functions on
Rn × (Rn \ {0}) are identified with either C ∞ (Rn × Sn−1 ) or C ∞ (Rn × ∂Rn ).
So one can correspondingly identify the principal symbol of A = qL (aL ),
aL ∈ C ∞ (Rn × Rn ), as a function on Rn × Sn−1 , or instead as a homogeneous
degree zero function on Rn × (Rn \ {0}).
Returning to σ , for
a = z
ζ m ã, ã ∈ C ∞ (Rn × Rn ),
one cannot simply restrict a to the boundary, though as (given

and m) a and
ã are in a bijective correspondence, one could restrict ã and call it the principal
symbol, that is the actual principal symbol, as we defined it, is given by any
C ∞ extension of this restriction times z
ζ m . In a more geometric context
this is not quite natural (it depends on the differentials of choices of boundary
defining functions, here z−1 and ζ −1 , at the boundary). Taking
= 0, as
it is the most common case, in terms of the R+ action on the second factor, it
is more common then to view the part of the principal symbol corresponding
to Rn × ∂Rn as a homogeneous degree m function on Rn × (Rn \ {0}). In
terms of ã and its identification with a homogeneous degree zero function on
Rn × (Rn \ {0}), the part of the principal symbol corresponding to Rn × ∂Rn is
σfiber, m,0 (A) = |ζ |m ã.
On the other hand, the part of the principal symbol corresponding to ∂Rn × Rn
can be described by simply restricting it to ∂Rn × Rn , with the result being
symbolic in the second variable:
σbase,m,0 (A) = ζ m ã|∂Rn ×Rn .

Concretely, if A is a differential operator, A = aα Dα , aα ∈ C ∞ (Rn ), then
the two parts of the principal symbol under this identification are

σfiber,m,0 (A)(z, ζ ) = aα (z)ζ α , (z, ζ ) ∈ Rn × (Rn \ {0}), (5.39)
|α|=m
and

σbase,m,0 (A)(z, ζ ) = aα (z)ζ α , (z, ζ ) ∈ ∂Rn × Rn . (5.40)
|α|≤m
As an example, if g is a Riemannian metric on Rn with gij ∈ C ∞ (Rn ), then

for V ∈ z−1 C ∞ (Rn ),
H = g + V − σ (5.41)
has principal symbol in these two senses given by

σfiber,2,0 = gij ζi ζj , σbase,2,0 = gij ζi ζj − σ .
In the case of σ (as opposed to σ∞ ), one could apply a similar construction

for the restriction of the symbol of A = qL (aL ) to ∂Rn × Rn ; it is then either
a homogeneous degree zero function on (Rn \ {0}) × Rn where the action is
in the first factor, or a function on Sn−1 × Rn ; the last version would be rarely
considered. Thus, two different point of views would be needed for describing
σ in terms of homogeneous functions, which is the reason for this being a less
useful point of view in this case than in that of σ∞ .
That the principal symbol captures the leading order behavior of pseudo-
differential operators is given in the following proposition.
Proposition 5.2 The sequences
m−1+2δ,
m,
m,
m−1+2δ,
0 → ∞,δ → ∞,δ → S∞,δ /S∞,δ → 0,
resp.
m−1+2δ,
−1+2δ m,
m,
m−1+2δ,
−1+2δ
0 → δ,δ → δ,δ → Sδ,δ /Sδ,δ → 0,
are short exact sequences of topological vector spaces.
260 A. Vasy
m−1+2δ,
−1+2δ m,

Here ι : δ,δ → δ,δ is the inclusion map and
m,
m,
m−1+2δ,
−1+2δ
σm,
: δ,δ → Sδ,δ /Sδ,δ
is the principal symbol map, with analogous definitions in the case of ∞,δ .
The analogous statements also hold if m = m,
= l are variable.
This is essentially tautological, given the short exact sequence
m−1+2δ,
−1+2δ m,
m,
m−1+2δ,
−1+2δ
0 → Sδ,δ → Sδ,δ → Sδ,δ /Sδ,δ → 0,

and the isomorphisms qL, m ,
: Sδ,δ m ,
→ δ,δm ,
with m = m, m − 1 + 2δ,

=
,
− 1 + 2δ , and that these are consistent with the inclusion ιS :
m−1+2δ,
−1+2δ m,
Sδ,δ → Sδ,δ , that is one has a commutative diagram qL,m,

◦ ιS =
ι ◦ qL,m−1+2δ,
−1+2δ .

5.3.5. The Operator Wavefront Set

We also define operator wavefront sets, for which variable orders are irrelevant.
We first start with the microsupport of symbols:
m,
Definition 5.3 Suppose a ∈ Sδ,δ (R ; R ). We say that α ∈ ∂(R ×R ) is not in

n n n n
esssupp(a) if there is a neighborhood U of α in R × R such that a|U∩(Rn ×Rn )

n n
is S = S−∞,−∞ (i.e. satisfies Schwartz estimates in U).

m,
Similarly, for a ∈ S∞,δ (Rn ; Rn ) we say that α ∈ Rn × ∂Rn is not in

esssupp∞,
(a) if there is a neighborhood U of α in Rn × Rn such that
−∞,
a|U∩(Rn ×Rn ) is S∞,δ (i.e. satisfies the corresponding symbol estimates in U).
In either case, esssupp is called the microsupport, or essential support, of a.
Now for operators we define the wavefront set in terms of the microsupport
of their left amplitudes aL .
m,
Definition 5.4 Suppose that A ∈ δ,δ , A = qL (aL ). We write
WF (A) = esssupp(a),
that is we say that α ∈ ∂(Rn × Rn ) is not in WF (A), the wavefront set of A,
if there is a neighborhood U of α in Rn × Rn such that aL |U∩(Rn ×Rn ) is S =
S−∞,−∞ (i.e. satisfies Schwartz estimates in U).
m,
Similarly, for A ∈ ∞,δ , we write WF ∞,

(A) = esssupp∞,
(A).
Note that directly from the definition, the complement of esssupp, and thus
the wavefront set, is open, that is the wavefront set itself is closed. Further,
even for WF ∞,
,
is only relevant for α ∈ ∂Rn × ∂Rn ; one commonly simply
writes WF ∞ , or indeed WF . While the principal symbol captures the leading
order behavior of a pseudodifferential operator, the (complement of the) wave

front set captures where it is (not) “trivial.”
As an example, if a ∈ C ∞ (Rn × Rn ), A = qL (a), then WF (A) ⊂ supp a ∩
∂(Rn × Rn ), since certainly in the complement of supp a, a vanishes, and is
thus a symbol of order −∞, −∞. However, notice that the containment is not
an equality, as for example a ∈ S(R2n ) which never vanishes on R2n (e.g.
a Gaussian) has support everywhere, but WF (qL (a)) = ∅. Thus, the more
precise statement is that α ∈/ WF (A) for such a, A, if α has a neighborhood U
in ∂(R × R ) on which the full Taylor series of a vanishes.
n n
Again, as in the case of the principal symbol, one could consider WF ∞,

a
subset of Rn × (Rn \ {0}) which is invariant under the R+ -action (dilations
in the second factor), that is which is conic; this is the standard point of view.
The corresponding statement for WF is, as in the case of the principal symbol,
more awkward, and is thus less common.
In view of Proposition 5.1, one could also use aR with A = qR (aR ) in place of
aL in the definition. Also, as ∂(Rn × Rn ) and Rn × ∂Rn are compact, so symbol
estimates corresponding to an open cover imply symbol estimates everywhere;
we have:
m,
Lemma 5.6 If A ∈ δ,δ −∞,−∞ .

and WF (A) = ∅, then A ∈
−∞,
m,
If A ∈ ∞,δ and WF ∞,
(A) = ∅, then A ∈ ∞ .
The analogues also hold in variable order spaces.
We also have from (5.21) that
m,
Proposition 5.3 If A ∈ ∞,δ then A∗ ∈ ∞,δ

m,
and
σ∞,m,
(A∗ ) = σ∞,m,
(A), WF ∞ (A∗ ) = WF ∞ (A).
m,
If A ∈ δ,δ ∗ m,
then A ∈ δ,δ and
σm,
(A∗ ) = σm,
(A), WF (A∗ ) = WF (A).

We can also strengthen the surjectivity part of Proposition 5.2:
m,
m,
Proposition 5.4 For a ∈ S∞,δ there exists A ∈ ∞,δ with σ∞,m,

(A) = [a] and

WF∞ (A) ⊂ esssupp∞ a.
m,
m,
Similarly, for a ∈ Sδ,δ there exists A ∈ δ,δ with σm,

(A) = [a] and
WF (A) ⊂ esssupp a.
Indeed, taking A = qL (a) or A = qR (a) will do the job.
262 A. Vasy
5.3.6. Composition and Commutators

The most important part of a treatment of pseudodifferential operators is their
properties under composition and commutators:
m,
m ,

m+m ,
+

Proposition 5.5 If A ∈ ∞,δ , B ∈ ∞,δ , then AB ∈ ∞,δ ,
σ∞,m+m ,
+
(AB) = σ∞,m,
(A)σ∞,m ,
(B),
and
WF ∞ (AB) ⊂ WF ∞ (A) ∩ WF ∞ (B).
m,
m ,
m+m ,
+

If A ∈ δ,δ , B ∈ δ,δ , then AB ∈ δ,δ , and
σm+m ,
+
(AB) = σm,
(A)σm ,
(B),
and
WF (AB) ⊂ WF (A) ∩ WF (B).
Thus, ∞ and are order-filtered ∗-algebras, and in case of ∞ , composi-
tion is commutative to leading order in terms of the differential order, m, while
in the case of , it is commutative to leading order in both the differential and
the growth orders m and
.
Proof This proposition is proved easily using Proposition 5.1, taking advan-
−∞,−∞
tage of (5.29) and (5.30). To do so, first assume a, b ∈ S∞ , then
(qL (a)qR (b)u)(z)

= (2π )−n eiζ ·z a(z, ζ ) FF −1 ζ → e−iz ·ζ b(z , ζ ) u(z ) dz dζ
R Rn
n

= (2π )−n eiζ ·(z−z ) a(z, ζ )b(z , ζ ) u(z ) dz dζ = (I(c)u)(z),
Rn Rn
with
c(z, z , ζ ) = a(z, ζ )b(z , ζ ) ∈ S∞
−∞,−∞,−∞
.
However, with c = c(a, b) so defined, the map
m,
m ,

,
,m+m
S∞,δ × S∞,δ (a, b) → c ∈ S∞,δ
is continuous, so as both trilinear maps
(a, b, u) → qL (a)qR (b)u, (a, b, u) → I(c(a, b))u
are continuous
m,
m ,

S∞,δ × S∞,δ ×S →S
for all m, m ,
,
, it follows that
qL (a)qR (b) = I(c(a, b)).

m,
Since qL , qR are isomorphisms, the closedness of ∞,δ under composition

is immediate, as is the continuity of composition. As for the principal
m ,

symbol, this statement follows since for B ∈ ∞,δ , if B = qR (b), then σ∞,m ,

m+m −1+2δ,
+

(B) = b, and then by (5.28), I(c(a, b)) = qL (cL ) with cL − ab ∈ S∞,δ .
The wavefront set statement is also immediate in view of (5.28).
In the case of , the same arguments go through, but corresponding to the
m+m −1+2δ,
+
−1+2δ
improvement in (5.28), cL − ab ∈ Sδ,δ .
Going one order farther in the asymptotic expansion of compositions, one

immediately obtains the principal symbol of the commutators. Here we recall
the Poisson bracket on Rnz × Rnζ , identified with T ∗ Rn :

n

{a, b} = (∂ζj a)(∂zj b) − (∂zj a)(∂ζj b) .
j=1
m,
m ,

m+m −1+2δ,
+

Proposition 5.6 If A ∈ ∞,δ , B ∈ ∞,δ , then [A, B] ∈ ∞,δ , and
1
σ∞,m+m −1+2δ,
+
(AB) = {σ∞,m,
(A), σ∞,m ,
(B)}.
i

m,
If A ∈ δ,δ m ,
m+m −1+2δ,
+
−1+2δ , and
, B ∈ δ,δ , then [A, B] ∈
1
σm+m −1+2δ,
+
−1+2δ (AB) = {σm,
(A), σm ,
(B)}.
i
5.3.7. Ellipticity
We now turn to the simplest consequences of the machinery we built up, such
as the parametrix construction for elliptic operators.
m,
m,
Definition 5.5 We say that A is elliptic in ∞,δ , resp. δ,δ , if [a]∞ , resp. [a],
−m,−
−m−1+2δ,−
is invertible, that is if there exists [b]∞ ∈ S∞,δ /S∞,δ , resp. [b] ∈

−m,−
−m−1+2δ,−
−1+2δ 0,0 −1+2δ,0
Sδ,δ /Sδ,δ with [a]∞ [b]∞ = [1] in S∞,δ /S∞,δ , resp.
0,0 −1+2δ,−1+2δ
[a][b] = [1] in Sδ,δ /Sδ,δ .
More generally, we make the analogous definition if m = m, l = l are
variable.
264 A. Vasy
These definitions are equivalent to the statements that there exist c > 0,
R > 0 such that
|a| ≥ cz
ζ m , c > 0, |ζ | > R, (5.42)
resp.
|a| ≥ cz
ζ m , c > 0, |ζ | + |z| > R; (5.43)
indeed, if a satisfies this, the reciprocal is easily seen to satisfy the appropriate
conditions in |ζ | > R, resp. |z| + |ζ | > R, and the multiplying by a cutoff,
identically 1 near infinity, in ζ , resp. (z, ζ ), gives b. Conversely, if b exists,
upper bounds for |b| give the desired lower bounds for |a|.

Concretely, if A = |α|≤m aα Dα as in (5.2), then under the identification of
the part of the principal symbol at Rn × ∂Rn with a homogeneous degree m
function on Rn ×(Rn \{0}), while identifying the principal symbol at ∂Rn ×Rn
as an mth order symbol on ∂Rn × Rn , ellipticity means:

z ∈ Rn , ζ = 0 ⇒ aα ζ α = 0,
|α|=m
and

z ∈ ∂Rn , ζ ∈ Rn ⇒ aα ζ α = 0.
|α|≤m
For H = g + V − σ as in (5.41), ellipticity means

(z, ζ ) ∈ Rn × (Rn \ {0}), ζ = 0 ⇒ gij (z)ζi ζj = 0,
(5.44)
(z, ζ ) ∈ ∂Rn × Rn ⇒ gij ζi ζj − σ = 0.
Now the first is just the statement that g is a Riemannian metric on Rn in the
uniform sense we discussed; the second holds if and only if σ ∈ / [0, ∞). Note
that if V ∈ S−ρ (Rn ) instead, ρ ∈ (0, 1), then V does affect the principal symbol
in the second sense, but it does not affect ellipticity.
m,
If A is elliptic in δ,δ (with the variable order case going through without
changes), say, then one can construct a parametrix B with a residual, or
−m,−
completely regularizing, error, that is B ∈ δ,δ such that
AB − Id, BA − Id ∈ −∞,−∞ .
Indeed, one takes any B0 with σ−m,−

(B0 ) being the inverse for σm,
(A), so
σ0,0 (AB0 − Id) = σm,

(A)σ−m,−
(B0 ) − 1 = 0,
−1+2δ,−1+2δ
thus E0 = AB0 − Id ∈ δ,δ . Now, AB0 = Id +E0 , so one wants to
invert Id +E0 approximately; this can be done by a finite Neumann series,
j j
Id + N j=1 (−1) E0 , then
⎛ ⎞

N
−(1−2δ)(N+1),−(1−2δ )(N+1)
(Id +E0 ) ⎝Id + (−1)j E0 ⎠ − Id
j
∈ δ,δ .
j=1
j
This can be improved by writing E0 = qL (ej ), then computing the asymptotic
sum
∞
−1+2δ,−1+2δ
ẽ ∼ (−1)j ej ∈ Sδ,δ ,
j=1
taking Ẽ = qL (ẽ), (Id +E0 )(Id +Ẽ) − Id ∈ −∞,−∞ , so B = B0 (Id +Ẽ)

provides a right parametrix: E = AB − Id ∈ −∞,−∞ . A left parametrix B
can be constructed similarly, and the standard identities showing the identity of
left and right inverses in a semigroup, as applied to the quotient by completely
regularizing operators, shows that B − B ∈ −∞,−∞ , so one may simply
replace B by B. Indeed, if B A = Id +E ,
B = B (AB − E) = (B A)B − B E = B − E B − B E,
(5.45)
B E, EB ∈ −∞,−∞ .
Notice that all of the constructions can be done uniformly as long as (5.43)
is satisfied for a fixed c and R, that is one can construct the maps A → B, E
−m,−
such that they are continuous from the set of elliptic operators to δ,δ resp.
−∞,−∞ .
m,
If A ∈ ∞,δ then the same argument only gains in the first order, m, so one
−m,−
−∞,0
obtains a parametrix B ∈ ∞,δ with errors E, E ∈ ∞ .
We have thus proved:
m,
−m,−
Proposition 5.7 If A ∈ δ,δ is elliptic then there exists B ∈ δ,δ such

−m,−
that AB − Id, BA − Id ∈ −∞,−∞ . Further, the maps A → B ∈ δ,δ and

A → AB − Id, BA − Id ∈ −∞,−∞ can be taken to be continuous from the set
m,
m,
m,
of elliptic operators in δ,δ (an open subset of δ,δ ), equipped with the δ,δ
topology.
m,
−m,−
If A ∈ ∞,δ is elliptic then there exists B ∈ ∞,δ such that AB − Id, BA −

−∞,0 −m,−
Id ∈ ∞,δ . Again, the maps A → B ∈ ∞,δ and A → AB − Id, BA − Id ∈

−∞,0 m,
∞,δ can be taken to be continuous from the set of elliptic operators in ∞,δ .
The analogous variable order statements also hold.
266 A. Vasy
m,
If A ∈ δ,δ elliptic is invertible in the weak sense that there exist G : S →

S continuous such that GA = Id : S → S and AG = Id : S → S then (i.e.
the left hand side, which a priori maps into S , actually maps into S with the
claimed equality), with B a parametrix for A, BA − Id = EL , AB − Id = ER ,
G = G(AB − ER ) = B − GER = B − (BA − EL )GER = B − BER + EL GER ,

−m,−
with the first two terms on the right in δ,δ , resp. −∞,−∞ , and the last
term is residual as well since it is a continuous linear map S → S, and thus
has Schwartz kernel in S(R2n ), thus is in −∞,−∞ . Hence G ∈ −m,−
, and
G − B ∈ −∞,−∞ . Thus, the inverses of actually invertible elliptic operators
are pseudodifferential operators themselves.
As a corollary we have elliptic regularity:
m,
m,l
Proposition 5.8 If A ∈ δ,δ (or more generally A ∈ δ,δ ) is elliptic and

Au ∈ S for some u ∈ S then u ∈ S.
Proof Let B be a parametrix for A with BA − Id = E ∈ −∞,−∞ . Then
u = Id u = (BA − E)u = B(Au) − Eu,
and Eu ∈ S as E is completely regularizing while Au ∈ S by assumption,

hence B(Au) ∈ S as well.
5.3.8. L2 and Sobolev Boundedness

We can now discuss Hörmander’s proof of L2 -boundedness of elements of
0,0 0,0
δ,δ , or indeed ∞,δ , via a square root construction.
Lemma 5.7 Suppose that A ∈ ∞,δ 0,0

is elliptic and symmetric (A∗ = A) with
principal symbol that has a positive (bounded below by a positive constant)
0,0
representative a. Then there exists B ∈ ∞,δ such that B is symmetric and
−∞,0 0,0
A = B2 + E with E ∈ ∞ . The maps A → B ∈ ∞,δ and A → E ∈
−∞,0
∞ can be taken continuously from the set of A satisfying these constraints
0,0
(equipped with the ∞,δ topology).
The same result holds with the (∞, δ) subscript replaced by (δ, δ ), but with
E ∈ −∞,−∞ .
√ 0,0 0,0
Proof Let b0 = a; one easily checks that b0 ∈ S∞ . Let B̃0 ∈ ∞,δ have
principal symbol b0 , and let B0 = 12 (B̃0 + B̃∗0 ), so B0 still has principal symbol
b0 and is symmetric. Then A − B20 has vanishing principal symbol, so E0 =
−1+2δ,0
A − B20 ∈ ∞,δ , providing the first step in the construction.
0,0
In general, for j ∈ N, suppose one has found Bj ∈ ∞,δ to be symmetric
−(1−2δ)(j+1),0
such that Ej = A − B2j ∈ ∞,δ ; we have shown this for j = 0. Let
−(1−2δ)(j+1),0
ej be the principal symbol of Ej , and let bj+1 = − 2b10 ej ∈ S∞,δ ;
−(1−2δ)(j+1),0
this uses b0 elliptic. Let B̃j+1 ∈ ∞,δ have principal symbol bj+1 ,
∗
Bj+1 = 2 (B̃j+1 + B̃j+1 ), Bj+1 = Bj + Bj+1 , so Bj+1 is symmetric. Further, the
1
principal symbol of
A − B2j+1 = A − (Bj + B j+1 )2 = A − B2j − Bj B j+1 − B j+1 Bj − (B j+1 )2
−(1−2δ)(j+1),0
= Ej − Bj B j+1 − B j+1 Bj − (B j+1 )2 ∈ ∞,δ
−(1−2δ)(j+2),0
is ej − 2b0 bj+1 = 0, so Ej+1 = A − B2j+1 ∈ ∞,δ , providing the
inductive steps. One can finish up by asymptotically summing, as in the elliptic
case.
0,0
Proposition 5.9 Elements A ∈ ∞,δ are bounded on L2 .
Further, if a is a representative of σ∞,0,0 (A) and C > infr∈S−1+2δ,0 sup |a + r|
∞,δ
−∞,0
then there exists E ∈ ∞ such that
AuL2 ≤ CuL2 + |Eu, u|.

−∞,0
Moreover, the map A → E ∈ ∞ can be taken to be continuous, and thus
0,0
the inclusion ∞,δ → L(L ) is continuous.
2
−∞,0
Proof We reduce the proof to the boundedness of elements of ∞ on L2 ,
which is an easy consequence of Schur’s lemma since by (5.38) the Schwartz
kernel of elements of this space is a bounded continuous function in z with
values in S(Rnz ) (hence with values in L1 (Rnz )), and similarly with z and z
interchanged.
0,0
Now, suppose that A ∈ ∞,δ , so its principal symbol has a bounded repre-
0,0
sentative a; let M > sup |a|. Then M 2 − |a|2 ∈ S∞,δ is bounded below by a
0,0
positive constant and is thus elliptic. By Lemma 5.7, there exists B ∈ ∞,δ
−∞,0
symmetric such that M 2 − A∗ A = B2 + E, E ∈ ∞ . Then, first for u ∈ S,
with inner products and norms the standard L2 ones,
M 2 u, u = Au2 + Bu2 + Eu, u,
that is with EL(L2 ) the L2 bound of E, which is finite as discussed above,
Au2 ≤ M 2 u2 + EL(L2 ) u2 .
Since S is dense in L2 , this implies that A has a unique continuous extension

to L2 ; one still denotes it by A. Since S is also dense in S , and the inclusion
268 A. Vasy
L2 → S is continuous, this extension is the restriction of A acting on S . This

proves the first part of the proposition.
−1+2δ,0
For the second part we simply replace a by a + r, choosing r ∈ S∞,δ
such that C > sup |a + r|, then we can take M = C in the argument above to
complete the proof.
0,0 0,0
are in ∞,δ for δ = 0 and are thus L -bounded, it is
While elements of δ,δ 2
useful to make the bound more explicit there as well, in addition to generalizing
to δ > 0:
0,0
Proposition 5.10 Elements A ∈ δ,δ 2
are bounded on L .
Further, if a is a representative of σ0,0 (A) and C > infr∈S−1+2δ,−1+2δ
δ,δ
sup |a + r| then there exists E ∈ −∞,−∞ such that
AuL2 ≤ CuL2 + |Eu, u|. (5.46)
Moreover, the map A → E ∈ −∞,−∞ can be taken to be continuous.
Concretely, if A = qL (a) with a ∈ C ∞ (Rn × Rn ), then for any

C > sup a|∂(Rn ×Rn ) ,
(5.46) holds.
0,0
Proof This is the same argument as above, but constructing B in δ,δ .
We now recall that the weighted Sobolev spaces are

H s,r = {u ∈ S : zr u ∈ H s }, uH s,r = zr uH s . (5.47)
Further, with
s = F −1 ζ s F ∈ s,0 ⊂ ∞
s,0
,
the standard Sobolev spaces are
H s = {u : s u ∈ L2 } with uH s = s uL2 .
We note here that
∪M,N∈R H M,N = S .
Thus, s,r = s zr : H s,r → L2 is an isometry, with inverse −s,−r =
z−r −s : L2 → H s,r . Hence, the boundedness of some A ∈ ∞,δ m,
as a map

H s,r → H s ,r is equivalent to the boundedness on L2 of s ,r A −s,−r as
A = −s ,−r (s ,r A −s,−r )s,r .

m+s −s,
+r −r
But s ,r A −s,−r ∈ ∞,δ , so we conclude that

m,
Proposition 5.11 An operator A ∈ ∞,δ is bounded H s,r → H s ,r if m = s−s

and
= r − r (thus if m ≤ s − s and
≤ r − r ).

This gives a quantified version of elliptic regularity:

Proposition 5.12 If A ∈ δ,δ m,
s,r for some u ∈ S then
is elliptic and Au ∈ H
u ∈ H s+m,r+
. In fact, for any M, N there is C > 0 such that
uH s+m,r+
≤ C(AuH s,r + uH M,N ).
m,
If A ∈ ∞,δ is elliptic and Au ∈ H s,r for some u ∈ H k,r+

, k ∈ R, then
u∈H s+m,r+
. In fact, for any k there is C > 0 such that
uH s+m,r+
≤ C(AuH s,r + uH k,r+
).
The point of the quantitative estimate is to allow M, N to be very negative,

so for example H s+m,r+
→ H M,N is compact. One thinks of uH M,N as a
“trivial” term correspondingly.
m,
In the case of ∞,δ ellipticity is too weak a notion to gain decay at infinity;
one simply has a uniform gain of Sobolev regularity.
m,
−m,−
Proof Suppose A ∈ δ,δ . Let B ∈ δ,δ be a parametrix for A with BA −

Id = E ∈ −∞,−∞ . Then
u = Id u = (BA + E)u = B(Au) + Eu,
and Eu ∈ S while Au ∈ H s,r by assumption, hence B(Au) ∈ H s+m,r+

, as
claimed. The bound in the proposition follows from E : H M,N → H s+m,r+
being bounded.
m,
−m,−
−∞,0
If A ∈ ∞,δ , and B ∈ ∞,δ is a parametrix, so BA − Id = E ∈ ∞
then the same argument gives, using E : H k,r+
→ H s+m,r+
bounded, the
conclusion that u ∈ H s+m,r+
, as well as the estimate.
An immediate corollary is:

m,
Proposition 5.13 Any elliptic A ∈ δ,δ is Fredholm as a map H

s,r →
H s−m,r−
for all m,
, s, r ∈ R, that is has a closed range, finite dimensional
nullspace and the range has finite codimension. Further, the nullspace is a
subspace of S, while the annihilator of the range in H s−m,r−
in the dual space
H −s+m,−r+
is also in S. Correspondingly, the nullspace of A as well as the
annihilator of its range is independent of r, s; if A is invertible for one value of
r, s, then it is invertible for all.
Proof If B is a parametrix for A, then B ∈ L(H s−m,r−
, H s,r ) and EL = BA −
Id, ER = AB − Id ∈ −∞,∞ . Thus EL , ER map H s,r , resp. H s−m,r−
, to S
270 A. Vasy
continuously, and are thus compact as maps in L(H s,r ), resp. L(H s−m,r−
).
Then standard arguments give the Fredholm property.
The property of the nullspace being in S is elliptic regularity. If v is in the
annihilator as stated, that is v, Au = 0 for all u ∈ H s,r then A∗ v, u = 0
for all u ∈ H s,r , so A∗ v = 0 in H −s,−r . As A∗ has principal symbol ā, elliptic
regularity shows that v ∈ S.
m,
Corollary 5.2 Suppose m,

> 0, A ∈ δ,δ 2
is symmetric on L and is elliptic.
Then A is self-adjoint with domain H m,

.
Proof It suffices to show that A − σ : H m,
→ L2 are invertible for σ ∈
C \ R. As m,
> 0, these are elliptic regardless of σ , thus Fredholm as stated,
with nullspace and cokernel being identified as the kernel of A∗ , in S. But
the symmetry of A shows that for u in the kernel, 0 = Im(A − σ )u, u =
− Im σ u2 , so u = 0, hence the kernel is trivial. Thus, the kernel of A∗ = A
is also trivial, so A is surjective, thus the desired invertibility follows.
m,
Corollary 5.3 Suppose m ≥ 0,

≥ 0, A ∈ δ,δ 2
is symmetric on L and
σfiber,m.
(A), resp. σbase,m,
(A), is elliptic if m > 0, resp.
> 0. Then A is
self-adjoint with domain H m,
.
Proof We have already dealt with m,
> 0; m,
= 0 is standard, so it remains
to deal with m > 0,
= 0 as m = 0,
> 0 is similar. Again, it suffices to show
that A − σ : H m,
→ L2 are invertible for σ ∈ C \ R. The principal symbol
has a real representative a (simply take the real part of any representative) and
by ellipticity at fiber infinity there exist c0 , R > 0 such that |a| ≥ c0 |ζ |m if
|ζ | > R. We claim that
|a − σ |2 = |a − Re σ |2 + | Im σ |2 ≥ cζ 2m , c > 0.
Indeed, for |a| ≥ 2| Re σ |, |a − Re σ |2 ≥ (|a| − | Re σ |)2 ≥ |a|2 /4, so for

|ζ | ≥ R with c0 |ζ |m > 2| Re σ | the inequality follows. On the other hand,
otherwise |ζ | ≤ max(R, (2c−1 0 | Re σ |)
1/m ), so ζ is bounded, and then the Im σ
term gives the desired inequality. So A − σ is elliptic when Im σ = 0, thus

Fredholm as stated, with nullspace and cokernel in S. Again, the symmetry of
A shows that for u in the kernel, 0 = Im(A − σ )u, u = − Im σ u2 , so u = 0,
hence the kernel of A − σ is trivial. Thus, the kernel of A∗ = A is also trivial,
so A is surjective, thus the desired invertibility follows.
We summarize our results so far for the Schrödinger operators:

Proposition 5.14 Let g be a Riemannian metric, gij ∈ C ∞ (Rn ), positive
definite on Rn , V ∈ S−ρ (Rn ) with ρ > 0. Let H = g + V. Then for
σ ∈ C \ [0, ∞), H − σ : H s,r → H s−2,r is Fredholm for all r, s, with nullspace

in S. If V is real-valued, then H is self-adjoint.
5.3.9. Variable Order Sobolev Spaces

We can now define variable order Sobolev spaces.
m,l
Definition 5.6 Let A ∈ δ,δ be elliptic, m ≥ m, l ≥
. Let H
m,l be a subspace
of H m,
given by
H m,l = {u ∈ H m,
: Au ∈ L2 },
with norm
u2H m,l = u2H m,
+ Au2L2 .
Then H m,l is easily seen to be a complete space, thus a Hilbert space, which

in the case of m, l being constant equal to m ,
, simply gives H m ,
. Indeed,
if {uj }∞ m,l
j=1 is Cauchy in H , then it is Cauchy in H
m,
, so it converges to some
u ∈ H m,
; in addition Auj is Cauchy in L2 so converges to some v ∈ L2 . But
A : S → S is continuous, so Auj → Au in S , so v = Au ∈ L2 , thus u ∈ H m,l .
Further, as Auj → Au in L2 , the completeness of H m,l follows.
Moreover, different choices of both A and (m,
) are equivalent in the sense
m,l
that they give the same space with equivalent norms: if Ã ∈ δ,δ is elliptic as
−m,−l −∞,−∞
well, writing B ∈ δ,δ as a parametrix, with E = BA − Id ∈ δ,δ ,
Ãu = Ã(BA) − ÃEu = (ÃB)Au − (ÃE)u

0,0 −∞,−∞
with ÃB ∈ ÃE ∈ δ,δ
δ,δ , , we deduce that Ãu ∈ L2 , and Ãu2 ≤
C(uH m,
+ AuL2 ), showing that the Ã-based norm is bounded by the
2 2
A-based norm. A similar argument gives the converse estimate, thus the
equivalence of norms.
We conclude:
m,l
Proposition 5.15 An operator A ∈ δ,δ is bounded H
s,r → H s ,r if

m = s − s and l = r − r (thus if m ≤ s − s and l ≤ r − r ).
m,l
Proof Let s, r be such that s ≤ s, r ≤ r and m ≥ m,
≥ l. Such an A ∈ δ,δ ⊂

m,
s,r to H s−m,r−
continuously. Further, if Ã ∈ s,r , Ã ∈ s ,r are
δ,δ maps H δ,δ δ,δ
−s,−r −∞,−∞
elliptic, then with B̃ ∈ δ,δ , B̃Ã − Id = Ẽ ∈ δ,δ , then
Ã Au = (Ã AB̃)Ãu − (Ã AẼ)u,

−∞,−∞
with Ã AB̃ ∈ δ,δ
0,0
and Ã AẼ ∈ δ,δ , thus bounded on L2 , giving the
conclusion.
272 A. Vasy
One then has a Fredholm and a self-adjointness statement as above for the
variable order setting.
5.3.10. Microlocalization
The elliptic parametrix construction can be microlocalized, that is if the
principal symbol of A is only assumed to be elliptic on (hence near) a closed
subset K of ∂(Rn × Rn ), one still can construct a microlocal parametrix B, that
is one whose errors BA−Id, AB−Id as a parametrix have wavefront set disjoint
from K. To make this precise, first we define microlocal ellipticity:
m,
Definition 5.7 We say that A ∈ δ,δ , σm,

(A) = [a], is elliptic at α ∈ ∂(R ×
n
R ) if α has a neighborhood U in R × R such that a|U∩Rn ×Rn is elliptic, that

n n n
is satisfies (5.43) on U. We say that A is elliptic on a subset K of ∂(Rn × Rn )

if it is elliptic at each point of K. The elliptic set Ell(A) is the set of points at
which A is elliptic; the characteristic set Char(A) is its complement.
m,
We say that A ∈ ∞,δ , σ∞,m,

(A) = [a], is elliptic at α ∈ Rn × ∂Rn if α has
a neighborhood U in Rn × Rn such that a|U∩Rn ×Rn is elliptic, that is satisfies
(5.42) on U. We say that A is elliptic on a subset K of Rn × ∂Rn if it is elliptic
at each point of K. One defines Ell∞ (A), Char∞ (A) analogously to the above
definition.
We also make the analogous definitions if m = m,
= l are variable.
m,
If A ∈ δ,δ is elliptic on a closed (hence compact) K, then a covering

argument shows that a satisfies (5.43) on a neighborhood of K. A similar
m,
statement holds for A ∈ ∞,δ .

m,
m,l
Proposition 5.16 If A ∈ δ,δ (or A ∈ δ,δ ) is elliptic on a compact set K
−m,−
−m,−l
then there exists B ∈ δ,δ (resp. B ∈ δ,δ ) such that EL = BA − Id,
ER = AB − Id satisfy WF (EL ) ∩ K = ∅, WF (ER ) ∩ K = ∅.

Proof If A is elliptic on K, there is a neighborhood U of K in Rn × Rn

such that a|U∩Rn ×Rn is elliptic, that is satisfies (5.43) on U. We may shrink
U so that |z| + |ζ | > R on U; thus |a|U | has a positive lower bound on all
of U. Let q ∈ C ∞ (Rn × Rn ) be identically 1 near K, be supported in U,
and let Q ∈ 0,0 be given by Q = qL (q). Thus, Q has principal symbol
σ0,0 (Q) = q|∂(Rn ×Rn ) , and WF (Q) ⊂ U, WF (Id −Q) ∩ K = ∅. Now let [a]
−m,−
be the principal symbol of A, let b0 = qa−1 ∈ Sδ,δ since a is elliptic

on U. Let B0 = qL (b0 ), so σ−m,−
(B0 ) = b0 and WF (B0 ) ⊂ U. Let
q0 ∈ C ∞ (Rn × Rn ) be identically 1 near K, have disjoint support from 1 − q,
so q0 (1 − q) = 0, and let Q0 = qL (q0 ). Note that WF (Id −Q0 ) ∩ K = ∅. Then
0,0 0,0
E0,L = Q0 (B0 A − Id) ∈ δ,δ , E0,R = (AB0 − Id)Q0 ∈ δ,δ have vanishing
−1+2δ,−1+2δ
principal symbols, so E0,L , E0,R ∈ δ,δ . As in the globally elliptic
j
case, one may asymptotically sum the amplitudes eL,j of (−1)j E0,L to obtain
j j −(1−2δ)(N+1),−(1−2δ )(N+1)
ẼL such that FN = ẼL − N j=1 (−1) E0,L ∈ δ,δ for
all N. Thus,
(Id +ẼL )Q0 B0 A = (Id +ẼL )(E0,L + Id) + (Id +ẼL )(Q0 − Id)
⎛ ⎞
N
= ⎝Id + (−1)j E0,L + FN⎠(Id +E0,L ) +(Id +ẼL )(Q0 − Id)
j
j=1
N+1
= Id +(−1)N+1 E0,L + FN (Id +E0,L ) + (Id +ẼL )(Q0 − Id).
Now,
N+1 −(1−2δ)(N+1),−(1−2δ )(N+1)
(−1)N+1 E0,L + FN (Id +E0,L ) ∈ δ,δ ,
and is independent of N since when adding the identity we get
⎛ ⎞
N
⎝Id + (−1)j E0,L + FN ⎠ (Id +E0,L ) = (Id +ẼL )(Id +E0,L ),
j
j=1
so it is in −∞,−∞ , and WF ((Id +ẼL )(Q0 − Id)) ⊂ WF (Q0 − Id), which is

disjoint from K. Thus, we may take
BL = (Id +ẼL )Q0 B0
as our microlocal left parametrix, and similarly obtain a microlocal right
parametrix BR . The parametrix identity (5.45) now shows that WF (BL − BR ) ∩
K = ∅, completing the proof.
The proof of the variable order case goes through without changes.
One corollary is the following.
Corollary 5.4 Suppose u ∈ S , A ∈ δ,δ
m,
, and Au ∈ H
s,r then for Q ∈ 0,0
δ,δ

with WF (Q) ∩ Char(A) = ∅, Qu ∈ H s+m,r+
. Further, for all M, N there exists
C > 0 such that
QuH s+m,r+
≤ C(AuH s,r + uH M,N ).
There is also an analogue with variable order spaces.
Proof Let B be a microlocal parametrix for A near WF (Q). Then BA − Id = E
with WF (E) ∩ WF (Q) = ∅. Thus,
Qu = Q(BA − E)u = QB(Au) − (QE)u.
274 A. Vasy
Now, WF (QE) = WF (Q) ∩ WF (E) = ∅, so QE ∈ −∞,−∞ , and thus

−m,−
QEu ∈ S, while QB ∈ δ,δ , so the proof is finished as for global elliptic

regularity.
Here the assumption Au ∈ H s,r is too strong; it only matters that Au is

microlocally near WF (Q). That is:
Corollary 5.5 (Microlocal elliptic regularity; operator version.) Suppose u ∈
S , A ∈ δ,δ
m,
0,0 0,0
, and for some Q ∈ δ,δ , Q (Au) ∈ H . Then for Q ∈ δ,δ with
s,r
WF (Q) ⊂ Ell(A) ∩ Ell(Q ), Qu ∈ H s+m,r+

. Further, for all M, N there exists
C > 0 such that
QuH s+m,r+
≤ C(Q AuH s,r + uH M,N ).
There is again an analogue with variable order spaces.

Proof We just note that Q A is elliptic on Ell(A) ∩ Ell(Q ), so the previous
corollary is applicable.
One can restate the corollary in terms of microlocalizing the distributions

instead of adding the microlocalizers explicitly as operators.
Definition 5.8 Suppose α ∈ ∂(Rn × Rn ), u ∈ S . We say that α ∈ / WFm,

(u) if
0,0
there exists A ∈ δ,δ elliptic at α such that Au ∈ H . We say that α ∈
m,
/ WF(u)
0,0
if there exists A ∈ δ,δ elliptic at α such that Au ∈ S.
For k,
, m ∈ R, u ∈ H k,
, WFm,
∞ (u) is defined similarly: if α ∈ R ×∂R , we

n n
m,
0,0
/ WF∞ (u) if there exists A ∈ ∞,δ elliptic at α such that Au ∈ H m,
.
say α ∈
0,0
/ WF∞,
(u) if there exists A ∈ ∞,δ
We say that α ∈ elliptic at α such that
∞,
Au ∈ H .
We also make the analogous definition for variable order spaces.
Notice that a priori the notion of WFm,

(u) depends on δ, δ , but in fact the
arguments below show that it in fact has no such dependence, see Lemma 5.8.
The most important property of WF and pseudodifferential operators is
microlocality:
m,
Proposition 5.17 If A ∈ δ,δ

and u ∈ S then
WFs,r (Au) ⊂ WF (A) ∩ WFs+m,r+

(u)
and
WF(Au) ⊂ WF (A) ∩ WF(u).
The variable order version of this statement also holds.
Proof We need to show that

WFs,r (Au) ⊂ WF (A) and WFs,r (Au) ⊂ WFs+m,r+
(u).
We start with the former, which is straightforward. Suppose α ∈ / WF (A). Let

Q ∈ be elliptic at α but with WF (Q) ∩ WF (A) = ∅; one can achieve this
0,0
as WF (A) is closed, so one simply needs to take q ∈ C ∞ (Rn × Rn ) equal to

1 near α and with essential support disjoint from WF (A). Then WF (QA) ⊂
WF (Q) ∩ WF (A) = ∅, so QA ∈ −∞,−∞ , thus QAu ∈ S.
Now for the second inclusion. Suppose α ∈ / WFs+m,r+
(u). Then there exists
0,0 0,0
B ∈ δ,δ elliptic at α such that Bu ∈ H s+m,r+
. Let G ∈ δ,δ be a microlocal
parametrix for B, so GB = Id +E with α ∈
/ WF (E). Then Au = AGBu − AEu,
and AG ∈ δ,δ m,
, so AGBu ∈ H . On the other hand, α ∈

s,r / WF (AE) ⊂ WF (E).
So let Q ∈ be elliptic at α but with WF (Q) ∩ WF (E) = ∅. Then QAE ∈
0,0
−∞,−∞ , and thus

QAu = Q(AG)(Bu) − (QAE)u ∈ H s,r ,
so α ∈/ WFs,r (u), completing the proof for WFs,r (Au). The proof for WF(Au)
is analogous.
Note that the last part of the proof shows more:
Lemma 5.8 If α ∈ / WFs,r (u) then there is a neighborhood U of α such that for
all Q ∈ δ,δ with WF (Q) ⊂ U, Qu ∈ H s,r .
0,0
Further, with the same U, for all Q ∈ δ,δ m,

with WF (Q) ⊂ U, Qu ∈
H s−m,r−
.
Thus, while the wavefront set definition is a “there exists” statement, in fact
it is equivalent to a “for all” statement, namely for all Q ∈ δ,δ0,0
with WF (Q)
in a sufficiently neighborhood of α, Qu ∈ H . (The other direction is simply
s,r
because these Q include those elliptic at α.)

Also, as immediate from the proof below, one can take U to be the elliptic
0,0
set of the B ∈ δ,δ , elliptic at α, with Bu ∈ H , whose existence is guaranteed
s,r
by α ∈ s,r
/ WF (u)
0,0
/ WFs,r (u). Then there exists B ∈ δ,δ
Proof Suppose α ∈ elliptic at α such
0,0
that Bu ∈ H s,r ; let G ∈ δ,δ be a microlocal parametrix for B, so GB = Id +E
with α ∈/ WF (E). Let U be the complement of WF (E); this is a neighborhood
of α. Then for any Q ∈ δ,δ 0,0 −∞,−∞ , so Qu =
with WF (Q) ⊂ U, QE ∈
0,0
QGBu − QEu ∈ H s,r as QG ∈ δ,δ .
m,
The second statement is proved the same way, noticing that QG ∈ δ,δ
now.
276 A. Vasy
An immediate consequence is:

Lemma 5.9 If u ∈ S and WFm,
(u) = ∅ then u ∈ H m,
.
If u ∈ H k,
and WFm,
∞ (u) = ∅ then u ∈ H
m,
.

Proof Suppose u ∈ S and WFm,
(u) = ∅. Then for all α ∈ ∂(Rn × Rn ) there
exists Uα open such that for all Q ∈ 0,0 with WF (Q) ⊂ Uα , Qu ∈ H m,
.
Now
{Uα : α ∈ ∂(Rn × Rn )}
is an open cover of the compact set ∂(Rn ×Rn ), so there is a finite subcover, say
{Uαj : j = 1, . . . , N}. Let Ũαj be open in Rn ×Rn with Ũαj ∩∂(Rn ×Rn ) = Uαj .
Then, with Ũα0 = Rn × Rn ,
{Ũαj : j = 0, 1, . . . , N}

is a finite open cover of Rn × Rn . Let N qj = 1 be a subordinate partition
N j=0
of unity, and let Qj = qL (qj ). Then j=0 Qj = Id, Q0 ∈ −∞,−∞ since q0 has
compact support, while for j = 1, . . . , N, WF (Qj ) ⊂ Uαj since supp qj ⊂ Ũαj .

Thus, Qj u ∈ H m,
for all j, and thus u = Qj u ∈ H m,
as claimed.
The argument for WF∞ is analogous.
The distributional version of microlocal elliptic regularity then is:
Corollary 5.6 (Microlocal elliptic regularity; distributional version.) Suppose
u ∈ S , A ∈ δ,δ
m,
. Then
WFs+m,r+
(u) ⊂ Char(A) ∪ WFs,r (Au).
/ Char(A)∪WFs,r (Au), we need to show α ∈
Proof Suppose α ∈ / WFs+m,r+
(u).
/ WFs,r (Au) there exists Q ∈ δ,δ elliptic at α such that Q Au ∈ H s,r .
As α ∈ 0,0
Let Q ∈ δ,δ0,0
be such that WF (Q) ⊂ Ell(A) ∩ Ell(Q ), note that the set on the
right is open and includes α. Then by Corollary 5.5, Qu ∈ H s+m,r+

. Taking Q
which is in addition elliptic at α completes the proof.
The consequence of what we have proved so far for Schrödinger operators is:
Proposition 5.18 Let g be a Riemannian metric, gij ∈ C ∞ (Rn ), positive
definite on Rn , V ∈ S−ρ (Rn ) with ρ > 0. Let H = g + V. Then for
σ ∈ [0, ∞), (H − σ )u ∈ H s,r implies

WFs+2,r (u) ⊂ (z, ζ ) ∈ ∂Rn × Rn : gij (z)ζi ζj = σ .
5.3.11. Diffeomorphism Invariance

Finally we note the diffeomorphism invariance of pseudodifferential operators.
Proposition 5.19 Suppose F : O → U is a diffeomorphism between bounded
m,
open subsets O and U of Rn . Suppose A ∈ ∞,δ (Rn ), with Schwartz kernel

supported in a compact subset of U × U. Then AF = F ∗ A(F −1 )∗ ∈ ∞,δ m,
.
Furthermore, with DF(z) the Jacobian matrix of F, that is with kj entry ∂j Fk (z),
and with † denoting Rn -adjoint (i.e. j, k reversed),
WF (AF ) = {(z, ζ ) : (F(z), (DF)† (z)−1 ζ ) ∈ WF (A)},
and
σ∞,m,
(AF )(z, ζ ) = σ∞,m,
(A)(F(z), (DF)† (z)−1 ζ ).
Remark 5.1 The principal symbol here shows why we had a single parameter
δ giving the losses in ζ upon differentiation in either z or ζ : differentiation
of the principal symbol of AF in z gives rise to ζ derivatives as well as in
that of A. Thus, to have the class diffeomorphism invariant, the losses under z
derivatives have to be at least as large as those under ζ -derivatives. Thus, the
ζ -derivatives (which are the derivatives tangent to the fibers of the cotangent
bundle of Rn , thus are invariantly defined) are necessarily better (in the sense
of “no worse”) behaved regarding these losses than the z-derivatives. If one
also wants Fourier-invariance, one needs the opposite inequality as well, hence
the equality.

Remark 5.2 Notice that if one writes a covector as η dw , then its
k k k
pull-back under the map F (with F(z) = w for clarity) is k ηk (∂j Fk )(z) dzj ,
that is

ζj = (∂j Fk )(z)ηk = ((DF)† (z)η)j ,
k
so ζ = (DF)† (z)η. This means that (DF)† (z)−1 ζ dw is the pull-back of ζ dz

by F −1 , that is the wavefront set and the principal symbol are well behaved
(invariant) if we regard them as subsets of T ∗ Rn \o, resp. functions on T ∗ Rn \o:
with F : TU∗ Rn → TO∗ Rn the map induced by pull-back of covectors by F, and
similarly for (F −1 ) : TO∗ Rn → TU∗ Rn , so ((F −1 ) )∗ maps functions on TU∗ Rn
to those on TO∗ Rn , then
σ∞,m,
(AF ) = ((F −1 ) )∗ σ∞,m,
(A),
and
WF (AF ) = ((F −1 ) )−1 (WF (A)).
278 A. Vasy
Proof Let G = F −1 to simplify the notation.

First we consider the off-diagonal behavior. To do so, suppose more
generally that A : S → S continuous linear with Schwartz kernel supported
in U × U (so A need not be a pseudo-differential operator). We claim that, with
KA the Schwartz kernel of A, the Schwartz kernel KAF of AF is the (compactly
supported) tempered distribution
KAF = ((F × F)∗ KA )(πR∗ | det(DF)|), (5.48)
where πR : Rn × Rn → Rn is the projection to the second factor. Indeed,

if KA is Schwartz (i.e. just C ∞ , in view of the support) then, with AF u also
considered as a distribution in the second expression,

KAF (u ⊗ v) = (AF u)(v) = (AF u)(z) v(z) dz

= A(G u)(F(z))v(z) dz = KA (F(z), w )G∗ u(w ) v(z) dw dz
∗

= KA (F(z), w )u(G(w )) v(z) dw dz

= KA (F(z), F(z ))u(z ) v(z) | det DF(z )| dz dz,
giving the above result for KAF . Since Schwartz functions with compact
support in O × O are dense in tempered distributions supported in O × O,
and since the operations in (5.48) are continuous, the result follows for general
tempered distributions KA .
Applying this to the case of pseudodifferential operators A, which have
C ∞ Schwartz kernel away from the diagonal, we conclude that AF has C ∞
Schwartz kernel away from the diagonal. In particular, when considering the
behavior near the diagonal, it suffices to work in a suitably small neighborhood
of the diagonal.
We have from the definition of A,
AF u(z) = (A(G∗ u))(F(z))

−n
= (2π ) ei(F(z)−w )·η a(F(z), w , η)u(G(w )) dw dη.
Letting z = G(w ), the change of variables formula for the integral gives

AF u(z) = (2π )−n ei(F(z)−F(z ))·η a(F(z), F(z ), η)u(z ) | det(DF)(z )| dz dη.
This is almost of the desired form, except for the appearance of F(z) − F(z )
instead of z − z in the exponent. To deal with this, we use the easiest case of
Taylor’s theorem (which really means the fundamental theorem of calculus in

this context),

n
Fk (z) − Fk (z ) = (zj − z j )Fkj (z, z )
j=1
with
1
Fkj (z, z ) = (∂j Fk )(tz + (1 − t)z ) dt,
0
so
Fkj (z, z) = ∂j Fk (z)
is the Jacobian matrix of F. More generally, let us write
(z, z ) = (∂j Fk (z, z ))kj
for this matrix. Thus, the exponent is

n
n
n
(zj − z j )Fkj (z, z )ηk = (zj − z j )ζj ,
k=1 j=1 j=1
where

n
ζj = ζj (z, z , η) = Fkj (z, z )ηk = († (z, z )η)j .
k=1
Note that the map

(z, z , η) → (z, z , ζ (z, z , η))
is a diffeomorphism, linear in η, if (z, z ) is close to the diagonal. Indeed, since

F is a diffeomorphism, (z, z) is invertible, and thus so is (z, z ) for (z, z )
near the diagonal, so the inverse of the above map is simply
(z, z , ζ ) → (z, z , † (z, z )−1 ζ ).
Changing the variable of integration from η to ζ gives, as
|dζ | = | det((z, z ))† | |dη| = | det (z, z )| |dη|,

AF u(z) = (2π )−n ei(z−z )·ζ a(F(z), F(z ), († (z, z ))−1 ζ )u(z )
| det (z, z )|−1 | det(DF)(z )| dz dζ

= (2π )−n ei(z−z )·ζ aF (z, z , ζ )u(z ) dz dζ
280 A. Vasy
with
aF (z, z , ζ ) = a(F(z), F(z ), († (z, z ))−1 ζ )| det (z, z )|−1 | det(DF)(z )|.
Thus, checking
m,
aF ∈ S∞,δ
completes the proof. For this purpose the two determinant factors are irrelevant
as they are C ∞ . Thus, it remains to note that Dζ applied to
a(F(z), F(z ), († (z, z ))−1 ζ )
again simply gives additional smooth factors, while Dz or Dz applied can either
correspond to derivatives of a in the first or second slot, in which case they are
harmless, or in the last slot when they give a factor in ζ , but also lower the
symbolic order by 1, thus preserving the estimates.
The principal symbol statement follows from the cancelation of the determi-
nant factors when one restricts to z = z , and that († (z, z ))−1 is (DF)† (z)−1
then; this also gives the wave front set statement.
In fact, the same proof gives:
Proposition 5.20 Suppose F : O → U is a diffeomorphism between open
m,
subsets O and U of Rn . Suppose A ∈ δ,δ , with Schwartz kernel supported
in a compact subset of U × U. Then AF = F ∗ A(F −1 )∗ ∈ δ,δ m,
. Furthermore,
with DF(z) the Jacobian matrix of F, that is with kj entry ∂j Fk (z), and with †
denoting Rn -adjoint (i.e. j, k reversed),
WF (AF ) = {(z, ζ ) : (F(z), (DF)† (z)−1 ζ ) ∈ WF (A)},
and
σm,
(AF )(z, ζ ) = σ∞,m,
(A)(F(z), (DF)† (z)−1 ζ ).
The point here is that for F as stated, DF is an elliptic symbol on O of
order 0, and thus the near-diagonal argument goes through: in fact, one even
gets the invertibility of (z, z ) for (z, z ) in a conic neighborhood of the diag-
onal (as follows by working with valid coordinates on the compactification,
and noting that a neighborhood in this compactified perspective gives a conic
neighborhood without the compactification). The Schwartz kernel of pseudo-
differential operators outside such a neighborhood is Schwartz, hence the off-
diagonal piece pulls back correctly as well.
We can now use our results to analyze Fredholm problems in geometric
settings. Note that the diffeomorphism invariance lets us define δm (X) when
X is a compact manifold:
Definition 5.9 For X a compact manifold (without boundary), δ ∈ [0, 1/2),

δm (X) consists of continuous linear maps A : C ∞ (X) → C ∞ (X), whose
Schwartz kernel is C ∞ away from the diagonal in X × X and with the property
that if U is a coordinate chart with : U → Ũ ⊂ Rn a diffeomorphism then
for χ ∈ Cc∞ (U), (−1 )∗ χ Aχ ∗ ∈ δ,0
m,0
.
m,

Notice that we could have used δ,δ in the definition for any δ ∈ [0, 1/2)
m,
and
∈ R, or instead ∞,δ , without changing δm (X) since the image of
supp χ under is a compact subset of Rn .
Notice also that if U, V are both coordinate charts with : U → Ũ, : V →
Ṽ and if supp χ ⊂ U ∩ V, then the statements that (−1 )∗ χ Aχ ∗ ∈ δm,0 and
(−1 )∗ χ Aχ ∗ ∈ δm,0 are equivalent since if for instance (−1 )∗ χ Aχ ∗ ∈
δm,0 , then so is

(−1 )∗ χ Aχ ∗ = ( ◦ −1 )∗ (−1 )∗ χ Aχ ∗ ( ◦ −1 )∗
as ◦ −1 : Ũ → Ṽ is a diffeomorphism of subsets of Rn so Proposition

5.19 is applicable. Thus the “for all” statement (i.e. for all coordinate charts)
in the definition can be replaced by an open cover and a subordinate partition
of unity.
Finally, notice that the C ∞ off-diagonal statement is reasonable because if
B ∈ δ,0 m,0
and ψ ∈ Cc∞ (Ũ) then ∗ ψBψ(−1 )∗ has C ∞ Schwartz kernel
away from the diagonal (since B has this property), so this is not an additional
restriction, and we may simply regard pseudodifferential operators on Rn with
support in a coordinate chart as pseudodifferential operators on X.
This also lets us define the principal symbol of A as a function on T ∗ X \ o:
if : U → Ũ is a coordinate chart, K ⊂ U is compact and χ ∈ Cc∞ (U)
with χ ≡ 1 on a neighborhood of K, then we let the principal symbol of
A on TK∗ X be the pull-back of the principal symbol of (−1 )∗ χ Aχ ∗ on
T ∗ Rn = Ũ × Rn by (−1 ) ; as a consequence of Remark 5.2 this is well
Ũ
defined independently of the choices of and χ . Here for A ∈ clm (X) one can
regard the principal symbol as a homogeneous degree m function on T ∗ X \ o,
or if m = 0 then on S∗ X = (T ∗ X \ o)/R+ (with the quotient corresponding
to R+ acting on the fibers of T ∗ X via dilations); in general it is an element
of Sδm (T ∗ X)/Sδm−1+2δ (T ∗ X), where the symbol space Sδm (T ∗ X) ⊂ C ∞ (T ∗ X) is
locally the pull-back of Sδ,0 m,0
(Rn ; Rn ) via (−1 ) ; again, different coordinate
charts give the same space in the overlap. Similarly, one defines WF (A) as
the inverse image of WF ((−1 )∗ χ Aχ ∗ ) under (−1 ) . In particular, the
notion of the principal symbol allows us to talk about elliptic operators; an
operator is elliptic if its principal symbol is invertible, or equivalently if the
282 A. Vasy
local coordinate version of the principal symbol is elliptic. One still has a short
exact sequence
0 → δm−1+2δ (X) → δm (X) → Sδm (T ∗ X)/Sδm−1+2δ (T ∗ X) → 0,
with the key point being the surjectivity of the penultimate map. This follows

by taking a ∈ Sδm (T ∗ X), using a partition of unity k χk = 1 subordinate to a
cover {Uk : k = 1, . . . , K} by coordinate charts, k : Uk → Ũk , and taking
the quantization

q(a) = ∗k ψk qL (−1 ∗ −1 ∗
k ) (χk a) ψk (k ) ,
k
where ψk ∈ Cc∞ (Ũk ) is identically 1 near the image of supp χk under k .

The statement q(a) ∈ δm (X) follows from our remarks regarding the C ∞ off-
diagonal behavior and that it suffices to check the pseudodifferential property
by a single cover by coordinate charts; the principal symbol is then easily seen

to be k ∗k (ψk2 )χk a = a.
Thus, if X is a compact manifold, and P ∈ δm (X) is an elliptic operator
(i.e. its principal symbol is invertible everywhere), then we can construct a
parametrix Q for P:
EL = QP − Id, ER = PQ − Id ∈ −∞ (X).
Indeed, one simply repeats the construction on Rn , by first inverting the

principal symbol p of P to get Q0 = p−1 , E0 = PQ0 − Id ∈ δ−1+2δ (X),

then considering the Neumann series ∞ j j
j=1 (−1) E0 . In order to sum it, use

a partition of unity k χk = 1 corresponding to an open cover {Uk : k =
1, . . . , K} of X by coordinate charts, let φk ∈ Cc∞ (Uk ) be identically 1 on
supp χk , so φk E0 χk is supported in Uk ×Uk and (−1 ∗ ∗
j j
k ) φk E0 χk k is an element
−j(1−2δ)
of δ . Then we can use asymptotic summation on Rn , that is write
−1 ∗
(k ) φk E0 χk ∗k = qL (ek,j ) and for each k asymptotically sum in j to get
j

ẽk ∼ ∞ ∞
j=1 (−1) ek,j , and let Ẽk = qL (ẽk ). Letting ψk ∈ Cc (Ũk ) (with Ũk the
j
image of Uk under k ), ψk identically 1 near supp φk , Ek = ∗k ψk Ẽk ψk (−1 ∗

k ) ,

K
Q = Q0 Id + Ek
k=1
provides a right parametrix. A left parametrix can be constructed similarly, and

their equality modulo −∞ (X) can be shown as on Rn .
Since −∞ (X) is bounded between any Sobolev spaces on X, we immedi-
ately obtain a Fredholm statement.
Proposition 5.21 Any elliptic A ∈ δm (X) is Fredholm as a map H s (X) →

H s−m (X) for all m, s ∈ R, that is has closed range, finite dimensional nullspace
and the range has finite codimension. Further, the nullspace is a subspace
of C ∞ (X), while the annihilator of the range in H s−m (X) in the dual space
H −s+m (X) is also in C ∞ (X). Correspondingly, the nullspace of A as well as
the annihilator of its range is independent of s; if A is invertible for one value
of s, then it is invertible for all.
There is an immediate analogue of all these results in the scattering algebra
on manifolds with boundary.
m,
Definition 5.10 For X a compact manifold with boundary, sc,δ,δ (X) consists
˙∞ ˙∞
of continuous linear maps A : C (X) → C (X), whose Schwartz kernel
is in C˙∞ away from the diagonal in X × X and with the property that if U
is a coordinate chart with : U → Ũ ⊂ Rn a diffeomorphism then for
χ ∈ Cc∞ (U), (−1 )∗ χ Aχ ∗ ∈ δ,δ
m,
m,
m,
. One writes sc (X) = sc,0,0 (X).
Note that this definition states that the Schwartz kernels of elements vanish
to infinite order, that is decay rapidly, away from the diagonal on X × X, in
particular near (y, y ) if y = y , y, y ∈ ∂X. Again, this is a reasonable definition,
m,
m,
have this property on R × R , and thus for B ∈ δ,δ ,

for elements of δ,δ n n
χ ∈ Cc∞ (Ũ), one has ∗ χ Bχ (−1 )∗ ∈ δ,δ m,
(X) automatically. (This also

uses the fact that, again in the overlap of coordinate charts, the pull-back
pseudodifferential operator statements are equivalent due to the same argument
as for the boundaryless case considered above.)
In this case the natural phase space is sc T ∗ X, which is locally, near a point
dy
on ∂X, spanned by dx , j if x is a local boundary defining function, and yj
x2 x
are coordinates on ∂X. Alternatively, this is locally simply the pull-back of the
bundle Rnz × Rnζ → Rnz via . Indeed, in local coordinates on Rn near a point
on ∂Rn , which can be taken as (x, y), x = |z|−1 = r−1 , y local coordinates
on Sn−1 , ζ dz is a smooth non-degenerate linear combination of dx x2
= −dr
dy
and x j = r dyj as is well-known, showing that locally Rnz × Rnζ is naturally
identified with sc T ∗ X.
m,
Then for A ∈ sc,δ,δ (X), the principal symbol is naturally an element of
m,
sc ∗ m−1+2δ,
−1+2δ sc ∗
Sδ,δ ( T X)/Sδ,δ ( T X).
One still has a short exact sequence.

s,r
One also has the scattering Sobolev spaces Hsc (X), defined naturally
as Hilbert spaces up to equivalence of norms, by saying that a tempered
distribution u ∈ C −∞ (X) is in Hsc
s,r
(X) if for all coordinate charts : U → Ũ,
284 A. Vasy
and for all χ ∈ Cc∞ (U), we have (−1 )∗ (χ u) ∈ H s,r . Equivalently, one may
s,r
require that for some (and hence for all) elliptic A ∈ sc (X), Au ∈ Lsc
2 (X),
where Lsc2 (X) is the scattering L2 -space, that is one given by a density Rn -
locally equivalent to the standard L2 density on Rn , and which can thus be

taken to be of the form x−n−1 ν where ν is a standard density on X, and x a
boundary defining function. (Notice that locally x−n−1 |dx dy1 . . . dyn−1 | =
rn−1 |dr dy1 , . . . dyn−1 |, showing the local equivalence to the Euclidean
version.)
The elliptic parametrix construction also goes through, resulting in the
Fredholm statement:
m,
s,r
Proposition 5.22 Any elliptic A ∈ sc,δ,δ (X) is Fredholm as a map Hsc (X) →
s−m,r−
Hsc (X) for all m,

, s, r ∈ R, that is has closed range, finite dimensional
nullspace and the range has finite codimension. Further, the nullspace is a
subspace of C˙∞ (X), while the annihilator of the range in Hsc s−m,r−
(X) in the
−s+m,−r+
˙∞
dual space Hsc (X) is also in C (X). Correspondingly, the nullspace of
A as well as the annihilator of its range is independent of r, s; if A is invertible
for one value of r, s, then it is invertible for all.
Further, tempered distributions u ∈ C −∞ (X) have wavefront sets WFsc (u),
WFs,r sc ∗
sc (u), which are subsets of ∂ T X, and can be defined either via local
/ WFs,r
identification with Rn , or again directly by saying α ∈ sc (u) if there exists
s,r
A ∈ sc (X), elliptic at α, such that Au ∈ Lsc (X).
2
An immediate application is to the Laplacian of Riemannian scattering

metrics (introduced by Melrose in [30]) which are Riemannian metrics g on
X ◦ which near ∂X have the form
dx2 h
g= + 2,
x4 x
where h is a symmetric 2-cotensor on X such that, at ∂X, h restricts to be
positive definite on T∂X. These generalize the Euclidean metric on Rn as
taking x = r−1 shows. Such g is a symmetric section on Sym2 sc T ∗ X,
and its dual gives a fiber metric on sc T ∗ X. Correspondingly, g = dg∗ d ∈
Diff2sc (X). For V ∈ S−ρ (X), ρ > 0, we then have g + V − σ elliptic if
σ ∈ C \ [0, ∞), and we have the following analogue of Proposition 5.14 and
Proposition 5.18:
Proposition 5.23 Let g be a Riemannian scattering metric on X, V ∈ S−ρ (X)
with ρ > 0. Let H = g + V.
s,r s−2,r
Then for σ ∈ C \ [0, ∞), H − σ : Hsc (X) → Hsc (X) is Fredholm for all
˙∞
r, s, with nullspace in C (X). If V is real-valued, then H is self-adjoint.
s,r
Further, for σ ∈ [0, ∞), (H − σ )u ∈ Hsc implies
∗
WFs+2,r
sc (u) ⊂ {(z, ζ ) ∈ sc T∂X X : g−1
z (ζ , ζ ) = σ }.
While we have not added vector bundles, this is straightforward using local
trivializations in the spirit of Definitions 5.9–5.10, that is a pseudodifferential
operator acting as a map between sections of two vector bundles is an operator
with a C ∞ , homomorphism valued, Schwartz kernel away from the diagonal
which, in local coordinates that at the same time are trivializations of the
bundles, is given by a matrix of pseudodifferential operators.
5.4. Propagation Phenomena

5.4.1. The Propagation of Singularities Theorem
We now understand elliptic operators in m,
; the next challenge is to deal with
non-elliptic operators. Let’s start with classical operators, and indeed let’s take
m =
= 0. Thus, A = qL (a), a ∈ C ∞ (Rn × Rn ), so σ0,0 (A) is just the restriction
of a to ∂(Rn × Rn ). Ellipticity is just the statement that a0 = a|∂(Rn ×Rn )
does not vanish. Thus, the simplest (or least degenerate/complicated) way an
operator can be non-elliptic is if a0 is real-valued and has a non-degenerate
zero set. As ∂(Rn × Rn ) is not a smooth manifold at the corner, ∂Rn × ∂Rn ,
one has to be a bit careful. Away from the corner non-degeneracy is the
statement that a0 (α) = 0 implies da0 (α) = 0; in this case the characteristic
set, Char(A) = a−1 ∞
0 ({0}), is a C codimension one embedded submanifold. At
the corner, for α ∈ ∂R × ∂Rn , one can consider the two smooth manifolds
n
with boundary Rn × ∂Rn and ∂Rn × Rn , and denoting by a0,fiber and a0,base
the corresponding restrictions, requiring that a0 (α) = 0 implies that da0,fiber
and da0,base are not in the conormal bundle of the corner; in this case in
both boundary hypersurfaces Char(A) is a smooth manifold transversal to the
boundary. In many cases, such as stationary (elliptic) Schrödinger operators,
the characteristic set is disjoint from the corner, Char(A) is disjoint from the
corner, but this is not the case for wave propagation.
More generally, if A is classical, that is a = z
ζ m ã, ã ∈ C ∞ (Rn × Rn ),
we impose the analogous condition on ã, that is that ã0 has a non-degenerate
zero set. Note that if b ∈ C ∞ (Rn × Rn ) is elliptic, and ã0 has a non-degenerate
zero set, then, with b0 denoting the restriction of b to the boundary, the same
holds for b0 ã0 as d(b0 ã0 ) = b0 dã0 + ã0 db0 , so when b0 ã0 = 0, that is when
ã0 = 0, d(b0 ã0 ) is a non-vanishing multiple of dã0 .
Let’s reinterpret this from the conic point of view, for instance corresponding
to Rn × ∂Rn , that is working on Rn × (Rn \ {0}). (Note that working with
286 A. Vasy
∂Rn × Rn , i.e. (Rn \ {0}) × Rn , is completely analogous.) Away from the

corner, we may drop the compactification from the first factor, and thus we may
assume
= 0. For A classical, then, σfiber,m,0 (A) is homogeneous of degree m,
given by am = |ζ |m ã0 , where ã0 is considered as a homogeneous degree zero
function. Now, d(|ζ |m ã0 ) = |ζ |m dã0 + ã0 d|ζ |m , and Char(A) is given by
ã0 = 0, so it is now a conic (invariant under the R+ -action) smooth
codimension one embedded submanifold of Rn × (Rn \ {0}).
The next relevant structure arises from the Hamilton vector field of A,

n

Ham = (∂ζj am )∂zj − (∂zj am )∂ζj .
j=1
Note that Ham am = 0, thus this vector field is tangent to Char(A), and thus
defines a flow on Char(A). Note that Ham is homogeneous of degree m − 1
in the sense that the push-forward of Ham under dilation in the fiber by
t > 0, Mt (z, ζ ) = (z, tζ ), is t−m+1 Ham since, using ∂zj am , resp. ∂ζj am , are
homogeneous of degree m, resp. m − 1,

n

(Ham Mt∗ f )(z, ζ ) = t1−m (∂ζj am )(z, tζ )(∂zj f )(z, tζ )

j=1

− t−m (∂zj am )(z, tζ )t(∂ζj f )(z, tζ ) = t1−m (Ham f )(z, tζ ).
In particular, integral curves of Ham through (z, ζ ) and (z, tζ ) are the “same” up
to reparameterization: denoting the first by γ , the second by γ̃ , and denoting
by M̃t dilations on R: M̃t (s) = ts,
γ̃ (s) = (z(γ (tm−1 s)), tζ (γ (tm−1 s))) = (Mt ◦ γ ◦ M̃tm−1 )(s),

d
since (M̃t )∗ ds = t ds
d
, so
d d
(Mt ◦ γ ◦ M̃tm−1 )∗ = tm−1 (Mt ◦ γ )∗ = tm−1 (Mt )∗ Ham = Ham ,
ds ds
as being an integral curve of Ham means exactly that the push-forward of
d
ds under the map is exactly Ham . In particular, up to reparameterization, the
integral curves can be considered curves on Rn × (Rn \ {0})/R+ , that is on this
quotient, the image of the curve is defined, but not the parameterization itself.
The exception is if m = 1, in which case even the parameterization is well-
defined on this quotient; in this case, Ham acts on homogeneous degree zero
functions, and is thus a vector field on the quotient. In general, if m is arbitrary,
one may consider instead the vector field |ζ |−m+1 Ham , which is homogeneous
of degree 0, and thus has well-defined integral curves on the quotient; this does
depend on the choice of a positive homogeneous degree 1 function, such as |ζ |,
but a different choice only multiplies |ζ |−m+1 Ham by a smooth non-vanishing

function, and thus simply reparameterizes the integral curves.
A different way of looking at this is that for m = 1, by the identification of
Rn × (Rn \ {0})/R+ with Rn × ∂Rn , one has a vector field on Rn × ∂Rn ; if
m = 1, then again this vector field is defined up to positive multiples.
Definition 5.11 The integral curves of Ham in Char(A) ⊂ Rn × (Rn \ {0}), as

well as their projections to Rn × (Rn \ {0})/R+ = Rn × Sn−1 = Rn × ∂Rn ,
are called null-bicharacteristics, or simply bicharacteristics.
The Hamiltonian version of classical mechanics corresponding to an energy

function p is just the statement that particles move on integral curves of Hp .
In physical terms, geometric optics for the wave equation is that at high
frequencies waves follow paths given by classical mechanics. Concretely,
including the time variable as part of our space, on Rnz = Rn−1 y × Rt , one has
a Lorentzian metric g, for instance a product-type metric, g = dt2 − h(y, dy),
and then the dual metric function G on Rnz × Rnζ , given by G(z, ζ ) = Gz (ζ , ζ ),
with Gz the inverse of gz . Note that G is homogeneous of degree 2. Then HG
is a vector field of homogeneity degree 1. The integral curves of HG are the
lifted geodesics; those inside {G = 0} are the lifted null-geodesics. Now, the
d’Alembertian

g = | det g|−1/2 Di | det g|1/2 Gij Dj
has principal symbol G, while waves are solutions of g u = 0. Mathe-

matically, the high frequency statement translates into singularities of solu-
tions u. Concretely, we have the following theorem due to Hörmander and
Duistermaat-Hörmander [28, 8]:
Theorem 5.1 Suppose that P ∈ m,0 and its principal symbol has a real homo-
geneous degree m representative p. Then in Rn × ∂Rn , WFs (u) \ WFs−m+1 (Pu)
is a union of maximally extended (null)-bicharacteristics of p.
Rather than giving a proof now, we prove a more general version.
We now turn to the general case, where the characteristic set Char(A) of A ∈
clm,
possibly intersects the corner. So first we define the rescaled Hamilton
vector field
Ha = Ha,m,
= z−
+1 ζ −m+1 Ha ,
and notice that as z∂zj and ζ ∂ζj are in Vb (Rnz ), resp. Vb (Rnζ ), and thus in
Vb (Rnz × Rnζ ),
Ha,m,
∈ Vb (Rnz × Rnζ ).
288 A. Vasy
As above, we note that if we replace z−

+1 ζ −m+1 by bz−
+1 ζ −m+1 ,
where b ∈ C ∞ (Rn × Rn ) is positive on ∂(Rn × Rn ), then Ha,m,
is only
multiplied by a positive factor, b, at Char(A) ⊂ ∂(Rn × Rn ). Since Ha,m,
is
tangent to the boundary, its integral curves are globally well-defined (i.e. they
are well-defined on R). We then make the definition:
Definition 5.12 The integral curves of Ha,m,

in Char(A) ⊂ ∂(Rn × Rn ) are
called null-bicharacteristics, or simply bicharacteristics, and we consider them
well-defined up to a direction-preserving reparameterization.
In general, we have the following theorem, which contains Theorem 5.1 as

a special case:
Theorem 5.2 Suppose that P ∈ clm,
, with real principal symbol p. Then in
∂(Rn × Rn ), WFs,r (u) \ WFs−m+1,r−
+1 (Pu) is a union of maximally extended
(null)-bicharacteristics of p.
A variant of this theorem in the variable order setting is the following:
, with real principal symbol p, and
suppose s ∈ C ∞ (Rn × ∂Rn ), r ∈ C ∞ (∂Rn × Rn ) are non-increasing along the
rescaled Hamilton flow, that is Hp,m,
s ≤ 0, Hp,m,
r ≤ 0. Then in ∂(Rn × Rn ),
WFs,r (u) \ WFs−m+1,r−

+1 (Pu)
is a union of maximally forward extended (null)-bicharacteristics of p.

The analogous conclusion holds if s, r are non-decreasing and “forward” is
replaced by “backward.”
Note that all these theorems have empty statements at points at which Hp,m,
vanishes (or is radial in the conic setting, i.e. is a multiple of the generator
of dilations in the conic variable), so there is nothing to prove at such points.
Thus, the key point is to understand what happens near points at which Hp,m,
is non-vanishing.
The proof of these theorems relies on positive commutators, that is con-
structing a pseudodifferential operator A such that i[P, A] is of the form B∗ B,
modulo terms that we can control by our assumptions. Such an estimate
actually gives bounds for the microlocal H s,r norm of u. However, the bound
can also be recovered from the regularity statement of the theorem via the
closed graph theorem as we show now.
with real principal symbol p. Suppose
that B, G, Q ∈ 0,0 , WF (B) ⊂ Ell(G) and for every α ∈ WF (B) ∩ Char(P),
there is a point α = γ (σ ) on the bicharacteristic γ through α with γ (0) = α
such that α ∈ Ell(Q) and such that for σ ∈ [0, σ ] (or σ ∈ [σ , 0] if σ < 0),
γ (σ ) ∈ Ell(G). Then for any M, N, there is C > 0 such that if Qu ∈ H s,r ,
GPu ∈ H s−m+1,r−
+1 then Bu ∈ H s,r and
BuH s,r ≤ C(QuH s,r + GPuH s−m+1,r−
+1 + uH M,N ).
The analogous conclusion also holds in the variable order setting if either
s, r are non-increasing along the Hamilton flow and σ < 0 or s, r are non-
decreasing along the Hamilton flow and σ > 0.
Proof Under the assumptions, by Theorem 5.2, Bu ∈ H s,r since WFs,r
(Bu) = ∅. Indeed, if α ∈ / WF (B), then α ∈ / WFs,r (Bu) automatically.
If α ∈ WF (B), then if α ∈ / Char(P) then GPu ∈ H s−m+1,r−
+1 implies

that Ell(G), thus WF (B), are disjoint from WFs−m+1,r−
+1 (Pu), and thus by
elliptic regularity, α ∈ / WFs+1,r+1 (u). If α ∈ WF (B) ∩ Char(P), then Theorem
5.2 states α ∈ s,r
/ WF (u), finishing the proof of the claim.
Now,
X = {u ∈ H M,N : Qu ∈ H s,r , GPu ∈ H s−m+1,r−
+1 }
is complete by the lemma below, as is
Y = {u ∈ H M,N : Bu ∈ H s,r }.
From the previous paragraph, the identity map on H M,N restricts to a map ι :
X → Y. Further, if uk → u in X and uk = ι(uk ) → v in Y then in particular
uk → u in H M,N and uk → v in H M,N , so ι(u) = u = v, that is the graph of ι
is closed. The closed graph theorem thus implies that ι is continuous, which is
exactly the estimate in the statement of the theorem.
Lemma 5.10 Suppose Aj ∈ mj ,
j , j = 1, . . . , N. Let
X = {u ∈ H r,s : Aj u ∈ H rj ,sj , j = 1, . . . , N} ⊂ H r,s ,
equipped with the norm

N
u2X = u2H r,s + Aj u2H rj ,sj .
j=1
Then X is complete.
Here all spaces and operators may have variable orders.
Proof Suppose that {uk }∞ k=1 is X -Cauchy. Then uk , resp. Aj uk , are H , resp.
r,s
H j j -Cauchy, and thus converge to some v ∈ H , resp. vj ∈ H j j . But Aj :

r ,s r,s r ,s
H r,s → H r−mj ,s−

j is continuous, so Aj uk → Aj v in H r−mj ,s−
j . Thus, Aj uk →
Aj v and Aj uk → vj in S , so vj = Aj v. Thus, Aj v ∈ H rj ,sj , so v ∈ X , and uk → v
in X .
290 A. Vasy
5.4.2. Overview of the Proof of the Propagation Theorem

To motivate the proof of Theorem 5.2, we compute (for A ∈ m ,
with A =
A∗ , with non-variable order for now)
Pu, Au − Au, Pu = (AP − P∗ A)u, u = ([A, P] + (P − P∗ )A)u, u.
Note that if, for instance, Pu = 0, then the left hand side vanishes, and so if
[A, P] + (P − P∗ )A has some definiteness properties, then we get an interesting
conclusion. Note that P − P∗ ∈ m−1,
−1 since its principal symbol in m,
is

p − p = 0. The principal symbol of i([A, P] + (P − P∗ )A) ∈ m+m −1,
+
−1 is
σm+m −1,
+
−1 (i([A, P] + (P − P∗ )A)) = −Hp a − 2p̃a,
where p̃ is the principal symbol of 1
2i (P − P∗ ) ∈ m−1,
−1 . Suppose we can
arrange that

Hp a + 2p̃a = −b2 + e, b ∈ S(m+m −1)/2,(
+
−1)/2 , e ∈ Sm+m −1,
+
−1 ,
(5.49)
with e supported in the region where we have a priori estimates on u. Then

letting B ∈ (m+m −1)/2,(
+
−1)/2 be such that its principal symbol is b, E ∈

m+m −1,
+
−1 such that its principal symbol is e,
i([A, P] + (P − P∗ )A) = B∗ B − E + F,

with F ∈ m+m −2,
+
−2 . Then we obtain
iPu, Au − iAu, Pu = Bu2 − Eu, u + Fu, u, (5.50)
that is we can control Bu in L2 by controlling Pu, Eu, and Fu. Since F is
lower order, it is dealt with inductively, gradually increasing m ,
, namely
one assumes that u already has some a priori regularity, so that Fu, u is
controlled (which is automatic for sufficiently low m ,
), and then one obtains
a bound for Bu in a stronger space than the a priori bound. Then one can
increase m ,
by 1 each to obtain new A , B , E , F ; the resulting F will give
F u, u controlled provided WF (F ) ⊂ Ell(B) ∪ Ell(E) (which follows if
WF (A ) ⊂ Ell(B) ∪ Ell(E)) by the microlocal elliptic estimate. Thus, B u
is controlled in L2 , which is a 1/2-order gain in terms of both orders over the
control provided by Bu, and so on. Here one really needs to regularize the
argument to make sense of its steps.
We remark that the sign in (5.49) is arbitrary;

Hp a + 2p̃a = b2 − e, b ∈ S(m+m −1)/2,(
+
−1)/2 , e ∈ Sm+m −1,
+
−1 , (5.51)
would also work, still giving rise to the control of Bu2 in terms of the other
quantities mentioned above.
5.4.3. The Commutant Construction

Before going through this in more detail, let’s see whether we can arrange
(5.49). Indeed, let’s work with m = 0,
= 0 to start with, for the additional
weights will not be an issue, and ignore p̃ as well. Thus, we essentially want
to find a ∈ C ∞ (Rn × Rn ) whose Hamilton derivative Hp,m,
a is negative (in
the sense of non-positive, with some definiteness in the region of interest),
apart from a region where e is supported: a is thus decreasing along the Hp,m,
orbits. But this is very simple to achieve locally when Hp,m,

is non-zero at α.
Since the statement is slightly different (in terms of numerology) depending on
whether α is at the corner or not, we consider these separately:
/ ∂Rn × ∂Rn , we can choose local coordinates
• If α ∈
q1 , q2 , . . . , q2n−1 , q2n
near α on Rn × Rn , with the chart O centered at α, such that
Hp,m,
= ∂q1 ,
and q2n is a boundary defining function (i.e. vanishes non-degenerately at
the unique boundary hypersurface, either ∂Rn × Rn or Rn × ∂Rn , containing
α; in the former case we can take ρbase = q2n in the latter ρfiber = q2n ). We
write q = (q2 , . . . , q2n−1 ).
• If α ∈ ∂Rn × ∂Rn , we can choose local coordinates
q1 , q2 . . . , q2n−2 , q2n−1 , q2n
near α on Rn × Rn , with the chart O centered at α, such that
Hp,m,
= ∂q1 ,
and q2n−1 , q2n are boundary defining functions (i.e. vanish non-degenerately
at ∂Rn × Rn , resp. Rn × ∂Rn ), so ρbase = q2n−1 , ρfiber = q2n locally are
acceptable choices. We write q = (q2 , . . . , q2n−2 ).
Recall that in either case this is achieved by choosing a local hypersurface
S transversal to Hp,m,
(α) through α, and using the fact that the Hp,m,
-
exponential map from S ×(−, ) ( > 0 small) to a neighborhood of α is a dif-
feomorphism. Since the Hp,m,
-flow is tangent to the boundary hypersurfaces,
if we choose coordinates on S of which the last, resp. last two, are boundary
defining functions, the pull-back under this diffeomorphism gives coordinates
near α which are boundary defining functions, with q2n−1 , resp. q2n−2 , being
the flow parameter in (−, ).
In fact, it does not really matter that Hp,m,
q2n = 0 (and Hp,m,
q2n−1 = 0 in
the second case); this vanishes at q2n = 0 in any case if q2n is any boundary
292 A. Vasy
Figure 5.4. The supports of a, b, e in the first case. Here Hp,m,

= ∂q1 , q are the
vertical directions, while the q2n axis points out of the page
defining function by the tangency of Hp,m,

to the boundary, so it is of the form
q2n h, with h smooth, and we can control such terms below just as we control
the terms coming from p̃.
In the first case, with q = (q2 , . . . , q2n−1 ), see Figure 5.4, for
a = χ0 (q1 )χ1 (q )2 χ2 (q2n )2 , Hp,m,

a = χ0 (q1 )χ1 (q )2 χ2 (q2n )2 ,
so we simply need to arrange that χ0 (q1 )χ1 (q )2 χ2 (q2n )2 is supported in the

region of validity of the coordinate chart, which is arranged by making χ0 , χ1 ,
and χ2 supported near 0, and such that χ0 ≤ 0, or more precisely it is the
negative of the square of a smooth function, away from the region where our
error e is supported. Here χ2 is not particularly important; being away from the
corner, this can be taken constant 1 near the boundary, so any contribution to
symbolic calculations is trivial.
Concretely, with γ the integral curve of Hp,m,
through α with γ (0) = α,
and γ (σ ) = β for some σ < 0 such that γ ([σ , 0]) ⊂ O, and further if we are
given a neighborhood U2 of β and U1 of γ ([σ , 0]), which we may assume is
a subset of O (otherwise replace these by their intersections with O), one can
construct χ0 and χ1 as above with supp χ0 χ12 χ22 ⊂ U1 , and with χ0 χ12 χ22 the
negative of the square of a smooth function outside a compact subset of U2 .
Indeed, one simply chooses > 0 such that in the coordinates qj ,
{q : |q | < , q2n < , q1 ∈ (σ − , )} ⊂ U1 ,

{q : |q | < , q2n < , q1 ∈ (σ − , σ + )} ⊂ U2 ,
then one considers χ1 supported in {q : |q | < }, identically 1 near 0, χ2 in

{q2n < }, identically 1 near 0, and lets with > 0 to be determined later as
convenient,
χ̃0 (t) = e/(t−/2) , t < /2, χ̃0 (t) = 0, t ≥ /2,

and
ψ0 (t) ∈ C ∞ (R), supp ψ0 ⊂ (σ − , ∞), supp(1 − ψ0 ) ⊂ (−∞, σ + ),
and
χ0 (t) = χ̃0 (t)ψ0 (t)2 .
Thus, ψ0 (q1 ) is constant outside {q : q1 ∈ (σ − , σ + )}, and χ̃0 ≤ 0, indeed

−χ̃0 is C ∞ . Writing

b = −χ̃0 (q1 )ψ0 (q1 )χ1 (q )χ2 (q2n ),
e = 2χ̃0 (q1 )ψ0 (q1 )ψ0 (q1 )χ1 (q )2 χ2 (q2n )2 ,
we have
Hp,m,
a = −b2 + e
as desired. Now, choosing large allows one to deal with p̃. The key point is
that one can make −χ̃0 dominate χ̃0 by choosing > 0 large. Indeed,
χ̃0 (t) = −−1 (t − /2)2 χ̃0 (t),
so as long as t is in a fixed compact set and χ̃0 is bounded by a small multiple
of −χ̃0 , provided is large. Since a strictly positive C ∞ function has a C ∞
−m+1
square root, given p̃ ∈ C ∞ , such as p̃ = ρfiber (2p̃), one can arrange that
Hp,m,
a + p̃a = −b2 + e
by taking large so that 1 − −1 (q1 − /2)2 p̃ > 1/2 on supp a. While we
assumed m = 0 =
, even if say m = 0 and α ∈ Rn × ∂Rn , we simply
−m
add a factor of q−m
2n as our weight, then Hp,m,
q2n = 0, so the weights do not
contribute to the commutator, completing the construction. We remark that one
only needs p̃ ∈ S0,0 , for then it is bounded, and for sufficiently large > 0,
1 − −1 (q1 − /2)2 p̃ > 1/2 on supp a, and thus

b = (1 − −1 (q1 − /2)2 p̃)1/2 −χ̃0 (q1 )ψ0 (q1 )χ1 (q )χ2 (q2n ) ∈ S0,0 . (5.52)
Now, the variable order version is similar, except we use a weight q−s 2n or
q−r
2n , with s ∈ C ∞ (Rn × ∂Rn ), r ∈ C ∞ (∂Rn × Rn ), depending on where α
lies. Here we need to extend s, r to functions defined at least locally in the
region of interest on Rn × Rn ; we do so by making them independent of
the boundary defining coordinate, q2n . (In fact, for the argument below even
s, r ∈ S0,0 work.) Then notice that Hp,m,
qs2n = qs2n (− log q2n )Hp,m,
s. Notice
294 A. Vasy
that − log q2n → +∞ as q2n → 0, so this term actually dominates all other
terms where Hp,m,
s does not vanish, so it cannot be estimated away using χ̃0 .
Correspondingly, we need to assume that Hp,m,
s has the correct sign, that is it
is non-positive, for then
Hp,m,
a = −b2 − b̃2 + e, b̃2 = qs2n (− log q2n )(−Hp,m,
s),
allows for a similar

argument as in the s constant case. (One either needs
to arrange that −Hp,m,
s is smooth, which is easily done since we have
flexibility in choosing s, or instead use the sharp Gårding theorem.)
The construction in the second case, with α in the corner, is completely
similar with q = (q2 , . . . , q2n−2 ) now, taking
a = χ0 (q1 )χ1 (q )2 χ2 (q2n−1 )2 χ2 (q2n )2 ,

Hp,m,
a = χ0 (q1 )χ1 (q )2 χ2 (q2n−1 )2 χ2 (q2n )2 .
Then with

b= −χ̃0 (q1 )ψ0 (q1 )χ1 (q )χ2 (q2n−1 )χ2 (q2n ),
and
e = 2χ̃0 (q1 )ψ0 (q1 )ψ0 (q1 )χ1 (q )2 χ2 (q2n−1 )2 χ2 (q2n )2 ,
we have
Hp,m,
a = −b2 + e.
More generally, with p̃ ∈ S0,0 ,

b = (1−−1 (q1 −/2)2 p̃)1/2 −χ̃0 (q1 )ψ0 (q1 )χ1 (q )χ2 (q2n−1 )χ2 (q2n ) ∈ S0,0 ,
(5.53)
gives
Hp,m,
a + p̃a = −b2 + e.
Again, variable orders are fine, but these orders s, r need to satisfy Hp,m,
s ≤ 0,
Hp,m,
r ≤ 0.
Note that if σ > 0, a similar construction works, but reversing the signs
corresponding to (5.51), so in the variable order case we need s and r to have
Hp,m,
s ≥ 0, Hp,m,
r ≥ 0 at the relevant boundary faces. Thus, with γ the
integral curve of Hp,m,
through α with γ (0) = α, and γ (σ ) = β for some
σ > 0 such that γ ([0, σ ]) ⊂ O, and further if we are given a neighborhood U2
of β and U1 of γ ([0, σ ]), which we may again assume is a subset of O, one can
construct χ0 , χ1 , and χ2 as above with supp χ0 χ12 χ22 ⊂ U1 , and with χ0 χ12 χ22
the square of a smooth function outside a compact subset of U2 . Namely, χ1 , χ2

are unchanged,
χ̃0 (t) = e−/(t+/2) , t > −/2, χ̃0 (t) = 0, t ≤ −/2,
and
ψ0 (t) ∈ C ∞ (R), supp ψ0 ⊂ (−∞, σ + ), supp(1 − ψ0 ) ⊂ (σ − , ∞),
and
χ0 (t) = χ̃0 (t)ψ0 (t)2 .
Thus, as above but with a sign switch, given a C ∞ function p̃ on Rn × Rn , or
indeed simply p̃ ∈ S0,0 , one can arrange that
Hp,m,
a + p̃a = b2 − e.
5.4.4. Regularization and Proof of the Propagation Theorem

Now, returning to σ < 0 for the sake of definiteness to deal with u that is not
a priori nice, we regularize the argument. Thus, we replace A by a bounded

family At , t ∈ [0, 1], [0, 1] → m ,
, continuous as a map [0, 1] → m +δ,
+δ

for δ > 0, such that for t > 0, At ∈ m −K,
−K (for sufficiently large K), and
,

such that A0 ∈ m is the operator denoted by A above. It will be convenient
to shift orders, and thus we take s,r ∈ s,r to be elliptic and invertible, with
principal symbol
ρs,r = ζ s zr .

Further, for future purposes, we arrange that, as operators in m ,
, the
principal symbols at satisfy for some M > 0
Hp at + 2p̃at = −b2t − M 2 ρm−1,
−1 at + et ,
√
b ∈ L∞ ([0, 1]t ; Ss,r ), e ∈ L∞ ([0, 1]t ; S2s,2r ), at ∈ L∞ ([0, 1]t ; Sm /2,
/2 );
(5.54)
where, to simplify the notation, we let
s = (m + m − 1)/2, r = (
+
− 1)/2.
In applications we also need some microsupport and ellipticity conditions for
the families, so we make the following definition:
Definition 5.13 For a bounded family a = {at : t ∈ [0, 1]} in Sm,
and for
α ∈ ∂(Rn × Rn ) we say that α ∈ / esssuppL∞ (a) if α has a neighborhood U in
Rn × Rn such that at |U∩(Rn ×Rn ) is bounded in S−∞,−∞ (i.e. each seminorm is
bounded).
296 A. Vasy
Further, we say that a bounded family a = {at : t ∈ [0, 1]} in Sm,

is elliptic
at α ∈ ∂(Rn × Rn ) if α has a neighborhood U in Rn × Rn such that for some
c > 0,

at |U∩(Rn ×Rn ) ≥ czm ζ
.
We then have the following lemma:

Lemma 5.11 For α0 ∈ ∂(Rn × Rn ) with Hp,m,
non-vanishing at α0 . Then
α0 has a neighborhood O in ∂(Rn × Rn ) on which Hp,m,
is non-vanishing,
and if α ∈ O, γ the integral curve of Hp,m,
through α with γ (0) = α, and
γ (σ ) = β for some σ < 0 such that γ ([σ , 0]) ⊂ O, and further if we are given
a neighborhood U2 of β and U1 of γ ([σ , 0]) contained in O, then there exists
at , bt , and et as in (5.54) such that
esssuppL∞ a, b ⊂ U1 , esssuppL∞ e ⊂ U2 ,
and b is elliptic on γ ([σ , 0]).
Proof We assume that α0 is in the corner ∂Rn × ∂Rn ; otherwise one of the
weights is irrelevant, and the corresponding coordinate can be taken to be one
of the q coordinates with no significant changes. We fix our coordinates so that
q2n−1 = ρbase , q2n = ρfiber .
We first define χ0 , χ1 , χ2 as above and let

ã = q−m −
2 2 2
2n q2n−1 χ0 (q1 )χ1 (q ) χ2 (q2n−1 ) χ2 (q2n ) .
For t ∈ [0, 1], let

φt (τ ) = (1 + tτ )−K/2 , τ ≥ 0, (5.55)
and note that
dφt
(τ ) = ft (τ )φt (τ ), ft (τ ) = t(−K/2)(1 + tτ )−1 , (5.56)
dτ
with | ft (τ )| ≤ K/2, |τ ft (τ )| ≤ K/2. We then let
at = φt (q−2 −2
2n−1 )φt (q2n )ã.

Note that a ∈ L∞ ([0, 1]; Sm ,
) satisfies at ∈ Sm −K,
−K for t > 0, esssuppL∞
√
(a) ⊂ U1 , and a is elliptic on γ ([σ , 0]). Further, a ∈ L∞ ([0, 1]; Sm /2,
/2 ).
Now, Hp,m,
(φt (q−2 −2
2n−1 )φt (q2n )) = 0, and

+
−1 m +m−1
q2n−1 q2n Hp at

2
= φt (q−2 −2
2n )φt (q2n−1 ) ψ0 (q1 )χ1 (q )χ2 (q2n−1 )χ2 (q2n ) χ̃0 (q1 )

2
+ 2φt (q−2 −2
2n )φt (q2n−1 )ψ0 (q1 )ψ0 (q1 ) χ1 (q )χ2 (q2n−1 )χ2 (q2n ) χ̃0 (q1 ).
Proceeding as above, we note that for > 0 sufficiently large

+
−1 m +m−1
q2n−1 q2n (Hp at + 2p̃at ) − M 2 ρm−1,
−1 at

2
= φt (q−2 −2
2n )φt (q2n−1 ) ψ0 (q1 )χ1 (q )χ2 (q2n−1 )χ2 (q2n ) (χ̃0 (q1 ) + rt χ̃0 (q1 ))
+ 2φt (q−2 −2 2
2n )φt (q2n−1 )ψ0 (q1 )ψ0 (q1 )χ1 (q )χ2 (q2n−1 )χ2 (q2n )) χ̃0 (q1 )
= −b2t + et ,
with the desired properties.
√
Let ǎt = at , and let Ǎt = 12 (qL (ǎt ) + qL (ǎt )∗ ), which is thus formally
self-adjoint. Let At = Ǎ2t . Then for t > 0
iPu, At u − iAt u, Pu
(5.57)
= Bt u2 + M 2 (m−1)/2,(
−1)/2 Ǎt u2 − Et u, u + Ft u, u,
where F ∈ L∞ ([0, 1]; 2s−1,2r−1 ). We write, with δ > 0,
|Pu, At u| = |(−1 ∗
(m−1)/2,(
−1)/2 ) Ǎt Pu, (m−1)/2,(
−1)/2 Ǎt u|
≤ (−1 ∗
(m−1)/2,(
−1)/2 ) Ǎt Pu (m−1)/2,(
−1)/2 Ǎt u
1 δ
≤ (−1 ∗ ∗
(m−1)/2,(
−1)/2 ) Ǎt Pu + 2 ((m−1)/2,(
−1)/2 ) Ǎt u .
2 2
2δ
One has a similar bound for At u, Pu. For δ > 0 sufficiently small (namely
δ < M 2 ), δ(m−1)/2,(
−1)/2 Ǎt u2 can be absorbed into M 2 (m−1)/2,(
−1)/2
Ǎt u2 , and thus we deduce from (5.57) that
Bt u2 ≤ |Et u, u| + |Ft u, u| + δ −1 (−1 ∗ 2
(m−1)/2,(
−1)/2 ) Ǎt Pu . (5.58)
Now we let t → 0; then assuming a priori control on the terms other than
Bt u2 we conclude that Bt u is bounded in L2 . But Bt u → B0 u in S , so using
the weak compactness of the unit ball in L2 , and thus that Bt u has a subsequence
Btj u converging L2 -weakly to some v ∈ L2 , and thus in S , we deduce that
B0 u = v ∈ L2 , yielding the desired regularity information.
In order to make the a priori control assumptions, we need a uniform version
of the operator wavefront set:
Definition 5.14 For a bounded family A = {At : t ∈ [0, 1]} in m,
and for
α ∈ ∂(Rn × Rn ) we say that α ∈ / WF L∞ (At ) if α has a neighborhood U in
Rn × Rn such that At = qL (at ) and at |U∩(Rn ×Rn ) is bounded in S−∞,−∞ (i.e.
each seminorm is bounded).
Note that if At = A0 for all t, then WF L∞ (A) = WF (A), that is this family
wavefront set is an appropriate generalization of the standard operator wave
front set.
298 A. Vasy
Lemma 5.12 If A and B are bounded families, then with AB = {At Bt : t ∈

[0, 1]},
WF L∞ (AB) ⊂ WF L∞ (A) ∩ WF L∞ (B).
In particular, if Q ∈ m,
then
WF L∞ (QA), WF L∞ (AQ) ⊂ WF (Q) ∩ WF L∞ (A).
/ WF L∞ (A) if and only if there exists Q ∈ 0,0 elliptic at α such

Further, α ∈
that QA is bounded in −∞,−∞ .
Proof The composition properties are automatic from the description of the
product as an asymptotic sum. In particular, if α ∈ / WF L∞ (A) and Q is elliptic

at α with WF (Q) ∩ WFL∞ (A) = ∅, then QA is bounded in −∞,−∞ . Now,

such a Q exists since WF L∞ (A) is closed, and α is in its complement.

Conversely, if a Q as stated exists, let B be a microlocal parametrix for Q, so
BQ = Id +E with α ∈ / WF (E). Then At = BQAt − EAt , with α ∈ / WF L∞ (EAt )
since α ∈
/ WF (E), while QA bounded in −∞,−∞ implies that BQA is also
bounded there, thus α ∈ / WF L∞ (A).
In summary, we have the following result, which is the basic microlocal

propagation estimate, propagating control on WF L∞ (E) to WF (B0 ):
Lemma 5.13 Suppose that (5.54) is satisfied. Let At = qL (at ), Bt = qL (bt ),
Et = qL (et ), and let Q1 , Q2 ∈ 0,0 such that Q1 is elliptic on WF L∞ (A),
Q2 is elliptic on WF L∞ (E). If Q1 u ∈ H s−1/2,r−1/2 , Q1 Pu ∈ H s−m+1,r−
+1 ,
Q2 u ∈ H s,r then B0 u ∈ L2 and for all M, N there is C > 0 such that
B0 uL2 ≤ C(Q2 uH s,r + Q1 PuH s−m+1,r−

+1
+ Q1 uH s−1/2,r−1/2 + uH M,N ).
Proof We note that as Q1 is elliptic on WF L∞ (A), hence on WF L∞ (F),
Ft uH −s+1/2,−r+1/2 ≤ C(Q1 uH s−1/2,r−1/2 + uH M,N ),
and thus with Q̃1 ∈ 0,0 with WF (Id −Q̃1 ) ∩ WF L∞ (A) = ∅, Q1 elliptic on
WF (Q̃1 ), we have (Id −Q̃1 )Ft is bounded in −∞,∞ and also
Q̃∗1 uH s−1/2,r−1/2 ≤ C (Q1 uH s−1/2,r−1/2 + uH M,N ),
so
|Ft u, u| ≤ |(Id −Q̃1 )Ft u, u| + |Ft u, Q̃∗1 u|

≤ C (u2H M,N + Q1 u2H s−1/2,r−1/2 ).
Similarly,
|Et u, u| ≤ C (u2H M,N + Q2 u2H s,r ).
Next, using m /2 − (m − 1)/2 = s − m + 1,
/2 − (
− 1)/2 = r −
+ 1,
(−1 ∗
(m−1)/2,(
−1)/2 ) Ǎt Pu ≤ C(Q1 PuH s−m+1,r−
+1 + uH M,N ).
Combining these estimates, we see that the right hand side of (5.58) remains
uniformly bounded, and thus B0 u ∈ L2 as claimed.
As a corollary:
Proposition 5.24 For α0 ∈ ∂(Rn × Rn ) with Hp,m,
α0 has a neighborhood O in ∂(Rn × Rn ) on which Hp,m,
is non-vanishing,
and if α ∈ O, γ the integral curve of Hp,m,
γ (σ ) = β for some σ < 0 such that γ ([σ , 0]) ⊂ O, and further if we are given
a neighborhood U2 of β and U1 of γ ([σ , 0]) contained in O, Q1 , Q2 ∈ 0,0
such that Q1 is elliptic on U1 , Q2 is elliptic on U2 , then there exists Q3 ∈
0,0 elliptic on γ ([σ , 0]) such that the following holds. If Q1 u ∈ H s−1/2,r−1/2 ,
Q1 Pu ∈ H s−m+1,r−
+1 , Q2 u ∈ H s,r then Q3 u ∈ H s,r and for all M, N there is
C > 0 such that
Q3 uH s,r ≤ C(Q2 uH s,r + Q1 PuH s−m+1,r−
+1
+ Q1 uH s−1/2,r−1/2 + uH M,N ).
The analogous result also holds for σ > 0.
Now propagation of singularities is an immediate consequence. Indeed, one
can iterate, improving half an order at a time, taking U1 = Ell(Q3 ) ∩ U1 and
U2 = Ell(Q3 ) ∩ U2 for the next iteration. This directly proves Theorem 5.4
when the bicharacteristic segment is contained in a single set O; in general
a compactness argument proves it in a finite number of steps from this local
version.
5.4.5. Complex Absorption

Although we required that P ∈ clm,
have real principal symbol, this is
not quite necessary. In particular, complex absorption has been a very useful
technical tool, see for instance the work of Nonnenmacher and Zworski [35]
(see also [5]): in this one “cuts off” a non-elliptic problem with real principal
symbol by adding a complex term to the principal symbol which has support
outside the region of interest.
So suppose instead that one is given a complex valued homogeneous
function, p−iq, with p, q real-valued. Let P ∈ clm,
, resp. Q ∈ clm,
, have (real)
300 A. Vasy
principal symbols p, resp. q. Note that when q = 0, p − iq is elliptic, while if q

vanishes near a point α, then the previous microlocal estimate works. Thus, the
key question is what happens at the characteristic set {p = 0, q = 0} at points
in supp q. The typical application here would involve a q that acts as microlocal
“absorption” along the Hamilton flow; absorbing any singularity propagating
along Hp,m,
in {p = 0} away from supp u. Thus, one should consider q a bump
function along the Hp,m,
-bicharacteristics, whose sign will be very important;
along a fixed Hp,m,
-bicharacteristic γ in {p = 0}, where q = 0, the problem
is elliptic, and thus if (P − iQ)u is regular, so is u, where q ≡ 0, singularities
propagate, so for a bump-function like q, the singularities are absorbed at the
boundary of supp q. Since q is mostly a technical tool here, it is reasonable to
make technically convenient assumptions on Q below.
Then for A formally self-adjoint,
i(P − iQ)u, Au − iAu, (P − iQ)u
= i([A, P] + (P − P∗ )A)u, u + (AQ + Q∗ A)u, u, (5.59)
∗ m+m −1,
+
−1 ∗ m+m ,
+

i([A, P] + (P − P )A) ∈ , AQ + Q A ∈ .
Now, if the principal symbol a of A is ≥ 0 as above, with the principal symbol
of i([A, P] + (P − P∗ )A) being
−(Hp a + 2p̃a) = b2 − e,
then for q ≥ 0, the principal symbol of the second term on the right hand side
of (5.59) has the same sign as that of the first. However, there is an issue with
such an argument since the second term is of higher order than the first one.
Thus, it is convenient to assume that Q = T 2 , with T ∈ m/2,
/2 , T = T ∗ .
Note that at the cost of changing P without changing its principal symbol, if
q = t2 for a non-negative symbol t, then this may always be arranged; simply
take T ∈ m/2,
/2 with principal symbol t and with T = T ∗ ; then Q − T 2 ∈
m−1,
−1 , so replacing P by P − i(Q − T 2 ), and Q by T 2 , we have the desired
form of Q. From now on we assume that
Q = T 2 , T ∈ m/2,
/2 , T = T ∗ .

When a = ǎ2 with ǎ ∈ Sm /2,
/2 , as arranged above, then with Ǎ ∈ m /2,
/2
of principal symbol ǎ and Ǎ∗ = Ǎ, let A = Ǎ2 (so A∗ = A) – we are simply a
bit more careful in our specification of A. Then
AQ + Q∗ A = Ǎ2 Q + QǍ2 = 2ǍQǍ + Ǎ[Ǎ, Q] + [Q, Ǎ]Ǎ = 2ǍT 2 Ǎ + [Ǎ, [Ǎ, Q]],

and [Ǎ, [Ǎ, Q]] ∈ m+m −2,
+
−2 , which is thus the same order as F above,
so is controlled by the a priori assumptions. On the other hand, ǍT 2 Ǎu, u =
T Ǎu2 , so we obtain
i(P − iQ)u, Au − iAu, (P − iQ)u

= Bu2 − Eu, u + (F + [Ǎ, [Ǎ, Q]])u, u + 2T Ǎu2 .
Regularizing Ǎ as Ǎt , all the previous arguments go through. Note also that if
instead σ > 0, that is we have
−(Hp a + 2p̃a) = −b2 + e,
then we need to change our requirements on Q, namely we need q ≤ 0, and for

the technically convenient setting we need q = −t2 , t ≥ 0.
Proposition 5.25 For α0 ∈ ∂(Rn × Rn ) with Hp,m,
non-vanishing at α0 .
Suppose also that Q = T 2 as above. Then α0 has a neighborhood O in
∂(Rn × Rn ) on which Hp,m,
is non-vanishing, and if α ∈ O, γ the integral
curve of Hp,m,
through α with γ (0) = α, and γ (σ ) = β for some σ < 0
such that γ ([σ , 0]) ⊂ O, and further if we are given a neighborhood U2 of β
and U1 of γ ([σ , 0]) contained in O, Q1 , Q2 ∈ 0,0 such that Q1 is elliptic on
U1 , Q2 is elliptic on U2 , then there exists Q3 ∈ 0,0 elliptic on γ ([σ , 0]) such
that the following holds. If Q1 u ∈ H s−1/2,r−1/2 , Q1 (P − iQ)u ∈ H s−m+1,r−
+1 ,
Q2 u ∈ H s,r then Q3 u ∈ H s,r and for all M, N there is C > 0 such that
Q3 uH s,r ≤ C(Q2 uH s,r + Q1 (P − iQ)uH s−m+1,r−

+1
+ Q1 uH s−1/2,r−1/2 + uH M,N ).
The analogous result also holds for σ > 0, provided Q = −T 2 .

Thus, for q ≥ 0, estimates propagate forward along the Hamilton flow; for
q ≤ 0, they propagate backwards. This means that singularities, as measured
by the wavefront set, propagate in the opposite direction: if q ≥ 0, and there is
WF at α, then for σ < 0, there is also WF at γ (σ ), for otherwise our estimate
(and the corresponding regularity statement) would give the absence of WF
at α.
While we used Q = T 2 here, this was for a purely technical point. If Q = Q∗
(which may always be arranged at the cost of changing P without changing its
principal symbol, namely replacing P by P − i(Q − Q∗ )/2, and replacing Q by
(Q + Q∗ )/2) and Q ≥ 0, then ǍQǍu, u = QǍu, Ǎu ≥ 0 still. In general,
just because q ≥ 0 and Q = Q∗ , we do not have Q ≥ 0, but by the sharp
Gårding inequality, this holds modulo a one order lower error term, that is
Qv, v ≥ −Cv2H (m−1)/2,(
−1)/2 . Applying this with v = Ǎu, we obtain a term
from the right hand side that is controlled by C (m−1)/2,(
−1)/2 Ǎu2 , which
can be controlled as above by choosing > 0 large to absorb this into Bu2 .
For proofs of the sharp Gårding inequality we refer to [27, Theorem 18.1.14]
302 A. Vasy
in the currently considered setting, and [58, Theorem 4.32] in the semiclassical
setting that is discussed in Section 5.5.
5.4.6. Fredholm Problems with Complex Absorption

This gives rise to the simplest non-elliptic Fredholm problem. So suppose that
P ∈ m,
with real homogeneous principal symbol. Suppose also that Q ∈
m,
is such that its principal symbol q satisfies q = t2 as above, and that
for all α ∈ ∂(Rn × Rn ) in (p) = { p = 0} the integral curve γ of Hp with
γ (0) = α reaches {q > 0} in finite time in both the forward and backward
direction, that is there exist T± > 0 such that q(γ (T± )) > 0. Let
X s,r = {u ∈ H s,r : (P − iQ)u ∈ H s−m+1,r−
+1 }, Y s,r = H s,r .
Then
P − iQ : X s,r → Y s−m+1,r−
+1 , P∗ + iQ∗ : X s,r → Y s−m+1,r−
+1
are Fredholm for all s, r. Note that this follows estimates of the kind
uH s,r ≤ C((P − iQ)uH s−m+1,r−
+1 + uH M,N ),
uH s,r ≤ C((P∗ + iQ∗ )uH s−m+1,r−
+1 + uH M,N ),
for sufficiently negative M, N, which in turn follow from the propagation of
singularities result as discussed beforehand in the real principal symbol setting.
Thus, we just need to check that if u ∈ S and (P − iQ)u ∈ H s−m+1,r−
+1
then u ∈ H s,r , that is WFs,r (u) = ∅. But this is straightforward: if either
p or q is elliptic at α, then α ∈ / WFs+1,r+1 (u) and thus α ∈ / WFs,r (u) by
microlocal elliptic regularity, and otherwise p vanishes at α, and thus the
backward bicharacteristic γ from α reaches q > 0, where we know there is
no WFs,r (u), so propagation of singularities gives α ∈ / WFs,r (u). A similar
argument works for the adjoint, using the forward bicharacteristic. This proves
the claim.
5.4.7. Radial Points and Generalizations

The final ingredient for scattering theory is radial points. These are the points
in the conic perspective where Hp is a multiple of the generator of dilations,
and from the compactification point of view the points α ∈ ∂(Rn × Rn ) where
Hp,m,
= ζ −m+1 z−
+1 Hp , as a smooth vector field on Rn × Rn , vanishes.
At such points the argument given above breaks down, since there are no local
coordinates on the boundary in which this rescaled Hamilton vector field would
be a coordinate vector field. Also, at such points, one cannot use the derivative
of a function on the boundary to dominate terms when Hp falls on weights or

regularizers, with the result that the weights are required to produce the correct
signs for the commutator argument.
Although here we are interested in rather large sets (higher dimensional
submanifolds) of radial points, which were explicitly considered in a relatively
general setting by Melrose in [30], as far as I am aware radial points made
their first appearance in the work of Guillemin and Schaeffer [16] in a setting
in which they are discrete. These discrete radial points also arise in scattering
theory due to scattering by a symbolic potential of order 0 with non-degenerate
critical points at infinity (i.e. at ∂Rn ), where they were studied by Herbst
and Skibsted [20, 21], and where microlocal techniques were applied by
Hassell, Melrose, and Vasy [18, 19] obtaining even precise iterated regularity
statements. In a less explicit way, such radial points have long played a role in
scattering theory; indeed, the propagation set of Sigal and Soffer [39], which
played a key role in their proof of the asymptotic completeness of short-
range N-body Hamiltonians, is an N-particle generalization of radial sets, see
[44, 45].
Following [48] it is convenient to proceed quite generally, using a general-
ization of this setting to a submanifold L of Char(P) ⊂ ∂(Rn × Rn ) closed
under the Hp,m,
flow, that is Hp,m,
is tangent to it, and which acts as a
source or sink for the flow in a neighborhood of L within Char(P). Since we
have corners here, we need to be more specific, and we require that L is a
submanifold of one of the two boundary hypersurfaces which is transversal to
the other boundary face (allowing of course that L does not intersect the other
boundary face at all), see Figure 5.5. For the sake of definiteness, we assume
that L is a subset of Rn × ∂Rn , that is is at fiber infinity, and is transversal to
∂Rn × ∂Rn , possibly via an empty intersection. Next, suppose that L is defined
by {ρ1,j = 0 : j = 1, . . . , k} in Char(P) within Rn × ∂Rn , that is the ρ1,j have
linearly independent differentials on their joint zero set, that is it is defined by
{ p̂ = 0, ρ1,j = 0, j = 1, . . . , k} within Rn × ∂Rn , where p̂ = ρfiber
m p. Let

k
ρ1 = 2
ρ1,j ,
j=1
which is thus a quadratic defining function of L, that is it vanishes there

quadratically in a non-degenerate manner. Since Hp,m,
is tangent to L,
Hp,m,
ρ1,j vanishes at L, that is is a linear combination of the ρ1,i with C ∞
coefficients, so Hp,m,
ρ1 vanishes quadratically at L. The source(+)/sink(−)
assumption is that there is a function β1 > 0 such that
∓Hp,m,
ρ1 = β1 ρ1 + F2 + F3 ,
304 A. Vasy
Figure 5.5. The Hamilton flow at radial sets L; in the examples both sinks. Left:
L ⊂ Rn × ∂Rn , that is away from base infinity (the corner). Right: L ⊂ Rn × ∂Rn
intersects ∂Rn × ∂Rn transversally
where F2 ≥ 0 and F3 vanishes cubically at L. Under this assumption, L is a

source or sink within Char(P), since ρ1 is an increasing (+), resp. decreasing
3/2
(−), function near L, as |F3 | ≤ Cρ1 ≤ β1 ρ1 /2 in a neighborhood of L. In
∞
particular, notice that if φ ∈ Cc ([0, ∞)) is such that φ is ≡ 1 near 0, φ ≥ 0
√ √
with φ smooth and φ ≤ 0 with −φ smooth, with support sufficiently
close to 0 then
±Hp,m,
φ(ρ1 ) = (−φ (ρ1 ))(β1 ρ1 + F2 + F3 ) ≥ 0,
and further the second factor on the right hand side is > 0 on supp φ , so indeed

φ1 = ±Hp,m,
φ(ρ1 )
is smooth, and vanishes near L.

To explain this source/sink condition note that if Hp,m,
actually vanishes
at L, then one can consider its linearization at each point α ∈ L, that is Hp,m,
then maps the ideal I of functions vanishing at α to itself, and thus I 2 to

itself, and thus acts on I/I 2 , which can be identified with the cotangent space
at α. Further, the conormal space of L at α is preserved by Hp,m,
as Hp,m,
ρ1,j
vanishes at L, that is Hp,m,
acts on this finite dimensional vector space. If
the ρ1,j are such that dρ1,j is an eigenvector with eigenvalue β1,j (α) at each

α ∈ L, then Hp,m,
ρ1 is given by kj=1 2β1,j ρ1,j 2 modulo cubically vanishing
functions, and thus if all β1,j have the same non-zero sign, ±, then letting β1
to be 2 minj |β1,j |, the above condition is satisfied (with ± corresponding to the
sign ± here). In fact, in applications often Hp,m,
is a multiple of the identity
operator on this finite dimensional vector space, and then all the β1,j are the
same, and any defining functions ρ1,j can be used to construct ρ1 .
At the possible radial points in L, the only way one can have a positive
commutator estimate is by taking into account the weights. Consider Hp,m,
as a vector field on Rn × Rn which is tangent to the boundary of this compact

space, and recall that ρbase = z−1 and ρfiber = ζ −1 are defining functions of
the boundary hypersurfaces. In particular, as already mentioned, Hp,m,
ρbase =
abase ρbase and Hp,m,
ρfiber = afiber ρfiber with abase , afiber well defined on ∂(Rn ×
Rn ) (though they depend on the choice of ρbase and ρfiber ). Moreover, if Hp,m,
actually vanishes at α, which is, say, at fiber infinity, then

Hp,m,
( f ρfiber ) = (Hp,m,
f )ρfiber + f (Hp,m,
ρfiber ) = (afiber f + Hp,m,
f )ρfiber ,
with (afiber f + Hp,m,
f )(α) = afiber (α)f (α), shows that afiber is actually well-
defined at such α, independent even of the choice of the boundary defining
function. We then require that L is a source or sink even taking into account the
infinitesimal “jet dynamics” at the boundary, that is afiber in our case (since we
are working at fiber infinity) has the same sign as Hp,m,
ρ1 : there is a defining
function ρ0 of Rn × ∂Rn such that
∓ Hp,m,
ρ0 = β0 ρ0 , β0 |L > 0. (5.60)
From the remarks above, if L consists of radial points, then one can simply
take our preferred boundary defining function ρ0 = ρfiber in checking whether
(5.60) holds, since all choices give the same result. Note that if (5.60) holds,
then β0 is thus bounded below by a positive constant in a neighborhood of L;
we can always restrict work to such a neighborhood below.
We now consider the possibility that L intersects the corner (which the reader
may ignore at first reading). We then also need to take care of weights in terms
of ρbase . Due to its importance in the analysis of Klein–Gordon type equations,
we assume that we have a defining function (a positive smooth multiple of
ρbase ) ρ2 such that
Hp,m,
ρ2 = 2β2 β0 ρ2 , β2 |L = 0;
in general β2 can easily be handled as well.
Further, as before, P − P∗ ∈ m−1,
−1 also plays a role in these arguments,
and at radial points it cannot be absorbed into other terms, just like the weight,
that is powers of ρ0 , could not be thus absorbed. We normalize this and write

1
p̃ = σm−1,
−1 (P − P∗ ) = ±β0 β̃ρ0−m+1 .
2i
Let φ0 ∈ Cc∞ (R) be identically 1 near 0, supported sufficiently close to 0; we
use φ0 (p̂) to localize near Char(P). Note that
Hp,m,
φ0 (p̂) = φ0 (p̂)Hp,m,
p̂
306 A. Vasy
has support where p̂ is in supp φ0 , that is away from the characteristic set. Then
we compute that with

a = φ(ρ1 )2 φ0 (p̂)2 ρ0−m ρ2−
the principal symbol −(Hp a + 2p̃a) of i([A, P] + (P − P∗ )A) is

∓ρ0−m −m+1 ρ2−

−
+1 φ(ρ1 )φ12 ± 2φ 2 φ0 φ0 Hp,m,
p̂ + β0 m φ(ρ1 )2

+ 2β2 β0
φ(ρ1 )2 + 2β0 β̃φ(ρ1 )2 .
The second term in the parentheses is supported away from Char(P), and
thus is controlled by microlocal elliptic estimates. The last three terms have
essential support at L itself, and they have a definite sign if m + 2β2
+ 2β̃ has
a definite sign. With s = (m + m − 1)/2, r = (
+
− 1)/2, this means that
s − (m − 1)/2 + β2 (r − (
− 1)/2) + β̃
needs to have a definite sign when we prove H s,r regularity. Further, the first
term in the parentheses has the same sign as these when the sign of s−(m−1)/
2 + β2 + β̃ is negative. Thus, when s − (m − 1)/2 + β2 (r − (
− 1)/2) + β̃ > 0,
then all signs agree (apart from the off-the-characteristic-set term), and we
have a result that has a different character from the standard propagation result
in so far as not having to assume a priori H s regularity anywhere to conclude
H s regularity near L. On the other hand, when s − (m − 1)/2 + β2 (r − (
−
1)/2) + β̃ < 0 we must assume regularity on supp φ , that is in a punctured
neighborhood of L, in order to conclude H s regularity at L. In summary, using
the fact that we can make φ have sufficiently small support so that the definite
sign of this expression just at L implies the analogous conclusion on supp φ,
modulo justifying the calculations via a regularization argument, we have
Proposition 5.26 Suppose that L is as above, and s − (m − 1)/2 + β̃ > 0 on L,
and indeed suppose that s − (m − 1)/2 + β̃ > 0 on L for some s ∈ [s − 1/2, s).
Suppose U a neighborhood of L. Then there exists Q ∈ 0,0 elliptic on L such

that for Q1 elliptic on WF (Q), with WF (Q) ⊂ U, and if Q1 u ∈ H s ,r−1/2 ,
Q1 Pu ∈ H s−m+1,r−
+1 , then Qu ∈ H s,r and for all M, N there is C > 0 such
that
QuH s,r ≤ C(Q1 PuH s−m+1,r−
+1 + Q1 uH s ,r−1/2 + uH M,N ).
On the other hand, suppose now that s − (m − 1)/2 + β̃ < 0 on L. Suppose

also that we are given neighborhoods U1 , U2 of L with U1 ⊂ U2 , and Q1 , Q2 ∈
0,0 with WF (Q2 ) ⊂ U2 \ U1 , WF (Q1 ) ⊂ U2 , Q1 elliptic on U1 , and such
that for every α ∈ (Char(P)∩U1 )\L, the bicharacteristic γ of p with γ (0) = α
enters Ell(Q2 ) while remaining in Ell(Q1 ). Then there exists Q ∈ 0,0 elliptic
on L such that if Q2 u ∈ H s,r , Q1 u ∈ H s−1/2,r−1/2 and Q1 Pu ∈ H s−m+1,r−

+1 ,
then Qu ∈ H s,r and for all M, N there is C > 0 such that
QuH s,r ≤ C(Q2 uH s,r + Q1 PuH s−m+1,r−
+1
+ Q1 uH s−1/2,r−1/2 + uH M,N ).
Remark 5.3 In the second part of the above proposition, the ‘there exists Q
elliptic on L’ statement can be replaced by ‘for all Q with WF (Q) ⊂ U1 ’.
In addition, the first part of the proposition can also be rephrased, see the
statement of Proposition 2.3 of [48].
Proof The first part of the proposition will follow immediately from the
construction given below since the support of a can be arranged to be in the
pre-specified set U, while the second part reduces to it taking into account that
by the hypotheses u is microlocally in H s,r on U1 \ L by microlocal elliptic
regularity and the (non-radial) propagation estimates, with an appropriate
estimate holding in this space.
Modulo justifying the pairing argument, the standard calculation, (5.50)
(with possibly a switch of the sign of the Bu2 term) takes care of the s −
(m − 1)/2 + β̃|L > 0 case, with no E term in the sense of the real principal type
result (since the main point there is that WF (E) ∩ Char(P) = ∅, i.e. one needs
to make a priori assumptions on the characteristic set to propagate estimates),
though with another potential error term Ẽ with WF (Ẽ) ∩ Char(P) = ∅ arising
from φ0 (p̂) above, so
ẽ = 2(∓Hp,m,
p̂)φ(ρ1 )2 φ0 (p̂)φ0 (p̂)ρ02s ρ22r ,

and B∗ B replaced by 2j=1 B∗j Bj , with Bj having symbol bj satisfying

b1 = 2β0 s − (m − 1)/2 + β2 (r − (
− 1)/2) + β̃ φ(ρ1 )φ0 (p̂)ρ0s ρ2r ,

b2 = φ(ρ1 )φ1 φ0 (p̂)ρ0s ρ2r .
Here we use the fact that one can make φ supported in a specified neighborhood
of L; in particular in one in which s − (m − 1)/2 + β2 (r − (
− 1)/2) + β̃ is
bounded below by a positive constant since it is s − (m − 1)/2 + β̃|L > 0 on L.
In a similar vein, but now regarding the φ1 term as part of the error e, and still
having an ẽ error term in the elliptic region, in the case s − (m − 1)/2 + β̃ < 0,
we can take

b = 2β0 s − (m − 1)/2 + β2 (r − (
− 1)/2) + β̃ φ(ρ1 )φ0 (p̂)ρ0s ρ2r ,
e = φ(ρ1 )φ12 φ0 (p̂)2 ρ02s ρ22r ,
ẽ = 2(±Hp,m,
p̂)φ(ρ1 )2 φ0 (p̂)φ0 (p̂)ρ02s ρ22r .
308 A. Vasy
In order to justify the argument, we use φt given in (5.55) with τ = ρ0−1 and
K = 2(s − s ). Then
∓Hp,m,
φt (ρ0−1 ) = −ρ0−1 ft (ρ0−1 )φt (ρ0−1 )β0 ,
with
0 ≤ −ρ0−1 ft (ρ0−1 ) ≤ s − s . (5.61)
Thus, with

at = φ(ρ1 )2 φ0 (p̂)2 ρ0−m ρ2−
φt (ρ0−1 )2
the principal symbol −(Hp at + 2p̃at ) of i([At , P] + (P − P∗ )At ) is

∓ρ0−m −m+1 ρ2−
−
+1 φt (ρ0−1 )2 (φ(ρ1 )φ12 + β0 m φ(ρ1 )2 + 2β2 β0
φ(ρ1 )2
+ 2β0 β̃φ(ρ1 )2 − 2(−ρ0−1 ft (ρ0−1 ))β0 ).
Then the above calculation of the commutator is unchanged provided either the
new term 2(−ρ0−1 ft (ρ0−1 ))β0 has the same sign as the previous bj or b terms,
which indeed happens when s − (m − 1)/2 + β̃|L < 0, or if it can be absorbed
in the bj terms. For the latter we need
s − (m − 1)/2 + β2 (r − (
− 1)/2) + β̃ − (−ρ0−1 ft (ρ0−1 )) > 0,
which is satisfied if
s − (m − 1)/2 + β2 (r − (
− 1)/2) + β̃ > 0
in view of (5.61). This proves the proposition, if the amount of regularization

provided suffices for the proof to go through for t > 0. There are no issues
with the argument provided we actually know that
Pu, At u − At u, Pu = (At P − P∗ At )u, u; (5.62)
for the rest of the argument it is straightforward to check that all steps work,
provided we take s = s − 1/2 in the second case, and have s as stated in the
first case. (In fact, in the second case we may simply take s = s − 1, and then
the whole argument goes through directly, including the proof of (5.62); the
same is true in the first case if (s − 1/2) − (m − 1)/2 + β̃ > 0 on L with the
notation of the statement of the proposition.)
So it remains to prove (5.62) for t > 0. The point is that while both sides
make sense by the a priori assumptions, they are not necessarily equal in
principle since for example P∗ At u, u is not defined for u with just the property

that WFs ,r−1/2 (u) is disjoint from WF L∞ ({At }) (a bit stronger assumption is
needed). This, however, is straightforward to overcome by using an additional

regularizer family τ , for then
Pu, At u − At u, Pu = lim (τ Pu, At u − τ At u, Pu)
τ →0
= lim (At τ Pu, u − P∗ τ At u, u)
τ →0
= lim (At τ P − P∗ τ At )u, u,
τ →0
so the argument is finished by showing that (At τ P − P∗ τ At ) → At P − P∗ At

strongly. But
(At τ P − P∗ τ At ) = τ (At P − P∗ At ) + [At , τ ]P − [P∗ , τ ]At ,

and At P − P∗ At ∈ 2s ,2r−1 , so τ (At P − P∗ At ) → (At P − P∗ At ) in

2s +δ,2r−1+δ for δ > 0 and is uniformly bounded in 2s ,2r−1 , while [At , τ ]P,
∗ 2s ,2r−1
[P , τ ]At are uniformly bounded in and converge to 0 in
2s +δ,2r−1+δ
for δ > 0, giving the desired strong convergence. This completes
the proof of the proposition.
As usual, one can iterate the argument and obtain:

Proposition 5.27 Suppose that L is as above, and s − (m − 1)/2 + β̃ > 0
on L, and s − (m − 1)/2 + β̃ > 0 on L and r, r ∈ R. Then there exist

Q ∈ 0,0 elliptic on L such that for Q1 elliptic on WF (Q), if Q1 u ∈ H s ,r ,
Q1 Pu ∈ H s−m+1,r−
+1 , then Qu ∈ H s,r and for all M, N there is C > 0 such
that
QuH s,r ≤ C(Q1 PuH s−m+1,r−

+1 + Q1 uH s ,r + uH M,N ).
On the other hand, suppose now that s − (m − 1)/2 + β̃ < 0 on L. Suppose

also that we are given neighborhoods U1 , U2 of L with U1 ⊂ U2 , and Q1 , Q2 ∈
0,0 with WF (Q2 ) ⊂ U2 \ U1 , WF (Q1 ) ⊂ U2 , and such that for every α ∈
(Char(p) ∩ U2 ) \ L, the bicharacteristic γ of p with γ (0) = α enters Ell(Q2 )
while remaining in Ell(Q1 ). Then there exist Q ∈ 0,0 elliptic on L such that if
Q2 u ∈ H s,r , and Q1 Pu ∈ H s−m+1,r−
+1 , then Qu ∈ H s,r and for all M, N there
is C > 0 such that
QuH s,r ≤ C(Q2 uH s,r + Q1 PuH s−m+1,r−

+1 + uH M,N ).
Notice that radial points are not structurally stable, unlike real principal type
points in the characteristic set, that is where Hp,m,
is non-zero as a vector
field on the boundary of the compactified phase space. Namely, if P ∈ m,
has real principal symbol p, Hp,m,

is non-radial at α then for any P̃ ∈ m,
with real principal symbol p̃ and close to P in the pseudodifferential topology
310 A. Vasy
(i.e. P, P̃ are locally quantizations which are close in a suitable symbol

seminorm), locally near α the characteristic set is near that of P and Hp̃,m,
is
also non-radial on it, and one has the real principal type propagation estimate.
On the other hand, if α is a radial point for P, not only need α not be a radial
point for P̃ (or be in the characteristic set): there may not even be radial points
nearby! Even the generalized version we stated above with a submanifold L
of normal sources/sinks need not apply. However, the estimates we proved are
stable in the sense that if in the proof of Proposition 5.26 a certain choice of φ,
φ0 and powers of ρ0 , ρ2 works, the same φ, φ0 and powers of ρ0 , ρ2 also give
an estimate for P̃ when P̃ is close to P in the above sense, with real principal
symbol. Proposition 5.27 is then similarly stable. See Remark 2.5 of [48].
5.4.8. Fredholm Problems with Radial Points

This gives rise to another non-elliptic Fredholm problem. So suppose that P ∈
m,
with real homogeneous principal symbol. Suppose also that Q ∈ m,
is
such that its principal symbol q satisfies q = t2 as above, WF (Q) ∩ L = ∅,
and that for all α ∈ ∂(Rn × Rn ) in (p) = {p = 0} the integral curve γ of
Hp with γ (0) = α reaches {q < 0} in finite time in the backward direction,
and tends to L in the forward direction. (Although we worked with Rn × Rn
as our compactified phase space, all our arguments extend directly to the full
manifold case, i.e. sc T ∗ X, discussed in Section 5.3.11.) Again let
X s,r = {u ∈ H s,r : (P − iQ)u ∈ H s−m+1,r−

+1 }, Y s,r = H s,r .
Then
P − iQ : X s,r → Y s−m+1,r−
+1 , P∗ + iQ∗ : X s̃,r̃ → Y s̃−m+1,r̃−
+1 (5.63)
are Fredholm for all r and for all s with
s − (m − 1)/2 + β̃|L > 0, s̃ − (m − 1)/2 − β̃|L < 0. (5.64)
Note that this is exactly the condition for “high regularity” H s,r estimates for
P at the radial points, as well as for “low regularity” H s̃,r̃ estimates for P∗ at
these. Further, for
s̃ = m − 1 − s, r̃ =
− 1 − r, (5.65)
the stated Sobolev spaces for P∗ + iQ∗ are the duals of those stated for P − iQ,
that is
H s̃,r̃ = (H s−m+1,r−
+1 )∗ , H s̃−m+1,r̃−
+1 = (H s,r )∗ .
Again, this Fredholm statement follows estimates of the kind
uH s,r ≤ C((P − iQ)uH s−m+1,r−

+1 + uH M,N ),
uH s̃,r̃ ≤ C((P∗ + iQ∗ )uH s̃−m+1,r̃−
+1 + uH M̃,Ñ ),
for M < s, N < r, M̃ < s̃, Ñ < r̃, which in turn follow from the propagation
of singularities and the result at the radial points. Thus, now we just need to

check that if u ∈ H s ,r for some r and with s as above, and (P − iQ)u ∈
H s−m+1,r−
+1 then u ∈ H s,r , that is WFs,r (u) = ∅. But if either p or q is
elliptic at α, then α ∈ / WFs+1,r+1 (u) and thus α ∈ / WFs,r (u) by microlocal
elliptic regularity. Further, there is a neighborhood U of α such that α ∈ U
implies α ∈ / WFs,r (u) by the high-regularity radial set result. Now propagating
the regularity backwards (which is what forces the sign of q), if p vanishes
at α, and thus the forward bicharacteristic γ from α reaches U, where we know
there is no WFs,r (u), so propagation of singularities gives α ∈ / WFs,r (u). For
the adjoint, again we do not need to be concerned with elliptic points. If p
vanishes at α but α ∈ / L, then the backward bicharacteristic from α reaches
q < 0, so the propagation of singularities result gives α ∈ / WFs̃,r̃ (u). Finally,
at L, we use the low-regularity radial set result to complete the proof of the
Fredholm claim.
This can be generalized when the characteristic set of p has two components
± , and in + the radial set L+ is a source, in − it is a sink L− . Namely,
suppose that P ∈ m,
with real homogeneous principal symbol, Q ∈ m,
is
such that its principal symbol q satisfies q = t+ 2 − t2 with supp t ∩ = ∅,
− + −
supp t− ∩ + = ∅, and that for all α ∈ ∂(Rn × Rn ) in + , resp. − , the
integral curve γ of Hp with γ (0) = α reaches {q > 0}, resp. {q < 0}, in finite
time in the forward, resp. backward, direction, and tends to L+ , resp. L− , in the
backward, resp. forward, direction. Then P − iQ, P∗ + iQ∗ are Fredholm as in
(5.63) subject to (5.64).
This setup suffices for instance for the analysis of scattering theory on
asymptotically hyperbolic spaces, which we will discuss in Section 5.5 as some
geometric preliminaries are required.
But first, we discuss the geometrically simpler asymptotically Euclidean
spaces in which no artificial complex absorption is needed, but for which
instead we need to use variable order Sobolev spaces. For simplicity, let us
take g to be a scattering metric actually asymptotic to the Euclidean metric,
that is

g= gij (z) dzi ⊗ dzj ,
ij
312 A. Vasy
with gij = C ∞ (Rn ), with dual metric

G= Gij (z) dzi ⊗ dzj ,
ij
with Gij = C ∞ (Rn ) still, and with

gij − δij ∈ z−1 C ∞ (Rn ).
As discussed earlier, the principal symbol of
P = g + V − σ ,
V ∈ z−1 C ∞ (Rn ) real-valued, is

Gij ζi ζj − σ ,
modulo S1,−1 . Thus, regardless of σ , P is elliptic at fiber infinity; all possible

issues are at the characteristic set, at finite ζ . If σ ∈ C \ [0, ∞), we have full
ellipticity, as we have already discussed, so consider σ > 0 (with σ = 0 being
a degenerate case). It is thus convenient to work in the homogeneous in the
base framework; see Figure 5.6.
Since Gij = δij at base infinity, the Hamilton vector field there is then

Hp = 2 ζj ∂zj ,
Figure 5.6. The characteristic set for σ > 0. From the compactified perspective, it
is a subset of base infinity, ∂Rn × Rn , denoted by : it is an (n − 1)-sphere bundle
over the sphere at infinity, ∂Rn , thus a torus if n = 2; if n = 1 it is disconnected as
shown, otherwise connected. From the homogeneous-in-the-base perspective, it is
a subset of (Rn \{0})×Rn , denoted by Char(P) in the picture, with the parentheses
at {0} × Rn showing that this, that is the fiber of the cotangent bundle at 0, is not
part of the homogeneous picture
Figure 5.7. The characteristic set and the bicharacteristics for σ > 0 in the
compactified picture; the figure is accurate for n = 2 when the characteristic set
is a torus, on the picture shown cut along two circles, so the left and right side,
resp. the top and bottom side, are identified. Here L− is a source in the normal
directions, while L+ is a sink
and Hp is radial on
L± = {(z, ζ ) : z = cζ , |ζ |2 = σ , ±c > 0}.
Since the bicharacteristics are straight lines in Rnz , t → (z0 + 2tζ0 , ζ0 ), with ζ
constant along them, it is easy to see that (in the compactified picture: recall
that parallel lines tend to the same point at infinity, i.e. ∂Rn !) they approach L±
as t → ±∞, and indeed that all the assumptions at L± required above hold;
see Figure 5.7.
Thus, if n ≥ 2 the bicharacteristics connect the two components of the radial
set L± (so L± lie inside the same component of the characteristic set); if n = 1
the characteristic set is disconnected. For the analysis, we note that, with
= 0,
β̃ = 0 (by the formal self-adjointness assumption due to the reality of V), if
r− − (
− 1)/2 + β̃|L− > 0, that is we are in the high regularity regime at

L− , then we can propagate H s,r− estimates out of L− for elements of H s ,r
with r − (
− 1)/2 + β̃|L− > 0. These can be propagated to a punctured
neighborhood of L+ . Now at L+ we can propagate then to L+ itself in H s+ ,r
provided r+ − (
− 1)/2 + β̃|L+ < 0, that is we are in the low regularity
regime. Thus, we need spaces with differing regularity at L± . Notice that the
adjoint operator, on the dual spaces corresponding to these microlocal Sobolev
spaces, will then proceed in the opposite direction, starting with high regularity
estimates from L+ , propagating them to a punctured neighborhood at L− , and
then propagating these to L− itself in the low regularity regime. (Of course,
we could have reversed the choice of the function spaces for the operator,
314 A. Vasy
making the regularity high at L+ ; then the high regularity would be at L− for
the adjoint.)
The geometric generalization is, as already discussed, Riemannian scattering
metrics, see Proposition 5.23, or more generally P = H − σ = g + V − σ ,
V ∈ S−1 (X) real valued. One can then compute the bicharacteristics in terms of
those of the boundary metric h; this was done by Melrose [30]. The conclusion
is that there are again radial sets L± , with the flow connecting them, that is
L+ is a sink, L− a source, in the normal directions. The threshold regularity
is (
− 1)/2 = −1/2. Choosing r so that r > −1/2 at L− , r < −1/2 at
L+ or vice versa, gives rise to Fredholm problems. In this case one can show
that in either case first Ker(H − σ ) ⊂ C˙∞ (X) (with the kernel understood as
s,r s−2,r
the operator acting Hsc (X) → Hsc (X)), and next that the kernel is actually
trivial; without this one would in principle have a finite dimensional kernel.
(The triviality of the kernel is called the “absence of embedded eigenvalues”
in scattering theory, corresponding to [0, +∞) being the continuous spectrum
of H, so the eigenvalues would be embedded in the continuous spectrum. We
do not discuss this here, but refer to [30, Sections 10–13] for a discussion in
the present setting. Conceptually, this absence is a combination of a positivity
property, described in [50] more generally, obtained via a boundary pairing
argument in [30], which allows one to conclude that the kernel is in C˙∞ (X), and
a unique continuation argument at infinity, obtained via a Carleman estimate,
as in [27, Theorem 17.2.8].) Thus,
Proposition 5.28 Let r ∈ C ∞ (sc T∂X ∗ X) be real valued, monotone in the
∗
characteristic set ⊂ T∂X X of P along the Hp flow and satisfying r > −1/2
sc
resp. r < −1/2 at precisely one of L+ and L− . Then

P : {u ∈ Hsc
s,r
(X) : Pu ∈ Hsc
s−2,r+1
(X)} → Hsc
s−2,r+1
(X) (5.66)
is invertible.
Further, if r > −1/2 at L− , then
P−1 = (H − (σ + i0))−1 = lim (H − (σ + i))−1 ,
→0
while if r > −1/2 at L+ then P−1 = (H − (σ − i0))−1 .

All statements except the triviality of the kernel (which we do not discuss
here) and the last claim have already been proved. As for the last claim, first
note that
H − (σ + i) : Hsc
s,r
(X) → Hsc
s−2,r
(X)
is invertible for > 0 by Proposition 5.23. Further, the propagation estimates
giving rise to the Fredholm statement in (5.66) also hold for P replaced by
H −(σ +i), with i acting as complex absorption, uniformly in ∈ [0, 1]. This
is how the limiting absorption principle corresponds to function spaces: > 0
necessitates propagating the estimates forward along the Hamilton flow of P,
that is r being larger than the threshold value at L− . Then a simple argument
shows that in fact lim→0 (H − (σ + i))−1 exists and is equal to P−1 on the
corresponding space. The case of (H − (σ − i0))−1 is analogous.
A very similar setup is applicable for the Klein–Gordon operator g − m2 ,
m > 0, on Minkowski space (with the signature of g being (1, n − 1)), or on
manifolds with Lorentzian scattering metrics, a notion introduced in [2]. (In [2]
the m = 0 problem was considered, which we discuss here in Section 5.6 when
the b-analysis tools are developed.)
We finally mention the Fredholm context of the stability statement made
after Proposition 5.27. Namely, the stability of the estimates implies the
stability of the Fredholm statement, that is that if P fits into any of the
Fredholm frameworks discussed above, then so does P̃ if P̃ is assumed to have
real principal symbol and is sufficiently close to P in the pseudodifferential
topology. In fact, it is not hard to show, see Section 2.7 of [48], that if the
inverse P−1 exists, and one has a continuous family w → P̃(w) of pseudod-
ifferential operators with P̃(w0 ) = P, then for w near w0 , P̃(w) is invertible
and the inverse depends continuously on w in the weak operator topology
of L(Y s−m+1,r−l+1 , X s,r ). Thus, our Fredholm framework is very stable in
the context of pseudodifferential operators with real principal symbols. In
particular, these remarks apply to the Klein–Gordon equation for Lorentzian
metrics g which are close to the Minkowski metric in a symbolic sense, and
provide a Fredholm framework even more generally than the setting of the
Lorentzian scattering metrics of [2] which needed extra properties for the more
precise asymptotic analysis presented there.
5.5. Conformally Compact Spaces

5.5.1. The Geometric Problem
We now show how the microlocal results of the previous sections give the
meromorphic extension of the resolvent of the Laplacian on even conformally
compact (asymptotically hyperbolic) spaces. Overall we follow [47] closely,
but we de-emphasize the estimates outside strips (in the “physical half plane,”
Im σ > 0). This has the benefit that conjugation by a σ -dependent power of
1 + μ becomes irrelevant, and thus can be dropped, simplifying the connection
between the conformally compact and the extended problems. The downside is
316 A. Vasy
that while one obtains a Fredholm family whether the geometry is trapping or
not, its invertibility at one point (thus the meromorphy of the inverse) follows
only in the non-trapping setting when the large σ (in a strip) estimates hold.
Thus, at the end of the section I present the additional conjugation, giving the
full Mazzeo–Melrose and Guillarmou [29, 15] result in the setting of even
metrics. In addition, I also refer to Zworski’s expository article [57] for a
particularly accessible treatment of part of the framework discussed below.
We start by recalling the definition of manifolds with even conformally
compact metrics. These are Riemannian metrics g0 on the interior of an n-
dimensional compact manifold with boundary X0 such that near the bound-
ary Y, with a product decomposition nearby and a defining function x, they are
of the form
dx2 + h
g0 = (5.67)
x2
where h is a family of metrics on Y = ∂X0 depending on x in an even
manner, that is only even powers of x show up in the Taylor series, that is
h = h(x2 , y, dy) with smooth dependence on x2 . (There is a much more natural
way to phrase the evenness condition due to Guillarmou [15].) It is convenient
to take x to be a globally defined boundary defining function. Then the dual
metric is
G0 = x2 (∂x2 + H),
with H the dual metric family of h (depending on x as a parameter), and

|dg0 | = | det g0 | dx dy = x−n | det h| dx dy
so
g0 = (xDx )2 + i(n − 1 + x2 γ )(xDx ) + x2 h , (5.68)
with γ even, and h the x-dependent family of Laplacians of h on Y. Below
we consider the spectral family
(n − 1)2
g0 − − σ2
4
of the Laplacian.
We start by noting that, regardless of the evenness of h as a function of x, g0
is a complete Riemannian metric on X0 , and thus we have, writing L02 (X0 ) for
the L2 space of g0 , due to Chernoff’s theorem:
2
Proposition 5.29 The operator g0 − (n−1)
4 is essentially self-adjoint on
˙ (n−1)2
C (X0 ), and thus R(σ ) : (g0 − 4 − σ 2 )−1 : L02 (X0 ) → L02 (X0 ),
∞
Im σ > 0, σ ∈
/ i[0, n−1
2 ].
As an aside, the natural Sobolev spaces for the analysis of g0 are
those corresponding to vector fields of finite length with respect to g0 with
appropriate smoothness properties. Thus, letting V0 (X0 ) = xV(X0 ) (recall that
x is a boundary defining function), for integer s ≥ 0,
H0s (X0 ) = {u ∈ L02 (X0 ) : ∀k ≤ s, V1 . . . Vk u ∈ L02 (X0 ), Vj ∈ V0 (X0 )},
where we again wrote L02 (X0 ) for the L2 space of g0 . For negative integer s
the spaces can be defined via dualization, and then for s ∈ R by interpolation.
An alternative way of defining H0s (X0 ) is via the 0-pseudodifferential algebra
developed by Mazzeo and Melrose in [29]. In any case, the domain of g0 then
is H02 (X0 ), and for σ as indicated above,
−1
(n − 1)2
R(σ ) : g0 − −σ 2
: H0s (X0 ) → H0s−2 (X0 ),
4
for all s ∈ R.
The question one asks is if there is a meromorphic extension of R(σ ) to C.
This was essentially proved by Mazzeo and Melrose [29], with an important
improvement by Guillarmou [15] regarding the potential essential singularities:
Theorem 5.5 ([29, 15]) Without assuming that h is even, the operator R(σ ),
2
given by (g0 − (n−1) 4 − σ 2 )−1 for Im σ > (n − 1)/2, has a meromorphic
continuation to C \ {−i(k + 1/2), k ∈ Z, k ≥ 0}, that is after one removes the
pure imaginary half-integer essential singularities, as an operator Cc∞ (X0◦ ) →
C ∞ (X0◦ ), thus its Schwartz kernel continues meromorphically as a distribution.
Further, if h is even, then actually there are no essential singularities, that is
the analytic continuation is meromorphic on C.
The goal of this section is to prove this theorem, and to provide uniform
control for R(σ ) as |σ | → ∞ in strips | Im σ | < C. The latter necessitates a
rather different approach to that of Mazzeo, Melrose, and Guillarmou, based
on extension of the operator across ∂X0 in an appropriate sense and using the
microlocal machinery that we have developed.
A particularly interesting point is the following: Mazzeo and Melrose [29]
in fact show that for f ∈ C˙∞ (X0 ), that is for f vanishing to infinite order at ∂X0 ,
R(σ )f is of the form
n−1
R(σ )f = x 2 −iσ u, u ∈ C ∞ (X0 ),
if Im σ > 0 and σ is not a pole of R(.), and indeed for arbitrary σ which is not
a pole of R(.) provided 2σ ∈/ iZ. (In the case of integer coincidences, 2σ ∈ iZ,
one has additional logarithmic terms; if g is even then the integer coincidence
318 A. Vasy
condition is σ ∈ iZ.) Note that if Im σ > 0, then R(σ )f ∈ L2 as Re(−iσ ) > 0,

but not if Im σ ≤ 0. Thus, for σ with Im σ > 0, u± = R(±σ )f both solve

(n − 1)2
g − − σ u± = f ;
2
4
the difference is that u+ is the unique L2 -solution. This shows the difficulty of
studying the analytic continuation because u− cannot be so easily characterized
without the full asymptotics: one cannot say that “the L2 -term is missing” since
it is dominated by the growing asymptotic, x(n−1)/2+iσ C ∞ (X0 ), and further if
Im σ changes sign the role of the two asymptotics (which is L2 and which is
not) reverses.
In the method we develop, the way of distinguishing between u+ and u−
is rather different. We extend a modification of the operator to an extension
X of the manifold X0 endowed with the “even smooth structure” (only even
functions on X0 are considered smooth) across the boundary, and we do the
same to u± after multiplying u± by x−(n−1)/2+iσ . Then x−(n−1)/2+iσ u+ extends
to be C ∞ across ∂X0 , but x−(n−1)/2+iσ u− does not. Thus, in the extended
picture we develop, the way of distinguishing between solutions is smoothness
or lack thereof, regardless of the sign of Im σ , that is smoothness rather than
decay. Ultimately this allows for obtaining the Fredholm framework that plays
a crucial role in the large parameter estimates.
Before proceeding with the analytic discussion we mention a closely related
geometric problem that plays a major role below at least indirectly. Even
asymptotically de Sitter spaces are manifolds with boundary X̃0 (with boundary
defining function x̃) equipped with a Lorentzian metric g̃0 in the interior for
which a neighborhood of the boundary has the form
dx̃2 − h̃
g̃0 = ,
x̃2
with h̃ = h̃(x̃2 , .) a family of metrics on Y = ∂ X̃0 , which further satisfy that
Y = Y+ ∪ Y− (disjoint union of connected components) and all null-geodesics
of g̃0 tend to either Y+ in the forward direction along the Hamilton flow and
to Y− in the backward direction, or vice versa. While not apparent from the
definition, these spaces are necessarily of the form Y+ × I, I = [−1, 1] a
compact interval, see [53] and the arguments of Geroch [13], and indeed the
spacetime is globally hyperbolic, so if the (time-like) variable on I is denoted
by T, with T = 1 − x̃ near Y+ , T = −1 + x̃ near Y− , then the Cauchy problem
at T = T0 ∈ (−1, 1) for the Klein–Gordon equation

(n − 1)2
g̃0 − −σ u=f
2
4
is well-posed. Further, for C ∞ Cauchy data, the solution has an asymptotic

expansion at T = ±1, as shown in [53]:
n−1 n−1
u = x̃ 2 +iσ u+ + x̃ 2 −iσ u− , u± ∈ C ∞ (X̃0 ),
and indeed u± are even, as long as σ ∈ / iZ (there are potential logarithmic
terms otherwise as for the asymptotically hyperbolic case above).
Now notice that if g is even asymptotically hyperbolic, and one replaces x
by ix formally (called Wick rotation), one obtains
dx2 − h(−x2 , y, dy)
,
x2
which is asymptotically de Sitter, and vice versa. Here of course h needs to
be extended smoothly to negative values of the argument. One way to think of
the analytic constructions in Section 5.5.3 is that they give a precise setting in
which Wick rotations are meaningful. Notice that if h is not even, but rather
has the form h (x, y, dy), one would get h (ix, y, dy), which would require an
analytic continuation, rather than a smooth extension, to make sense.
5.5.2. Large Parameter and Semiclassical

Pseudodifferential Operators
As discussed above, in addition to working with finite σ , or σ in a compact
set, we also want to consider σ → ∞, mostly in strips, with | Im σ | bounded.
In that case we should consider σ as a “large parameter.” In general, on Rn ,
the large parameter setting just means that one has a family of operators given
by symbols a on Rnz × Rnz × Rnζ × σ , ⊂ C, satisfying, for example in the
scattering setting, estimates
|Dαz Dz Dζ Dkσ a| ≤ Cαβγ ,k z
1 −|α| z
2 −|β| (ζ , σ )m−|γ |−k ,
β γ
that is there is joint symbolic behavior in (ζ , σ ), with differentiation in either

giving rise to joint decay. This is natural, as a typical way such a family of
operators arises is by Mellin transforming the normal operator of a b-operator;
then σ is the b-dual of the boundary defining function. For instance, starting
with the elliptic b-operator on [0, ∞)x × X
L = (xDx )2 + g ,
with g a Riemannian metric on X, the Mellin transform gives rise to the family
σ 2 + g , and the (joint symbolic) large parameter behavior is a consequence
of the fact that the b-principal symbol of L is a symbol (concretely a quadratic
polynomial).
320 A. Vasy
Since σ is simply a parameter as far as the action of pseudodifferential

operators is concerned, in order to have a well-behaved algebra (so that e.g.
left and right quantization are equivalent), one simply needs to make sure that
the various asymptotic expansions are now asymptotic in this large parameter
sense (i.e. jointly in (ζ , σ )), which is straightforward to check. Thus, the
composition of order (m,
) and (m ,
) large parameter operators is order
(m + m ,
+
) as a large parameter operator, with principal symbol given
by the product of the principal symbols, and so on Notice that the natural
compactification on which the symbols are well behaved is not Rnz × Rnζ × C
(with the image of inserted into the last factor), but Rnz × Rnζ × C, that is
(ζ , σ ) are jointly compactified. Correspondingly, in the manifold setting, one
needs to take T ∗ X × C, best thought of as the vector bundle (T ∗ ⊕ C)X over X,
and then radially compactifying the fibers to obtain T ∗ ⊕ CX.
Below we are interested in the setting of compact manifolds, which means
that the behavior as |z| → ∞ is irrelevant, so we could just as well use ∞ -
type estimates in our definition of the large parameter algebra. To make things
concrete for differential operators, if

P(σ ) = aα (z)σ β Dαz
|α|+|β|≤m
is an order m differential operator depending on a large parameter σ , then the

large-parameter symbol is denoted by

σfull (P(σ )) = aα (z)σ β ζ α .
|α|+|β|=m
It is often convenient to relax the requirements a bit on the joint symbolic

behavior, and consider the semiclassical operator algebra, which we at first
simply pull out of a hat. (We refer to [58] for a thorough treatment.) We
adopt the convention that h̄ denotes semiclassical objects, while h is the actual
semiclassical parameter. This algebra, ∪m,
h̄m,
(Rn ), is given by

−n
Ah = Oph̄ (a); Oph̄ (a)u(z) = (2πh) ei(z−z )·ζ /h a(z, ζ , h) u(z ) dζ dz ,
Rn ×Rn
∞
u ∈ S(R ), a ∈ C ([0, 1)h ; Sm,
(Rnz ; Rnζ ));
n
(5.69)
its classical subalgebra,h̄,cl (Rn ) corresponds to a ∈ C∞ ([0, 1)h ; Scl m,
(Rnz ; Rnζ )).

More generally, we write

−n
Ih̄ (a)u(z) = (2πh) ei(z−z )·ζ /h a(z, z , ζ , h) u(z ) dζ dz ,
Rn
a ∈ C ∞ ([0, 1)h ; Sm,

1 ,
2 (Rnz , Rnz ; Rnζ )). A straightforward computation gives,
with ι : Rnz × Rnζ → Rnz × Rnz × Rnζ the inclusion map as the diagonal (z = z )
in the first two factors, the left reduction formula
i|α|
aL ∼ ι∗ Dαz (hDζ )α a, (5.70)
α
α!
and the right reduction formula

(−i)|α|
aR ∼ ι∗ Dαz (hDζ )α a,
α
α!
where the factors of h arise in front of Dζ since one has
Ih̄ ((z − z )α a) = Ih̄ ((hDζ )α a)
as

(z − z )ei(z−z )·ζ /h = hDζ ei(z−z )·ζ .
This provides that the αth term in (5.70) is in h|α| Sm−|α|,

1 +
2 −|α| , that is the
terms not only become lower order as symbols, but also have higher order
vanishing as h → 0.
The semiclassical principal symbol is now σh̄,m,
(A) = a|h=0 ∈ Sm,
(Rn ×
R ); there is still the standard principal symbol [a] ∈ C ∞ ([0, 1)h ; Sm,
/Sm−1,
−1 )
n
which is now a function depending on the parameter h. As usual, the classical

setting is best encoded in terms of a compactification (or bordification if h = 1
is not added!)
[0, 1)h × Rnz × Rnζ ,
which has now three boundary hypersurfaces:
{0}h × Rnz × Rnζ ,
carrying the semiclassical principal symbol,
[0, 1)h × ∂Rnz × Rnζ ,
carrying the scattering symbol at base-infinity (which is irrelevant when we

transfer to compact manifolds without boundary) and
[0, 1)h × Rnz × ∂Rnζ ,
carrying the usual principal symbol at fiber infinity.

In the setting of a general manifold X, Rn × Rn is replaced by T ∗ X.
Correspondingly, WF h̄ (A) and Ellh̄ (A) are subsets of T ∗ X. We can add an extra
322 A. Vasy
parameter λ ∈ O, so a ∈ C ∞ ([0, 1)h ; Sm (Rn × O; Rnζ )); then in the invariant

setting the principal symbol is a|h=0 ∈ Sm (T ∗ X × O).
In order to motivate the definition a posteriori, notice that differential
operators now take the form

Ah,λ = aα (z, λ; h)(hDz )α . (5.71)
|α|≤m
Indeed, if a is actually a polynomial in ζ , depending on a parameter λ, in (5.69),

namely

ah,λ = aα (z, λ; h)ζ α ,
|α|≤m
then letting ζ̃ = ζ /h, we have

(Oph̄ (a)u)(z) = (2π )−n ei(z−z )·ζ̃ aα (z, λ; h)(hζ̃ )α u(z ) dz dζ̃
Rn ×Rn |α|≤m

= α
aα (z)((hDz ) u)(z).
|α|≤m
The two principal symbols (ignoring base-infinity) are the standard one (but
taking into account the semiclassical degeneration, i.e. based on (hDz )α rather
than Dαz ), which depends on h and is homogeneous, and the semiclassical one,
which is at h = 0, and is not homogeneous:

σm (Ah,λ ) = aα (z, λ; h)ζ α ,
|α|=m

σh̄ (Ah,λ ) = aα (z, λ; 0)ζ α .
|α|≤m
However, the restriction of σm (Ah,λ ) to h = 0 is the principal symbol of

σh̄ (Ah,λ ). In the special case in which σm (Ah,λ ) is independent of h (which is
true in the setting considered below), one can simply regard the usual principal
symbol as the principal part of the semiclassical symbol.
In terms of the fiber compactification of T ∗ X, in the semiclassical context
one considers T ∗ X × [0, 1) and notes that “classical” semiclassical operators
of order 0 are given locally by Oph̄ (a) with a extending to be smooth up to
the boundaries of this space, with the semiclassical symbol given by restriction
to T ∗ X × {0}, and the standard symbol given by restriction to S∗ X × [0, 1).
Thus, the claim regarding the limit of the semiclassical symbol at infinity is
simply a matching statement of the two symbols at the corner S∗ X × {0} in this
compactified picture.
We can convert a large parameter operator into a semiclassical one (but

not conversely) as follows. If P(σ ) = β α
|α|+|β|≤m aα (z)σ Dz is an order m
differential operator depending on a large parameter σ , then letting σ = h−1 λ,
where h−1 ∼ σ (e.g. one may take h−1 = |σ |, but this is often too restrictive),
so λ is in a compact set, and where we restrict to |σ | > 1, say,

Ph̄,λ = hm P(σ ) = hm−|α|−|β| aα (z)λβ (hDz )α
|α|+|β|≤m
is a semiclassical differential operator with semiclassical symbol

σh̄ (Ph̄,λ ) = aα (z)λβ ζ α .
|α|+|β|=m
Note that the full large-parameter symbol and the semiclassical symbol are “the
same,” that is they are simply related to each other.
The semiclassical Sobolev norms are defined analogously to the standard
ones. Namely, on Rn , they are defined for h-dependent families of elements of
the standard Sobolev space as follows. For s ∈ R, uh ∈ H s , h ∈ (0, h0 ),
uHh̄s = Fh̄−1 (.s (Fh̄ u)(.))L2 = (2π )−n/2 .s (Fh̄ u)(.)L2 ,
where Fh̄ is the semiclassical Fourier transform

−n/2
(Fh̄ u)(ζ ) = h e−iz·ζ /h u(z) dz,
Rn
with inverse

(Fh̄−1 v)(z) = (2π )−n h−n/2 eiz·ζ /h v(ζ ) dζ .
Rn
Thus, for s ≥ 0 integer, the squared Hh̄s -norm is equivalent to

(hD)α u2L2 .
|α|≤s
Thus, the semiclassical norms are very similar to the standard ones, but the
derivative estimates become weaker as h → 0 in that each derivative comes
with a factor of h. The weighted norms are then
uH s,r = zr uHh̄s .

h̄
We say that a family u = (uh ), h ∈ (0, h0 ), is in Hh̄s if uh Hh̄s is uniformly

bounded for h ∈ (0, h0 ). Variable order semiclassical Sobolev spaces as well
as semiclassical Sobolev spaces on manifolds can be introduced analogously
to the standard case.
324 A. Vasy
There are complete analogues of the microlocal results discussed so far in

the semiclassical setting. For instance, the analogue of Proposition 5.24 is
Proposition 5.30 For α0 ∈ Rn × Rn with Hp,m,
α0 has a neighborhood O in Rn × Rn on which Hp,m,
is non-vanishing, and
if α ∈ O, γ the integral curve of Hp,m,
γ (t) = β for some t < 0 such that γ ([t, 0]) ⊂ O, and further if we are given a
neighborhood U2 of β and U1 of γ ([t, 0]) contained in O, Q1 , Q2 ∈ h̄0,0 such
that Q1 is elliptic on U1 , Q2 is elliptic on U2 , then there exists Q3 ∈ h̄0,0
s−1/2,r−1/2
elliptic on γ ([t, 0]) such that the following holds. If Q1 u ∈ Hh̄ ,
Q1 Pu ∈ Hh̄s−m+1,r−
+1 , Q2 u ∈ Hh̄s,r then Q3 u ∈ Hh̄s,r and for all M, N, K there
are C > 0, h0 > 0 such that
Q3 uH s,r ≤ C(Q2 uH s,r + h−1 Q1 PuH s−m+1,r−
+1
h̄ h̄ h̄
+ hQ1 uH s−1/2,r−1/2 + hK uH M,N ).

h̄ h̄
The analogous result also holds for t > 0.

The proof is completely analogous to that of Proposition 5.24, with the
point being that the commutator of two operators has an additional h vanishing
relative to the product, hence the h−1 in front of Pu in the estimate, while the
error term has an additional order of vanishing, hence the h in front of the
s−1/2,r−1/2
Hh̄ norm of Q1 u.
This in turn gives rise to
m,
Theorem 5.6 Suppose that P ∈ h̄,cl with real principal symbol p. Suppose
that B, G, Q ∈ h̄0,0 , WF h̄ (B) ⊂ Ellh̄ (G) and for every α ∈ WF h̄ (B)∩Charh̄ (P),
there is a point α = γ (t ) on the bicharacteristic γ through α with γ (0) = α
such that α ∈ Ellh̄ (Q) and such that for t ∈ [0, t ] (or t ∈ [t , 0] if t < 0),
γ (t) ∈ Ellh̄ (G). Then for any M, N, K, there is C > 0 such that if Qu ∈ Hh̄s,r ,
GPu ∈ Hh̄s−m+1,r−
+1 then Bu ∈ Hh̄s,r and
BuH s,r ≤ C(QuH s,r + h−1 GPuH s−m+1,r−

+1 + hK uH M,N ).
h̄ h̄ h̄ h̄
The analogous conclusion also holds in the variable order setting if either
s, r are non-increasing along the Hamilton flow and t < 0 or s, r are non-
decreasing along the Hamilton flow and t > 0.
The estimates involving complex absorption and radial points change
analogously, for similar reasons.
5.5.3. From the Laplacian to the Extended Operator

We show now that if we change the smooth structure on X0 by declaring that
only even functions of x are smooth, that is introducing μ = x2 as the boundary
defining function, then after a suitable conjugation and division by a vanishing
factor the resulting operator smoothly and non-degenerately continues across
the boundary, that is continues to X−δ0 = (−δ0 , 0)μ × Y # X0,even , where X0,even
is the manifold X0 with the new smooth structure. At the level of the principal
symbol, that is the dual metric, the conjugation is irrelevant, so we can easily
see what happens: changing to coordinates (μ, y), μ = x2 , as x∂x = 2μ∂μ ,
G0 = 4μ2 ∂μ2 + μH = μ(4μ∂μ2 + H),
so after dividing by μ, we obtain
μ−1 G0 = 4μ∂μ2 + H.
This is a quadratic form that is positive definite for μ > 0, is Lorentzian for
μ < 0, and has a transition at μ = 0 that as we shall see involves radial points.
To see that the full spectral family of the Laplacian is well behaved, first,
changing to coordinates (μ, y), μ = x2 , we obtain
g0 = 4(μDμ )2 + 2i(n − 1 + μγ )(μDμ ) + μh . (5.72)
Now we conjugate by μ−iσ/2+(n+1)/4 to obtain

(n − 1)2
μiσ/2−(n+1)/4 g0 − − σ 2 μ−iσ/2+(n+1)/4
4

= 4(μDμ − σ/2 − i(n + 1)/4)2 + 2i(n − 1 + μγ ) μDμ − σ/2 − i(n + 1)/4
(n − 1)2
+ μh − − σ2
4
= 4(μDμ )2 − 4σ (μDμ ) + μh − 4i(μDμ ) + 2iσ − 1

+ 2iμγ μDμ − σ/2 − i(n + 1)/4 .
Next we multiply by μ−1/2 from both sides to obtain

(n − 1)2
μ−1/2 μiσ/2−(n+1)/4 g0 − − σ 2 μ−iσ/2+(n+1)/4 μ−1/2
4
= 4μD2μ − μ−1 − 4σ Dμ − 2iσ μ−1 + h − 4iDμ + 2μ−1 + 2iσ μ−1 − μ−1

+ 2iγ μDμ − σ/2 − i(n − 1)/4

= 4μD2μ − 4σ Dμ + h − 4iDμ + 2iγ μDμ − σ/2 − i(n − 1)/4 .
(5.73)
326 A. Vasy
This operator is in Diff2 (X0,even ) and now can be continued smoothly across
the boundary to an operator P̃σ , by extending h and γ in an arbitrary smooth
manner. This form suffices for analyzing the problem for σ in a compact set,
or indeed for σ going to infinity in a strip near the reals. Later we discuss a
modification that is necessary for semiclassical ellipticity when σ is far away
from the reals in Im σ > 0. Since the details of the operator (5.73) do not
matter much, it is convenient to write
P̃σ = 4(1 + ã1 )μD2μ − 4(1 + ã2 )σ Dμ − ã3 σ 2 + h
(5.74)
− 4iDμ + b̃1 μDμ + b̃2 σ + c̃1
with ãj smooth, real, and vanishing at μ = 0, b̃j and c̃1 smooth. In fact, we have
ã1 ≡ 0, but it is sometimes convenient to have more flexibility in the form of
the operator since this means that we do not need to start from the relatively
rigid form (5.68). For the purposes of finite σ behavior this can be further
simplified to
P̃σ = 4(1 + ã)μD2μ − 4(σ + i + μb̃(σ ))Dμ + h + c̃(σ ), (5.75)
with ã smooth, real, and vanishing at μ = 0, independent of σ , b̃, c̃ being

smooth, including in σ -dependence, which in fact is holomorphic.
Writing covectors as
ξ dμ + η dy,
the principal symbol of P̃σ ∈ Diff2 (X−δ0 ), including in the high energy sense
(σ → ∞), is
p̃full = 4(1 + ã1 )μξ 2 − 4(1 + ã2 )σ ξ − ã3 σ 2 + |η|2μ,y , (5.76)
and is real for σ real. The Hamilton vector field is

Hp̃full = 4(2(1 + ã1 )μξ − (1 + ã2 )σ )∂μ + H̃|η|2μ,y

∂|η| 2
∂ ã1 ∂ ã2 ∂ ã3 μ,y
− 4 1 + ã1 + μ ξ2 − 4 σξ + σ2 + ∂ξ
∂μ ∂μ ∂μ ∂μ
∂ ã ∂ ã2 ∂ ã3 2
1
− 4 μξ 2 − 4 σξ − σ ∂η ,
∂y ∂y ∂y
(5.77)
where H̃ indicates that this is the Hamilton vector field in T ∗ Y, that is with
μ considered a parameter. Correspondingly, the standard, “classical,” principal
symbol is
p̃ = σ2 (P̃σ ) = 4(1 + ã1 )μξ 2 + |η|2μ,y , (5.78)
which is real, independent of σ , while the Hamilton vector field is
Hp̃ = 8(1 + ã1 )μξ ∂μ + H̃|η|2μ,y

∂ ã1 ∂|η|2μ,y ∂ ã1 2 (5.79)
− 4 1 + ã1 + μ ξ +
2
∂ξ − 4 μξ ∂η .
∂μ ∂μ ∂y
In order to proceed in computing the principal symbol of P̃σ − P̃∗σ , it is useful

to keep in mind that as g0 − σ 2 − (n − 1)2 /4 is formally self-adjoint relative
to the metric density |dg0 | for σ real, so the same holds for μ−1/2 (g0 − σ 2 −
(n − 1)2 /4)μ−1/2 (as μ is real), and indeed for its conjugate by μ−iσ/2 for
σ real since this is merely unitary conjugation. As for f real, A formally self-
adjoint relative to |dg0 |, f −1 Af is formally self-adjoint relative to f 2 |dg0 |, we
then deduce that for σ real, P̃σ is formally self-adjoint relative to
1
μ(n+1)/2 |dg0 | = |dh| |dμ|,
2
as x−n dx = 12 μ−(n+1)/2 dμ. Note that μ(n+1)/2 |dg0 | thus extends to a C ∞

density to X−δ0 , and we deduce that with respect to the extended density,
σ1 ( 2i1 (P̃σ − P̃∗σ ))|μ≥0 vanishes when σ ∈ R. Since in general P̃σ − P̃Re σ
differs from −4i(1 + a2 ) Im σ Dμ by a zeroth order operator, we conclude that

1
σ1 (P̃σ − P̃σ )
∗
= −4(Im σ )ξ . (5.80)
2i μ=0
In fact, it is not important that P̃σ is related to g0 − σ 2 − (n − 1)2 /4 in

this manner as shown by an explicit computation, or simply by noting that the
coefficients ãj do not affect the result of this computation.
In the interior, away from the region of validity of the product decomposition
(5.67) (where we had no requirements so far on μ), one can choose μ
arbitrarily, without affecting our arguments. Indeed, the classical principal
symbol is unaffected by the choice, and for the semiclassical principal symbol,
for real z, with σ = z/h, one simply has a conjugation by e−iσ φ where
eφ = μ1/2 , which does not affect any of the relevant properties.
While we have not considered vector bundles over X0 , for instance for the
Laplacian on the differential form bundles, it is straightforward to check that
slightly changing the power of μ in the conjugation the resulting operator
extends smoothly across ∂X0 , has scalar principal symbol of the form (5.76),
and the principal symbol of 2i1 (P̃σ − P̃∗σ ), which plays a role below, is also as
in the scalar setting, so all the results in fact go through; see [46].
328 A. Vasy
Notice that if one starts with an even asymptotically de Sitter space (X̃0 , g̃0 ),
with boundary defining function x̃, and considers
(n − 1)2
g̃0 − − σ 2,
4
the same construction gives rise to a family of operators P̃ σ on X̃, the extension
of X̃0,even across its boundary. Namely, with μ̃ = x̃2 ,

(n − 1)2
μ̃−1/2 μ̃iσ/2−(n+1)/4 g0 − − σ 2 μ̃−iσ/2+(n+1)/4 μ̃−1/2
4

= 4μ̃D2μ̃ − 4σ Dμ̃ − h̃ − 4iDμ̃ + 2iγ μ̃Dμ̃ − σ/2 − i(n − 1)/4 ,
(5.81)
and again one just needs to extend h and γ in an arbitrary smooth manner to
obtain the operator P̃ σ . Thus, μ̃ > 0 is the Lorentzian region and μ̃ < 0 the
Riemannian one.
Furthermore, if one has a fixed tangential metric h defined near μ = 0 in R,
the two constructions are closely related in the sense that with h̃(μ̃, y, dy) =
h(−μ̃, y, dy), P̃σ can be chosen the same as the pull-back of P̃ σ by the map
μ → −μ = μ̃, or in other words P̃σ and the pull-back of P̃ σ are extensions of
each other (from μ > 0, resp. μ < 0).
5.5.4. Local Dynamics Near the Radial Set

First, p̃ does not vanish in μ > 0, and in general p̃ = 0 and Hp̃ = 0 at a point
imply ξ = 0, 4(1 + ã1 )μ = −ξ −2 |η|2μ,y together with H̃|η|2μ,y = 0, so η = 0, so
μ = 0, that is the point is in
N ∗ S \ o = {(μ, y, ξ , η) : μ = 0, η = 0, ξ = 0}, S = {μ = 0},
so S ⊂ X−δ0 can be identified with Y = ∂X0 (= ∂X0,even ); but then
Hp̃ |N ∗ S = −4ξ 2 ∂ξ , (5.82)
which is non-zero at N ∗ S \ o, a contradiction. Thus, if p̃ = 0 then Hp̃ never

vanishes, and thus dp̃ never vanishes; in particular dp̃ = 4ξ 2 dμ at N ∗ S. Thus,
the characteristic set = {p̃ = 0} is smooth. Note that p̃ = 0 at N ∗ S\o indeed,
and Hp̃ is radial there in view of (5.82). We further divide
N ∗ S \ o = + ∪ − , ± = N ∗ S ∩ {±ξ > 0}.
Let L± be the image of ± in S∗ X−δ0 . Next we analyze the Hamilton flow

at ± . First,
∂ ã1
Hp̃ |η|2μ,y = 8(1 + ã1 )μξ ∂μ |η|2μ,y − 4μξ 2 ·h η (5.83)
∂y
and
Hp̃ μ = 8(1 + ã1 )ξ μ. (5.84)
In terms of linearizing the flow at N ∗ S, p̃ and μ are equivalent as dp̃ = 4ξ 2 dμ
there, so one can simply use p̂ = p/|ξ |2 (which is homogeneous of degree 0,
like μ), in place of μ. Finally,
Hp̃ |ξ | = −4 sgn(ξ ) + b̃, (5.85)
with b̃ vanishing at ± .
It is convenient to rehomogenize (5.83) in terms of η̂ = η/|ξ |. This can
be phrased more invariantly by working with S∗ X−δ0 = (T ∗ X−δ0 \ o)/R+
(Figure 5.8). Let L± be the image of ± in S∗ X−δ0 . Homogeneous degree zero
functions on T ∗ X−δ0 \ o, such as p̂, can be regarded as functions on S∗ X−δ0 .
For semiclassical purposes, it is best to consider S∗ X−δ0 as the boundary at
∗
fiber infinity of the fiber-radial compactification T X−δ0 of T ∗ X−δ0 . Then at
fiber infinity near N ∗ S, we can take (|ξ |−1 , η̂) as (projective, rather than polar)
coordinates on the fibers of the cotangent bundle, with ρ̃ = |ξ |−1 defining
∗
S∗ X−δ0 in T X−δ0 . Then W = |ξ |−1 Hp̃ is a C ∞ vector field in this region and
|ξ |−1 Hp̃ |η̂|2μ,y = 2|η̂|2μ,y Hp̃ |ξ |−1 + |ξ |−3 Hp̃ |η|2μ,y = 8(sgn ξ )|η̂|2 + ã, (5.86)
where ã vanishes cubically at N ∗ S. In similar notation we have
Hp̃ ρ̃ = 4 sgn(ξ ) + ã , ρ̃ = |ξ |−1 , (5.87)
and
|ξ |−1 Hp̃ μ = 8(sgn ξ )μ + ã , (5.88)
with ã smooth (indeed, homogeneous degree zero without the compactifi-
cation) vanishing at N ∗ S, and ã is also smooth, vanishing quadratically at
N ∗ S. As the vanishing of η̂, |ξ |−1 and μ defines ∂N ∗ S, we conclude that
L− = ∂− is a sink, while L+ = ∂+ is a source, in the sense that all nearby
bicharacteristics (in fact, including semiclassical (null) bicharacteristics, since
Hp̃ |ξ |−1 contains the additional information needed; see (5.97)) converge to
L± as the parameter along the bicharacteristic goes to ∓∞. In particular, the
quadratic defining function of L± in given by
ρ0 = p∂ , where p∂ = |η̂|2 ,
330 A. Vasy
Figure 5.8. The cotangent bundle of X−δ0 near S = {μ = 0}, which is drawn in
a fiber-radially compactified view. The boundary of the fiber compactification is
the cosphere bundle S∗ X−δ0 ; it is the surface of the cylinder shown. ± are the
components of the (classical) characteristic set containing L± . They lie in μ ≤ 0,
∗
only meeting SS∗ X−δ0 at L± . Semiclassically, that is in the interior of T X−δ0 ,
for z = h−1 σ > 0, only the component of the semiclassical characteristic set
containing L+ can enter μ > 0. This is reversed for z < 0
satisfies, on ,
3/2
(sgn ξ )Wρ0 ≥ 8ρ0 + O(ρ0 ). (5.89)
We also need information on the principal symbol of 2i1 (P̃σ − P̃∗σ ) at the
radial points. At L± this is given by

1
σ1 (P̃σ − P̃∗σ ) |N ∗ S = −(4 sgn(ξ )) Im σ |ξ |; (5.90)
2i
here (4 sgn(ξ )) is pulled out due to (5.87), namely its size relative to Hp̃ |ξ |−1
matters. This corresponds to the fact that (μ ± i0)iσ , which are Lagrangian
distributions associated with ± , solve the PDE (5.101) modulo an error that
is two orders lower than what one might a priori expect, that is P̃σ (μ ±
i0)iσ ∈ (μ ± i0)iσ C ∞ (X−δ0 ). Note that P̃σ is second order, so one should
lose two orders a priori, that is get an element of (μ ± i0)iσ −2 C ∞ (X−δ0 ); the
characteristic nature of ± reduces the loss to 1, and the particular choice
of exponent eliminates the loss. This has much in common with eiλ/x x(n−1)/2
being an approximate solution in asymptotically Euclidean scattering.
5.5.5. Global Behavior of the Characteristic Set

By (5.78), points with ξ = 0 cannot lie in the characteristic set. Thus, with
± = ∩ {±ξ > 0},
Figure 5.9. The cotangent bundle near S = {μ = 0}, which is drawn in a fiber-
radially compactified view, as in Figure 5.8. The circles on the left show the
support of q; it has opposite signs on the two disks corresponding to the opposite
directions of propagation relative to the Hamilton vector field
= + ∪ − and ± ⊂ ± . Further, the characteristic set lies in μ ≤ 0, and

intersects μ = 0 only in ± .
Moreover, as Hp̃ μ = 8(1 + ã1 )ξ μ and ξ = 0 on , and μ only vanishes
at + ∪ − there, for 0 > 0 sufficiently small the C ∞ function μ provides
a negative global escape function on μ ≥ −0 which is decreasing on + ,
increasing on − . Correspondingly, bicharacteristics in − travel from μ =
−0 to L− , while in + they travel from L+ to μ = −0 .
5.5.6. Complex Absorption

The final step of fitting P̃σ into our general microlocal framework is moving
the problem to a compact manifold and adding a complex absorbing second
order operator. We thus consider a compact manifold without boundary X for
which Xμ0 = {μ > μ0 }, μ0 = −0 < 0, with 0 > 0 as above, and which is
identified as an open subset with smooth boundary; it is convenient to take X
to be the double of Xμ0 , so there are two copies of X0,even in X.
In the case of hyperbolic space, this doubling process can be realized from
the perspective of (n + 1)-dimensional Minkowski space as discussed in
[46, 54]. Then the Poincaré model shows up in two copies, namely in the
interior of the future and past light cone inside the sphere at infinity, while
de Sitter space shows up as the “equatorial belt,” that is the exterior of the light
cone at the sphere at infinity. One can take the Minkowski equatorial plane,
t = 0, as μ = μ0 , and place the complex absorption there, thereby decoupling
the future and past hemispheres.
At first we consider the bounded σ problem; in this case the potential
trapping for g0 is irrelevant.
332 A. Vasy
We then introduce a “complex absorption” operator Qσ ∈ cl2 (X) with real

principal symbol q supported in, say, μ < −1 , with the Schwartz kernel also
supported in the corresponding region (i.e. in both factors on the product space
this condition holds on the support) such that p̃ ± iq is elliptic near ∂Xμ0 , that is
near μ = μ0 , and which satisfies that ±q ≥ 0 near ± (Figure 5.9). This can
easily be done since ± are disjoint, and away from these p̃ is elliptic, hence so
is p̃ ± iq regardless of the choice of q; we simply need to ensure q has support
sufficiently close to ± , elliptic on ± at μ = −0 , with the appropriate sign
near ± . Having done this, we extend p̃ and q to X in such a way that p̃ ± iq
are elliptic near ∂Xμ0 ; the region we added is thus irrelevant at the level of
bicharacteristic dynamics (of p̃) in so far as it is decoupled from the dynamics
in X0 , and indeed also for analysis as we see shortly (in so far as we have two
essentially decoupled copies of the same problem). This is accomplished, for
instance, by using the doubling construction to define p̃ on X \Xμ0 (in a smooth
fashion at ∂Xμ0 , as can be easily arranged; the holomorphic dependence of P̃σ
on σ is still easily preserved), and then, noting that the characteristic set of
p̃ still has two connected components, making q elliptic on the characteristic
set of p̃ near ∂Xμ0 , with the same sign in each component as near ∂Xμ0 . (An
alternative would be to make q elliptic on the characteristic set of p̃ near X\Xμ0 ;
it is just slightly more complicated to write down such a q when the high energy
behavior is taken into account. With the present choice, due to the doubling,
there are essentially two copies of the problem on X0 : the original and the
one from the doubling.) Finally we take Qσ to be any operator with principal
symbol q with Schwartz kernel satisfying the desired support conditions and
which depends on σ holomorphically. For instance, we may take q and Qσ to
be ± a square, as convenient for the complex absorption discussion. We may
even choose Qσ to be independent of σ so Qσ is indeed holomorphic.
In view of Section 5.5.5 we have arranged the following. For α ∈ S∗ X ∩ ,
let γ+ (α), resp. γ− (α), denote the image of the forward, resp. backward, half-
bicharacteristic of p̃ from α. We write γ± (α) → L± (and say γ± (α) tends to
L± ) if given any neighborhood O of L± , γ± (α) ∩ O = ∅; by the source/sink
property this implies that the points on the curve are in O for sufficiently large
(in absolute value) parameter values. Then, with Ell(Qσ ) denoting the elliptic
set of Qσ ,
α ∈ − \ L− ⇒ γ+ (α) → L− and γ− (α) ∩ Ell(Qσ ) = ∅,

α ∈ + \ L+ ⇒ γ− (α) → L+ and γ+ (α) ∩ Ell(Qσ ) = ∅. (5.91)
That is, all forward and backward half-(null) bicharacteristics of P̃σ either
enter the elliptic set of Qσ , or go to ± , that is L± in S∗ X. The point of
the arrangements regarding Qσ and the flow is that we are able to propagate
estimates forward near where q ≥ 0, backward near where q ≤ 0, so by our
hypotheses we can always propagate estimates for P̃σ − iQσ from ± towards
the elliptic set of Qσ . On the other hand, for P̃∗σ + iQ∗σ , we can propagate
estimates from the elliptic set of Qσ towards ± . This behavior of P̃σ − iQσ
vs. P̃∗σ + iQ∗σ is important for duality reasons as discussed in Section 5.4.8.
An alternative to the complex absorption would be simply adding a boundary
at μ = μ0 ; this is easy to do since this is a space-like hypersurface. This is less
convenient from the perspective of microlocal analysis, but is straightforward
to do, and is indeed very useful for non-linear problems, see [26, 24, 22].
5.5.7. Fredholm Theory

An immediate consequence of (5.91) is
Proposition 5.31 Let r ∈ R, and let s ≥ 1
2 + r. Let r = {σ ∈ C : Im σ > −r},
and let
X s = {u ∈ H s (X) : (P̃σ − iQσ ) ∈ H s−1 (X)}, Y s−1 = H s−1 (X).
For σ ∈ C, the operator P̃σ − iQσ : X s → Y s−1 is Fredholm, and r σ →

P̃σ − iQσ ∈ L(X s , Y s−1 ) is a holomorphic Fredholm family. Furthermore,

Ker(P̃σ − iQσ ) ⊂ C ∞ (X) for all σ , while Ker(P̃∗σ + iQ∗σ ) ⊂ H s (X) for all
s < 12 − Im σ .
Remark 5.4 Note here that the principal symbol of P̃σ − iQσ is p̃ − iq,
independent of σ . On the other hand, X s only depends on the principal symbol
of P̃σ − iQσ , since elements of Diff1 (X) map H s (X) to H s−1 (X). Thus, any
fixed value σ0 is σ and can be used to define X s , and thus the meromorphic
statement actually makes sense as a family of maps between fixed spaces.
Proof For the proof we simply need to observe the non-trapping statement
(5.91) together with the threshold regularity at the radial points given by 1/2 −
Im σ due to (5.90) and that the order of the operator P̃σ − iQσ is m = 2.
Thus, Section 5.4.8 is applicable for each fixed σ , and indeed in the analytic
Fredholm sense in the indicated half-spaces of σ . The regularity statement for
Ker(P̃σ −iQσ ) is a consequence of propagation of singularities, since s > 12 +r
ensures that the regularity of elements of X s is above the threshold at the radial
set, and so (P̃σ − iQσ )u = 0 implies WF(u) ∩ (L+ ∪ L− ) = ∅, and then by
(5.91), WF(u) = ∅. Similarly, the regularity statement for Ker(P̃∗σ + iQ∗σ )
follows from propagation of singularities, this time from Ell(Qσ ), taking into
account that s is below the threshold regularity.
334 A. Vasy
What we do not have at this point is that P̃σ − iQσ is invertible for any fixed
value σ0 of σ ; if it is we conclude that
(P̃σ − iQσ )−1 : Y s−1 → X s
is a meromorphic family. Note that for any fixed σ invertibility is independent
of the choice of s satisfying s > 12 − Im σ in view of the regularity of the kernel
of P̃σ − iQσ and its adjoint, so one does not need to specify s explicitly as long
as it is understood that it satisfies the constraint.
Correspondingly we have:
Proposition 5.32 Suppose that (X0 , g0 ) is an n-dimensional manifold with
boundary with an even conformally compact metric and boundary defining
function x. Let X0,even denote the even version of X0 , that is with the boundary
defining function replaced by its square with respect to a decomposition
in which g0 is even, and let P̃σ be the extended operator in the sense of
Section 5.5.3, and Qσ be the complex absorption in the sense of Section 5.5.6.
Finally suppose P̃σ − iQσ : X s → Y s−1 is invertible for some σ0 ∈ r , s, r, r ,
X s , Y s−1 as in Proposition 5.31.
Then (P̃σ − iQσ )−1 is a meromorphic family in r ; further the inverse of

n−1 2
g0 − − σ 2,
2
written as R(σ ) : L2 → L2 , has a meromorphic continuation from Im σ $ 0
to r ,
R(σ ) : C˙∞ (X0 ) → C −∞ (X0 ),
with poles with finite rank residues, and for f ∈ C˙∞ (X0 ),
R(σ )f = x−iσ x(n+1)/2 x−1 Rs (P̃σ − iQσ )−1 Es−1 xiσ x−(n+1)/2 x−1 f , (5.92)
where
Es : H s (X0,even ) → H s (X)
is a continuous extension operator, Rs : H s (X) → H s (X0,even ) the restriction
map. Furthermore, for s ≥ 12 + r, σ ∈ r not a pole of R(.),
x−(n−1)/2+iσ R(σ )f H s (X0,even ) ≤ Cx−(n+3)/2+iσ f H s−1 (X0,even ) , (5.93)
where C depends on σ . If f is supported in X0◦ , the s − 1 norm on f can be
replaced by the s − 2 norm.
Proof By Proposition 5.31, (P̃σ − iQσ )−1 is meromorphic on r with values
in L(Y s−1 , X s ). On the other hand, by self-adjointness and positivity of g0
and as C˙∞ (X0 ) is in its domain,
2
n−1
g0 − σ − 2
u = f ∈ C˙∞ (X0 )
2
has a unique solution u = R(σ )f ∈ L2 (X0 , |dg0 |) when Im σ $ 0. On the other

hand, with f̃0 = xiσ −(n+1)/2−1 f in μ ≥ 0, and f̃0 still vanishes to infinite order
at μ = 0, let f̃ be an arbitrary smooth extension of f̃0 to the compact manifold
X on which P̃σ − iQσ is defined; for instance we can take f̃ = Es−1 f̃0 . Let
ũ = (P̃σ − iQσ )−1 f̃ ; this satisfies (P̃σ − iQσ )ũ = f̃ and ũ ∈ C ∞ (X). Thus,
u = x−iσ +(n+1)/2−1 Rs ũ satisfies u ∈ x(n−1)/2 x−iσ C ∞ (X0 ), and
2
n − 1
g0 − σ 2 − u = f
2
by (5.101) and (5.110) (as Qσ is supported in μ < 0). Since u ∈ L2 (X0 , |dg0 |)
for Im σ > 0, by the aforementioned uniqueness, u = u .
While, for the sake of simplicity, Qσ is constructed in Section 5.5.6 in such
a manner that it is not holomorphic in all of Im σ > −C due to a cut in the
upper half-plane, this cut can be moved outside any fixed compact subset, so
taking into account that R(σ ) is independent of the choice of Qσ , the theorem
follows immediately from our earlier microlocal results.
Thus, in order to prove the theorem of Mazzeo and Melrose [29] and
Guillarmou [15], it remains to show the invertibility of P̃σ − iQσ for some σ0 .
We do this by considering the high energy asymptotics of the P̃σ − iQσ , and as
a byproduct we obtain high energy resolvent estimates on even asymptotically
hyperbolic spaces.
We now proceed to make some remarks under this invertibility hypothesis
at some σ0 . Our proof of Proposition 5.32 implies that every pole of R(σ )
is a pole of (P̃σ − iQσ )−1 (for otherwise (5.92) would show R(σ ) does not
have a pole either), but it is possible for (P̃σ − iQσ )−1 to have poles which
are not poles of R(σ ). However, in the latter case, the Laurent coefficients of
(P̃σ − iQσ )−1 would be annihilated by multiplication by Rs from the left, that
is the resonant states (which are smooth) would be supported in μ ≤ 0, in
particular vanish to infinite order at μ = 0.
In fact, a stronger statement can be made: as remarked at the end of Section
5.5.3, in μ < 0, P̃σ is a conjugate (times a power of μ) of a Klein–Gordon-
type operator on n-dimensional de Sitter space with μ = 0 being the boundary
(i.e. where time goes to infinity). Thus, if σ is not a pole of R(σ ) and
(P̃σ − iQσ )ũ = 0 then one would have a solution u of this Klein–Gordon-
type equation near μ = 0, that is infinity, that rapidly vanishes at infinity. As
shown in [53], one can use a Carleman-type estimate to show that this cannot
336 A. Vasy
happen; alternatively one can use an argument due to Zworski [57, Lemma 1]
which makes the usual energy estimate quantitative in moving the initial slice
μ = −, → 0. Thus, if Qσ is supported in μ < c, c < 0, then ũ is also
supported in μ < c. This argument can be iterated for Laurent coefficients
of higher order poles; their range (which is finite dimensional) contains only
functions supported in μ < c.
Remark 5.5 We now return to our previous remarks regarding the fact that our
solution disallows the conormal singularities (μ ± i0)iσ from the perspective
of conformally compact spaces of dimension n. Recalling that μ = x2 , the two
indicial roots on these spaces correspond to the asymptotics μ±iσ/2+(n−1)/4 in
μ > 0. Thus for the operator

−1/2 iσ/2−(n+1)/4 (n − 1)2
μ μ g0 − − σ μ−iσ/2+(n+1)/4 μ−1/2 ,
2
4
or indeed P̃σ , they correspond to

−1
μ−iσ/2+(n+1)/4 μ−1/2 μ±iσ/2+(n−1)/4 = μiσ/2±iσ/2 .
Here the indicial root μ0 = 1 corresponds to the smooth solutions we construct

for P̃σ , while μiσ corresponds to the conormal behavior we rule out. Back
to the original Laplacian, thus, μ−iσ/2+(n−1)/4 is the allowed asymptotic and
μiσ/2+(n−1)/4 is the disallowed one. Notice that Re iσ = − Im σ , so the
disallowed solution is growing at μ = 0 relative to the allowed one, as expected
in the physical half-plane, and the behavior reverses when Im σ < 0. Thus,
in the original asymptotically hyperbolic picture one has to distinguish two
different rates of growth, whose relative size changes. On the other hand, in our
approach, we rule out the singular solution and allow the non-singular (smooth
one), so there is no change in behavior at all for the analytic continuation.
5.5.8. High Energy, or Semiclassical, Asymptotics

We are thus interested in the high energy behavior, as |σ | → ∞. For the
associated semiclassical problem one obtains a family of operators
P̃h,z = h2 P̃h−1 z ,
with h = |σ |−1 , and z corresponding to σ/|σ | in the unit circle in C. Then

the semiclassical principal symbol p̃h̄,z of P̃h,z is a function on T ∗ X−δ0 , whose
asymptotics at fiber infinity of T ∗ X−δ0 is given by the classical principal
symbol p̃. We are interested in Im σ ≥ −C, which in semiclassical notation
corresponds to Im z ≥ −Ch. However, the present form of P̃σ is only suitable

for analysis in the strip | Im σ | ≤ C, that is | Im z| ≤ Ch; the more general case
is taken up in Section 5.5.11.
It is often convenient to think of p̃h̄,z , and its rescaled Hamilton vector field,
∗
as objects on T X−δ0 . Thus,
p̃h̄,z = σ2,h̄ (P̃h,z ) = 4(1 + ã1 )μξ 2 − 4(1 + ã2 )zξ − ã3 z2 + |η|2μ,y ; (5.94)
when | Im z| ≤ Ch, one can simply replace z by Re z since one obtains an

equivalent representative of the semiclassical principal symbol.
Explicitly, if we introduce for instance
(μ, y, ν, η̂), ν = |ξ |−1 , η̂ = η/|ξ |, (5.95)

∗
as valid projective coordinates in a (large!) neighborhood of L± in T X−δ0 ,
then
ν 2 p̃h̄,z = 4(1 + ã1 )μ − 4(1 + ã2 )(sgn ξ )zν − ã3 z2 ν 2 + |η̂|2y,μ .
We assume that Re z > 0 for the sake of definiteness. Observe that the
semiclassical Hamilton vector field is
Hp̃h̄,z = 4(2(1 + ã1 )μξ − (1 + ã2 )z)∂μ + H̃|η|2μ,y

∂|η| 2
∂ ã1 ∂ ã2 ∂ ã 3 μ,y
− 4 1 + ã1 + μ ξ2 − 4 zξ + z2 + ∂ξ
∂μ ∂μ ∂μ ∂μ

∂ ã1 2 ∂ ã2 ∂ ã3 2
− 4 μξ − 4 zξ − z ∂η ; (5.96)
∂y ∂y ∂y
∗
here we are concerned about z real. Near S∗ X−δ0 = ∂T X−δ0 , using the
coordinates (5.95) (which are valid near the characteristic set)
Wh̄ = νHp̃h̄,z = 4(2(1 + ã1 )μ(sgn ξ ) − (1 + ã2 )zν)∂μ + ν H̃|η|2μ,y

∂ ã1 ∂ ã2
+ (sgn ξ ) 4 1 + ã1 + μ −4 z(sgn ξ )ν
∂μ ∂μ

∂ ã3 2 2 ∂|η̂|2μ,y
+ z ν + (ν∂ν + η̂∂η̂ )
∂μ ∂μ

∂ ã1 ∂ ã2 ∂ ã3 2 2
− 4 μ − 4(sgn ξ ) zν − z ν ∂η̂ , (5.97)
∂y ∂y ∂y
338 A. Vasy
∂Hij
with ν H̃|η|2μ,y = ij Hij η̂i ∂yj − ijk ∂yk η̂i η̂j ∂η̂k smooth. Thus, Wh̄ is a smooth
∗
vector field on the compactified cotangent bundle, T X−δ0 which is tangent
to its boundary, S∗ X−δ0 , and Wh̄ − W = νW (with W considered as a
homogeneous degree zero vector field) with W smooth and tangent to S∗ X−δ0 .
In particular, by (5.87) and (5.89), using the fact that ρ̃ 2 + ρ0 is a quadratic
defining function of L± ,
(sgn ξ )Wh̄ (ρ̃ 2 + ρ0 ) ≥ 8(ρ̃ 2 + ρ0 ) − O((ρ̃ 2 + ρ0 )3/2 )
shows that there is 1 > 0 such that in ρ̃ 2 + ρ0 ≤ 1 , ξ > 0, ρ̃ 2 + ρ0 is strictly
increasing along the Hamilton flow except at L+ (where it is constant), while
in ρ̃ 2 + ρ0 ≤ 1 , ξ < 0, ρ̃ 2 + ρ0 is strictly decreasing along the Hamilton
flow except at L− . Indeed, all null-bicharacteristics in this neighborhood of L±
except the constant ones at L± tend to L± in one direction and to ρ̃ 2 + ρ0 =
1 in the other direction. In particular, the local semiclassical bicharacteristic
geometry near L± consists of sources/sinks, and the semiclassical versions of
the radial point estimates are applicable.
In order to proceed, we need to understand the global structure of the
bicharacteristics. Namely, we would like to be able to say that if the original
metric is non-trapping, that is for all points in T ∗ X0◦ , the projection to X0
of the bicharacteristic through that point tends to ∂X0 in both the forward
and backward directions, then in fact with suitable complex absorption near
μ = −0 , the problem is globally non-trapping in the sense that all semi-
classical bicharacteristics except those contained in L+ ∪ L− either reach the
complex absorption in finite time or tend to the radial set L+ ∪ L− . Since there
is an asymmetry between L± , namely one of them is a source, another is a
sink, and correspondingly there is a sign difference for the complex absorption
along the bicharacteristics, we need to be able to separate these, which is done
by observing that, semiclassically, for Im z = O(h), that is z almost real,
the characteristic set can be divided into two components h̄,± , with L± in
different components.
Indeed, consider the hypersurface given by
ξ = − Re z/4,
that is sgn ξ = −ν Re z/4 in coordinates valid near fiber infinity away from
ξ = 0, on which, by (5.94), p̃h̄,z cannot vanish where |μ| is small as it is

1
p̃h̄,z = (1 + ã1 )μ + 1 + ã2 − ã3 (Re z)2 + |η|2 > 0,
4
and ã2 , ã3 vanish at μ = 0. Moreover, if ξ ≤ − Re z/4 and μ > 0 then p̃h̄,z
is a decreasing function of ξ , so it cannot vanish for ξ < − Re z/4 either; if
μ = 0, the only place where it can vanish in ξ < − Re z/4 is at fiber infinity.
Thus, for small 2 > 0, in |μ| < 2 the characteristic set h̄ can be written as
the disjoint union of two non-empty components:
h̄ ∩ {|μ| < 2 } = (h̄,+ ∩ {|μ| < 2 }) ∪ (h̄,− ∩ {|μ| < 2 }),
h̄,± ∩ {|μ| < 2 } = h̄ ∩ {±(ξ + Re z/4) > 0} ∩ {|μ| < 2 },
and
h̄,− ∩ {0 < μ < 2 } = ∅.
We finally need more information about the global semiclassical dynamics.

Lemma 5.14 There exists 0 > 0 such that the following holds. Suppose
Re z > 0. Then only h̄,+ enters μ > 0, and all semiclassical null-
bicharacteristics in (h̄,+ \ L+ ) ∩ {−0 ≤ μ ≤ 0 } go to either L+ or to
μ = 0 in the backward direction and to μ = 0 or μ = −0 in the forward
direction, with the limit being at μ = 0 in at most one of the cases, while all
semiclassical null-bicharacteristics in (h̄,− \ L− ) ∩ {−0 ≤ μ ≤ 0 } go to
L− in the forward direction and to μ = −0 in the backward direction.
For Re z < 0 the analogous conclusion holds with the μ = 0 possibilities
added to the case of h̄,− instead of h̄,+ .
Proof Let 1 > 0 be as above, so in ρ̃ 2 + ρ0 ≤ 1 , ξ > 0, ρ̃ 2 + ρ0 is strictly
increasing along the Hamilton flow except at L+ (where it is constant), while
in ρ̃ 2 + ρ0 ≤ 1 , ξ < 0, ρ̃ 2 + ρ0 is strictly decreasing along the Hamilton flow
except at L− .
∗
Choosing 0 > 0 sufficiently small, the characteristic set in T X−δ0 ∩{−0 ≤
μ ≤ 0 } is disjoint from S∗ X−δ0 \ {ρ̃ 2 + ρ0 ≤ 1 }, and indeed only contains
points in h̄,+ as Re z > 0. Since Hph̄,z μ = 4(2(1 + ã1 )μξ − (1 + ã2 )z), it
∗
is negative on T {μ=0} X−δ0 \ S∗ X−δ0 . In particular, there is a neighborhood U
of μ = 0 in h̄,+ \ S∗ X−δ0 on which the same sign is preserved; since the
∗
characteristic set in T X−δ0 \ {ρ̃ 2 + ρ0 < 1 } is compact, and is indeed a
subset of T ∗ X−δ0 \ {ρ̃ 2 + ρ0 < 1 }, we deduce that |μ| is bounded below on
\ (U ∪ {ρ̃ 2 + ρ0 < 1 }), say |μ| ≥ 0 > 0 there, so with 0 = min(0 , 0 ),
Hph̄,z μ < 0 on h̄,+ ∩ {−0 ≤ μ ≤ 0 } \ {ρ̃ 2 + ρ02 < 1 }. As Hph̄,z μ < 0 at
μ = 0, bicharacteristics can only cross μ = 0 in the outward direction.
Thus, if γ is a bicharacteristic in h̄,+ , there are two possibilities. If γ is
disjoint from {ρ̃ 2 + ρ0 < 1 }, it has to go to μ = 0 in the backward direction
and to μ = −0 in the forward direction. If γ has a point in {ρ̃ 2 + ρ0 < 1 },
then it has to go to L+ in the backward direction and to ρ̃ 2 + ρ0 = 1 in the
forward direction; if |μ| ≥ 0 by the time ρ̃ 2 + ρ0 = 1 is reached, the result
340 A. Vasy
is proved, and otherwise Hph̄,z μ < 0 in ρ̃ 2 + ρ0 ≥ 1 , |μ| ≤ 0 , shows that the

bicharacteristic goes to μ = −0 in the forward direction.
If γ is a bicharacteristic in h̄,− , only the second possibility exists, and the
bicharacteristic cannot leave {ρ̃ 2 +ρ0 < 1 } in |μ| ≤ 0 , so it reaches μ = −0
in the backward direction (as the characteristic set is in μ ≤ 0).
If we assume that g0 is a non-trapping metric we have the following stronger

conclusion:
Lemma 5.15 Suppose g0 is non-trapping, that is all the bicharacteristics of g0
in T ∗ X0◦ \ o tend to ∂X0 in both the forward and the backward directions. Then
for sufficiently small 0 > 0, and for Re z > 0, any bicharacteristic in h̄,+ in
−0 ≤ μ has to go to L+ in the backward direction, and to μ = −0 in the
forward direction (with the exception of the constant bicharacteristics at L+ ),
while in h̄,− , all bicharacteristics in −0 ≤ μ lie in −0 ≤ μ ≤ 0, and go to
L− in the forward direction and to μ = −0 in the backward direction (with
the exception of the constant bicharacteristics at L− ).
Proof The conclusion follows from Lemma 5.14 in the case of h̄,− , so it
suffices to consider h̄,+ .
We first observe that by the non-trapping hypothesis all bicharacteristics
in h̄,+ ∩ {μ ≥ −0 } enter −0 ≤ μ ≤ 0 . Now, for a bicharacteristic
with a point in −0 ≤ μ ≤ 0 , the conclusion of this lemma holds by
Lemma 5.14 unless the μ = 0 possibility holds in its conclusion either in the
forward or backward direction. Suppose for instance this μ = 0 possibility
holds in the backward direction. Then by the non-trapping hypothesis, the
backward bicharacteristic segment continued from μ = 0 enters μ < 0
again, and then applying Lemma 5.14 it must approach L+ (otherwise the local
segment in |μ| ≤ 0 would have both endpoints at |μ| = 0 ). Similarly, if the
μ = 0 possibility holds in the forward direction, the analogous argument
shows the further forward extend bicharacteristic in fact reaches μ = −0 .
This completes the proof.
For applications to gluing constructions as in [5] it is also useful to remark

that μ is bicharacteristically strictly convex (in the sense that the super-level
sets are convex) for μ > 0 small. More precisely, for sufficiently small 0 > 0,
and for α ∈ T ∗ X0 ,
0 < μ(α) < 0 , p̃h̄,z (α) = 0 and (Hp̃h̄,z μ)(α) = 0 ⇒ (H2p̃h̄,z μ)(α) < 0.
(5.98)
Indeed, as Hp̃h̄,z μ = 4(2(1 + ã1 )μξ − (1 + ã2 )z), the hypotheses imply z =
2(1 + ã1 )(1 + ã2 )−1 μξ and
0 = p̃h̄,z
= 4(1 + ã1 )μξ 2 − 8(1 + ã1 )μξ 2 − 4(1 + ã1 )2 (1 + ã2 )−2 ã3 μ2 ξ 2 + |η|2μ,y
= −4(1 + ã1 )μξ 2 − 4(1 + ã1 )2 (1 + ã2 )−2 ã3 μ2 ξ 2 + |η|2μ,y ,
so |η|2μ,y = 4(1 + b)μξ 2 , with b vanishing at μ = 0. Thus, at points where

Hp̃h̄,z μ vanishes, writing ãj = μâj ,
H2p̃h̄,z μ = 8(1 + ã1 )μHp̃h̄,z ξ + 8μ2 ξ Hp̃h̄,z â1 − 4zμHp̃h̄,z â2

= 8(1 + ã1 )μHp̃h̄,z ξ + O(μ2 ξ 2 ). (5.99)
Now

∂ ã1 ∂ ã2 ∂ ã3 2 ∂|η|2μ,y
Hp̃h̄,z ξ = − 4 1 + ã1 + μ ξ −4
2
zξ + z + .
∂μ ∂μ ∂μ ∂μ
Since zξ is O(μξ 2 ) due to Hp̃h̄,z μ = 0, z2 is O(μ2 ξ 2 ) for the same reason, and
|η|2 and ∂μ |η|2 are O(μξ 2 ) due to p̃h̄,z = 0, we deduce that Hp̃h̄,z ξ < 0 for
sufficiently small |μ|, so (5.99) implies (5.98). Thus, μ can be used for gluing
constructions.
5.5.9. Complex Absorption at High Energies

For the semiclassical problem, when z is almost real we need to increase the
requirements on Qσ , and what we need to do depends on whether g0 is non-
trapping.
If g0 is non-trapping, we choose Qσ such that h2 Qh−1 z ∈ h̄,cl 2 (X) with
semiclassical principal symbol qh̄,z , and in addition to the above requirement

for the classical symbol, we need semiclassical ellipticity near μ = μ0 , that is
p̃h̄,z −iqh̄,z and its complex conjugate are elliptic near ∂Xμ0 , that is near μ = μ0 ,
and which satisfies that for z real ±qh̄,z ≥ 0 on h̄,± . Again, we extend P̃σ and
Qσ to X in such a way that p̃ − iq and p̃h̄,z − iqh̄,z (and thus their complex
conjugates) are elliptic near ∂Xμ0 ; the region we added is thus irrelevant. This
is straightforward to arrange if one ignores the fact that one wants Qσ to be
holomorphic: one easily constructs a function qh̄,z on T ∗ X (taking into account
the disjointness of h̄,± ), and defines Qh−1 z to be h−2 times the semiclassical
quantization of qh̄,z (or any other operator with the same semiclassical and
standard principal symbols). However, it is not hard to have an example of a
holomorphic family Qσ in a strip at least: in view of (5.112) we can take (with
C > 0 sufficiently large, depending on the strip in C we are working in)
qh,z = (4ξ + z)(ξ 2 + |η|2 + z2 + C2 h2 )1/2 χ (μ),
342 A. Vasy
where χ ≥ 0 is supported near μ0 ; the corresponding full symbol is

σfull (Qσ ) = (4ξ + σ )(ξ 2 + |η|2 + σ 2 )1/2 χ (μ),
and Qσ is taken as a quantization of this full symbol. Here the square root
is defined on C \ [0, −∞), with the real part of the result being positive,
and correspondingly qh,z is defined and is holomorphic away from h−1 z ∈
±i[C, +∞). We remark for future use that ξ 2 +|η|2 +σ 2 is an elliptic symbol in
(ξ , η, Re σ , Im σ ) in the larger region | Im σ | < C | Re σ |, so the corresponding
statement also holds for its square root.
If g0 is trapping, we need to add complex absorption inside X0 as well,
at μ = 0 , so in addition to Qσ being supported in μ < −0 /2, we add
Q σ supported in μ > 0 /2, but we require in addition to the other classical
requirements that p̃h̄,z − i(qh̄,z + q h̄,z ) and its complex conjugate are elliptic
near μ = ±0 , and which satisfies ±qh̄,z ≥ 0 on h̄,± . This can be achieved as
above for μ near μ0 . Again, we extend P̃σ and Qσ to X in such a way that p̃−iq
and p̃h̄,z − iqh̄,z (and thus their complex conjugates) are elliptic near ∂Xμ0 .
In either of these semiclassical cases we have arranged that for sufficiently
small δ0 > 0, p̃h̄,z − iqh̄,z and its complex conjugate are semiclassically non-
trapping for | Im z| < δ0 , namely the bicharacteristics from any point in h̄ \
(L+ ∪ L− ) flow to Ell(qh̄,z ) ∪ L− (i.e. either enter Ell(qh̄,z ) at some finite time,
or tend to L− ) in the forward direction, and to Ell(qh̄,z ) ∪ L+ in the backward
direction. Here δ0 > 0 arises from the particularly simple choice of qh̄,z for
which semiclassical ellipticity is easy to check for Im z > 0 (bounded away
from 0) and small; a more careful analysis would give a specific value of δ0 ,
and a more careful choice of qh̄,z would give a better result.
5.5.10. Meromorphic Continuation of the Resolvent

We now state our results in the original conformally compact setting. Without
the non-trapping estimate, these are a special case of a result of Mazzeo and
Melrose [29], with improvements by Guillarmou [15], with “special” meaning
that evenness is assumed. If the space is asymptotic to actual hyperbolic space,
the non-trapping estimate is a stronger version of the estimate of Melrose,
Sá Barreto, and me [32], where it is shown by a parametrix construction;
here conformal infinity can have arbitrary geometry. The point is thus that,
first, we do not need the machinery of the zero calculus here, second, we do
have non-trapping high energy estimates in general (and without a parametrix
construction), and third, we add the semiclassically outgoing property which
is useful for resolvent gluing, including for proving non-trapping bounds
microlocally away from trapping, provided the latter is mild, as shown by
Datchev and Vasy [5]. Since the original work in [48, 47] (somewhat weaker,
but the full result should follow with some effort), high energy estimates have
been obtained by zero calculus parametrix methods by Chen and Hassell [4]
for real σ and by Wang [55] in general.
Theorem 5.7 Suppose that (X0 , g0 ) is an n-dimensional non-trapping mani-
fold with boundary with an even conformally compact metric and boundary
defining function x. Let X0,even denote the even version of X0 , that is with
the boundary defining function replaced by its square with respect to a
decomposition in which g0 is even. Then the inverse of

n−1 2
g0 − − σ 2,
2

to C,
R(σ ) : C˙∞ (X0 ) → C −∞ (X0 ),
with poles with finite rank residues. Furthermore, non-trapping estimates hold
in every region | Im σ | < r, | Re σ | $ 0: for s > 12 + r,
x−(n−1)/2 xiσ R(σ )f H s (X0,even ) ≤ C̃|σ |−1 x−(n+3)/2 xiσ f H s−1 (X0,even ) .
|σ |−1 |σ |−1
(5.100)
If f is supported in X0◦ ,
the s − 1 norm on f can be replaced by the s − 2 norm.
Furthermore, for Re z > 0, Im z = O(h), the resolvent R(h−1 z) is
semiclassically outgoing with a loss of h−1 in the sense that if f has compact
support in X0◦ , α ∈ T ∗ X is in the semiclassical characteristic set and if
WFs−1h̄ ( f ) is disjoint from the backward bicharacteristic from α, then α ∈ /
s −1
WFh̄ (hR(h z)f ).
We remark that, although in order to go through without changes our
methods require the evenness property, it is not hard to deduce more restricted
results without this. Essentially one would have operators with coefficients that
have a conormal singularity at the event horizon; as long as this is sufficiently
mild relative to what is required for the analysis, it does not affect the results.
The problems arise for the analytic continuation, when one needs strong
function spaces (H s with s large); these are not preserved when one multiplies
by the singular coefficients.
Proof First note that semiclassical estimates hold for u in terms of (P̃h,z −
iQh,z )u plus a positive power of h times a weak norm of u by the same
arguments as in the standard setting, using Lemma 5.15 to control u by
344 A. Vasy
propagating estimates within h̄ from the radial sets L+ ∪ L− : for all r, R,
for {(h, z) : −r < h−1 Im z < R}, for s > 12 + r, and for all k > 0, N,
uHh̄s ≤ C(h−1 (P̃h,z − iQh,z )uH s−1 + hk uH N ).
h̄ h̄
Similarly, semiclassical estimates hold for v in terms of (P̃∗h,z + iQ∗h,z )v plus a

positive power of h times a weak norm of v, this time using Lemma 5.15 to
control u by propagating estimates within h̄ from the complex absorption
Ellh̄ (Qh,z ). This gives that in fact P̃h,z − iQh,z is invertible for sufficiently
small h.
Now, Proposition 5.32 gives the non-high energy part of the statement
provided P̃σ − iQσ is invertible for some σ ; that this is the case follows
from the semiclassical behavior in strips | Im σ | < C, since for large Re σ
these guarantee invertibility. Further, the semiclassical estimates also yield
high energy bounds for (P̃σ − iQσ )−1 . By (5.92) this gives the desired high
energy estimates for R(σ ) in strips, as stated.
5.5.11. High Energy Estimates in Imσ > 0 Farther from

the Real Axis
It is instructive to modify our construction of P̃σ so that the resulting operator
is semiclassically elliptic when σ is away from the reals. We achieve this via
conjugation by an additional power of a smooth function, with the exponent
depending on σ . The latter would make no difference even semiclassically
in the real regime as it is conjugation by an elliptic semiclassical Fourier
integral operator, which simply rearranges the cotangent bundle. However,
in the non-real regime (where we would like ellipticity) it does matter; P̃σ
is not semiclassically elliptic at the zero section. So finally we conjugate by
(1 + μ)iσ/4 to obtain
Pσ = 4(1 + a1 )μD2μ − 4(1 + a2 )σ Dμ − (1 + a3 )σ 2 + h
(5.101)
− 4iDμ + b1 μDμ + b2 σ + c1
with aj smooth, real, and vanishing at μ = 0, bj and c1 smooth. In fact, we
have a1 ≡ 0, but it is sometimes convenient to have more flexibility in the
form of the operator since this means that we do not need to start from the
relatively rigid form (5.68). Note that the only change compared to (5.74) is in
the coefficient of σ 2 ; this is exactly what is needed for the far-from-real-axis
ellipticity.
Now the full (including high energy) principal symbol of Pσ ∈ Diff2 (X−δ0 ) is
pfull = 4(1 + a1 )μξ 2 − 4(1 + a2 )σ ξ − (1 + a3 )σ 2 + |η|2μ,y . (5.102)
The Hamilton vector field is
Hpfull = 4(2(1 + a1 )μξ − (1 + a2 )σ )∂μ + H̃|η|2μ,y

∂a1 ∂a2 ∂a3 2 ∂|η|2μ,y
− 4 1 + a1 + μ ξ −4
2
σξ + σ + ∂ξ
∂μ ∂μ ∂μ ∂μ

∂a1 2 ∂a2 ∂a3 2
− 4 μξ − 4 σξ − σ ∂η , (5.103)
∂y ∂y ∂y
where H̃ still indicates that this is the Hamilton vector field in T ∗ Y, that is with
μ considered a parameter. Correspondingly, essentially unchanged relative
to P̃σ , the standard, “classical,” principal symbol is
p = σ2 (Pσ ) = 4(1 + a1 )μξ 2 + |η|2μ,y , (5.104)
which is real, independent of σ , while the Hamilton vector field is
Hp = 8(1 + a1 )μξ ∂μ + H̃|η|2μ,y

∂a1 ∂|η|2μ,y ∂a1 2 (5.105)
− 4 1 + a1 + μ ξ +
2
∂ξ − 4 μξ ∂η .
∂μ ∂μ ∂y
We still have the fact that as g0 − σ 2 − (n − 1)2 /4 is formally self-adjoint

relative to the metric density |dg0 | for σ real, the same holds for the conjugate
of μ−1/2 (g0 −σ 2 −(n−1)2 /4)μ−1/2 by μ−iσ/2 (1+μ)iσ/4 for σ real since this
is merely unitary conjugation. This gives finally that for σ real, Pσ is formally
self-adjoint relative to
1
μ(n+1)/2 |dg0 | = |dh| |dμ|,
2
as x−n dx = 12 μ−(n+1)/2 dμ. We deduce that with respect to the extension of
the density μ(n+1)/2 |dg0 | to X−δ0 , σ1 ( 2i1 (Pσ − P∗σ ))|μ≥0 vanishes when σ ∈ R.
Since in general Pσ − PRe σ differs from −4i(1 + a2 ) Im σ Dμ by a zeroth order
operator, we conclude that

1
σ1 (Pσ − Pσ )
∗
= −4(Im σ )ξ . (5.106)
2i μ=0
We now need to choose μ appropriately in the interior of X0 far from ∂X0 ;

this choice matters greatly for the far-from-real-axis ellipticity. Thus, near μ =
0, but μ bounded away from 0, the only semiclassically non-trivial action we
have done was to conjugate the operator by e−iσ φ where eφ = μ1/2 (1+μ)−1/4 ;
we need to extend φ into the interior. But the semiclassical principal symbol of
346 A. Vasy
the conjugated operator is, with σ = z/h,
(ζ − z dφ, ζ − z dφ)G0 − z2 = |ζ |2G0 − 2z(ζ , dφ)G0 − (1 − |dφ|2G0 )z2 . (5.107)
For z non-real this is elliptic if |dφ|G0 < 1. Indeed, if (5.107) vanishes then
from the vanishing imaginary part we get
2 Im z((ζ , dφ)G0 + (1 − |dφ|2G0 ) Re z) = 0, (5.108)
and then the real part is
|ζ |2G0 − 2 Re z(ζ , dφ)G0 − (1 − |dφ|2G0 )((Re z)2 − (Im z)2 )

(5.109)
= |ζ |2G0 + (1 − |dφ|2G0 )((Re z)2 + (Im z)2 ),
which cannot vanish if |dφ|G0 < 1. But, reading off the dual metric from the
principal symbol of (5.72),
2 2
1 1
= 1− μ
d(log μ − log(1 + μ)) <1
4 2
G0 2(1 + μ)
for μ > 0, with a strict bound as long as μ is bounded away from 0.
Correspondingly, μ1/2 (1 + μ)−1/4 can be extended to a function eφ on all
of X0 so that semiclassical ellipticity for z away from the reals is preserved,
and we may even require that φ is constant on a fixed (but arbitrarily large)
compact subset of X0◦ . Then, after conjugation by e−iσ φ ,
Ph,z = eizφ/h μ−(n+1)/4−1/2 (h2 g0 − z)μ(n+1)/4−1/2 e−izφ/h (5.110)
is semiclassically elliptic in μ > 0 (as well as in μ ≤ 0, μ near 0, where this

is already guaranteed), as desired.
We again want to think of ph̄,z , and its rescaled Hamilton vector field, as
∗
objects on T X−δ0 . Thus,
ph̄,z = σ2,h̄ (Ph,z ) = 4(1 + a1 )μξ 2 − 4(1 + a2 )zξ − (1 + a3 )z2 + |η|2μ,y ,

(5.111)
so
Im ph̄,z = −2 Im z(2(1 + a2 )ξ + (1 + a3 ) Re z). (5.112)
In particular, for z non-real, Im ph̄,z = 0 implies 2(1+a2 )ξ +(1+a3 ) Re z = 0,

so
Re ph̄,z = ((1 + a1 )(1 + a3 )2 (1 + a2 )−2 μ + (1 + 2a2 )(1 + a3 ))(Re z)2
+ (1 + a3 )(Im z)2 + |η|2μ,y > 0
(5.113)
near μ = 0, that is ph̄,z is semiclassically elliptic on T ∗ X−δ0 , but not at

fiber infinity, that is at S∗ X−δ0 (standard ellipticity is lost only in μ ≤ 0, of
course). In μ > 0 we have semiclassical ellipticity (and automatically classical
ellipticity) by our choice of φ following (5.107). Explicitly, if we introduce for
instance
(μ, y, ν, η̂), ν = |ξ |−1 , η̂ = η/|ξ |, (5.114)
∗
as valid projective coordinates in a (large!) neighborhood of L± in T X−δ0 ,
then
ν 2 ph̄,z = 4(1 + a1 )μ − 4(1 + a2 )(sgn ξ )zν − (1 + a3 )z2 ν 2 + |η̂|2y,μ
so
ν 2 Im ph̄,z = −4(1 + a2 )(sgn ξ )ν Im z − 2(1 + a3 )ν 2 Re z Im z
which automatically vanishes at ν = 0, that is at S∗ X−δ0 . Thus, for σ large
and pure imaginary, the semiclassical problem adds no complexity to the
“classical” quantum problem, but of course it does not simplify it. In fact,
we need somewhat more information at the characteristic set, which is thus
at ν = 0 when Im z is bounded away from 0:
ν small, Im z ≥ 0 ⇒ (sgn ξ ) Im ph̄,z ≤ 0 ⇒ ± Im ph̄,z ≤ 0 near h̄,± ,
ν small, Im z ≤ 0 ⇒ (sgn ξ ) Im ph̄,z ≥ 0 ⇒ ± Im ph̄,z ≥ 0 near h̄,± ,
(5.115)
which means that for Ph,z with Im z > 0 one can propagate estimates forwards
along the bicharacteristics where ξ > 0 (in particular, away from L+ , as
the latter is a source) and backwards where ξ < 0 (in particular, away from
L− , as the latter is a sink), while for P∗h,z the directions are reversed since its
semiclassical symbol is ph̄,z . The directions are also reversed if Im z switches
sign. This is important because it gives invertibility for z = i (corresponding to
Im σ large positive, i.e. the physical half-plane), but does not give invertibility
for z = −i negative.
We now give a different perspective on what we have shown already (for P̃σ ,
but as it is related to Pσ by conjugation by a smooth function, this distinction
does not matter), namely that semiclassically, for z almost real (i.e. when z
is not bounded away from the reals; we are not fixing z as we let h vary!),
when the operator is not semiclassically elliptic on T ∗ X−δ0 as mentioned
above, the characteristic set can be divided into two components h̄,± , with
L± in different components. The vanishing of the factor following Im z in
(5.112) gives a hypersurface that separates h̄ into two parts. Indeed, this is
the hypersurface given by
348 A. Vasy
2(1 + a2 )ξ + (1 + a3 ) Re z = 0,
on which, by (5.113), Re ph̄,z cannot vanish, so
h̄ = h̄,+ ∪ h̄,− , h̄,± = h̄ ∩ {±(2(1 + a2 )ξ + (1 + a3 ) Re z) > 0}.
Farther in μ > 0, the hypersurface is given, due to (5.108), by
(ζ , dφ)G0 + (1 − |dφ|2G0 ) Re z = 0,
and on it, by (5.109), the real part is |ζ |2G0 +(1−|dφ|2G0 )((Re z)2 +(Im z)2 ) > 0;
correspondingly
h̄ = h̄,+ ∪ h̄,− , h̄,± = h̄ ∩ {±((ζ , dφ)G0 + (1 − |dφ|2G0 ) Re z) > 0}.
In fact, more generally, the real part is
|ζ |2G0 − 2 Re z(ζ , dφ)G0 − (1 − |dφ|2G0 )((Re z)2 − (Im z)2 )
= |ζ |2G0 − 2 Re z((ζ , dφ)G0 + (1 − |dφ|2G0 ) Re z)
+ (1 − |dφ|2G0 )((Re z)2 + (Im z)2 ),
so for ± Re z > 0, ∓((ζ , dφ)G0 + (1 − |dφ|2G0 ) Re z) > 0 implies that ph̄,z does
not vanish. Correspondingly, only one of the two components of h̄,± enter
μ > 0, namely for Re z > 0, only h̄,+ enters, while for Re z < 0, only h̄,−
enters.
This completes the characteristic set and bicharacteristic geometry discus-
sion even for Im σ large. Correspondingly, we obtain high energy estimates in
cones c1 | Re σ | < Im σ < c2 | Re σ |, 0 < c1 < c2 , where the limitation given
by c2 arises due to the construction of the complex absorption in Section 5.5.9.
This then yields the invertibility of Pσ − iQσ for large |σ | in these cones. This
shows that the invertibility assumption of Proposition 5.32 is satisfied, and thus
we have the meromorphic continuation theorem in general:
Theorem 5.8 (Special case of the theorem of Mazzeo and Melrose [29] and
Guillarmou [15].) Suppose that (X0 , g0 ) is an n-dimensional manifold with
boundary with an even conformally compact metric and boundary defining
function x. Let X0,even denote the even version of X0 , that is with the boundary
defining function replaced by its square with respect to a decomposition in
which g0 is even, and let Pσ be the extended operator in the sense of Section
5.5.3, and Qσ be the complex absorption in the sense of Section 5.5.6. Let
s, r, r be as in Proposition 5.31.
Then (Pσ − iQσ )−1 is a meromorphic family in r , further the inverse of

n−1 2
g0 − − σ 2,
2

to r ,
R(σ ) : C˙∞ (X0 ) → C −∞ (X0 ),
with poles with finite rank residues, and for f ∈ C˙∞ (X0 ),
R(σ )f = x−iσ +(n−1)/2 Rs (Pσ − iQσ )−1 Es−1 xiσ −(n−3)/2 f , (5.116)
where
Es : H s (X0,even ) → H s (X)
is a continuous extension operator, Rs : H s (X) → H s (X0,even ) the restriction

map. Furthermore, for s > 12 + r, σ ∈ r not a pole of R(.),
x−(n−1)/2+iσ R(σ )f H s (X0,even ) ≤ Cx−(n+3)/2+iσ f H s−1 (X0,even ) , (5.117)
where C depends on σ . If f is supported in X0◦ , the s − 1 norm on f can be

replaced by the s − 2 norm.
5.6. Microlocal Analysis in the b-Setting

We now consider the geometric settings of b- and (complete) cusp analysis.
For a much more detailed description of the former, including parametrix
constructions for resolvents of elliptic operators, I refer to [34] and to [14]; here
we proceed via estimates as these generalize easily to non-elliptic settings.
Recall that Vb (M) is the set of smooth vector fields on a manifold with
boundary of corners; this does not depend on any choices. Fixing a boundary
defining function x modulo x2 C ∞ (M), we also let Vcu (M) be the set of
vector fields V tangent to ∂M such that Vx ∈ x2 C ∞ (M). This means that

V = b0 (x∂x ) + bj ∂yj and Vx = b0 x ∈ x2 C ∞ (M), so b0 ∈ xC ∞ (M), so

V = a0 (x2 ∂x ) + aj ∂yj , aj ∈ C ∞ (M).
These two structures are closely connected to each other (Figure 5.10).
Indeed, given a manifold with boundary with a C ∞ defining function x, choose
a local product decomposition near ∂M, so a neighborhood U of ∂M is
identified with [0, )x × ∂M; one gets local coordinates on this space via local
coordinates yj on ∂M. If we now introduce ρ(x) = e−1/x , x > 0, ρ(0) = 0, the
map ρ × id : [0, ) × ∂M → [0, e−1/ ) × ∂M is C ∞ , and the push forward of
a cusp vector field V is

a0 ρ∂ρ + aj ∂yj .
350 A. Vasy
Figure 5.10. The relationship between the b- (right), cusp (left), and cylindrical
(center) structures. The vertical direction is that of O ⊂ Rn−1 , that is the
y-variables
Note that the aj here need not be smooth after the push-forward: they are
smooth functions of (x, y), y ∈ ∂M, that is of (− log1 ρ , y), which is a much
weaker statement than smoothness in ρ.
Conversely, starting with Vb (M), with a boundary defining function ρ and a
local product decomposition of a neighborhood U of ∂M as [0, )ρ × ∂M, we
can let x = −1/(log ρ), which then gives rise to a cusp structure. The push-
forward of a C ∞ b-vector field is then a C ∞ cusp vector field, since smoothness
in ρ implies smoothness in x.
Letting t = x−1 , we have t = − log ρ, and the identification of (0, )ρ × ∂M
is with (− log , +∞)t × ∂M (the boundary is pushed out to infinity). Then

the push-forward of a0 ρ∂ρ + aj ∂yj is −a0 ∂t + aj ∂yj , and a coordinate
chart O in ∂M with coordinates y gives an identification of (0, ) × O with
a cylinder (− log , +∞)t × O ⊂ R × Rn−1 = Rn . Now, the smoothness
of the coefficients aj in [0, ) × O is equivalent to the statement that ∂ρk ∂yα aj is
bounded for all k, α on (0, )×O (with aj (0, y) = limρ→0 aj (ρ, y), and the limit
automatically existing under these conditions), and thus (as the push-forward
of ∂ρ = ρ −1 (ρ∂ρ ) is −et ∂t ) to the statement that (−et ∂t )k ∂yα aj is bounded on
(− log , +∞)t × O. This is a rather stringent requirement on the t derivatives.
A weaker, and more convenient, requirement on the aj is conormality. In
general, on a manifold with boundary with a boundary defining function ρ,
one says that a function a is conormal to ∂M of order s and writes a ∈ As ,
if ρ −s V1 . . . V
a is bounded for all
and all Vj ∈ Vb (M). Note that such a
statement is independent of the choice of ρ (since it is unchanged by inserting a
positive smooth factor into ρ), and locally, in coordinates (ρ, y), it is equivalent
to ρ −s (ρ∂ρ )k ∂yα a being bounded for all k, α, as can be seen (in the less trivial
direction) by expanding the Vj in terms of ρ∂ρ and ∂yj with smooth coefficients,
and using the product rule iteratively. Notice that this translates to est (−∂t )k ∂yα a
being bounded for all k, α, that is for s = 0 simply to the boundedness of all
(standard, constant coefficient) derivatives in the cylindrical region.
This means that conormal b-vector fields, that is elements of A0 ⊗C ∞ Vb (M),
that is vector fields of the form

a0 (ρ∂ρ ) + aj ∂yj , aj ∈ A0 ,
are exactly vector fields in the cylindrical region of Rn with coefficients that
are bounded with all derivatives; if they have support inside [t0 , +∞) × K,
K ⊂ O compact, t0 > − log , as one would arrange when using a partition
of unity on M, then they can also be regarded as vector fields on all of Rn
with coefficients bounded with all derivatives. But this means that they are
elements of the ∞ 1 = 1,0 space, and differential operators based on them
∞
are elements of ∞ m = m,0 . Correspondingly, the analysis of conormal b-
∞
differential operators, that is elements of A0 Diffb (M), is locally equivalent to
the analysis of ∞m.
Notice that Diffb (M) ⊂ A0 Diffb (M), so this in particular includes at least
aspects of the Diffb (M)-analysis (which however has stronger aspects as well).
In fact, what matters for us most right now is that if P ∈ Diffb (M) then

one can write in the local coordinate decomposition P = N−1 j=0 ρ Pj + ρ P̃,
j N
where Pj ∈ Diffb (M) have coefficients independent of ρ, and P̃ ∈ Diffb (M).

Correspondingly, translated to Rn , we have

N−1
P= e−jt Pj + e−Nt P̃,
j=0
with P̃ ∈ ∞ , and Pj having t-independent coefficients.

Turning to the cusp vector fields, as t = 1/x, the bordification of ( , ∞)t as
[0, 1/ ) via t → 1/t simply undoes the change of variables x → t, so elements
of C ∞ (M̃) are locally equivalently in C ∞ (Rt × Oy ).
In order to do the analysis on Rn , it is more convenient to bordify all
of Rn , not just a cylindrical subset, as Rt × Rn−1y , with the bordification of
the first factor given by t → t in the cusp setting, and t → e−t in the
−1
b-setting; in the latter case we write the compactification as Rexp × Rn−1 .

Then for P ∈ Diffm b (M), one has a principal symbol in S∞ (R ; R ) in fact
m n n
lying in C ∞ (Rexp × Rn−1 ; Sm (Rn )). Correspondingly, it is convenient to regard

Rexp × Rn−1 × Rn as the locus of b-microlocal analysis, with σb,m (P) defined
on this space. Thus, one defines the elliptic set Ell(P), the characteristic set
Char(P), and the operator wavefront set WF b (P) as subsets of this space.
Similarly, for cusp analysis the locus of microlocalization is R × Rn−1 where
352 A. Vasy
the lack of the subscript exp means that the reciprocal map t → t−1 is
used to construct the bordification. Since the connection between these two
compactifications is ρ = e−1/x , resp. x = −1/(log ρ), the two smooth structures
differ, with the extension of the identity map R → R to R → Rexp being a
C ∞ map, but its inverse is only continuous, not differentiable, that is R × Rn−1
and Rexp × Rn−1 are naturally homeomorphic (in the sense of the extension
of the identity map on Rn ), that is the same as topological manifolds, but not
diffeomorphic, though the identity map is (in the sense of “extends to”) C ∞ as
a map R × Rn−1 → Rexp × Rn−1 .
Concretely then define the local cusp algebra lcum , resp. the extended local
b-algebra elb , as quantizations of symbols a ∈ Slcu

m m (R × Rn−1 ; Rn ), resp.
m (R × Rn−1 ; Rn ), where a ∈ Sm (R × Rn−1 ; Rn ) if
a ∈ Selb lcu
|(t2 ∂t )k ∂yβ ∂ζα a| ≤ Cαβ,k ζ m ,
and a ∈ Selb
m (R × Rn−1 ; Rn ) if
|(eφ(t) ∂t )k ∂yβ ∂ζα a| ≤ Cαβ,k ζ m ,
where φ ∈ C ∞ (R) with φ(t) = |t| for |t| ≥ 1. We also define the extended
m . This means that
local symbolic b-algebra as ∞
∞
m
Slcu (R × Rn−1 ; Rn ) = C∞ (R × Rn−1 ; Sm (Rn ))
and
∞
m
Selb (R × Rn−1 ; Rn ) = C∞ (Rexp × Rn−1 ; Sm (Rn )),
where the subscript ∞ refers to uniform bounds in the non-compact variable y.

The reason for the term “extended” is the slow off-diagonal decay that we
discuss below; we in fact want to impose faster decay conditions.
The membership of elb m is equivalent to the statement that P ∈ m and
∞
for all N there exist R-translation invariant operators Pj,± ∈ ∞
m , j = 0, 1, . . . ,
N − 1, and P̃N,± ∈ ∞ m such that

N−1
P= e−jφ± (t) Pj,± + e−Nφ± (t) P̃N,± , (5.118)
j=0
where φ± (t) = ±t for ±t > 1, φ± (t) = 0 for ±t < −1, and φ± ∈ C ∞ (R); note
that the + sign is making a non-trivial statement over membership in ∞ m only
as t → +∞, while the − sign is making a non-trivial statement as t → −∞.

Thus, the Pj,± are Taylor series of P at t = ±∞ in Rexp .
In view of the algebra property of ∞ m , it is easy to check that the other
spaces mentioned also form algebras. Further, as ∞ m is bounded between
weighted Sobolev spaces H s,r → H s−m,r , so are its subalgebras. For the cusp
algebra this is the best boundedness statement one would like (essentially
because smoothness in x rather than in ρ = e−1/x is considered from the
perspective described at the beginning of this section), but in the b-setting one
would like to improve this to boundedness between exponentially weighted
Sobolev spaces:
Hbs,r = e−rφ H s (Rn ).
The obstacle for this is that the Schwartz kernels of elements of ∞ only
decay as CN z − z −N ; to have action on the exponentially weighted spaces,
with exponential weight in t, one needs superexponential decay. Note that if
A ∈ ∞m , and K is its Schwartz kernel, then with ψ ∈ C ∞ (R), identically 1
A c
near 0,
KR = (1 − ψ(t − t ))KA
defines an element R of ∞ −∞ , that is if we let K = ψ(t − t )K , then
B A
−∞
R = A − B ∈ ∞ , and the Schwartz kernel of B is supported in a region
where |t − t | < C. Correspondingly, for any r,
e−rφ Berφ ∈ ∞
m
,
since it has Schwartz kernel

e−r(φ(t)−φ(t )) KB ,
and in view of the boundedness of t − t , φ(t) − φ(t ) is also bounded, with
all derivatives. Now, we do not actually need compact supports in t − t ,
only superexponential decay, so it is natural to consider operators R with C ∞
Schwartz kernel KR satisfying estimates

|∂tk ∂tl ∂yα ∂y KR (t, y, t , y )| ≤ CklαβNM y − y −N e−Mφ(t−t )
β
(5.119)
−∞ , whose elements
for all α, β, k, l, M, N. These are of course elements of ∞
have similar estimates, except the decay is superpolynomial in t − t .
Definition 5.15 The local symbolic b-algebra ∪m lbc m consists of operators L

of the form L = B + R , with KB = ψ(t − t )KA , A ∈ ∞ m , ψ ∈ C ∞ (R)
c
identically 1 near 0, while R is as in (5.119).
The local b-algebra ∪m lbm consists of operators L ∈ m of the form B+R ,
lbc

with KB = ψ(t − t )KA , A ∈ elb m , ψ ∈ C ∞ (R) identically 1 near 0, while R
c
in addition to (5.119) has an expansion

N−1
R = e−jφ± (t) R j,± + e−Nφ± (t) R̃ N,± , (5.120)
j=0
354 A. Vasy
where R j,± are R-translation invariant (i.e. the Schwartz kernels depend on t, t
only via t − t ) satisfying the estimates (5.119) and R̃ N,± also satisfies (5.119).
Thus, both B and R have expansions in terms of translation-invariant operators
as in (5.118).
Notice that for R as in (5.119), for any r,
e−rφ R erφ ∈ ∞
−∞
,
since for ±r ≥ 0,

e−r(φ(t)−φ(t ))∓rφ(t−t )

is bounded: it is essentially e−r(|t|−|t |±|t−t |) , and |t| − |t | + |t − t | ≥ 0 (the
case r ≥ 0), while |t| − |t | − |t − t | ≤ 0 (the case r ≤ 0). In particular, such R

is bounded on Hbs,r → Hbs ,r for all s, s , r since this boundedness is equivalent

to the boundedness of erφ R e−rφ : H s → H s . Similarly, for B ∈ ∞ m arising
by localizing the Schwartz kernel of A ∈ ∞ near the diagonal as above,

m
B : Hbs,r → Hbs−m,r , so we conclude that elements of lb m map H s,r → H s−m,r

b b
continuously for all s, r.
For L ∈ lbc m we can now define the principal symbol σ (L) and the
b
wavefront set WF b (L) as the corresponding object in the decomposition of
Definition 5.15 for A. However, the subtlety is that we use a different notion
from that of the standard case at ∂Rn . Indeed, the partial compactification of
Rn we want to use is Rexp × Rn−1 (the compactification is partial as Rn−1 is
non-compact, but this is not important here). For A ∈ ∞ m we thus take the
standard notion of the principal symbol, by writing A = qL (a), and taking the
class of a in S∞ m /Sm−1 . However, the notion of the wavefront set we want is
∞
based on the notion of essential support in Rexp × Rn−1 × Rn ; thus a point
α ∈ Rexp × Rn−1 × Rn is not in esssuppb (a) if there exists a neighborhood U
of α in Rexp ×Rn−1 ×Rn such that a|U∩(Rn ×Rn ) is in S∞ −∞ . Similarly, we define
the elliptic and characteristic sets as subsets of Rexp × Rn−1 × Rn . Note that
given L, A is defined modulo ∞ −∞ since B is also so determined, and A and B
−∞ . Thus, WF (L) is a subset of R
exp × R
differ by an element of ∞ n−1 × ∂Rn ,
b
and is well-defined as such.
In order to explain the notion of compactification we have used, recall that
for L ∈ lbm we may regard the principal symbol as a function on R
exp ×R
n−1 ×
Rn which is symbolic in the last variable of order m, modulo such symbols

of order m − 1. Furthermore, if L has a homogeneous principal symbol, for
instance if it is classical, we may regard the principal symbol as a homogeneous
degree m function on Rexp × Rn−1 × (Rn \ {0}). In particular, if L ∈ lbc 0
is classical of order 0 (in the sense that A is such, which property is again
independent of the particular choice of A) then σb,0 (L) can be regarded as a

C ∞ function on Rexp × Rn−1 × ∂Rn .
We also need the notion of the wavefront set of a distribution. Since the
principal symbol captures b-pseudodifferential operators modulo one order
lower pseudodifferential operators, but without giving any additional decay
properties, the b-wavefront set microlocally captures the regularity (differen-
tiability) of distributions, but does not imply any additional decay properties
beyond what they a priori possess, much like the case of WFm,
∞ . Indeed, the
difference between WFb and WF∞ is the space on which the wavefront set
lives, reflecting the analogous difference between elliptic sets. Concretely:
Definition 5.16 Suppose k,

, m ∈ R, u ∈ Hbk,
. For α ∈ Rexp × Rn−1 × Rn , we
say α ∈/ WFm,
0,0 m,
b (u) if there exists A ∈ lbc elliptic at α such that Au ∈ Hb . We

/ WFb (u) if there exists A ∈ lbc elliptic at α such that Au ∈ Hb∞,
.
say that α ∈ ∞,
0,0
We also make the analogous definition for variable order spaces.

When
= 0, we sometimes drop the weight from the notation and simply
write WFm b (u).
Now, if p ∈ Slcu
m is homogeneous of degree m (outside a neighborhood of the
zero section) then writing ζ = (τ , η),

∂p ∂ ∂p ∂ ∂p ∂ ∂p ∂

Hp = − + − (5.121)
∂τ ∂t ∂t ∂τ ∂ηj ∂yj ∂yj ∂ηj
j
is homogeneous of degree m − 1 away from o, and it is smooth on R × Rn−1 ×

Rn , tangent to the boundary, ∂R × Rn−1 × Rn . Note that ∂t = −x2 ∂x shows
that the first two terms even vanish as b-vector fields at this boundary, that is
the lack of extra decay there arises from the y and η derivatives.
Similarly, if p ∈ Selb
m is homogeneous of degree m (outside a neighborhood
of the zero section) then Hp is homogeneous of degree m − 1 away from o, and

it is smooth on Rexp × Rn−1 × Rn , tangent to the boundary ∂Rexp × Rn−1 × Rn .
Correspondingly, we define the b-Hamilton vector field
Hp,b,m = ζ −m+1 Hp ∈ Vb (Rexp × Rn−1 × Rn ),
well defined modulo ρfiber Vb (Rexp × Rn−1 × Rn ). Now ∂t = −ρ∂ρ , so unlike

in the cusp setting none of the terms in (5.121) vanish in general at the
boundary, ρ = 0, as b-vector fields. Then the analogues of our results from
the scattering setting all go through, keeping in mind that we only have gains
in differentiability.
Note that the additional content in the next propositions is only what happens
at ∂Rexp × Rn−1 × Rn in the homogeneous setting. First, the elliptic result is
356 A. Vasy
m , except that if P ∈ m then one can also work

essentially the result for ∞ lbc
with exponentially weighted spaces.
Proposition 5.33 Suppose P ∈ elbc m = m is elliptic at α ∈ R
∞ exp × R
n−1 ×
−N
∂Rn . If u ∈ H for some N, α ∈ s−m
/ WFb (Pu), then α ∈ s
/ WFb (u).
On the other hand, suppose P ∈ lbc is elliptic at α ∈ Rexp × Rn−1 × ∂Rn .
m
If u ∈ Hb−N,r for some N, r, α ∈

/ WFs−m,r
b (Pu), then α ∈/ WFs,r
b (u).
For the non-elliptic setting, we assume that P has an expansion at infinity,

that is P ∈ elb m ; this assures a well-behaved Hamilton flow. In fact, one
should think of the expansion of P as being analogous to the requirement

of the operator possessing a homogeneous principal symbol in the standard
microlocal propagation of the singularities theorem. Then the proof of the
propagation of the singularities theorem presented in Section 5.4 goes through
without essential changes; one just needs to keep in mind that Rexp ×Rn−1 ×Rn
is the location of the analysis.
Proposition 5.34 Suppose P ∈ elb m . If u ∈ H −N for some N, then
s−m+1
WFb (u) \ WFb
s (Pu) is a union of maximally extended bicharacteristics,
that is integral curves of Hp,b,m in Char(P).
On the other hand, suppose P ∈ lb m . If u ∈ H −N,r for some N, r, then
b
WFs,r
b (u) \ WF s−m+1,r
b (Pu) is a union of maximally extended bicharacteristics,
that is integral curves of Hp,b,m in Char(P).
Proposition 5.35 Suppose P ∈ elb
m , L ⊂ ∂R
exp × R
n−1 × ∂Rn is a compact
smooth embedded submanifold, and
∓Hp,b,m ρfiber = β0 ρfiber , β0 |L > 0,

1 ∗
p̃ = σb,m−1 (P − P )
2i
satisfies
1−m
p̃ = ±β0 β̃ρfiber ,
and there is a quadratic defining function ρ1 of L in Char(P) ⊂ Rexp × Rn−1 ×

∂Rn such that
∓Hp,b,m ρ1 = β1 ρ1 + F2 + F3 , β1 > 0,
where F2 ≥ 0 and F3 vanishes cubically at L.

(i) Suppose that s − (m − 1)/2 + β̃ > 0 on L, and s − (m − 1)/2 + β̃ > 0 on

L. Then u ∈ H −N for some N, L ∩ WFs−m+1
b (Pu) = ∅, WFsb (u) ∩ L = ∅
imply WFsb (u) ∩ L = ∅.
(ii) Suppose that s − (m − 1)/2 + β̃ < 0 on L. If u ∈ H −N and L has a neigh-

borhood U such that (U \ L) ∩ WFsb (u) = ∅ and U ∩ WFs−m+1
b (Pu) = ∅,
then L ∩ WFb (u) = ∅.
s
The exponentially weighted version is:

m , L ⊂ ∂R
Proposition 5.36 Suppose P ∈ lb exp × R
n−1 × ∂Rn is a compact
smooth embedded submanifold, with

and
∓Hp,b,m ρbase = β2 β0 ρbase , β2 |L > 0.
Suppose further that

1
p̃ = σb,m−1 (P − P∗ )
2i
satisfies
1−m
∂Rn such that
∓Hp,b,m ρ1 = β1 ρ1 + F2 + F3 , β1 > 0,
where F2 ≥ 0 and F3 vanishes cubically at L.
(i) Suppose that s + β2 r − (m − 1)/2 + β̃ > 0 on L, and s + β2 r − (m − 1)/
2 + β̃ > 0 on L. Then u ∈ Hb−N,r for some N, r, L ∩ WFs−m+1,r
b (Pu) = ∅,

WFsb ,r (u) ∩ L = ∅ imply WFs,r
b (u) ∩ L = ∅.
(ii) Suppose that s + β2 r − (m − 1)/2 + β̃ < 0 on L. If u ∈ H −N,r and
L has a neighborhood U such that (U \ L) ∩ WFs,r
b (u) = ∅ and U ∩
s−m+1,r s,r
WFb (Pu) = ∅, then L ∩ WFb (u) = ∅.
We eventually consider the case when L is a saddle point of the flow, but
what is stated so far suffices for elliptic problems, as well as for the Minkowski
analysis.
While often not strictly necessary, it is convenient to globalize the boundary
pseudo-differential operators algebra to simplify various statements. The
reason this is often not necessary is that the symbolic considerations above are
almost all local, except that the radial set results are global within the radial set.
Thus, as long as the radial sets can be thought to lie over a single coordinate
chart, no globalization is needed. In fact, the radial set results themselves can
be localized, as was shown by Haber and Vasy [17], but we do not pursue this
358 A. Vasy
here. On the other hand, the normal operator considerations are global, but
for this we use the global Mellin transform on a collar neighborhood of the
boundary of our manifold M, and thus one need not explicitly globalize the
pseudo-differential operator algebra to deal with Diffb (M).
For the next definition recall that C˙∞ (M) is the subset of C ∞ (M) consisting
of functions vanishing with all derivatives at ∂M.
Definition 5.17 Suppose M is a compact manifold with boundary. We define

the symbolic (or conormal) b-algebra ∪m bcm (M), in analogy with standard
pseudodifferential operators on manifolds without boundary by requiring for

A ∈ bc
m (M) that
(i) A : C˙∞ (M) → C˙∞ (M) continuously (so in particular A has a distribu-
tional Schwartz kernel),
(ii) if U is any coordinate chart with : U → Ũ ⊂ Rn a diffeomorphism,
then for χ ∈ Cc∞ (U), (−1 )∗ χ Aχ ∗ ∈ lbc m ,
∞
(iii) and if χ , φ ∈ C (M) have disjoint support in local coordinate charts U,
resp. V, then χ Aφ has a Schwartz kernel which is conormal on M × M
(i.e. has iterated regularity relative to b-vector fields) relative to bounded
functions which decay rapidly as ρ/ρ tends to 0 or +∞, that is have
bounds CN (ρ/ρ )N in ρ/ρ < 1, CN (ρ /ρ)N in ρ/ρ > 1, where ρ stands
for the pull-back of ρ from the first factor, while ρ for the pull-back from
the second factor.
Specifically, (iii) means (see Figure 5.11)
(i) if U, V are both disjoint from ∂M then χ Aφ has a C ∞ Schwartz kernel,
(ii) if U is disjoint from ∂M and V is not, or vice versa, then χ Aφ has a C ∞
Schwartz kernel on M × M vanishing to infinite order M × ∂M, resp.
∂M × M,
(iii) if both U and V intersect ∂M, with a product decomposition [0, )ρ × U0 ,
resp. [0, )ρ × V0 , and with coordinate maps and then the Schwartz
kernel Kχ,ψ of (−1 )∗ χ Aφ ∗ satisfies, with e−t being the pull-back of

ρ from the first factor, e−t from the second factor, y, resp. y , coordinates
on U0 , resp. V0 , estimates

|∂tk ∂t
∂yα ∂y Kχ,ψ | ≤ CklαβM e−M|t−t | .
β
Notice that these are exactly the estimates (5.119), apart from the y − y
factor, which would not make sense here as U0 , V0 are unrelated coordi-
nate charts and which is irrelevant even for related coordinate charts, as
we have compact support in both y and y .
Figure 5.11. The product space M × M, with the two factors being horizontal and
vertical, respectively, and regions of the kind (i), (ii), and (iii) corresponding to
product sets U × V. Note that U ∩ V = ∅ means that (U × V) ∩ diag = ∅
Although we have not discussed geometric blow-ups, see [34] for a discus-
sion in precisely this setting, these are exactly the statement that the Schwartz
kernel of χ Aφ is conormal on [[0, )ρ × [0, )ρ ; {0} × {0}] × U0 × V0 , relative
to bounded functions with infinite order of vanishing at the lift of ρ = 0 and
ρ = 0: the various regions above describe various regions of this blow-up.
Note that the terms in part (iii) of the definition, corresponding to R
−∞
in the local algebra, are in bc (M), and thus do not affect the symbolic
statements/estimates. Here we emphasize that elements of the global algebra
−∞
bc (M) preserve the Sobolev spaces Hbs,r (M), by the same arguments as
above for R . Correspondingly, the above propositions are all valid in the global
algebra ∪m bc m (M), provided one also assumes smoothness in the coefficients
in the non-elliptic settings in the following sense.
Definition 5.18 The smooth subalgebra b (M) of bc m (M) consists of ele-
ments P of bc (M) for which there exist Pj ∈ bc ([0, ∞) × ∂M) which
m m
are dilation invariant and such that for all k, using a collar neighborhood

identification of a neighborhood of ∂M with [0, ) × ∂M, P − j<k Pj ∈
m (M).
ρ k bc
The one-step nature is not very important here; more general expansions
would also work. However, if one wants to ensure diffeomorphism invariance,
one must include all powers ρ α+j , j a positive integer, if one includes ρ α in the
expansion.
First consider elliptic problems. Suppose that L ∈ bc m (M) where M is a
compact manifold with boundary, and suppose that L is elliptic. Then, by
360 A. Vasy
the elliptic regularity (plus the closed graph theorem, or doing the estimate
directly), for all s, r ∈ R, and for all s ∈ R, there exists C such that
uH s,r ≤ C(LuH s−m,r + uH s ,r ), (5.122)

b b b
and one also has a similar estimate for L∗ as it is also elliptic. Since the

inclusion Hbs,r → Hbs ,r is non-compact, this is insufficient for Fredholm
analysis. This is where the normal operators enter.
So let us assume that L ∈ bm (M), that is L is smooth in the above sense.
Let N(L) = L0 denote the normal operator of L; L̂(σ ) be its conjugate by
the Mellin transform (sometimes called the indicial operator). Then L̂(σ ) is
elliptic, even in the large parameter sense, when σ lies in strips | Im σ | < C.
Correspondingly, L̂(σ )−1 exists in such strips when | Re σ | is large, and thus
L̂(σ )−1 is a meromorphic family in any such strip with finite rank poles.
Choose r such that −r is not the imaginary part of any of the poles. Then
one has the estimate
vH s ≤ CL̂(σ )vH s−m (5.123)
for compact regions of σ ; notice that this is just the standard elliptic estimate
with the usual weak error term vH −N dropped due to invertibility, as a
standard functional analysis argument allows one to do, see [27, Proof of
Theorem 26.1.7] or [47, Section 4.3]. Furthermore, in strips | Im σ | < C , one
has the semiclassical/large parameter estimates (recall |σ |−m L̂(σ ) = L̂h,z )
vH s ≤ C|σ |−m L̂(σ )vH s−m , |σ | $ 1. (5.124)

|σ |−1 |σ |−1
Now, the Mellin transform is an isomorphism between Hbs,r ([0, ∞) × ∂M) and
the large parameter Sobolev space, which for s ≥ 0 integer corresponds to the
squared norm

σ k V α v(σ )2L2 (∂M) dσ ,
|α|+k≤s Im σ =−r
where V = (V1 , . . . , Vl ) are vector fields on ∂M spanning V(∂M), with the

general case following by interpolation and duality. Thus, for v = v(σ ) =
(Mũ)(σ ), (5.123)–(5.124) is exactly an estimate in Hbs,r , if one performs the
inverse Mellin transform integral on the contour line Im σ = −r (which, recall,
is disjoint from poles). Thus, one has the estimate
ũH s,r ≤ CN(L)ũH s−m,r .

b b
Let χ ∈ C ∞ (M)
be supported in a collar neighborhood of ∂M as above in
which we have the stated expansion and the identification with [0, )ρ × ∂M,
also assume χ ≡ 1 near ∂M. We apply our normal operator estimate with s
replaced by s , and with ũ = χ u, so:
uH s ,r ≤ χuH s ,r +(1−χ )uH s ,r ≤ CN(L)(χ u)H s −m,r +(1−χ )uH s ,r .

b b b b b
Since 1 − χ is compactly supported in M◦, for any r , (1 − χ )uH s ,r ≤

b
CuH s ,r . On the other hand,
b
N(L)(χ u) = [N(L), χ ]u + χ Lu + χ (N(L) − L)u

m−1
so, using χ (N(L) − L) ∈ ρbm (M), [N(L), χ ] ∈ bc supported away from

∂M, for all r ,
N(L)(χ u)H s −m,r ≤ C(uH s −1,r + LuH s −m,r + uH s ,r−1 ).

b b b b
Thus,
uH s ,r ≤ CN(L)(χ u)H s −m,r +(1−χ )uH s ,r ≤ C(LuH s −m,r +uH s ,r−1 ).
b b b b b
Combining with (5.122),
uH s,r ≤ C(LuH s−m,r + uH s ,r )

b b b
≤ C(LuH s−m,r + uH s ,r−1 ).

b b

We now have the compact inclusion Hbs,r → Hbs ,r−1 that we desired.
A completely analogous argument for L∗ gives for any s̃, s̃ , r̃ with −r̃ not
being the imaginary part of any pole of L∗ (σ ),
uH s̃,r̃ ≤ C(L∗ uH s̃−m,r̃ + uH s̃ ,r̃ )

b b b
∗
≤ C(L uH s̃−m,r̃ + uH s̃ ,r̃−1 ).
b b
Note that for a product-type b-metric used to define the inner product L∗ (σ ) =
(L̂(σ ))∗ as shown by a simple computation, so the requirement on r̃ is that r̃ is
not the imaginary part of any pole of L̂(σ ).
Now take r̃ = −r, s̃ = m − s, so s̃ − m = −s, and notice the expected duality
between the spaces in the estimates for L and L∗ , as before. This yields that
L : Hbs,r → Hbs−m,r
is Fredholm. Notice that this is not quite an invertibility statement, though it

gives an inverse on a finite codimensional subspace.
362 A. Vasy
For the Laplacian g of a Riemannian b-metric g, if g is product type at ∂M,

that is
dρ 2
g= + h + g1 , g1 ∈ ρC ∞ (M; b T ∗ M ⊗s b T ∗ M), (5.125)
ρ2
h a metric on ∂M, extended via the local collar neighborhood identification,
and L = g − λ, N(L) = (ρDρ )2 + h − λ, so N̂(L) = h + σ 2 − λ. In
particular, this is invertible as long as λ − σ 2 is not an eigenvalue of h , in
particular if λ − σ 2 ∈ / [0, ∞). So for λ ∈ C \ [0, ∞), σ ∈ R, one has the
desired invertibility, which then gives the Fredholm statement for L as a map
Hbs,0 → Hbs−2,0 , as well as for Hbs,r → Hbs−2,r for |r| small. Since any element
of Ker L and Ker L∗ would then lie in Hb∞,r for some r > 0, it follows from the
symmetry of L that these kernels are trivial, thus the operator is invertible. We
state this as a theorem:
Theorem 5.9 Let g be a Riemannian b-metric of the type (5.125), and let λj ,
√ λ ∈ C \ [0, ∞),
j ≥ 1, be the eigenvalues of the boundary Laplacian h . For
g − λ is invertible as a map Hbs,r → Hbs−2,r for |r| < Im λ.

Furthermore, with σj = λ − λj , with the square root being in the lower

half-plane, for r > 0 such that r = σj for any j, if f ∈ Hbs−2,r then there exist
aj,k,
∈ C ∞ (∂M), j ∈ N+ , k ∈ N, N
≤
j,k such that
(g − λ)−1 f |ρ</2

= ρ iσj +k (− log ρ)
aj,k,
+ u , u ∈ Hbs,r .
j:0<− Im σj <r k:− Im σj +k<r
≤
j,k
Remark 5.6 Note that if Im σj becomes more negative, then ρ iσj vanishes
faster as ρ → 0.
Proof The first part is already

√ proved, taking into account that λj ≥ 0 implies

that Im λ − λj ≥ Im λ if both square roots are taken to have positive
imaginary parts, so the poles σj , which satisfy σj2 + λj = λ, lie at least distance
√
Im λ from the real axis.
For the second half, note that we have u = (g − λ)−1 f ∈ Hbs,0 . Thus, if
χ ≡ 1 where ρ ≤ /2, supported in ρ ≤ , then
N(g − λ)(χ u) = f + [g , χ ]u − (g − λ − N(g − λ))(χ u),
and [g , χ] and thus [g , χ ]u are compactly supported, thus in Hbs−1,∞ , while
(g − λ − N(g − λ))(χ u) ∈ Hbs−2,1 . Thus,
N(g − λ)(χ u) = f1 ∈ Hbs−2,1 .
Mellin transforming both sides we obtain holomorphic functions in Im σ > 0

with a continuous extension to Im σ ≥ 0 in the Sobolev sense discussed
above. Moreover the right hand side is actually holomorphic in Im σ > −1
with continuous extension to Im σ ≥ −1. On the other hand, N̂(g − λ)−1 is
meromorphic in C, and thus (N̂(g −λ)(σ ))−1 (Mf1 )(σ ) is such in Im σ > −1
(with continuous extension to Im σ ≥ −1 if there is no resonance with
imaginary part = −1). Letting r1 ∈ (0, 1] (so we may assume r1 is arbitrarily
close to 1) be such that no resonance lies in Im σ = −r1 , we have by Cauchy’s
theorem (crucially using the high energy estimates for (N̂(g − λ)(σ ))−1 to
ensure that there are no issues at infinity for the contour shifting)

ρu = (2π )−1 ρ iσ N̂(g − λ)(σ )−1 (Mf1 )(σ ) dσ
Im σ =0

−1
= (2π ) ρ iσ N̂(g − λ)(σ )−1 (Mf1 )(σ ) dσ
Im σ =−r1

+ aj,0,
ρ iσj (− log ρ)
aj,0,
,
j:0<− Im σj ≤r1
and the first term is in Hbs,r1 . This proves the theorem if r < 1.
In general, we iterate, using the already obtained partial expansion for ρu.
This gives that

(g − λ − N(g − λ))χ u = f2 + aj,0,
ρ iσj +1 (− log ρ)
aj,0,
,
j:0<− Im σj ≤r1
where now f2 ∈ Hbs−2,r1 +1 , and correspondingly its Mellin transform is

meromorphic in Im σ > −r1 with poles at σj − i with Im σj > −r1 . Taking r2
now with r2 ∈ (r1 , r1 + 1] and with no resonance on Im σj = −r2 (so again
r2 can be arbitrarily close to r1 + 1) we repeat the above argument, except that
Cauchy’s theorem now gives contributions from the positive integer shifted
poles σj − i, Im σj > −r1 , in addition to the poles σj with − Im σj ∈ (r1 , r2 ).
This gives the statement of the theorem if r < 2; in general it follows by
induction.
Note that the same argument gives in general the asymptotic expansion
of approximate solutions, with “approximate” understood in the sense that
(g − λ)u decays more than u does:

Proposition 5.37 Suppose u ∈ Hbs,r , r ∈ R, and (g − λ)u = f ∈ Hbs−2,r ,

r > r and r = − Im σj,± for any j. Let σj,± = ± λ − λj be the resonances.
Then
364 A. Vasy

u|ρ</2 = ρ iσj,± +k (− log ρ)
aj,k,
j,±:r≤− Im σj,± <r k:Im σj,± +k<r

≤
j,k

+ u , u ∈ Hbs,r .
Indeed, the same expansion works whenever one has a meromorphic inverse
normal operator family with large parameter estimates.
For a Riemannian scattering metric g on an n-dimensional space, one
computes quite easily that
L = ρ −(n−2)/2 ρ −2 g ρ (n−2)/2 ∈ Diff2b
has normal operator

(n − 2)2
N(L) = (ρDρ )2 + h + ,
4
so
(n − 2)2
L̂(σ ) = h + σ 2 + .
4
2
Thus the resonances σ satisfy σ 2 + (n−2) 4 = −λj , where λj ≥ 0 are the
eigenvalues of h . In particular, for n ≥ 3, there are no poles of L̂(σ )−1 in the
strip | Im σ | < (n − 2)/2, so for any |r| < (n − 2)/2,
L : Hbs,r → Hbs−2,r
is Fredholm. Moreover, any element of Ker L or Ker L∗ would necessarily be

in Hbs ,r for all s and for all r < (n − 2)/2. This means for u ∈ Ker L that

ρ (n−2)/2 u ∈ Ker g , and ρ (n−2)/2 u ∈ Hbs ,r for all r < n − 2. Now, in view of
ρ 2 g being a b-metric, the L2 space of the metric, Lg2 , which is the scattering L2
2 , is ρ n/2 L2 . Thus, for n−2 > n/2, that is n > 4, such a u is in L2 , with
space, Lsc b g
indeed all b-derivatives being in this space (so the scattering derivatives are in
better spaces). This suffices to use that g = dg∗ d to conclude that du2 = 0,
so du = 0, and thus u = 0. As a similar argument also applies to L∗ (which,
recall, was the adjoint with respect to a b-metric, such as ρ 2 g). This yields that
g is invertible as a map
s,r+(n−2)/2 s−2,r+(n+2)/2
g : H b → Hb
if r = 0, and thus if |r| < (n − 2)/2 by the Fredholm property. One also gets
an expansion of solutions of g u = f , with, say, f ∈ C˙∞ , in terms of the
resonances of N̂(L).
The cases n = 3, 4 require just a bit more work. Indeed, with λj still denoting
the eigenvalues of h , the resonances in the lower half-plane are

σj = −i λj + (n − 2)2 /4,
the square root being the positive one. Thus for any r, any element of Ker L (or
Ker L∗ ) on Hbs,r with |r| < (n − 2)/2 has an asymptotic expansion

u|ρ</2 = ρ iσj +k (− log ρ)
aj,k,
+ u , u ∈ Hbs,r .
j:0<− Im σj <r k:Im σj +k<r
≤
j,k
Thus, ρ (n−2)/2 u ∈ Ker g is C ∞ on M ◦ , and it tends to 0 at infinity, so

it vanishes by the maximum principle. Thus, in all dimensions ≥ 3, g is
s,r+(n−2)/2 s−2,r+(n+2)/2
invertible as a map Hb → Hb for |r| < (n − 2)/2. We
now state this as a theorem:
Theorem 5.10 Let g be a Riemannian scattering metric on M, dim M ≥ 3,
and let λj , j ≥ 1, be the eigenvalues of the boundary Laplacian h . Then g
s,r+(n−2)/2 s−2,r+(n+2)/2
is invertible as a map Hb → Hb for |r| < n−2
2 .

Furthermore, with σj = −i λj + (n − 2)2 /4, with the square root being the

positive one, for r > 0 such that r = σj,+ for any j, if f ∈ Hbs−2,r then there
exist aj,k,
∈ C ∞ (∂M), j ∈ N+ , k ∈ N, N
≤
j,k such that

−1
g f |ρ</2 = ρ iσj +k+(n−2)/2 (−log ρ)
aj,k,
+ u ,
j:0<− Im σj <r k:Im σj +k<r
≤
j,k
s,r +(n−2)/2
where u ∈ Hb .
The Fredholm argument goes through with minor changes if L is non-
elliptic:
Theorem 5.11 Suppose P ∈ Diffm b (M), with real principal symbol, and the
bicharacteristic flow is non-trapping in the sense that all the (null) bichar-
acteristics, apart from the ones contained in the radial sets, tend to radial
points of the kind described above (sources/sinks) in both the forward and in
the backward direction along the Hamilton flow.
Let r ∈ R. Suppose also that in each component of the characteristic set we
choose either the sources or the sinks as the locations to propagate estimates
from by demanding that s is larger than a threshold regularity, and demand that
at the other radial set in each component of the characteristic set s is smaller
than the threshold regularity, where s is a variable order monotone along the
bicharacteristics. Suppose that L̂(σ ) in invertible on the induced spaces on ∂M
when Im σ = −r. Then
L : X s,r → Y s−m+1,r
366 A. Vasy
is Fredholm with

Y s ,r = Hbs ,r , X s ,r = u ∈ Hbs ,r : Lu ∈ Hbs −m+1,r .
In particular, this applies to
L = ρ −(n−2)/2 ρ −2 g ρ (n−2)/2 ∈ Diff2b (M)
for non-trapping Lorentzian scattering metrics g (generalizing spaces asymp-
totic to Minkowski space), as shown by Baskin, Vasy, and Wunsch [2],
extended to a long-range setting in [1]. Lorentzian scattering metrics are
symmetric bilinear forms (non-positive definite inner products) of signature
(1, n − 1) on sc TM, with sc TM having been discussed in Section 5.3.11,
which have a certain form near ∂M generalizing that of Minkowski space,
see [2, Definition 3.1]. The main point of this form is that the radial set for
the Hamilton flow is the conormal bundle of a codimension 1 submanifold
S of ∂M, given by the zero set of a function v, with a specific structure
of the linearization of the Hamilton flow important for the more precise
results (in terms of the precise expansion at the radial set) shown in this
chapter. The non-trapping hypothesis is that S = S+ ∪ S− (disjoint union),
with all bicharacteristics tending to the radial sets over S± as their affine
parameter tends to ±∞, together with an additional topological condition that
{v > 0} = C+ ∪ C− (disjoint union of open sets), with S± = ∂C± , with
the latter assumption playing a role in the identification of the poles of the
indicial operator below. The C± are called the future/past “hyperbolic caps,”
while C0 = {v < 0} is the “asymptotically de Sitter region,” because in a
natural manner g induces asymptotically hyperbolic metrics on C± and an
asymptotically de Sitter metric on C0 .
Concretely then, in the case of non-trapping Lorentzian scattering metrics,
the Fredholm framework applies with any fixed r ∈ R, r = − Im σj for any
resonance σj , with variable order s equal to s = s+ ∈ R at the future radial sets
(those over S+ ) with s+ + r < 1/2, and s = s− at the past radial sets (those
over S− ) with s− + r > 1/2. Indeed, L̂(σ ) is meromorphic in
{σ ∈ C : s+ − 1/2 > Im σ > s− − 1/2},
with finitely many poles in strips when Im σ is restricted to compact subin-
tervals, so r can be chosen as desired, and proves the Fredholm property of L,
giving the forward solution if L is actually invertible (and otherwise the forward
solution on a finite codimensional subspace). If instead one reverses these
inequalities, one obtains the backward solution operator, which is the adjoint of
the forward operator. There are also Feynman propagators, when one always
propagates forward, resp. always propagates backwards, along the Hamilton
flow, that is where s is greater than the threshold value on one component of
the radial set over S+ (source for forward, sink for backward propagation) and
smaller on the other one, and similarly for the radial set over S− . These were
studied by Gell-Redman, Haber, and Vasy [12].
While we have not discussed actual invertibility, that is the triviality of
Ker L and Ker L∗ , this in fact can be done quite easily for the backward and
forward operators, using energy estimates on spaces with |r| small, assuming
that the poles of L̂(σ )−1 have an appropriate structure, and if one makes a
time-like assumption for a boundary defining function near the closed interior
of the future and past light cones at infinity. Indeed, in the case of the forward
operator one first shows that the resonances of L̂(σ ) are of three (potential)
kinds: induced by resonances of the resolvent of the asymptotically hyperbolic
Laplacian on the future cap, negatives of the resonances of the resolvent of
the asymptotically hyperbolic Laplacian on the past cap, as well as possibly
non-zero pure imaginary integers. Furthermore, the resonant state structure of
L̂(σ ) is such that the polar parts of the resonances induced by the resolvent
of the asymptotically hyperbolic Laplacian on the future cap are all supported
in the future cap, see [2], and the same holds for the potential resonances in
−iN+ which do not correspond to asymptotically hyperbolic resonances (they
are in fact differentiated delta distributions supported at the light cone). Thus,
assuming that all the resonances of the past hyperbolic cap lie in the open lower
half plane (so their negatives lie in the open upper half-plane), the argument of
the proof of Proposition 5.37 shows that elements of Ker L on the 0-weighted
space decay rapidly (faster than any power of ρ) near the closed past light cone,
and then standard energy estimates show that it in fact vanishes near there, and
thus globally, see [26]. Note that the resonances of the (n − 1)-dimensional
hyperbolic space actually satisfy this, indeed they lie in −i( n−2 2 + N) (they
actually do not exist when n is even, in which case however L̂(σ ) has the
differentiated delta distribution resonances mentioned above; this corresponds
to the strong Huygens principle), and thus the conclusion holds for Minkowski
space (since for Minkowski space the induced asymptotically hyperbolic space
is hyperbolic space) and its perturbations. Note also that a priori there are
at most finitely many resonances of an asymptotically hyperbolic space in
the closed upper half-plane corresponding to either L2 eigenvalues or the
bottom of the continuous spectrum (which would barely miss being in L2 ).
A similar result holds for Ker L∗ if one assumes that the resonances of the
future hyperbolic cap lie in the open lower half plane. Thus one has:
Theorem 5.12 ([2, 26]) Suppose g is a non-trapping Lorentzian scattering

metric, with M having a time-like boundary defining function near the closed
368 A. Vasy
interior of the future and past light cones at infinity and the asymptotically
hyperbolic Laplacians on the future and past caps have all resonances in
the open lower half-planes. Then all the hypotheses of Theorem 5.11 are
satisfied for |r| small and a suitable choice of s above threshold regularity
at the past, below threshold regularity at the future, so the wave operator g is
s,r+(n−2)/2 s−2,r+(n+2)/2
invertible as a map Hb (M) → Hb (M), that is on forward-
in-time (retarded) function spaces. Similar results hold for the backward-in-
time (advanced) function spaces, with the role of future and past reversed.
On the other hand, for the Feynman propagators on Rn , one can find the
2 + N), so there are no poles
poles of L̂(σ ) directly, namely they are in ±i( n−2
near the real axis. Then the perturbation stability of the b-analysis means that
this conclusion also holds for perturbations of Minkowski space, see [12],
meaning that L : X s,r → Y s−1,r is Fredholm for |r| small. In fact, the complex
scaling (Wick rotation) argument of [12] also shows that Ker L, Ker L∗ are
trivial for Minkowski space for |r| small, so L is actually invertible, and thus
the same holds for perturbations of Minkowski space.
Finally, for the Feynman propagator in general, a generalization of the
positivity argument of [50] carried out in [51], as applied to L̂(σ ) when σ is
real, so L̂(σ ) is a symmetric operator, shows that any element of the nullspace,
which is a priori a conormal distribution associated to the light cones, is
necessarily in C ∞ , so in particular one has an element of the nullspace for
the operator L̂(σ ) on the forward and backward function spaces as well, thus a
resonance for L. Thus, if we assume that the resonances of the asymptotically
hyperbolic caps lie in the open lower half-plane as for the forward/backward
problems, we deduce that any element of Ker L̂(σ ) is trivial, with the analogous
statement for the kernel of the adjoint, so in fact there cannot be any Feynman
resonances on the real axis. This gives the Fredholm statement for L, and thus
g , for |r| small. Furthermore, a variation of this positivity argument, again
carried out in [51], shows that any element of Ker L is in Hb∞,r , and thus is in
the nullspace of L acting on the forward/backward spaces. Correspondingly,
if one also assumes that there are time-like boundary defining functions
near the asymptotically hyperbolic caps using Theorem 5.12 one concludes
that all elements of Ker L are trivial, and the same holds for Ker L∗ . We
thus have:
Theorem 5.13 ([12, 51]) Suppose g is a non-trapping Lorentzian scattering
metric, with M having a time-like boundary defining function near the closed
interior of the future and past light cones at infinity and the asymptotically
hyperbolic Laplacians on the future and past caps have all resonances in the
open lower half planes. Then all the hypotheses of Theorem 5.11 are satisfied
for |r| small and a suitable choice of s with above threshold regularity at the
sources, below threshold regularity at the sinks, so the wave operator g is
s,r+(n−2)/2 s−2,r+(n+2)/2
invertible as a map Hb (M) → Hb (M), i.e. on Feynman
function spaces. A similar result holds on anti-Feynman function spaces, with
the role of sources and sinks reversed.
There are three additional complications for Kerr–de Sitter space, two of
which also hold for de Sitter-like spaces from a similar perspective (as opposed
to the perspective of extension across the boundary discussed in Section 5.5).
The first requires a simple modification of Proposition 5.36 to saddle points
of the Hamilton flow; in this case these are given by the conormal bundle of
the event horizon at the boundary. The other is that the global structure of
the spacetime is more complex in that the null-bicharacteristics do not tend to
sources/sinks beyond the event horizon. One way to handle this is to impose
complex absorption which gives rise to desired Fredholm statements, but
without giving the precise support properties required for the identification of
the forward/backward solutions unless one imposes some additional structure
as in [48] where the Mellin transform properties far from the real axis were
used very strongly. Another way is to add Cauchy hypersurfaces, which is
less microlocal, but imposes the vanishing properties automatically; this was
the approach of [26]. Finally, the third is trapping (this is the part that is not
∗ M which
relevant for de Sitter-like spaces): there is an additional set in b S∂M
acts as a saddle manifold of the flow in a more complicated way than the
radial set, for instance there are both stable and unstable directions within the
boundary which means that it is a trapped set at high energies (semiclassical
setting) even for the Mellin transformed normal operator. More precisely,
this is normally hyperbolic trapping in the appropriate sense. Fortunately, the
behavior of such trapping in the standard microlocal setting was extensively
studied by Wunsch and Zworski [56], Nonnenmacher and Zworski [36], and
Dyatlov [9, 10] recently, and one can apply their results here.
The propagation of singularities at saddles which are sources/sinks within
the boundary, but have a stable/unstable manifold transversal to the boundary,
is the statement that above the threshold regularity one can propagate estimates
from outside the boundary into the boundary, while below the threshold
regularity one can propagate estimates from the boundary into the interior.
Concretely, the result is
m , L ⊂ ∂R
Proposition 5.38 ([26, Proposition 2.1]) Suppose P ∈ lb exp ×
Rn−1 × ∂Rn is a compact smooth embedded submanifold, with
370 A. Vasy
and
±Hp,b,m ρbase = β2 β0 ρbase , β2 |L > 0.
Suppose further that

1
p̃ = σb,m−1 (P − P∗ )
2i
satisfies
1−m

∂Rn such that
∓Hp,b,m ρ1 = β1 ρ1 + F2 + F3 , β1 > 0,
where F2 ≥ 0 and F3 vanishes cubically at L. Let L be the stable (+) or

unstable (−) submanifold of L, automatically transversal to ∂Rexp × Rn−1 ×
∂Rn in Rexp × Rn−1 × ∂Rn ; the unstable, resp. stable, manifold of L on the
other hand lies in ∂Rexp × Rn−1 × ∂Rn .
(i) Suppose that s − β2 r − (m − 1)/2 + β̃ > 0 on L, and s + β2 r − (m − 1)/2 +
β̃ > 0 on L. Suppose u ∈ Hb−N,r for some N, r, L ∩ WFs−m+1,r
b (Pu) = ∅,

WFsb ,r (u) ∩ L = ∅, and suppose that L has a neighborhood U such that
(U \ L) ∩ L ∩ WFs,r s,r
b (u) = ∅. Then WFb (u) ∩ L = ∅.
(ii) Suppose that s − β2 r − (m − 1)/2 + β̃ < 0 on L. If u ∈ H −N,r and L has
a neighborhood U such that
(U \ L) ∩ (∂Rexp × Rn−1 × ∂Rn ) ∩ WFs,r

b (u) = ∅
and U ∩ WFs−m+1,r
b (Pu) = ∅, then L ∩ WFs,r
b (u) = ∅.
This at once yields the Fredholm statement in the presence of saddle points
of the flow and complex absorption (though not the trapping):
Theorem 5.14 (cf. [26, Proposition 2.3]) Suppose P ∈ Diffm b (M), with real
principal symbol, and the bicharacteristic flow is non-trapping in the sense
that all the (null) bicharacteristics, apart from the ones contained in the radial
sets, which are saddle points of the kind described above (so exactly one of
the stable and unstable manifolds is a subset of the boundary while the other
is transversal to it), either reach complex absorption in finite time, or tend
to radial points of the kind described above, in both the forward and in the
backward direction along the Hamilton flow.
Let r ∈ R, and suppose also that s is a variable order (possibly constant)
monotone along the bicharacteristics with the property that if s is above the
threshold regularity in a component of the characteristic set then estimates are

propagated into ∂M (i.e. s is decreasing along bicharacteristics which tend
to the radial set from outside ∂M), and if s is below threshold regularity in a
component of the characteristic set then estimates are propagated out of ∂M.
Suppose that L̂(σ ) is invertible on the induced spaces on ∂M when Im σ = −r.
Then
L : X s,r → Y s−m+1,r
is Fredholm with

Y s ,r = Hbs ,r , X s ,r = {u ∈ Hbs ,r : Lu ∈ Hbs −m+1,r }.
This now describes the microlocal analysis in a neighborhood of the static
patch of de Sitter space, and indeed on the analogous region in asymptotically
de Sitter spaces. The only missing ingredient for Kerr–de Sitter space is the
analysis of trapping, which, as discussed above, is normally hyperbolic in
this case, and was analyzed by Wunsch and Zworski [56], Nonnenmacher
and Zworski [36], and Dyatlov [9, 10]. Note that on growing function spaces
trapping is not an issue, and the standard non-trapping-type estimates hold, see
[25], where borderline (no decay, but no growth) estimates are also discussed.
These can be combined with Theorem 5.14 using gluing methods obtained by
Datchev and Vasy [5, 6]. We refer to the discussion around Definition 2.16–
Theorem 2.17 in [48] for details.
References
[1] D. Baskin, A. Vasy, and J. Wunsch. Asymptotics of scalar waves on long-range
asymptotically Minkowski spaces. Preprint, arXiv:1602.04795, 2016.
[2] Dean Baskin, András Vasy, and Jared Wunsch. Asymptotics of radiation fields in
asymptotically Minkowski space. Amer. J. Math., 137(5):1293–1364, 2015.
[3] Michael Beals and Michael Reed. Microlocal regularity theorems for nonsmooth
pseudodifferential operators and applications to nonlinear problems. Trans. Amer.
Math. Soc., 285(1):159–184, 1984.
[4] Xi Chen and Andrew Hassell. Resolvent and spectral measure on non-trapping
asymptotically hyperbolic manifolds I: Resolvent construction at high energy.
Comm. Partial Differential Equations, 41(3):515–578, 2016.
[5] Kiril Datchev and András Vasy. Gluing semiclassical resolvent estimates via
propagation of singularities. Int. Math. Res. Not. IMRN, (23):5409–5443, 2012.
[6] Kiril Datchev and András Vasy. Propagation through trapped sets and semiclas-
sical resolvent estimates. Ann. Inst. Fourier (Grenoble), 62(6):2347–2377 (2013),
2012.
[7] J. J. Duistermaat. On Carleman estimates for pseudo-differential operators. Invent.
Math., 17:31–43, 1972.
372 A. Vasy
[8] J. J. Duistermaat and L. Hörmander. Fourier integral operators. II. Acta Math.,
128(3-4):183–269, 1972.
[9] Semyon Dyatlov. Resonance projectors and asymptotics for r-normally hyper-
bolic trapped sets. J. Amer. Math. Soc., 28(2):311–381, 2015.
[10] Semyon Dyatlov. Spectral gaps for normally hyperbolic trapping. Ann. Inst.
Fourier (Grenoble), 66(1):55–82, 2016.
[11] Semyon Dyatlov and Maciej Zworski. Dynamical zeta functions for Anosov flows
via microlocal analysis. Ann. Sci. Éc. Norm. Supér. (4), 49(3):543–577, 2016.
[12] Jesse Gell-Redman, Nick Haber, and András Vasy. The Feynman propagator on
perturbations of Minkowski space. Comm. Math. Phys., 342(1):333–384, 2016.
[13] Robert Geroch. Domain of dependence. J. Mathematical Phys., 11:437–449,
1970.
[14] D. Grieser. Basics of the b-calculus. Preprint, arXiv:math/0010314, 2000.
[15] Colin Guillarmou. Meromorphic properties of the resolvent on asymptotically
hyperbolic manifolds. Duke Math. J., 129(1):1–37, 2005.
[16] Victor Guillemin and David Schaeffer. On a certain class of Fuchsian partial
differential equations. Duke Math. J., 44(1):157–199, 1977.
[17] Nick Haber and András Vasy. Propagation of singularities around a Lagrangian
submanifold of radial points. Bull. Soc. Math. France, 143(4):679–726, 2015.
[18] A. Hassell, R. B. Melrose, and A. Vasy. Spectral and scattering theory for
symbolic potentials of order zero. Advances in Mathematics, 181:1–87, 2004.
[19] A. Hassell, R. B. Melrose, and A. Vasy. Microlocal propagation near radial points
and scattering for symbolic potentials of order zero. Analysis and PDE, 1:127–
196, 2008.
[20] Ira Herbst and Erik Skibsted. Quantum scattering for potentials independent of |x|:
asymptotic completeness for high and low energies. Comm. Partial Differential
Equations, 29(3-4):547–610, 2004.
[21] Ira Herbst and Erik Skibsted. Absence of quantum states corresponding to
unstable classical channels. Ann. Henri Poincaré, 9(3):509–552, 2008.
[22] P. Hintz and A. Vasy. Global analysis of quasilinear wave equations on asymptot-
ically Kerr–de Sitter spaces. Int. Math. Res. Notices, 2016(17):5355–5426, 2016.
[23] P. Hintz and A. Vasy. The global non-linear stability of the Kerr–de Sitter family
of black holes. Preprint, arXiv:1606.04014, 2016.
[24] Peter Hintz. Global analysis of quasilinear wave equations on asymptotically de
Sitter spaces. Ann. Inst. Fourier (Grenoble), 66(4):1285–1408, 2016.
[25] Peter Hintz and Andras Vasy. Non-trapping estimates near normally hyperbolic
trapping. Math. Res. Lett., 21(6):1277–1304, 2014.
[26] Peter Hintz and András Vasy. Semilinear wave equations on asymptotically de
Sitter, Kerr–de Sitter and Minkowski spacetimes. Anal. PDE, 8(8):1807–1890,
2015.
[27] L. Hörmander. The analysis of linear partial differential operators, vol. 1-4.
Springer-Verlag, Berlin and Heidelberg, 1983.
[28] Lars Hörmander. On the existence and the regularity of solutions of linear pseudo-
differential equations. Enseignement Math. (2), 17:99–163, 1971.
[29] Rafe R. Mazzeo and Richard B. Melrose. Meromorphic extension of the resolvent
on complete spaces with asymptotically constant negative curvature. J. Funct.
Anal., 75(2):260–310, 1987.
[30] R. B. Melrose. Spectral and scattering theory for the Laplacian on asymptotically
Euclidian spaces. Marcel Dekker, New York, 1994.
[31] R. B. Melrose. Lecture notes for ‘18.157: Introduction to microlocal analysis’.
Available at http://math.mit.edu/∼rbm/18.157-F09/18.157-F09.html, 2009.
[32] R. B. Melrose, A. Sá Barreto, and A. Vasy. Analytic continuation and semiclas-
sical resolvent estimates on asymptotically hyperbolic spaces. Comm. in PDEs,
39(3):452–511, 2014.
[33] Richard B. Melrose. Transformation of boundary problems. Acta Math., 147(3-4):
149–236, 1981.
[34] Richard B. Melrose. The Atiyah-Patodi-Singer index theorem, volume 4 of
Research Notes in Mathematics. A K Peters Ltd., Wellesley, MA, 1993.
[35] Stéphane Nonnenmacher and Maciej Zworski. Quantum decay rates in chaotic
scattering. Acta Math., 203(2):149–233, 2009.
[36] Stéphane Nonnenmacher and Maciej Zworski. Decay of correlations for normally
hyperbolic trapping. Invent. Math., 200(2):345–438, 2015.
[37] Cesare Parenti. Operatori pseudo-differenziali in Rn e applicazioni. Ann. Mat.
Pura Appl. (4), 93:359–389, 1972.
[38] Xavier Saint Raymond. A simple Nash–Moser implicit function theorem.
Enseign. Math. (2), 35(3-4):217–226, 1989.
[39] I. M. Sigal and A. Soffer. N-particle scattering problem: asymptotic completeness
for short range systems. Ann. Math., 125:35–108, 1987.
[40] Plamen Stefanov, Gunther Uhlmann, and Andras Vasy. Boundary rigidity with
partial data. J. Amer. Math. Soc., 29(2):299–332, 2016.
[41] M. A. Šubin. Pseudodifferential operators in Rn . Dokl. Akad. Nauk SSSR,
196:316–319, 1971.
[42] Gunther Uhlmann and András Vasy. The inverse problem for the local geodesic
ray transform. Invent. Math., 205(1):83–120, 2016.
[43] André Unterberger. Résolution d’équations aux dérivées partielles dans des
espaces de distributions d’ordre de régularité variable. Ann. Inst. Fourier (Greno-
ble), 21(2):85–128, 1971.
[44] A. Vasy. Propagation of singularities in many-body scattering. Ann. Sci. École
Norm. Sup. (4), 34:313–402, 2001.
[45] A. Vasy. Propagation of singularities in many-body scattering in the presence of
bound states. J. Func. Anal., 184:177–272, 2001.
[46] András Vasy. Analytic continuation and high energy estimates for the resolvent
of the Laplacian on forms on asymptotically hyperbolic spaces. Adv. Math.,
306:1019–1045, 2017.
[47] A. Vasy. Microlocal analysis of asymptotically hyperbolic spaces and high energy
resolvent estimates, volume 60 of MSRI Publications. Cambridge University
Press, Cambridge, 2012.
[48] A. Vasy. Microlocal analysis of asymptotically hyperbolic and Kerr–de Sitter
spaces. Inventiones Math., 194:381–513, 2013. With an appendix by S. Dyatlov.
[49] A. Vasy. Some recent advances in microlocal analysis. In Proceedings of the ICM,
pp. 915–939, 2014.
[50] András Vasy. On the positivity of propagator differences. Ann. Henri Poincaré,
18(3):983–1007, 2017.
374 A. Vasy
[51] A. Vasy and M. Wrochna. Quantum fields from global propagators on asymptot-
ically Minkowski and extended de Sitter spacetime. Preprint, arXiv:1512.08052,
2015.
[52] András Vasy. Propagation of singularities for the wave equation on manifolds with
corners. Ann. of Math. (2), 168(3):749–812, 2008.
[53] András Vasy. The wave equation on asymptotically de Sitter-like spaces. Adv.
Math., 223(1):49–97, 2010.
[54] András Vasy. Resolvents, Poisson operators and scattering matrices on asymptot-
ically hyperbolic and de Sitter spaces. J. Spectr. Theory, 4(4):643–673, 2014.
[55] Yiran Wang. The resolvent of the Laplacian on non-trapping asymptotically
hyperbolic manifolds. ProQuest LLC, Ann Arbor, MI, 2015. Thesis (Ph.D.)–
Purdue University, arXiv:1410.6936.
[56] Jared Wunsch and Maciej Zworski. Resolvent estimates for normally hyperbolic
trapped sets. Ann. Henri Poincaré, 12(7):1349–1385, 2011.
[57] Maciej Zworski. Resonances for asymptotically hyperbolic manifolds: Vasy’s
method revisited. J. Spectr. Theory, 6(4):1087–1114, 2016.
[58] Maciej Zworski. Semiclassical analysis, volume 138 of Graduate Studies in
Mathematics. American Mathematical Society, Providence, RI, 2012.
Department of Mathematics, Building 380, Stanford University, 450 Serra Mall,

Stanford CA 94305-2125, USA.

Asymptotic Analysis in General Relativity

Uploaded by

Copyright:

Available Formats

Asymptotic Analysis in General Relativity

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Asymptotic Analysis in General Relativity

Uploaded by

Copyright:

Available Formats

L O N D O N M AT H E M AT I C A L S O C I E T Y L E C T U R E N OT E S E R I E S

Asymptotic Analysis in General Relativity

Cambridge University Press is part of the University of Cambridge.

1 Introduction to Modern Methods for Classical and

2 Geometry of Black Hole Spacetimes 9

3 An Introduction to Conformal Geometry and Tractor

3.3 Lecture 2: Conformal Transformations and Conformal Covariance 94

4 An Introduction to Quantum Field Theory on

5 A Minicourse on Microlocal Analysis for Wave Propagation 219

1.1. Geometry of Black Hole Spacetimes

1.2. Quantum Field Theory on Curved Spacetimes

1.3. Conformal Geometry and Conformal Tractor Calculus

1.4. A Minicourse in Microlocal Analysis

[28] D. C. Robinson, Four decades of black hole uniqueness theorems. In D.L.

Laboratoire AGM, Département de Mathématiques, Université de Cergy-Pontoise,

Abstract. These notes, based on lectures given at the summer school on

non-linear stability results without symmetry assumptions, including the sta-

The discussion of black hole spacetimes is started in Section 2.3 with a

4 2 = (g ) dxa dxb , and metrics, eg. (g ) ,

Figure 2.1. Causal diagram of Minkowski space

Then, with 2 = 2 cos U cos V, the conformally transformed metric g̃ab =

which we recognize as the metric on the cylinder R × S3 . This spacetime is

2.2.2. Lorentzian Geometry and Causality

product × R. In the following, unless otherwise stated, we shall consider

We remark that for a globally hyperbolic spacetime, which is extendible, the

2.2.3. Conventions and Notation

where ac b = 1 gbd (∂ g + ∂ g − ∂ g ) is the Christoffel symbol. In order

the first Bianchi identity R[abc]d = 0. Here square brackets [· · · ] denote

Rabcd = − 12 gad gbc R + 12 gac gbd R + 2 gbd Sac

2.2.4. Einstein Equation

∇ a Fab = 0, ∇[a Fbc] = 0

imply that Tab is conserved, ∇ a Tab = 0. The contracted second Bianchi

where Lmatter is the Lagrangian describing the matter content in spacetime. In

2.2.5. The Cauchy Problem

where H and J are the densitized left-hand sides of (2.11). If we consider

By standard results, the equation (2.15) has a locally well-posed Cauchy

Figure 2.2. A partial Cauchy surface which touches the boundary of ∂U

for a construction of the maximal globally hyperbolic extension which does

2.2.6.1. Bianchi Identities as a Hyperbolic System

2.2.6.2. Null Condition

ab = − 2 g ∂c ∂d gab + Sab (g, ∂g)

2.2.6.3. Gauge Source Functions

The field equations can be written as a system of equations for tetrad

2.2.6.4. Asymptotically Flat Data

where dσ i denotes the hypersurface area element of a family of spheres (which

2.2.6.5. Killing Initial Data

2.2.6.6. Komar Integrals

is the conservation of the charge integrals S Fab , S abcd F , which again

If we consider asymptotically flat spacetimes, we have in the stationary case,

2.3. Black Holes

with f = 1 − 2M/r. Here d

curvature diverges as 1/r3 . The Schwarzschild metric is asymptotically flat and

This solves dr∗ = f −1 dr, r∗ (4M) = 4M. As r  2M, r∗ diverges

where W is the principal branch of the Lambert W function.7 We can now

On the exterior region in Schwarzschild, (u, v) take values in the range

u = − 4M log(− tan U ), U ∈ (−π/2, 0)

t = 12 (v + u) = 4M log (− tan V tan U )

r = 2MW(−e−1 tan U tan V) + 2M (2.22)

In particular, the location r = 2M of the coordinate singularity in the

where ac b = 1 gbd (∂ g + ∂ g − ∂ g ) is the Christoffel symbol. In order

is the conservation of the charge integrals S Fab , S abcd F , which again

This solves dr∗ = f −1 dr, r∗ (4M) = 4M. As r 2M, r∗ diverges

ka = 12 δka log(A(S)) (2.25)

Here R = R(r, ∂t , ) where

The inverse AB of AB is defined by AB CB = δA C , AB AC = δC B . As with

For s = 1/2, this is the Dirac–Weyl equation ∇A A ϕA = 0, for s = 1, we

I {1, 1, 1, 1} ABCD = α(A βB γC δD) I

2 = − la mb m̄d nc Cabcd . (2.50)

(ABC F κD···E)F = 0 (2.52)

(ABC E κD)E = 0. (2.53)