SEMINAIRE BOURBAKI: Volume 2009/2010. Exposes 1012–1026 2856293263, 9782856293263

As in the preceding volumes of this seminar, which now counts more than one thousand talks, one finds here fifteen surve

179 44 13MB

English; French Pages 409 [448] Year 2011

Report DMCA / Copyright

DOWNLOAD FILE

Polecaj historie

SEMINAIRE BOURBAKI: Volume 2009/2010. Exposes 1012–1026
 2856293263, 9782856293263

Table of contents :
Introduction
1. Courbes sur les corps p-adiques : résultats classiques
2. Surfaces sur les corps p-adiques : quelques applications de la méthode de Bloch
3. LE THÉORÈME DE S. SAITO ET K. SATO SS2
4. Résultats récents et questions ouvertes
Références
Introduction
1. Modular forms, Hecke algebras, and Galois representations
2. p-adic families of systems of Hecke eigenvalues
References
Introduction
1. Description et premiers exemples
1.1. Description
1.2. Premier exemple
1.3. Algèbres amassées de rang 2
2. Algèbres amassées associées aux carquois
2.1. Mutation des carquois
2.2. Mutation des graines, algèbres amassées
2.3. Y-graines, application à la conjecture de périodicité
3. Algèbres amassées à coefficients
3.1. Définition
3.2. Exemple : le cône sur la grassmannienne des plans d'un espace vectoriel
3.3. Exemple : le sous-groupe unipotent maximal de SL(n+1,C)
3.4. Exemple d'application aux bases duales semi-canoniques
4. Catégorifications
4.1. Catégorification additive : la catégorie amassée
4.2. Catégorification additive : modules sur les algèbres préprojectives
4.3. Catégorification monoïdale
4.4. Catégorification via les carquois à potentiel
Remerciements
References
1. Introduction
2. Initial value problem
3. Stability of Kerr
4. Stability of Minkowski Space
5. Linear stability of the Kerr family
6. Vector field method for the wave equation
7. Red Shift
8. Boundedness results
9. Decay Mechanism
10. Appendix
References
Introduction
1. 2d NS Equations
2. Invariant measure
3. Dissipation and smoothing
4. Asymptotic strong Feller property
5. Hypoellipticity
6. Malliavin matrix
7. Low mode control
8. Turbulence
References
Introduction
1. Représentations en caractéristique p
2. Représentations en caractéristique 0
3. La série principale unitaire
4. La correspondance pour GL2(Qp)
5. Applications
Références
Introduction
1. Géométrie kählérienne torique
2. Démonstration du théorème
Références
Introduction
1. Matrices gaussiennes et lois déterminantales
2. Universalité pour un certain ensemble de matrices de Wigner
3. Universalité au bord du spectre
4. Universalité à l'intérieur du spectre; le théorème des quatre moments
5. Universalité à l'intérieur du spectre ; une approche dynamique
Références
Introduction
1. La structure des groupes p-compacts
2. La classification des groupes de pseudo-réflexions
3. La classification des groupes p-compacts
4. Décompositions homotopiques
5. Éléments de la démonstration pour p impair
6. Les démonstrations dans le cas p=2
7. Quelques applications de la classification
8. Espaces de lacets finis rationnellement exotiques
Références
Introduction
1. GROUPES PSEUDO-RÉDUCTIFS
2. STRUCTURE DES GROUPES ALGÉBRIQUES CONNEXES
3. CLASSIFICATION DES GROUPES PSEUDO-RÉDUCTIFS
4. PREMIÈRES APPLICATIONS DE LA CLASSIFICATION
5. UNIRATIONALITÉ, FINITUDES ET COMPACITÉ
Références
Introduction
1. Restrictions
2. Examples
3. The Bieri–Neumann–Strebel invariant of a Kähler group
References
Introduction
1. Pourquoi étudier ces groupes ?
2. Espaces virtuellement classifiants
3. Structures des sous-groupes
4. Géométrie asymptotique
Références
1. Problèmes variationnels invariants conformes en dimension 2
2. Les ingrédients utilisés
3. Les systèmes antisymétriques
4. Les surfaces de Willmore
Références
Introduction
1. Introduction to log canonical thresholds
2. Formula for the log canonical threshold
3. The log canonical threshold of an ideal of functions
4. Approximation of the log canonical threshold
5. Generic limits of ideals
6. ACC for log canonical thresholds on smooth varieties
References
1. Introduction to profinite groups
2. Cartesian products and their images
3. Subgroups of finite index
4. Some notation
5. Baire's category theorem and verbal subgroups
6. Width of words; Serre's theorem
7. Subgroups of finite index and d-finite words
8. Other closed verbal subgroups
9. Uniform bounds
10. Hensel's lemma for groups
11. Twisted commutators in simple groups
12. Applications to model theory
References
Table par noms d'auteurs

Citation preview

339

ASTÉRISQUE

XXXXXXXXXXFIRSTPAGEXXXXXXXXXX

2011

SÉMINAIRE BOURBAKI VOLUME 2009/2010 EXPOSÉS 1012-1026

SOCIÉTÉ MATHÉMATIQUE DE FRANCE Publié avec le concours du CENTRE NATIONAL DE LA RECHERCHE SCIENTIFIQUE

Astérisque est un périodique de la Société Mathématique de France. Numéro 339, octobre 2011

Comité de rédaction Bob Oliver Fabrice Planchon Pierre Schapira Éric Vasserot Wendelin Werner

Ahmed Abbes Viviane Baladi Laurent Berger Gérard Besson Hélène Esnault Damien Gaboriau Yves André (dir.) Diffusion Maison de la SMF Case 916 - Luminy 13288 Marseille Cedex 9 France [email protected]

Hindustan Book Agency O-131, The Shopping Mall Arjun Marg, DLF Phase 1 Gurgaon 122002, Haryana Inde

AMS P.O. Box 6248 Providence RI 02940 USA www.ams.org

Tarifs Abonnement

Vente au numéro : 70 e ($ 105) Europe : 454 e, hors Europe : 503 e ($ 754)

Des conditions spéciales sont accordées aux membres de la SMF. Secrétariat : Nathalie Christiaën Astérisque Société Mathématique de France Institut Henri Poincaré, 11, rue Pierre et Marie Curie 75231 Paris Cedex 05, France Tél : (33) 01 44 27 67 99 • Fax : (33) 01 40 46 90 96 [email protected]



http://smf.emath.fr/

© Société Mathématique de France 2011 Tous droits réservés (article L 122–4 du Code de la propriété intellectuelle). Toute représentation ou reproduction intégrale ou partielle faite sans le consentement de l’éditeur est illicite. Cette représentation ou reproduction par quelque procédé que ce soit constituerait une contrefaçon sanctionnée par les articles L 335–2 et suivants du CPI.

ISSN 0303-1179 ISBN 978-2-85629-326-3 Directeur de la publication : Bernard HELFFER

339

ASTÉRISQUE 2011

SÉMINAIRE BOURBAKI VOLUME 2009/2010 EXPOSÉS 1012-1026

SOCIÉTÉ MATHÉMATIQUE DE FRANCE Publié avec le concours du CENTRE NATIONAL DE LA RECHERCHE SCIENTIFIQUE

Association des collaborateurs de Nicolas Bourbaki. École normale supérieure, 45, rue d’Ulm, F-75230 Paris Cedex 05. URL : http://www.bourbaki.ens.fr

Mots-clefs et classification mathématique par sujet (2000) Exposé no 1012. — Groupes de Chow, cycles algébriques, zéro-cycles, corps p-adiques — 11G25, 14C25, 14G20, 14C35. Exposé no 1013. — Algèbres de Hecke p-adiques, familles ordinaires, courbe de Hecke, représentations galoisiennes p-adiques — 11F33, 11F80. Exposé no 1014. — Algèbre amassée, théorie de Lie, base canonique, positivité totale, représentation de carquois, système dynamique discret — 16S99, 05E15, 22E46, 16G20, 18E30. Exposé no 1015. — Trou noir, stabilité, Kerr, Schwarzschild, linéaire, méthode du champ de vecteurs — 35J10. Exposé no 1016. — Équations de Navier-Stokes, ergodicité, turbulence — 37A25, 37A60, 37N10, 37L55, 76F55, 76F20, 60H15, 60H07, 35R60, 60H15, 60H30, 60H07, 76F20, 76B03, 35J60. Exposé no 1017. — Correspondance de Langlands, représentations galoisiennes p-adiques, théorie de Hodge p-adique, (ϕ, Γ)-modules — 11Fxx, 11Sxx, 22Exx. Exposé no 1018. — Métrique extrémale, variété torique, K-stabilité — 53C55, 32Q26. Exposé no 1019. — Matrices aléatoires — 15B52. Exposé no 1020. — Groupes de Lie compacts, espaces classifiants, espaces de lacets, p-complétion, groupes de pseudo-réflexions — 55R35, 55P35, 20F55. Exposé no 1021. — Groupe algébrique linéaire, groupe (pseudo-) réductif, restriction des scalaires, conjugaison, structure, classification — 20Gxx, 10G07, 20G15, 14L15, 20G30, 20G35. Exposé no 1022. — Groupe fondamental, variété kählérienne, groupe résoluble, invariant de BieriNeumann-Strebel — 14F35, 20F65, 32J27. Exposé no 1023. — Groupe d’automorphismes, groupe libre, groupe de surface, groupe spécial linéaire, action de groupe sur les arbres, espace de Teichmüller, outre-espace de Culler-Vogtmann, géométrie asymptotique des groupes — 20E08, 20E36, 20E05, 20F69, 20G20. Exposé no 1024. — Applications harmoniques, lois de conservation, régularité, suites de Palais-Smale, systèmes antisymétriques, surfaces de Willmore — 53C42, 35J60, 35J40, 35D10, 35J60, 58E20. Exposé no 1025. — Conjecture des modèles minimaux, seuil log-canonique, condition de chaîne ascendante, approximation m-adique, théorème de connexité de Shokurov — 14B05, 14E30, 14E15. Exposé no 1026. — Groupes profinis, sous-groupes d’indice fini, sous-groupes verbaux, valeurs des mots — 20E18, 20F12, 20F10, 20D99.

SÉMINAIRE BOURBAKI VOLUME 2009/2010 EXPOSÉS 1012–1026

Résumé. — Comme les précédents volumes de ce séminaire, qui compte maintenant plus de mille exposés, celui-ci contient quinze exposés de synthèse sur des sujets d’actualité : cinq exposés concernant les groupes dans différents contextes, trois de physique mathématique, deux liés au programme de Langlands, deux de géométrie algébrique, un de géométrie différentielle, un sur les algèbres amassées et un sur les matrices aléatoires. Abstract (Séminaire Bourbaki, volume 2009/2010, exposés 1012–1026) As in the preceding volumes of this seminar, which now counts more than one thousand talks, one finds here fifteen survey lectures on topics of current interest : five lectures around group theory, three about mathematical physics, two related to Langlands’ program, two on algebraic geometry, one about differential geometry, one on clusters algebras, and one about random matrices.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2010

TABLE DES MATIÈRES

Résumés des exposés . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

v

vii

NOVEMBRE 2009 1012 1013 1014 1015 1016

J.-L. COLLIOT-THÉLÈNE — Groupe de Chow des zéro-cycles sur les variétés p-adiques (d’après S. Saito, K. Sato et al.) . . . . . . . . .

1

M. EMERTON — p-adic families of modular forms (after Hida, Coleman, and Mazur) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

31

B. KELLER — Algèbres amassées et applications (d’après FominZelevinsky, ...) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

63

S. KLAINERMAN — Linear stability of black holes (d’après M. Dafermos et I. Rodnianski) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

91

A. KUPIAINEN — Ergodicity of two dimensional turbulence (after Hairer and Mattingly) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

137

MARS 2010 1017 1018 1019 1020 1021

L. BERGER — La correspondance de Langlands locale p-adique pour GL2 (Qp ) (d’après C. Breuil et P. Colmez) . . . . . . . . . . . . . . . . . . . . .

157

O. BIQUARD — Métriques kählériennes extrémales sur les surfaces toriques (d’après S. Donaldson) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

181

A. GUIONNET — Grandes matrices aléatoires et théorèmes d’universalité (d’après Erdős, Schlein, Tao, Vu et Yau) . . . . . . . .

201

B. OLIVER — La classification des groupes p-compacts (d’après Andersen, Grodal, Møller, et Viruel) . . . . . . . . . . . . . . . . . . . . . . . . . . .

237

B. RÉMY — Groupes algébriques pseudo-réductifs et applications (d’après J. Tits et B. Conrad-O. Gabber-G. Prasad) . . . . . . . . . . .

257

JUIN 2010 1022 1023 1024

M. BURGER — Fundamental groups of Kähler manifolds and geometric group theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

303

F. PAULIN — Sur les automorphismes de groupes libres et de groupes de surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

321

S. SERFATY — Lois de conservation et régularité par compensation pour les systèmes antisymétriques et les surfaces de Willmore (d’après Tristan Rivière) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

357

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2010

vi

1025 1026

TABLE DES MATIÈRES

B. TOTARO — The ACC conjecture for log canonical thresholds (after de Fernex, Ein, Mustaţă, Kollár) . . . . . . . . . . . . . . . . . . . . . . . .

371

J. S. WILSON — Finite index subgroups and verbal subgroups in profinite groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

387

ASTÉRISQUE 339

RÉSUMÉS DES EXPOSÉS

vii

J.-L. COLLIOT-THÉLÈNE — Groupe de Chow des zéro-cycles sur les variétés p-adiques (d’après S. Saito, K. Sato et al.) À toute variété algébrique projective et lisse sur un corps k, on associe le groupe de Chow CH0 (X), quotient du groupe des cycles de dimension zéro par l’équivalence rationnelle. Pour X de dimension 1, ce groupe est contrôlé par les points rationnels d’une variété abélienne ; sur un corps p-adique k, cela mène à des théorèmes classiques de dualité (Tate, Lichtenbaum) entre CH0 (X) et le groupe de Brauer de la courbe X. Ces théorèmes sont en défaut en dimension supérieure. Dans l’article qui fait l’objet de ce séminaire, S. Saito et K. Sato montrent que la situation est plus contrôlable si l’on s’intéresse à un groupe qui couvre le groupe CH0 (X), à savoir le groupe des cycles de dimension 1 modulo l’équivalence rationnelle sur un modèle régulier convenable de X au-dessus de l’anneau des entiers du corps p-adique k. Pour ce faire, ils ont recours à diverses techniques développées par U. Jannsen et S. Saito. M. EMERTON — p-adic families of modular forms (after Hida, Coleman, and Mazur) We describe the theory of p-adic families of modular Hecke eigenforms, as developed by Hida, Coleman and Mazur, and others. We also describe the relationships (both known and conjectured) between this theory and the theory of two-dimensional p-adic representations of the absolute Galois group of Q. B. KELLER — Algèbres amassées et applications (d’après Fomin-Zelevinsky, ...) Sergey Fomin et Andrei Zelevinsky ont inventé les algèbres amassées (cluster algebras) au début des années 2000 dans le but de fournir un cadre algébrique à l’étude des bases canoniques dans les groupes quantiques et de la positivité totale dans les groupes algébriques. Il s’est avéré rapidement que la combinatoire des algèbres amassées intervenait également dans de nombreux autres sujets et notamment en théorie des représentations des carquois et des algèbres de dimension finie. Dans cet exposé, nous donnons une introduction concise aux algèbres amassées et esquissons deux applications significatives portant sur l’étude de certains systèmes dynamiques discrets et la construction de bases duales semi-canoniques. S. KLAINERMAN — Linear stability of black holes (d’après M. Dafermos et I. Rodnianski) Nonlinear stability of black holes is one of the central open problem in General Relativity today. Heuristic, physics type, arguments have been advanced to establish the first necessary step, i.e. linear stability; yet none of them were either convincing or sufficiently robust to apply to the nonlinear setting. This situation has changed dramatically in the last few years through new geometric methods introduced by a number of authors. In my talk I will try to present the main ideas behind these results and focus, in particular, on the remarkable new results of M. Dafermos and I. Rodnianski on the boundedness and decay of solutions to the wave equation in a Kerr background. A. KUPIAINEN — Ergodicity of two dimensional turbulence (after Hairer and Mattingly) The phenomenon of turbulence in fluid flow can be posed in mathematical terms as a property of a stationary state of the flow generated by Navier-Stokes equations subjected to a spatially smooth random source term. A basic question concerns the existence and uniqueness of such a stationary state. We discuss this question in the context of two dimensional flows and in particular we describe the work of Hairer and Mattingly who prove ergodicity of the flow in a very degenerate situation.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2010

viii

RÉSUMÉS DES EXPOSÉS

L. BERGER — La correspondance de Langlands locale p-adique pour GL2 (Qp ) (d’après C. Breuil et P. Colmez) La correspondance de Langlands locale p-adique pour GL2 (Qp ) est une bijection entre certaines représentations de dimension 2 de Gal(Qp /Qp ) et certaines représentations de GL2 (Qp ). Cette bijection peut en fait être construite en utilisant la théorie des (ϕ, Γ)-modules et des résultats d’analyse p-adique. On déduit alors des propriétés de cette construction quelques applications intéressantes en arithmétique. O. BIQUARD — Métriques kählériennes extrémales sur les surfaces toriques (d’après S. Donaldson) Un des grands problèmes de la géométrie différentielle complexe est l’existence de métriques kählériennes à courbure scalaire constante sur les variétés complexes. Une conjecture de Yau-Tian-Donaldson relie cette existence à une forme de stabilité algébrique de la variété. Donaldson a démontré cette conjecture dans le cas des surfaces toriques. On expliquera cette première confirmation de la conjecture. A. GUIONNET — Grandes matrices aléatoires et théorèmes d’universalité (d’après Erdős, Schlein, Tao, Vu et Yau) L’étude du spectre de matrices aléatoires dont la taille tend vers l’infini est apparue dans de nombreux problèmes de physique et de mathématique depuis les travaux de Wishart et Wigner il y a plus d’un demi-siècle. Dans le cas de matrices hermitiennes gaussiennes, le comportement local des valeurs propres (espacements typiques des valeurs propres au centre du spectre et fluctuations des valeurs propres extrêmes) est bien compris depuis une quinzaine d’années. Nous discuterons de l’extension très récente de ces résultats à des modèles bien plus généraux, en suivant les travaux de Soshnikov, Johansson, Erdős, Schlein, Yau, Tao et Vu. B. OLIVER — La classification des groupes p-compacts (d’après Andersen, Grodal, Møller, et Viruel) Un groupe p-compact (p un nombre premier fixé) est la donnée d’un espace X dont la cohomologie modulo p est finie et d’un « espace classifiant » BX qui est « p-complet ». Le prototype d’un tel objet est le p-complété d’un groupe de Lie compact connexe. Dwyer et Wilkerson ont montré (vers 1990) que les groupes p-compacts connexes possèdent des « tores maximaux » et des « groupes de Weyl », ces derniers étant engendrés par des pseudoréflexions p-adiques. La liste de tous les groupes engendrés par des pseudo-réflexions p-adiques a été faite par Clarke et Ewing dans les années 1970. Ce qu’ont montré les quatre auteurs cités dans le titre est qu’il y a une correspondance bijective entre ces groupes de pseudo-réflexions (à isomorphisme près) et les groupes p-compacts connexes (à homotopie près), au moins pour p impair ; pour p = 2, la correspondance qu’ils établissent est un peu plus subtile. B. RÉMY — Groupes algébriques pseudo-réductifs et applications (d’après J. Tits et B. Conrad-O. Gabber-G. Prasad) Les groupes algébriques réductifs forment une classe naturelle de groupes de matrices. Celle-ci contient les groupes dits classiques, notamment des groupes d’automorphismes de formes bilinéaires. Les groupes réductifs ont été analysés et classés par C. Chevalley sur les corps algébriquement clos, et par A. Borel et J. Tits sur les corps quelconques. Au début des années 90, J. Tits a entamé l’étude des groupes pseudo-réductifs : dans le cas d’un corps de base non parfait, il s’agit d’une généralisation non triviale des groupes réductifs. Cette étude

ASTÉRISQUE 339

RÉSUMÉS DES EXPOSÉS

ix

vient d’être menée à bien par B. Conrad, O. Gabber et G. Prasad. En retour, ce travail et un complément de B. Conrad contiennent et impliquent des résultats généraux de structure et de finitude pour les groupes algébriques quelconques en caractéristique positive. M. BURGER — Fundamental groups of Kähler manifolds and geometric group theory The general topic of the talk concerns the problem of which are the additional restrictions imposed on a finitely presented group if one assumes that it is the fundamental group of a compact Kähler manifold (Kähler group). Since Gromov’s seminal work showing that such a group cannot be a nontrivial free product, methods of geometric group theory have become important in that field. As an illustration, we will explain Delzant’s recent result that a solvable Kähler group contains a finite index nilpotent subgroup. F. PAULIN — Sur les automorphismes de groupes libres et de groupes de surface Les groupes spéciaux linéaires entiers, les groupes modulaires de surfaces et les groupes des automorphismes extérieurs de groupes libres apparaissent dans de nombreux domaines. Leurs analogies, soulignées en particulier par les travaux de K. Vogtmann, font couler beaucoup d’encre. Dans ce rapport, nous nous concentrerons sur les espaces contractiles sur lesquels ces groupes agissent de manière analogue, sur les propriétés communes de leurs sous-groupes et sur les propriétés semblables (ou envisagées semblables) de leur géométrie asymptotique. S. SERFATY — Lois de conservation et régularité par compensation pour les systèmes antisymétriques et les surfaces de Willmore (d’après Tristan Rivière) Dans des travaux récents, Tristan Rivière a découvert que les problèmes variationnels correspondant à des lagrangiens quadratiques invariants conformes en dimension deux, tels que le problème des applications harmoniques et l’équation de courbure moyenne prescrite appartiennent à une classe plus large de systèmes de Schrödinger antisymétriques, que l’on peut écrire sous forme d’une loi de conservation. Ces lois de conservation font apparaître une structure particulière de compensation : équations elliptiques avec non-linéarité de forme jacobienne, ce qui permet de déduire la régularité höldérienne des solutions et la compacité des suites de Palais-Smale. Ceci place dans un cadre unique et étend tous les résultats antérieurs de régularité. Il a également démontré des résultats de même nature pour la fonctionnelle de Willmore. B. TOTARO — The ACC conjecture for log canonical thresholds (after de Fernex, Ein, Mustaţă, Kollár) The minimal model conjecture in algebraic geometry tries to produce a “simplest” projective variety which is birational to any given variety. Shokurov showed that the minimal model conjecture would follow if certain invariants of singularities satisfied the ascending chain condition (ACC). These invariants are rational numbers, with milder singularities corresponding to larger invariants. Thus the ACC conjectures say that singularities cannot be “improved” infinitely many times. Although the general ACC conjectures remain open, de Fernex, Ein and Mustaţă (with contributions by Kollár) gave an elegant proof of ACC for log canonical thresholds of hypersurfaces in smooth varieties of any dimension.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2010

x

RÉSUMÉS DES EXPOSÉS

J. S. WILSON — Finite index subgroups and verbal subgroups in profinite groups In the 1970s, after proving a weaker statement, J-P. Serre commented that he did not know whether all subgroups of finite index in finitely generated profinite groups are open. A result of Nikolov and Segal concerning finite groups proved in 2003 shows that this is indeed the case. This result is the most spectacular achievement in a type of group theory whose importance is being increasingly recognised: the study of the distribution of the values of a fixed group word in finite groups.

ASTÉRISQUE 339

339

ASTÉRISQUE 2011

SÉMINAIRE BOURBAKI VOLUME 2009/2010 EXPOSÉS 1012-1026 (1012) Groupe de Chow des zéro-cycles sur les variétés p-adiques Jean-Louis COLLIOT-THÉLÈNE

SOCIÉTÉ MATHÉMATIQUE DE FRANCE Publié avec le concours du CENTRE NATIONAL DE LA RECHERCHE SCIENTIFIQUE

Séminaire BOURBAKI 62e année, 2009-2010, no 1012, p. 1 à 30

Novembre 2009

GROUPE DE CHOW DES ZÉRO-CYCLES SUR LES VARIÉTÉS p-ADIQUES [d’après S. Saito, K. Sato et al.] par Jean-Louis COLLIOT-THÉLÈNE

INTRODUCTION Soient k un corps et X une k-variété algébrique projective, lisse et géométriquement irréductible (cette dernière hypothèse sera souvent tacitement faite). On note Z0 (X) le groupe des zéro-cycles sur X, c’est-à-dire le groupe abélien libre sur les points fermés de X (points x du schéma X dont le corps résiduel κ(x) est une extension finie de k). On dispose d’une application degré deg : Z0 (X) → Z définie par linéarité à partir de l’application envoyant un point fermé x sur le degré [κ(x) : k]. À tout couple formé d’une courbe fermée intègre C ⊂ X et d’une fonction rationnelle non nulle f ∈ k(C)× , on associe un zéro-cycle, le diviseur de f . Celui-ci est ainsi défini : on considère la normalisation C˜ → C de la courbe C, et le morphisme composé π : C˜ → C → X. On définit alors div(f ) = π∗ (divC˜ (f )). Le groupe de Chow des zéro-cycles sur X est par définition le quotient de Z0 (X) par le sous-groupe engendré par tous les div(f ) pour tous les couples (C, f ). Comme la k-variété X est projective, l’application degré induit un homomorphisme deg : CH0 (X) → Z. On note A0 (X) le noyau de cette application. On dispose donc d’une suite exacte 0 → A0 (X) → CH0 (X) → Z. L’image de la flèche degré est un sous-groupe Z.IX ⊂ Z d’indice fini. La suite est scindée si X possède un zéro-cycle z0 de degré 1, ce qu’on suppose désormais dans cette introduction.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

2

J.-L. COLLIOT-THÉLÈNE

À la k-variété X on associe sa variété de Picard Pic0X/k,red , qui est une variété abélienne. On a une suite exacte 0 → Pic0X/k,red (k) → PicX → NS(X) → 0 où NS(X) est un groupe abélien de type fini. La variété abélienne duale de Pic0X/k,red est la variété d’Albanese AlbX/k de X. À la donnée de z0 est associé un k-morphisme albX : X → AlbX/k induisant un isomorphisme sur les variétés de Picard de ces deux variétés. Ce morphisme induit un homomorphisme de groupes abéliens albX : A0 (X) → AlbX/k (k) qui ne dépend pas du choix de z0 . Lorsque dim(X) = 1, c’est-à-dire lorsque X est une courbe (projective, ' lisse), on a un isomorphisme PicX → CH0 (X) qui induit un isomorphisme ' Pic0X/k,red (k) → A0 (X). La flèche albX : A0 (X) → AlbX/k (k) est un isomorphisme. Les propriétés des groupes de points rationnels de variétés abéliennes donnent alors des théorèmes sur la structure des groupes CH0 (X) et A0 (X). En particulier, pour X/k une courbe de genre g avec ou sans zéro-cycle de degré 1, on a les propriétés suivantes : (1) Si k est un corps de type fini sur le corps premier, le groupe CH0 (X) est un groupe abélien de type fini (Mordell-Weil). (2) Si k est un corps fini, le groupe A0 (X) est fini. (3) Si k est un corps p-adique (ce qui dans cet exposé signifie extension finie du corps p-adique Qp ), le groupe A0 (X) est extension d’un groupe fini par un sous-groupe isomorphe à une somme directe de g exemplaires de l’anneau des entiers de k (Lutz, Mattuck). En conséquence, (3.1) Le groupe A0 (X) est somme directe d’un groupe fini (d’ordre premier à p) et d’un groupe p0 -divisible (c’est-à-dire divisible par tout entier premier à p). (3.2) Pour tout entier n > 0, le quotient CH0 (X)/n est fini. (3.3) Pour presque tout premier l, on a A0 (X)/l = 0. (3.4) Le sous-groupe de torsion de CH0 (X) est fini. On peut en outre détecter les classes dans CH0 (X) au moyen de la cohomologie étale sur X (voir le paragraphe 1 ci-après). Il est naturel de se demander si certaines parmi ces propriétés du groupe CH0 (X) valent encore pour une k-variété projective lisse X de dimension quelconque. Dans la situation (1), même pour k le corps des rationnels, en dehors des cas qui se réduisent formellement au théorème de Mordell-Weil, on n’a aucun résultat non

ASTÉRISQUE 339

(1012)

ZÉRO-CYCLES DES VARIÉTÉS p-ADIQUES

3

trivial sur la finitude de CH0 (X)/n pour n > 1 ou sur la finitude de la dimension du Q-vectoriel CH0 (X) ⊗Z Q. Dans la situation (2), qui porte sur le cas des corps finis, la finitude de A0 (X) est un théorème de K. Kato et S. Saito [27]. On en sait beaucoup plus : voir à ce sujet l’exposé récent de T. Szamuely [42] sur le corps de classes de dimension supérieure. Le présent exposé porte sur le cas des corps p-adiques. Depuis les années 1980, une méthode de K-théorie algébrique inventée par S. Bloch et reposant sur un théorème de Merkur0 ev et Suslin a permis d’obtenir un certain nombre de résultats, en particulier pour les surfaces. On évoquera ces résultats au paragraphe 2. En 2006, S. Saito et K. Sato [35] réalisèrent que pour obtenir des énoncés généraux il vaut mieux considérer non le groupe de Chow des zéro-cycles sur une variété projective et lisse X sur un corps p-adique k, mais le groupe de Chow des 1-cycles sur un modèle régulier et projectif de X au-dessus de l’anneau des entiers de k (lorsqu’un tel modèle existe). On décrira en détail leur travail au paragraphe 3. Le théorème principal est le théorème 3.17. En voici deux applications (Théorèmes 3.21 et 3.25). Théorème 0.1. — Soit X une variété projective, lisse, géométriquement connexe sur un corps p-adique k. Si X a bonne réduction Y sur le corps résiduel fini F , alors la flèche de spécialisation A0 (X) → A0 (Y ), qui est une surjection sur le groupe fini A0 (Y ), a un noyau p0 -divisible, c’est-à-dire divisible par tout entier premier à p. Théorème 0.2. — Soit X une variété projective, lisse, géométriquement connexe sur un corps p-adique k. (i) Pour presque tout premier l, le quotient A0 (X)/l est nul. (ii) Pour tout premier l 6= p, le quotient A0 (X)/l est fini

(1)

.

Les résultats de Saito et Sato [35] furent ensuite combinés par Asakura et Saito [2] à des techniques de théorie de Hodge pour établir l’existence de surfaces X de degré au moins 5 dans P3k dont les sous-groupes de torsion l-primaire (l 6= p) sont infinis (voir le paragraphe 4 ci-après). Je remercie Tamás Szamuely pour de nombreuses discussions sur le théorème de Saito et Sato et pour ses commentaires critiques sur une première version du présent texte. (1)

Dans [35], l’énoncé (ii) est établi pour les k-variétés qui admettent un modèle quasi-semistable sur l’anneau des entiers ; comme on verra, le cas général s’y ramène grâce à un théorème récent de Gabber.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

4

J.-L. COLLIOT-THÉLÈNE

Notations Soit A un groupe abélien. Pour n > 0 un entier, on note A[n] ⊂ A le sous-groupe formé des éléments annulés par n. Pour l un nombre premier, on note A{l} ⊂ A le sous-groupe de torsion l-primaire.

1. COURBES SUR LES CORPS p-ADIQUES : RÉSULTATS CLASSIQUES Comme mentionné dans l’introduction, pour une courbe projective et lisse X sur un corps p-adique, on peut détecter les classes dans CH0 (X) ' Pic(X) au moyen de la cohomologie étale. Expliquons plus précisément ce que nous entendons par là. Théorème 1.1 (Tate (1958) [43]). — Soient k un corps p-adique, A une variété abélienne sur k et Aˆ la variété abélienne duale. Il y a une dualité parfaite ˆ → Brk = Q/Z A(k) × H 1 (k, A) entre le groupe abélien compact A(k) des points rationnels de A et le groupe discret défini par le premier groupe de cohomologie galoisienne de k à valeurs dans le groupe ˆ des points de A. En s’appuyant sur ce théorème, on montre : Théorème 1.2 (Roquette (1966), Lichtenbaum (1969) [29]) Soient k un corps p-adique et X une k-courbe projective, lisse, géométriquement connexe. a) Il y a un accouplement naturel PicX × BrX → Brk = Q/Z, et cet accouplement est non dégénéré des deux côtés. b) Le noyau de la flèche Q/Z = Brk → BrX induite par le morphisme structural est Z/I, où I est l’index de X, c’est-à-dire le pgcd des degrés, sur k, des points fermés sur X. Ainsi, pour X une courbe projective, lisse, géométriquement connexe sur un corps k p-adique, possédant un k-point, le groupe A0 (X) est isomorphe au groupe de Lie p-adique JX (k), et l’accouplement CH0 (X) × BrX → Brk = Q/Z est non dégénéré à gauche et à droite.

ASTÉRISQUE 339

(1012)

ZÉRO-CYCLES DES VARIÉTÉS p-ADIQUES

5

Théorème 1.3 (Artin (1966) [17], §3). — Soit R un anneau de valuation discrète hensélien excellent de corps résiduel F . Soit X un R-schéma fidèlement plat, projectif, intègre, régulier, de dimension relative 1. Soit Xs /F sa fibre spéciale. (a) La flèche de restriction BrX → BrXs est un isomorphisme (de groupes de torsion). (b) Si F est un corps sépablement clos ou un corps fini, BrX ' BrXs = 0. (c) Sous les mêmes hypothèses qu’en (b), pour tout entier n > 0, l’injection naturelle PicX/n ,→ Hf2ppf (X, µn ) est un isomorphisme. L’énoncé (c) provient de (b) et de la suite de Kummer x7→xn

1 → µn → Gm −→ Gm → 1 sur X, considérée comme suite de faisceaux pour la topologie étale sur X si n est inversible, et comme suite de faisceaux pour la topologie fppf en général. La partie première à p = car(F ) des énoncés est plus facile à établir, elle ne nécessite pas l’hypothèse d’excellence. Le théorème principal de Saito et Sato (théorème 3.17 ci-dessous) généralise l’énoncé (c) du théorème 1.3 pour n premier à p.

2. SURFACES SUR LES CORPS p-ADIQUES : QUELQUES APPLICATIONS DE LA MÉTHODE DE BLOCH Soit X un schéma noethérien de dimension finie. Pour tout entier i ≥ 0, le groupe Zi (X) des cycles de dimension i est le groupe abélien libre sur les points (schématiques) de X de dimension i (ou si l’on préfère les sous-schémas fermés intègres de dimension i) Zi (X) = ⊕x∈Xi Z. Notons κ(x) le corps résiduel en un point x. On sait définir une application « diviseur » div : ⊕x∈Xi+1 κ(x)× → ⊕x∈Xi Z qui généralise la notion de diviseur d’une fonction rationnelle (voir [16]). Par définition, le groupe de Chow CHi (X) est le conoyau de cette application. Supposons X intègre et équidimensionnel de dimension d. On note alors CH r (X) = CHd−r (X). Pour i = d, CH0 (X) = Z. Pour i = d − 1, CHd−1 (X) = CH 1 (X). On a une flèche naturelle PicX → CH 1 (X)

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

6

J.-L. COLLIOT-THÉLÈNE

qui est un isomorphisme si X est régulier. On note K i le faisceau pour la topologie de Zariski sur X associé au préfaisceau qui à un ouvert affine U associe le groupe de K-théorie Ki (U ) défini par Quillen. En combinant les théorèmes sur la conjecture de Gersten, tant en K-théorie (Quillen) qu’en cohomologie étale (Bloch et Ogus [6], 1974) et le théorème de Merkur0 ev et Suslin ([30], 1982) sur le symbole de restes normique, Spencer Bloch a établi le résultat suivant ([3, 4, 5], voir aussi [7]). Théorème 2.1 (Bloch). — Soient k un corps, X une k-variété lisse intègre, k(X) son corps des fonctions rationnelles, et n un entier non nul dans k. On a une suite exacte naturelle de groupes abéliens 1 3 2 0 → HZar (X, K 2 )/n → N Hét (X, µ⊗2 n ) → CH (X)[n] → 0, 3 où N Hét (X, µ⊗2 n ) est le noyau de la flèche de restriction de groupes de cohomologie 3 3 ⊗2 étale Hét (X, µ⊗2 n ) → Hét (k(X), µn ). 1 (X, K 2 ) a été identifié avec d’autres Dans la littérature récente, le groupe HZar groupes : le groupe de Chow supérieur CH 2 (X, 1) de Bloch d’une part, le groupe d’hypercohomologie H3Zar (X, Z(2)) du complexe motivique Z(2) d’autre part.

Il existe des analogues de cette suite exacte pour les schémas lisses au-dessus d’un anneau de valuation discrète. Cette suite exacte a eu de nombreuses applications, qu’on ne saurait décrire ici de façon exhaustive. On l’a utilisée, conjointement avec le théorème de Deligne établissant les conjectures de Weil, pour donner des démonstrations alternatives du théorème de Kato et Saito sur le corps de classes non ramifié pour les variétés projectives et lisses sur un corps fini mentionné dans l’introduction ([13, 42]). Sur les corps p-adiques et sur les corps de nombres, on l’a utilisée pour obtenir des résultats de finitude pour la torsion du groupe de Chow de codimension 2, et aussi du groupe de Chow des zéro-cycles, pour certaines classes de variétés. Pour des résultats sur les corps de nombres, je renvoie le lecteur à [11, 34, 38, 31, 36] et aux rapports [7, 8]. Théorème 2.2 ([13]). — Soit k un corps p-adique. Soit X une k-variété lisse. (i) Pour tout entier n > 0, le groupe CH 2 (X)[n] est un groupe fini. (ii) Pour tout l premier, le groupe de torsion l-primaire CH 2 (X){l} est un groupe de cotype fini (somme d’un groupe fini l-primaire et d’un groupe (Ql /Zl )N ).

ASTÉRISQUE 339

(1012)

ZÉRO-CYCLES DES VARIÉTÉS p-ADIQUES

7

Démonstration. — D’après le théorème 2.1, le groupe CH 2 (X)[n] est un sous3 quotient du groupe Hét (X, µ⊗2 n ). La finitude de ce groupe pour un corps local est 3 bien connue. Elle implique que le groupe Hét (X, Ql /Zl (2)) est un groupe de cotype fini, et donc aussi tout sous-quotient. Théorème 2.3. — Soient k un corps p-adique et X une k-variété projective, lisse, géométriquement intègre. (i) Pour tout l premier, l 6= p, et tout entier n > 0, l’application naturelle A0 (X){l}/ln → A0 (X)/ln est un isomorphisme. (ii) Supposons que X/k a bonne réduction Y /F. Alors pour tout premier l 6= p, l’application de réduction induit une surjection A0 (X){l} → A0 (Y ){l}. Démonstration. — (i) Que l’application A0 (X){l}/ln → A0 (X)/ln soit une injection est clair. Pour établir que c’est une surjection, on se ramène par le théorème de Bertini au cas où X est une courbe projective lisse géométriquement intègre, et l’assertion résulte alors de la structure du groupe des points d’une variété abélienne sur un corps p-adique. (ii) On dispose d’une application de spécialisation A0 (X) → A0 (Y ) qui est surjective (lemme de Hensel). Soit m > 0. D’après (i) l’application A0 (X){l}/lm → A0 (X)/lm est surjective. L’application composée A0 (X){l}/lm → A0 (X)/lm → A0 (Y )/lm est donc surjective. Le groupe A0 (Y ) est fini (théorème de Kato et Saito [27]). Prenant alors m tel que lm annule la partie l-primaire de A0 (Y ), on obtient l’énoncé. Théorème 2.4. — Soient k un corps p-adique et X une k-surface projective, lisse et géométriquement intègre. (i) Pour tout entier positif n, le groupe A0 (X)[n] est un groupe fini. (ii) Pour tout l premier, le groupe de torsion l-primaire A0 (X){l} est un groupe de cotype fini (somme d’un groupe fini l-primaire et d’un groupe (Ql /Zl )N ). (iii) (Saito et Sujatha) Pour l premier, l 6= p, le groupe A0 (X) est somme de son sous-groupe l-divisible maximal et d’un groupe fini l-primaire. (iv) (Saito et Sujatha) Pour n > 0 premier à p, le groupe A0 (X)/n est fini. Démonstration (voir [8]). — Pour X une surface projective lisse géométriquement connexe, on a l’égalité CH0 (X) = CH 2 (X). Les énoncés (i) et (ii) sont des cas particuliers du théorème 2.2. D’après (ii) on peut écrire A0 (X){l} = (Ql /Zl )N ⊕ Fl avec Fl un groupe abélien fini annulé par une puissance lt de l, et N ≥ 0. En utilisant la proposition 2.3(i) on voit alors que l’application composée Fl → A0 (X) → A0 (X)/lt est un isomorphisme. On en déduit que dans la suite exacte 0 → Dl → A0 (X) → A0 (X)/lt → 0

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

8

J.-L. COLLIOT-THÉLÈNE

définissant Dl , la projection A0 (X) → A0 (X)/lt est scindée, d’où A0 (X) ' Dl ⊕ Fl , avec Dl /l = 0. Le groupe Dl est donc le sous-groupe l-divisible maximal de A0 (X). Ceci établit (iii), et l’énoncé (iv) suit. Remarque 2.5. — On verra au paragraphe 3 que pour presque tout premier l le groupe fini Fl est nul, en d’autres termes le groupe A0 (X) est l-divisible. La méthode ci-dessus ne permet pas d’obtenir ce résultat. Le théorème suivant, détaillé dans [9], regroupe des travaux des années 1985 à 1991, dus à Raskind et au rédacteur [10, 11], à Salberger [38] et à S. Saito [34]. Théorème 2.6. — Soient k un corps p-adique et X une k-surface projective et lisse, géométriquement intègre. Supposons H 2 (X, OX ) = 0. Alors (i) Le groupe A0 (X)tors est fini. (ii) Si la flèche naturelle A0 (X) → AlbX (k) (sur une clôture algébrique k de k) est un isomorphisme (ce qui résulterait de H 2 (X, OX ) = 0 suivant une conjecture de S. Bloch, connue pour les surfaces qui ne sont pas de type général), alors le groupe A0 (X) est extension d’un sous-groupe ouvert de AlbX (k) par un groupe fini. En particulier, le groupe A0 (X) est somme directe d’un groupe fini d’ordre premier à p et d’un groupe p0 -divisible. (iii) Si la variété d’Albanese de X a bonne réduction, alors l’accouplement A0 (X)tors × Br(X) → Q/Z est non dégénéré à gauche. (iv) Si la flèche naturelle A0 (X) → AlbX (k) est un isomorphisme et si la variété d’Albanese de X a bonne réduction, alors l’accouplement A0 (X) × Br(X) → Q/Z est non dégénéré à gauche. L’énoncé (iii) peut être établi sous des hypothèses un peu plus larges ([38], [34], [39], [36]), mais on ne peut totalement ignorer les hypothèses dans (iii) et (iv). Parimala et Suresh [31] construisent une surface X fibrée en coniques lisses au-dessus d’une courbe C (la surface satisfait donc H 2 (X, OX ) = 0), la conique ayant mauvaise réduction sur un corps p-adique k de caractéristique résiduelle impaire, surface pour laquelle : (a) le noyau à gauche de l’accouplement A0 (X){2} × BrX → Q/Z n’est pas nul ; 4 (b) l’application cycle CH 2 (X)/2 → Hét (X, Z/2) n’est pas injective ; (c) le noyau de l’application A0 (X) → Hom(BrX, Q/Z) n’est pas le sous-groupe divisible maximal de A0 (X).

ASTÉRISQUE 339

(1012)

ZÉRO-CYCLES DES VARIÉTÉS p-ADIQUES

9

La plupart des théorèmes précédents furent obtenus en étudiant l’action du groupe de Galois absolu de k sur divers groupes ( K -cohomologie, cohomologie étale) attachés aux variétés (après passage à une clôture algébrique du corps k). Une autre méthode consiste à considérer des modèles des k-variétés au-dessus d’un ouvert de l’anneau des entiers de k. Limitons-nous ici au cas de bonne réduction. Soit R l’anneau des entiers d’un corps p-adique k, de corps résiduel F. Soit X un R-schéma intègre, projectif et lisse, de fibre générique X/k géométriquement intègre, et soit Y /F la fibre spéciale. Pour un tel X , on a la suite exacte de localisation H 1 (X, K 2 ) → Pic(Y ) → CH 2 ( X ) → CH 2 (X) → 0. Notons δ : H 1 (X, K 2 ) → Pic(Y ). Le groupe Pic(Y ) est un groupe de type fini (théorème de Néron-Severi et finitude du groupe des points rationnels d’une variété abélienne sur un corps fini). L’énoncé suivant, détaillé dans [9], regroupe des résultats de Raskind [33], Raskind et l’auteur [11], Spieß [41]. Théorème 2.7. — Avec les notations ci-dessus, supposons dim(X) = 2, et supposons (H) L’application H 1 (X, K 2 ) ⊗ Q → Pic(Y ) ⊗ Q est surjective. Alors (i) L’application de spécialisation CH0 (X) → CH0 (Y ) induit un isomorphisme sur les sous-groupes de torsion première à p. (ii) Le groupe A0 (X) est la somme directe d’un groupe fini d’ordre premier à p et d’un groupe uniquement divisible par tout entier n premier à p. Pour tout l premier distinct de p, le groupe Dl ci-dessus est uniquement l-divisible. (iii) Pour n > 0 premier à p, l’application cycle 4 CH 2 (X)/n → Hét (X, µ⊗2 n )

est injective. (iv) L’accouplement A0 (X) × Br(X) → Q/Z a son noyau à gauche formé d’éléments divisibles par tout entier n > 0 premier à p. Voici des cas où l’hypothèse (H) a été établie. a) Le cas où H 2 (Y, OY ) = 0 (et donc aussi H 2 (X, OX ) = 0). C’est le cas le plus simple. Dans ce cas le conoyau de la flèche composée Pic(X) ⊗ k × → H 1 (X, K 2 ) → Pic(Y ) est nul, car la flèche de restriction Pic( X ) → Pic(Y ) est surjective. Ce cas fut considéré par Raskind [33], Coombes, CT-Raskind ([11]). b) Le cas où le rang du groupe de Néron-Severi géométrique ne grandit pas par spécialisation (Raskind [33]).

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

10

J.-L. COLLIOT-THÉLÈNE

c) Le cas où X est le produit fibré de deux courbes elliptiques avec bonne réduction (Spieß [41]). Dans le cas c) il faut, pour établir (H), trouver des éléments « indécomposables » dans H 1 (X, K 2 ), i.e. d’autres éléments que ceux provenant de Pic(X) ⊗ k × . Spieß utilise certaines correspondances entre courbes elliptiques provenant de travaux de Frey et Kani. Dans les années 1990 à 2000, il y eut dans cette direction (recherche d’éléments indécomposables) une série de travaux (Flach, Mildenhall, S. Saito, Langer, Raskind, Otsubo) reposant souvent sur une arithmétique très fine des variétés considérées (par exemple des produits de courbes modulaires). Je renvoie ici le lecteur à l’article récent [36] de S. Saito et K. Sato, tant pour la liste de références que pour le lien entre la finitude du groupe CH 2 (X)tors et le comportement d’applications « régulateurs » de source le groupe H 1 (X, K 2 ), applications déjà considérées par Salberger [38]. Le récent théorème de Asakura et Saito (Théorème 4.1 ci-dessous) montre que l’hypothèse (H) ne vaut pas toujours : elle est en défaut pour des surfaces lisses dans P3k de degré au moins 5, suffisamment génériques. Asakura et Sato ont posé la question de sa validité lorsque la surface sur le corps p-adique provient d’une surface définie sur un corps de nombres.

3. LE THÉORÈME DE S. SAITO ET K. SATO [35] Dans tout ce paragraphe, on adopte les notations suivantes. On note R un anneau de valuation discrète, F son corps résiduel et k son corps des fractions. On note B = Spec R, s = Spec F , η = Spec k. qp On note Schqp B la catégorie des schémas quasi-projectifs sur B. Pour X ∈ SchB , on note δ(X) ∈ N la dimension de Krull d’une compactification de X au-dessus de B. Pour X ∈ Ob(Schqp B ), on note Xs /F sa fibre spéciale et Xη /k sa fibre générique. Si X est intègre et Xη est vide, donc X = Xs , alors δ(X) est égal au degré de transcendance sur F du corps des fonctions F (Xs ). Si X est intègre et Xη est non vide, alors δ(X) − 1 est égal au degré de transcendance du corps des fonctions k(Xη ) sur k. On note C la sous-catégorie pleine de Schqp B dont les objets satisfont Xs 6= ∅. Pour X irréductible dans C , on a δ(X) = dim(X). Un objet X ∈ C est appelé quasi-semistable s’il satisfait les conditions : (QS1) X est régulier, équidimensionnel, plat et de type fini sur B. (QS2) Le diviseur réduit Xs,red sur X est à croisements normaux stricts. On note Q S ⊂ C , resp. Q S P ⊂ C , la sous-catégorie pleine dont les objets sont les objets quasi-semistables, resp. les objets quasi-semistables et projectifs sur B.

ASTÉRISQUE 339

(1012)

ZÉRO-CYCLES DES VARIÉTÉS p-ADIQUES

11

Une Q S -paire est un couple (X, Y ) de schémas dans Ob( Q S ) pour lequel Y est un diviseur sur X et le diviseur Xs,red ∪ Y sur X est à croisements normaux stricts. Posant U := X \ Y , on note indifféremment (X, Y ) = (X, Y ; U ). Une Q S P -paire est une Q S -paire (X, Y ) pour laquelle X et Y sont projectifs sur B. Une Q S P -paire ample est une Q S P -paire (X, Y ; U ) pour laquelle U est affine. 3.1. Groupes de Chow, homologie étale, suite spectrale de niveau, complexe de Bloch-Ogus et Kato, application cycle Pour les variétés algébriques sur un corps, les résultats exposés au paragraphe 2 utilisent diverses propriétés de la cohomologie et de l’homologie étale, la dualité de Poincaré, les filtrations par le niveau et par le coniveau et les suites spectrales associées. Ces théories sont développées dans [17] III, §10.1, [28], [6]. Il a fallu étendre ces théories aux schémas de type fini au-dessus d’un anneau de valuation discrète. Sauf mention du contraire, la cohomologie employée est la cohomologie étale. Pour X dans C , et q ≥ 0 entier, on note Xq l’ensemble des points x ∈ X dont l’adhérence {x} ⊂ X satisfait δ({x}) = q. Pour q ≥ 0, et X ∈ C , on définit le groupe de Chow de dimension q par la formule usuelle : CHq (X) = Coker [div : ⊕x∈Xq+1 κ(x)∗ → ⊕x∈Xq Z]. On fixe un premier l 6= car F . On note µln le faisceau pour la topologie étale associé au B-schéma en groupes des racines ln -ièmes de l’unité. Pour m ∈ N, on a le faisceau n n n Z/ln (m) = µ⊗m ln . Pour −m ∈ N on note Z/l (m) le faisceau Hom(Z/l (−m), Z/l ). Pour m ∈ Z on note Ql /Zl (m) = lim Z/l(m). Lorsque l’on voudra considérer simul−→ n

tanément le cas Z/ln et le cas Ql /Zl , on utilisera la notation Λ. Pour X ∈ Ob(Schqp B ) de morphisme structural f : X → B et pour q ∈ Z et m ∈ Z, suivant Grothendieck, Artin, Verdier, Deligne, on définit Hq (X, Λ(m)) := H2−q (X, Rf ! Λ(m)) où Rf ! est le foncteur image inverse extraordinaire ([1, XVIII, Thm. 3.1.4]). Ceci définit une théorie homologique sur la catégorie Schqp B ayant toutes les propriétés voulues : fonctorialité covariante par morphisme propre, fonctorialité contravariante par morphisme étale, existence pour toute immersion fermée Y ,→ X de complémentaire U ,→ X d’une longue suite exacte · · · → Hq (Y, Λ) → Hq (X, Λ) → Hq (U, Λ) → Hq−1 (Y, Λ) → · · · , fonctorialité de cette suite exacte. Le lecteur se reportera à [28, 20, 23]. Cette théorie homologique à la Borel-Moore se relie à la cohomologie étale grâce à un théorème de « dualité de Poincaré » :

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

12

J.-L. COLLIOT-THÉLÈNE

Proposition 3.1. — Soit X ∈ Ob(Schqp B ) un schéma régulier intègre. Posons d = δ(X) − 1. Pour tout sous-schéma fermé Y ⊂ X, et tout q ∈ Z, il existe un isomorphisme canonique ∼ =

→ Hq (Y, Λ). HY2d−q+2 (X, Λ(d)) − (Par convention H r = 0 pour r < 0.) Cet isomorphisme satisfait une série de propriétés fonctorielles (cf. [35]). Pour X ∈ Ob(Schqp B ), x un point de X, et q ∈ Z on note Hq (x, Λ) := lim Hq (U, Λ), −→ U ⊂{x}

où {x} est l’adhérence de x dans X et U parcourt les ouverts non vides de {x}. La longue suite exacte de localisation donnée ci-dessus et la filtration par le niveau donnent naissance à une suite spectrale de type homologique 1 Ea,b (X, Λ) = ⊕x∈Xa Ha+b (x, Λ) =⇒ Ha+b (X, Λ)

dont les différentielles de niveau r sont de degré (−r, r − 1). Pour n ∈ Z et m ∈ Z, notons H n (x, Λ(m)) le groupe de cohomologie galoisienne H (κ(x), Λ(m)), groupe qui par définition est nul pour n < 0. Par passage à la limite dans les isomorphismes dans la proposition 3.1 (pour Y = X = U ) on obtient la n

Proposition 3.2 (Suite spectrale de niveau). — Pour X ∈ Ob(Schqp B ), il y a une suite spectrale homologique 1 Ea,b (X, Λ) = ⊕x∈Xa H a−b (x, Λ(a − 1)) =⇒ Ha+b (X, Λ).

Cette suite exacte est fonctorielle covariante par rapport aux morphismes propres et contravariante par rapport aux morphismes étales. La ligne b = 0 de la suite spectrale est un complexe M M M 0← H 0 (x, Λ(−1)) ← H 1 (x, Λ) ← · · · H a (x, Λ(a − 1)) ← · · · x∈X0

x∈X1

x∈Xa

(la somme sur les points de dimension a étant placée en degré a). Pour X ∈ C intègre, de corps des fonctions κ(X), avec d = δ(X) − 1 = dim(X) − 1, le complexe commence (à droite) par · · · ← H d+1 (κ(X), Λ(d)) ← 0. On note ce complexe KC(X, Λ). On note KHa (X, Λ) le groupe d’homologie en degré a. D’après Jannsen, Saito et Sato [23], les flèches dans ce complexe sont les opposées des flèches de bord en cohomologie galoisienne utilisées par Kato [26].

ASTÉRISQUE 339

(1012)

ZÉRO-CYCLES DES VARIÉTÉS p-ADIQUES

13

Remarque 3.3. — Lorsque l’on étudie ([26], [19], [22]) le corps de classes de dimension supérieure sur le corps des fonctions d’une variété intègre X de dimension d + 1 sur un corps fini F, le principal complexe considéré va de H d+2 (κ(X), Λ(d + 1)) à L 1 x∈X0 H (x, Λ). 1 Dans la suite spectrale de niveau, on a clairement Ea,b (X, Λ) = 0 pour a ∈ / [0, δ(X)] et pour a − b < 0. On a en particulier des applications 1 2 ∞ E1,1 → E1,1 → E1,1 ,→ H2 (X, Λ). 2 1 1 Le groupe E1,1 est le conoyau de la flèche d12,1 : E2,1 → E1,1 , donc (en utilisant la suite de Kummer) 2 E1,1 (X, Λ) = Coker [d12,1 : ⊕x∈X2 κ(x)× ⊗ Λ → ⊕x∈X1 Λ].

Les auteurs identifient cette application avec l’opposée de l’application div ⊗ Λ. Pour X ∈ C , le conoyau s’identifie donc avec le groupe CH1 (X)⊗Λ. L’application composée 2 ∞ E1,1 → E1,1 ,→ H2 (X, Λ) définit un homomorphisme ρX : CH1 (X) ⊗ Λ → H2 (X, Λ) dont on vérifie qu’il coïncide avec l’application cycle (sur les 1-cycles de X). Cette application jouit des propriétés fonctorielles attendues : elle est covariante par morphismes propres dans C , et contravariante par morphismes étales dans C . La fonctorialité suivante est particulièrement importante. Pour une immersion fermée Y ⊂ X dans C dont le complémentaire U ,→ X est dans C , on a un diagramme commutatif de suites de localisation CH1 (Y ) ⊗ Λ ρY

 H2 (Y, Λ)

/ CH1 (X) ⊗ Λ ρX

 / H2 (X, Λ)

/ CH1 (U ) ⊗ Λ

/0

ρU

 / H2 (U, Λ)

dont les lignes sont exactes. Proposition 3.4. — Soit X ∈ Ob( C ). (1) Si δ(X) = 1, la flèche ρX : CH1 (X) ⊗ Λ → H2 (X, Λ) est un isomorphisme. (2) Si δ(X) = 2, on a une suite exacte 0 → CH1 (X) ⊗ Λ → H2 (X, Λ) → KH2 (X, Λ) → 0. (3) Si δ(X) = 3, on a une suite exacte H3 (X, Λ) → KH3 (X, Λ) → CH1 (X) ⊗ Λ → H2 (X, Λ). (4) Si δ(X) ≤ 3, on a KH3 (X, Z/ln ) = KH3 (X, Ql /Zl )[ln ].

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

14

J.-L. COLLIOT-THÉLÈNE

Démonstration. — Les énoncés (1) à (3) résultent immédiatement de la forme de la suite spectrale de niveau. L’énoncé (4) résulte de la fonctorialité en les coefficients de la suite spectrale, du théorème 90 de Hilbert, et du théorème de Merkur0 ev-Suslin [30]. Théorème 3.5. — Supposons R hensélien et F fini ou séparablement clos. Soit X un R-schéma régulier, projectif et plat sur R, de dimension 2. Soit Λ = Z/ln ou Λ = Ql /Zl . Alors (a) L’application cycle ρX : CH1 (X) ⊗ Λ → H2 (X, Λ) est bijective. (b) Le groupe d’homologie KH2 (X, Λ) = 0. Démonstration. — On peut supposer X intègre et Λ = Z/ln . Pour dim(X) = 2, la suite exacte 0 → CH1 (X)/ln → H2 (X, Z/ln ) → KH2 (X, Z/ln ) → 0 de la proposition 3.4 s’identifie à la suite exacte 0 → Pic(X)/ln → H 2 (X, µln ) → Br(X)[ln ] → 0 déduite de la suite de Kummer en cohomologie étale. L’énoncé est alors le théorème d’Artin (Théorème 1.3). 3.2. Deux conjectures Soient X → B dans C et x ∈ Xa . Si p(x) = s, alors cdl (κ(x)) = cdl (F ) + a. Si p(x) = η, alors cdl (κ(x)) = cdl (k) + a − 1, ce qui si R est hensélien implique cdl (κ(x)) = cdl (F ) + a. Ceci implique immédiatement les énoncés (i) et (ii) dans la proposition suivante. Cette proposition n’est pas utilisée dans la démonstration du théorème principal 3.17, elle ne sert que dans la démonstration du théorème 3.29, qui établit les conjectures 3.7 et 3.8 en bas degré. Proposition 3.6. — Supposons R hensélien. Soit X ∈ Ob( C ). 1 (i) Si F est séparablement clos, alors Ea,b (X, Z/ln ) = 0 pour b < 0 et a quelconque. 1 (ii) Si F est un corps fini, alors Ea,b (X, Z/ln ) = 0 pour b < −1 et a quelconque. 1 (iii) Si F est un corps fini, alors Ea,−1 (X, Ql /Zl ) = 0 pour tout a. L’énoncé (iii), qui est un cas particulier d’un énoncé d’annulation pour la cohomologie galoisienne à coefficients Ql /Zl (m) pour certaines torsions m, est moins classique. Un tel énoncé avait été obtenu par B. Kahn [25]. Dans les situations de (i) et (iii), les termes non nuls de la suite spectrale sont tous dans le premier quadrant. Dans le cas (ii), il y a une ligne supplémentaire en dessous du premier quadrant.

ASTÉRISQUE 339

(1012)

ZÉRO-CYCLES DES VARIÉTÉS p-ADIQUES

15

Supposons R hensélien excellent et F séparablement clos. Pour X ∈ C , régulier, projectif, le complexe KC(X, Z/ln ) est exact en degré a = 0, 1. Ceci résulte de la forme de la suite spectrale, de l’isomorphisme H n (X, Z/ln ) ' H n (Xs , Z/ln ) (théorème de changement de base) et de la nullité de H n (Xs , Z/ln ) pour n > 2d = 2dim(Xs ). Inspirés par des conjectures de Kato [26] sur les variétés sur les corps finis, Saito et Sato [35] suggèrent : Conjecture 3.7. — Supposons R hensélien excellent et F séparablement clos. Pour tout X ∈ Ob( Q S P ) le complexe KC(X, Z/ln ) est exact. Conjecture 3.8. — Supposons R hensélien excellent et F fini. Notons I(Xs ) l’ensemble des composantes irréductibles de Xs,red . Pour tout X ∈ Ob( Q S P ) le complexe KC(X, Ql /Zl ) est exact sauf en degré a = 1, où l’homologie KH1 (X, Ql /Zl ) s’idenI(X ) tifie naturellement à Ql /Zl s . La forme de la suite spectrale et des résultats classiques de cohomologie étale [1] montrent ici encore que cette conjecture vaut en degré a = 0, 1. 3.3. Théorème de Lefschetz affine affiné Théorème 3.9. — Soit (X, Y ; U ) une Q S P -paire ample avec dim(X) = d + 1 ≥ 2. Supposons R hensélien. (i) Si F est séparablement clos, alors Hq (U, Z/ln ) = 0 pour q ≤ d + 1. (ii) Si F est fini, alors Hq (U, Z/ln ) = 0 pour q ≤ d. (iii) Si F est fini et dim(X) ≥ 3, alors Hd+1 (U, Ql /Zl ) = 0. Démonstration (esquisse). — Le schéma U est régulier, plat et de dimension relative d sur B. D’après la proposition 3.1, pour tout entier q, on a un isomorphisme Hq (U, Λ) ' H 2d+2−q (U, Λ(d)). Le schéma U est affine et R est hensélien. Les théorèmes de Lefschetz affines de Artin et Gabber (voir Fujiwara [15]) donnent la nullité de ces groupes pour q ≤ d si F est séparablement clos et pour q ≤ d − 1 si F est fini. C’est un de moins que ce qui est affirmé dans (i) et (ii). Pour aller plus loin, il faut utiliser le fait que (X, Y ; U ) est une Q S P -paire, en particulier que Xs,red ∪ Y est un diviseur à croisements normaux stricts. En utilisant le théorème de pureté cohomologique absolu de Gabber, on étend le théorème de changement de base de Rapoport et Zink [32] aux Q S P -paires, et l’on montre :

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

16

J.-L. COLLIOT-THÉLÈNE

Proposition 3.10. — Soit (X, Y ; U ) une Q S P -paire sur B. Soit V := Us,red . Si R est hensélien, pour tous entiers q et j, la flèche image réciproque H q (U, Z/ln (j)) → H q (V, Z/ln (j)) est un isomorphisme. Combinant cette proposition et l’isomorphisme ci-dessus, pour établir le théorème on est ramené à établir l’annulation des groupes H 2d+2−q (V, Λ(d)) pour q comme dans le théorème. Les énoncés (i) et (ii) sont des conséquences du théorème de Lefschetz affine ([1, XIV, Cor. 3.2]). Pour l’énoncé (iii), on montre H d+1 (V, Ql /Zl (d)) = 0 pour F fini et d ≥ 2. En tenant compte de l’énoncé (ii) pour V , on est ramené à montrer H d+1 (V, Ql (d)) = 0 puis à voir que l’opérateur de Frobenius sur H d+1 (V ×F F , Ql (d)) n’a pas la valeur propre 1. Ceci résulte des hypothèses et des théorèmes de Deligne sur la cohomologie des variétés sur les corps finis. 3.4. Théorèmes de Bertini relatifs (Jannsen et Saito) La littérature contient diverses versions du théorème de Bertini sur les sections hyperplanes d’une variété X projective et lisse sur un corps k. Dans un premier temps, on affirme l’existence de sections hyperplanes X ∩ H lisses sur k. Une référence sur le sujet est un livre de J.-P. Jouanolou. Ceci vaut si le corps k est infini. Si le corps est un corps fini F, l’énoncé vaut encore si l’on remplace le plongement projectif donné par un plongement de Veronese convenable. B. Poonen et O. Gabber ont donné des énoncés précis dans cette direction. Pour les problèmes ici considérés, on peut aussi garder le plongement donné et remplacer le corps fini F donné par deux extensions finies de F suffisamment grandes et de degrés premiers entre eux. Dans un second temps, on se donne une sous-k-variété fermée Z ⊂ X et l’on demande s’il existe une section hyperplane lisse de X qui contient Z. Il faut bien sûr imposer des conditions à Z, par exemple Z lisse et 2dim(Z) < dim(X). A. Altman et S. Kleiman ont consacré un article à cette question. Sur un corps fini, Poonen a obtenu les énoncés adéquats. Quand on étudie les schémas arithmétiques, pour faire des démonstrations par récurrence sur la dimension, on a besoin d’adapter ces théorèmes au-dessus d’un anneau de valuation discrète (ou mieux d’un schéma de Dedekind). L’énoncé suivant (Jannsen et Saito [21]) correspond à la première situation. Théorème 3.11 (Théorème de Bertini relatif, version 1). — Soit R un anneau de valuation discrète. Soit X ∈ Ob( Q S P ) de dimension dim(X) ≥ 2. Il existe un

ASTÉRISQUE 339

(1012)

ZÉRO-CYCLES DES VARIÉTÉS p-ADIQUES

17

R-plongement X ,→ PnR et un R-hyperplan H = Pn−1 ⊂ PnR tels que l’intersecR tion schématique X ∩ H soit dans Ob( Q S P ) et que le couple (X, X ∩ H) soit une Q S P -paire (ample). La seconde situation mène à un énoncé plus technique, dû aussi à Jannsen et Saito ([35], Thm. 4.2). On impose en particulier des conditions de transversalité entre le sous-schéma régulier Z ⊂ X considéré et les composantes de la fibre spéciale réduite. Je ne reproduis pas cet énoncé ici. Dans ce rapport, je me contente de l’appeler « Théorème de Bertini relatif, version 2 ».

3.5. Lemme de déplacement ; homologie d’un éclatement La démonstration du lemme de déplacement suivant ne pose pas de difficulté particulière. Proposition 3.12. — Soit X un schéma régulier intègre dans Ob( C ). Soit Y ⊂ X un sous-schéma fermé propre. Soit U : X \ Y . Pour tout entier q ≥ 0, l’application ⊕x∈Xq ∩U Z → CHq (X) est surjective. La proposition suivante n’est pas non plus trop surprenante. Elle décrit le comportement du groupe H2 (•, Λ) dans un éclatement. Proposition 3.13. — Soit X ∈ Ob( C ) régulier de dimension d + 1. Soient ˜ → X l’éclaté de X en un point fermé x ∈ Xs , et E ,→ X ˜ le diviseur π : X d exceptionnel. Supposons cdl (F ) ≤ 1. Soit W ⊂ E ' Pκ(x) un sous-schéma fermé intègre de dimension 1. Si Λ = Ql /Zl , ou si Λ = Z/ln et le degré sur κ(x) de W ,→ Pdκ(x) est premier à l, alors ˜ Λ) → H2 (X, Λ)] = Λ(W ). Ker [π∗ : H2 (X, ˜ Z/ln ) Pour Λ = Z/ln , le groupe Λ(W ) est par définition le sous-groupe de H2 (X, engendré par ρX ([W ]). Pour Λ = Ql /Zl , c’est par définition la limite inductive de ces groupes pour n tendant vers l’infini. Pour établir la proposition, on utilise les isomorphismes H2 (E, Z/ln ) ' H 2d−2 (E, Z/ln (d − 1)) ' Z/ln , où le second résulte de l’hypothèse cdl (F ) ≤ 1 et du calcul bien connu de la cohomologie d’un espace projectif.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

18

J.-L. COLLIOT-THÉLÈNE

3.6. Démonstration du théorème principal Théorème 3.14. — Supposons R hensélien et F fini ou séparablement clos. Soit X un R-schéma dans Q S P . Pour Λ = Z/ln ou Λ = Ql /Zl , l’application cycle ρX : CH1 (X) ⊗ Λ → H2 (X, Λ) est surjective. Démonstration. — On peut supposer X intègre et Λ = Z/ln . Pour dim(X) = 1, resp. dim(X) = 2, la proposition 3.4, resp. 3.5, montre que l’application cycle est un isomorphisme. Supposons dim(X) ≥ 3. Pour établir l’énoncé, on procède par récurrence sur la dimension. Par le théorème de Bertini relatif, version 1, il existe un diviseur ample Y ⊂ X qui définit une Q S P -paire ample. Soit U := X \ Y . Par la fonctorialité de l’homologie et de l’application cycle, on a un diagramme commutatif CH1 (Y )/ln ρY



H2 (Y, Z/ln )

/ CH1 (X)/ln ρX

 / H2 (X, Z/ln )

/ H2 (U, Z/ln )

où la suite inférieure est exacte. Par hypothèse de récurrence, ρY est surjective. Par le théorème 3.9 (Lefschetz affine affiné) pour la Q S P -paire ample sur l’anneau de valuation discrète hensélien R, comme la dimension de X est au moins 3, on a H2 (U, Z/ln ) = 0. Ainsi ρX est surjectif. Théorème 3.15. — Supposons R hensélien excellent et F fini ou séparablement clos. Soit X un R-schéma dans Q S P de dimension 3. L’application cycle ρX : CH1 (X) ⊗ Ql /Zl → H2 (X, Ql /Zl ) est bijective. Démonstration. — On a déjà établi la surjectivité. Soit α dans le noyau de ρX . Par P le lemme de déplacement 3.12, on peut représenter α sous la forme 1≤j≤m [Cj ] ⊗ λj avec λj ∈ Ql /Zl et des courbes fermées intègres Cj ⊂ X non contenues dans la fibre spéciale Xs . Par une version de la résolution plongée des singularités des courbes due à U. Jannsen (appendice de [35], qui utilise l’hypothèse que X est excellent), on ˜ → X en des points fermés de la peut trouver un composé de N éclatements π : X fibre spéciale de façon que les transformés propres C˜j des Cj soient des courbes régulières sans point commun deux à deux et satisfassent des conditions de transversalité ˜s. convenables par rapport aux composantes réduites de la fibre spéciale X ˜i ⊂ X, ˜ i = 1, . . . , N les transformés propres dans X ˜ des divers diviseurs Notons E exceptionnels.

ASTÉRISQUE 339

(1012)

ZÉRO-CYCLES DES VARIÉTÉS p-ADIQUES

19

˜ Y) Par le théorème de Bertini relatif, version 2, il existe une Q S P -paire ample (X, ˜ ˜ telle que le diviseur ample ι : Y ⊂ X contienne tous les Cj . On a α = π∗ ι∗ β, avec P β = j [C˜j ] ⊗ λj ∈ CH1 (Y ) ⊗ Ql /Zl . Comme Y est ample, pour chaque i, le sous˜i ∩ Y du schéma X ˜ (de dimension 3) est un fermé non vide, intègre et schéma Wi := E ˜i0 pour i0 6= i. Par une version itérée de dimension 1, et Wi n’est inclus dans aucun E de la proposition 3.13 (homologie d’un éclatement), on a ˜ Ql /Zl ) → H2 (X, Ql /Zl )] = Ker [π∗ : H2 (X,

N X

Ql /Zl (Wi ).

i=1

Dans le diagramme CH1 (Y ) ⊗ Ql /Zl

ι∗

˜ ⊗ Ql /Zl / CH1 (X) ρX ˜

ρY

 / H2 (Y, Ql /Zl )

˜ \ Y, Ql /Zl ) H3 (X

 ˜ Ql /Zl ) / H2 (X,

ι∗

la ligne médiane est la suite exacte de localisation, et le carré est commutatif. Le théorème 3.9 (théorème de Lefschetz affine affiné) assure la nullité du terme H3 . Comme Y est dans Ob( C ) et de dimension 2, on sait que la flèche ρY est injective. Combinant ces résultats avec le diagramme commutatif ˜ ⊗ Ql /Zl CH1 (X) 

ρX ˜

π∗

CH1 (X) ⊗ Ql /Zl

˜ Ql /Zl ) / H2 (X, π∗

ρX

 / H2 (X, Ql /Zl )

on voit que β est dans le groupe engendré par les Wi dans CH1 (Y ) ⊗ Ql /Zl . Comme ˜ → X de chaque Wi est un point fermé, on conclut l’image par π : X α = π∗ ι∗ (β) = 0 ∈ CH1 (X) ⊗ Ql /Zl . Théorème 3.16. — Supposons R hensélien excellent et F fini ou séparablement clos. Soit X un R-schéma dans Q S P de dimension 3. (a) L’application cycle ρX : CH1 (X) ⊗ Z/ln → H2 (X, Z/ln ) est bijective. (b) Pour tout n > 0 on a KH3 (X, Z/ln ) = KH3 (X, Ql /Zl )[ln ] = 0. Démonstration. — La surjectivité dans (a) a été établie au théorème 3.14. Il reste à établir l’injectivité. D’après la proposition 3.4 (dont la preuve utilise le théorème de Merkur0 ev et Suslin), pour établir (a) et (b) il suffit de montrer KH3 (X, Ql /Zl ) = 0. On fixe un diviseur Y ⊂ X tel que (X, Y ; U ) soit une Q S P -paire ample. Comme

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

20

J.-L. COLLIOT-THÉLÈNE

X est de dimension 3, la restriction KH3 (X, Ql /Zl ) → KH3 (U, Ql /Zl ) est injective (c’est purement formel, ceci vaut pour tout ouvert U dense dans X). Il suffit donc de montrer KH3 (U, Ql /Zl ) = 0. On a le diagramme commutatif de suites de localisation CH1 (Y ) ⊗ Ql /Zl ρY

 H2 (Y, Ql /Zl )

/ CH1 (X) ⊗ Ql /Zl ρX

 / H2 (X, Ql /Zl )

/ CH1 (U ) ⊗ Ql /Zl

/0

ρU

 / H2 (U, Ql /Zl ).

Nous avons déjà établi que ρY et ρX sont bijectives. Ainsi ρU est injective. Mais H2 (U, Ql /Zl ) = 0 (Thm. 3.9, Lefschetz affine affiné). Donc CH1 (U ) ⊗ Ql /Zl = 0. Par ailleurs H3 (U, Ql /Zl ) = 0 (Thm. 3.9, Lefschetz affine affiné). En appliquant la proposition 3.4(3) à U , on conclut KH3 (U, Ql /Zl ) = 0. Le théorème principal en dimension quelconque s’énonce : Théorème 3.17. — Supposons R hensélien excellent et F fini ou séparablement clos. Soit X un R-schéma dans Q S P . L’application cycle ρX : CH1 (X) ⊗ Z/ln → H2 (X, Z/ln ) est un isomorphisme de groupes finis. Démonstration. — La surjectivité a été établie (Thm. 3.14). Pour établir l’injectivité, on fait une récurrence sur la dimension, dans le même esprit que celle du théorème 3.15, avec Λ = Z/ln , en utilisant le théorème de Bertini relatif, version 2. Je renvoie ici le lecteur à [35]. Des isomorphismes H2 (X, Z/ln ) ' H 2d (X, Z/ln (d)) (Proposition 3.1) et H 2d (X, Z/ln (d)) ' H 2d (Xs , Z/ln (d)) (changement de base propre, R est hensélien) on déduit que pour F séparablement clos ou fini les groupes H2 (X, Z/ln ) sont finis. Remarque 3.18. — K. Sato [40] a défini des applications cycle ρX,pr sur les quotients CH1 (X)/pr , à valeurs dans la cohomologie étale de certains complexes de Z/pr -faisceaux étales. Lorsque X/R est propre, régulier et semistable, Saito et Sato montrent dans [37] que ces applications sont surjectives. 3.7. Applications aux groupes de Chow Le théorème suivant est une sorte de théorème de Lefschetz pour le groupe de Chow des 1-cycles modulo un entier. Théorème 3.19. — Soit R un anneau de valuation discrète hensélien excellent de corps résiduel F . Soit X un R-schéma dans Q S P . Soit n ≥ 1 un entier. Soit i : Y ,→ X un diviseur tel que (X, Y ) soit une Q S P -paire ample. Soit U = X \ Y .

ASTÉRISQUE 339

(1012)

ZÉRO-CYCLES DES VARIÉTÉS p-ADIQUES

21

(a) Supposons F séparablement clos et dim(X) = d + 1 ≥ 2. Alors CH1 (U )/ln = 0 et l’application i∗ = CH1 (Y )/ln → CH1 (X)/ln est surjective. Si d ≥ 2, elle est bijective. (b) Supposons F fini et dim(X) = d + 1 ≥ 3. Alors CH1 (U )/ln = 0 et l’application i∗ = CH1 (Y )/ln → CH1 (X)/ln est surjective. Si d ≥ 3, elle est bijective. Démonstration. — Considérons le diagramme commutatif à lignes exactes CH1 (Y )/ln

H3 (U, Z/ln )



i∗

ρY

/ CH1 (X)/ln

/ CH1 (U )/ln

ρX

/ H2 (Y, Z/ln )

 / H2 (X, Z/ln )



/0

ρU

/ H2 (U, Z/ln ).

D’après le théorème 3.17, ρY et ρX sont des isomorphismes. Le théorème résulte alors du théorème de Lefschetz affine affiné (Théorème 3.9). Théorème 3.20. — Soit R un anneau de valuation discrète hensélien excellent de corps résiduel F séparablement clos. Soit X un objet de Q S P , et soient Y1 , . . . , YN les composantes de Xs,red . Alors pour tout entier n > 0 premier à car(F ), l’intersection avec les composantes induit un isomorphisme '

CH1 (X)/n −→

N M

Z/nZ.

i=1

Démonstration. — Soit d = dim(X) − 1. On a des isomorphismes naturels H2 (X, Z/ln ) ' H 2d (X, Z/ln (d)) '

' H 2d (Xs , Z/ln (d)) −→

N M

'

H 2d (Yj , Z/ln (d)) −→

i=1

N M

Z/ln ,

i=1

et d’après le théorème principal 3.17 on a un isomorphisme CH1 (X)/ln ' H2 (X, Z/ln ). On vérifie que pour chaque composante Yi la flèche associée CH1 (X)/ln → Z/ln associe à un 1-cycle la classe modulo ln de son nombre d’intersection avec le diviseur vertical Yi ⊂ X. Théorème 3.21. — Soit R un anneau de valuation discrète hensélien excellent, de corps des fractions k, de corps résiduel F fini ou séparablement clos. Soit p l’exposant caractéristique de F . Soit X un R-schéma projectif et lisse à fibres géométriquement connexes. Soit i : Xs ,→ X la fibre spéciale, et j : V = Xη ,→ X la fibre générique. (1) Pour tout entier n > 0 premier à p = car(F ), on a des isomorphismes de groupes abéliens finis '

'

CH0 (Xs )/n ←− CH1 (X)/n −→ CH0 (V )/n,

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

22

J.-L. COLLIOT-THÉLÈNE

où CH0 (V ) désigne le groupe de Chow de dimension zéro usuel de la k-variété V , la flèche de gauche est i∗ et la flèche de droite j ∗ . (2) Supposons F séparablement clos. Le groupe A0 (V ) est divisible par tout entier n premier à p. (3) Supposons F fini. (a) Le groupe A0 (V ) est extension du groupe fini A0 (Xs ) par un groupe p0 -divisible. (b) Pour tout entier n premier à p, l’application cycle CH0 (V )/n → H 2d (V, µ⊗d n ) est injective. (c) Soit Br0 (V ) le sous-groupe de Br V de torsion première à p. Le noyau à gauche de l’accouplement A0 (V ) × Br0 V → Br0 k ⊂ Q/Z est divisible par tout entier n premier à p. Démonstration. — Soit d la dimension de V . Soit n un entier premier à p. Les applications cycle sur X et sur Xs s’inscrivent dans un diagramme commutatif CH1 (X)/n

/ H 2d (X, Z/n(d))

clX

i∗

 CH0 (Xs )/n



clXs

i∗

/ H 2d (Xs , Z/n(d)).

La flèche verticale de droite est un isomorphisme (changement de base propre). D’après le théorème principal 3.17, la flèche clX est un isomorphisme. Comme X est lisse sur R et R hensélien, la flèche de restriction i∗ : CH1 (X) → CH0 (Xs ) est surjective. Ceci suffit à assurer que toutes les flèches dans le diagramme ci-dessus sont des isomorphismes. Il en est donc ainsi de i∗ : CH1 (X)/n → CH0 (Xs )/n. [Lorsque F est un corps fini, on reconnaît dans l’isomorphisme obtenu CH0 (Xs )/n ' H 2d (Xs , Z/n(d)) un cas particulier du théorème du corps de classes non ramifié pour les variétés projectives et lisses sur un corps fini, voir [42].] Une suite de localisation élémentaire fournit la suite exacte j∗

i

∗ CH1 (Xs )/n −→ CH1 (X)/n −→ CH0 (Xη )/n → 0.

Comme Xs est un diviseur principal sur X, l’application composée i

i∗

∗ CH1 (Xs )/n −→ CH1 (X)/n −→ CH0 (Xs )/n

est nulle. Comme on a établi que i∗ est un isomorphisme, ceci implique que i∗ est nul. Ainsi j ∗ est un isomorphisme. Ceci établit le point (1). Le point (2) en résulte, puisque le groupe A0 (Xs ) est divisible.

ASTÉRISQUE 339

(1012)

ZÉRO-CYCLES DES VARIÉTÉS p-ADIQUES

23

Supposons F fini. Le théorème de Kato et Saito [27] assure que le groupe A0 (Xs ) est fini. L’application de réduction A0 (Xη ) → A0 (Xs ) est surjective (lemme de Hensel). En utilisant (1) (voir aussi le théorème 2.3) on obtient (3)(a). Comme X/R est lisse, une version connue de la conjecture de Gersten assure que l’application de restriction 2d ⊗d H 2d (X, µ⊗d n ) → H (V, µn ) est injective. Du diagramme commutatif CH1 (X)/n clX

 H 2d (X, Z/n)

j∗

/ CH0 (V )/n clV

 / H 2d (V, Z/n),

où la flèche supérieure est surjective, on déduit (3)(b). L’énoncé (3)(c) résulte alors du fait connu que l’accouplement naturel 2 ⊗d 2d+2 H 2d (V, µ⊗d (V, µn⊗d+1 ) ' Z/n n ) × H (V, µn ) → H

est une dualité parfaite de groupes finis, et que cet accouplement induit un accouplement CH0 (V )/n × Br(V )[n] → Z/n. Lemme 3.22. — Soit A un groupe abélien. (i) Les propriétés suivantes sont équivalentes (a) Le groupe A est la somme directe d’un groupe fini d’ordre premier à p et d’un groupe p0 -divisible. (b) Pour presque tout premier l 6= p, le quotient A/l est nul, et pour tout premier l 6= p, il existe un entier nl > 0 tel que la projection A/ln+1 → A/ln soit un isomorphisme de groupes finis pour tout n ≥ nl . (ii) Si un groupe A possède ces propriétés, il en est de même de tout quotient de A. Théorème 3.23. — Soit R un anneau de valuation discrète hensélien excellent, de corps des fractions k, de corps résiduel F fini ou séparablement clos. Soit p l’exposant caractéristique de F . Soit V une variété projective, lisse, géométriquement connexe sur k. Soit X un R-schéma dans Q S P et soit V la k-variété Xη , supposée géométriquement intègre. Le groupe A0 (V ) est isomorphe à la somme directe d’un groupe fini d’ordre premier à p et d’un groupe p0 -divisible. En particulier A0 (V )/l = 0 pour presque tout premier l, et A0 (V )/n est fini pour tout n > 0 premier à p. Démonstration. — D’après le théorème 3.19 il existe une R-courbe Y ,→ X qui est dans Q S P et pour laquelle pour tout l premier, l 6= p, et pour tout n > 0 l’application CH1 (Y )/ln → CH1 (X)/ln est surjective. Les restrictions à la fibre générique CH1 (X) → CH0 (Xη ) et CH1 (Y ) → CH0 (Yη ) sont clairement surjectives, et elles sont compatibles. Ainsi les applications naturelles CH0 (Yη )/ln → CH0 (Xη )/ln sont

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

24

J.-L. COLLIOT-THÉLÈNE

surjectives. Cela implique le même énoncé pour A0 (Yη )/ln → A0 (Xη )/ln . Pour la courbe projective et lisse Yη , on montre que le groupe A0 (Yη ) est la somme directe d’un groupe fini d’ordre premier à p et d’un groupe p0 -divisible. Le lemme 3.22 donne alors l’énoncé sur A0 (Xη ). Remarque 3.24. — La finitude de A0 (V )/n et de CH0 (V )/n résulte directement de la finitude de CH1 (X)/n ' H2 (X, Z/ln ). Théorème 3.25. — Soit R un anneau de valuation discrète hensélien excellent, de corps des fractions k, de corps résiduel F fini ou séparablement clos. Soit p l’exposant caractéristique de F . Soit V une variété projective, lisse, géométriquement connexe sur k. Le groupe A0 (V ) est isomorphe à la somme directe d’un groupe fini d’ordre premier à p et d’un groupe p0 -divisible. En particulier A0 (V )/l = 0 pour presque tout premier l, et A0 (V )/n est fini pour tout n > 0 premier à p. Démonstration. — Le théorème d’uniformisation de de Jong [24], dans la version raffinée de Gabber, théorème décrit par Illusie dans [18], implique que, pour tout premier l 6= p, il existe un k-morphisme propre f : V 0 → V , génériquement fini de degré d premier à l, et une extension finie d’anneaux de valuation discrète R0 /R tels que la variété V 0 soit lisse et géométriquement intègre sur le corps des fractions de R0 et admette un modèle X 0 /R0 qui soit Q S P . Les propriétés usuelles des groupes de Chow des variétés lisses impliquent que le composé f∗

f∗

A0 (V )−→A0 (V 0 )−→A0 (V ) est la multiplication par d. Le théorème résulte alors du théorème 3.23. Remarque 3.26. — Le théorème originel de de Jong [24] combiné avec le théorème 3.23 suffit à établir la trivialité de A0 (V )/l pour presque tout premier l. C’est cet énoncé que l’on trouve dans [35]. Corollaire 3.27. — Soient R, p et k comme dans le théorème ci-dessus. Soit V une variété projective, lisse, géométriquement connexe sur k. Si pour tout corps algébriquement clos Ω contenant k on a A0 (V ×k Ω) = 0, alors le groupe A0 (V ) est la somme directe d’un groupe fini et d’un groupe d’exposant une puissance de p. Démonstration. — Sur un corps quelconque, on montre en effet que le groupe A0 (V ) de toute telle variété V est annulé par un entier positif. L’énoncé du corollaire résulte alors du théorème. Le résultat s’applique en particulier aux k-variétés géométriquement rationnellement connexes (au sens de Kollár, Miyaoka, Mori). Ainsi le groupe A0 (V ) d’une variété rationnellement connexe sur le corps k = C((t)) est un groupe fini. C’est une question ouverte de savoir si dans ce cas le groupe A0 (V ) est nul.

ASTÉRISQUE 339

(1012)

ZÉRO-CYCLES DES VARIÉTÉS p-ADIQUES

25

3.8. Applications aux conjectures 3.7 et 3.8 Théorème 3.28. — Soit R un anneau de valuation discrète hensélien excellent de corps résiduel F . Soit X un R-schéma dans Q S P . Soit n ≥ 1 un entier. (a) Si F est un corps séparablement clos, alors KH2 (X, Z/ln ) = 0 et KH2 (X, Ql /Zl ) = 0. (b) Si F est un corps fini, on a KH2 (X, Ql /Zl ) = 0. Démonstration. — Dans chacun des cas considérés, la proposition 3.6 assure que la suite spectrale de niveau est concentrée dans le premier quadrant. La forme de cette suite spectrale donne alors des suites exactes CH1 (X) ⊗ Λ → H2 (X, Λ) → KH2 (X, Λ) → 0. Du théorème 3.14 on déduit KH2 (X, Λ) = 0. Théorème 3.29. — Soit R un anneau de valuation discrète hensélien excellent de corps résiduel F . Soit X un R-schéma dans Q S P . Soit n ≥ 1 un entier. (a) Si F est un corps séparablement clos, alors KH3 (X, Z/ln ) = 0 et KH3 (X, Ql /Zl ) = 0. (b) Si F est un corps fini et dim(X) ≤ 4, alors KH3 (X, Z/ln ) = 0. (c) Si F est un corps fini, on a KH3 (X, Ql /Zl ) = 0. Démonstration. — L’énoncé est trivial pour dim(X) ≤ 2, et il a été démontré pour dim(X) = 3 (Thm. 3.16). Supposons donc dim(X) ≥ 4 et établissons le résultat par récurrence sur la dimension. Soit (X, Y ; U ) une Q S P -paire ample. Pour Λ = Z/ln si F est séparablement clos et Λ = Ql /Zl si F est fini, pour tout X ∈ C , la proposition 3.6 assure que la suite spectrale de niveau est concentrée dans le premier quadrant (c’est ici qu’on se limite à Λ = Ql /Zl lorsque F est un corps fini). Appliquant ceci à U , on trouve une suite exacte H3 (U, Λ) → KH3 (U, Λ) → CH1 (U ) ⊗ Λ. Lorsque dim(U ) = 4, la forme de la suite spectrale assure que l’on a encore cette suite exacte pour F fini et Λ = Z/ln . D’après le théorème 3.19, on a CH1 (U ) ⊗ Λ = 0. D’après le théorème de Lefschetz affine affiné 3.9, comme on a dim(U ) ≥ 4, on a H3 (U, Λ) = 0. On conclut donc KH3 (U, Λ) = 0 dans chacun des trois cas (a), (b), (c). Par ailleurs on a une longue suite exacte · · · → KH3 (Y, Λ) → KH3 (X, Λ) → KH3 (U, Λ) → · · · Par hypothèse de récurrence, KH3 (Y, Λ) = 0. Ainsi KH3 (X, Λ) = 0.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

26

J.-L. COLLIOT-THÉLÈNE

Remarque 3.30. — Dans [35], les auteurs demandent si l’énoncé (b) vaut en toute dimension. En établissant le théorème 3.16, ils montrent que c’est le cas pour dim(X) = 3. Dans l’argument ci-dessus, ils commencent la récurrence en dimension 3, et n’observent pas le résultat pour dim(X) = 4.

4. RÉSULTATS RÉCENTS ET QUESTIONS OUVERTES Le théorème de Saito et Sato joue un rôle important dans la démonstration du théorème suivant, dont la démonstration, qui utilise la théorie de Hodge et est très élaborée, ne peut être évoquée ici. Théorème 4.1 (M. Asakura et S. Saito [2]). — Soient k un corps p-adique, R son anneau d’entiers et F son corps résiduel. Soit X ⊂ P3R une R-hypersurface lisse de degré au moins 5. Supposons la fibre générique Xk très générale. Soit r le rang du groupe de Picard de la fibre spéciale XF . Alors le sous-groupe de torsion l-primaire de A0 (X) est somme d’un groupe fini et de (Ql /Zl )r−1 . Il est facile de donner des exemples de telles surfaces avec r > 1. Le théorème suivant généralise une partie du théorème 1.2, qui porte sur les courbes, pour lesquelles BrX = 0 (Théorème 1.3). Théorème 4.2 (S. Saito et K. Sato [37]). — Soient k un corps p-adique, R son anneau d’entiers, X un R-schéma propre, connexe et régulier. Le sous-groupe BrX ⊂ BrXk est dans le noyau à droite de l’accouplement CH0 (Xk ) × BrXk → Q/Z. Si le théorème de pureté vaut pour le groupe de Brauer de X, alors BrX ⊂ BrXk est le noyau à droite de cet accouplement. La partie première à p de ce théorème est déjà dans [12]. La partie p-primaire est beaucoup plus délicate. Le théorème de pureté vaut pour le groupe de Brauer si la dimension de X est au plus 3 (Gabber), il vaut pour la torsion non p-primaire du groupe de Brauer (Gabber [15]), il vaut aussi pour la torsion p-primaire dans un certain nombre de cas (voir [37]).

ASTÉRISQUE 339

(1012)

ZÉRO-CYCLES DES VARIÉTÉS p-ADIQUES

27

Quelques questions Soient k un corps p-adique et V une k-variété projective, lisse, géométriquement connexe. (1) Pour V de dimension au moins 3 et n > 0 entier, le groupe A0 (V )[n] est-il fini ? (2) Le quotient A0 (V )/p est-il fini ? (3) Les noyaux des applications albX : A0 (V ) → AlbV (k) et A0 (V ) → Hom(BrV, Q/Z) sont-ils chacun extension d’un groupe fini par un groupe divisible ? (4) Supposons que V /k est la fibre générique de X/R projectif quasisemistable sur l’anneau des entiers R de k. Pour n entier premier à p, le théorème principal 3.17 donne une formule pour le quotient CH1 (X)/n. Peut-on en déduire une formule pour CH0 (V )/n ? Un cas particulier est étudié dans [14]. (5) Peut-on comprendre de façon « invariante » l’exemple de Parimala et Suresh [31] ?

RÉFÉRENCES [1] M. Artin, A. Grothendieck & J.-L. Verdier (éds.) – Théorie des topos et cohomologie étale des schémas. Tome 3, Lecture Notes in Math., vol. 305, Springer, 1973, Séminaire de Géométrie Algébrique du Bois-Marie 1963–1964 (SGA 4). [2] M. Asakura & S. Saito – Surfaces over a p-adic field with infinite torsion in the Chow group of 0-cycles, Algebra & Number Theory 1 (2007), p. 163–181. [3] S. Bloch – Torsion algebraic cycles, K2 , and Brauer groups of function fields, Bull. Amer. Math. Soc. 80 (1974), p. 941–945. [4]

, Lectures on algebraic cycles, Duke University Mathematics Series, vol. IV, 1980, Second edition, New mathematical monographs vol. 16, Cambridge University Press (2010).

[5]

, Torsion algebraic cycles, K2 , and Brauer groups of function fields, in The Brauer group, Lecture Notes in Math., vol. 844, Springer, 1981, p. 75–102.

[6] S. Bloch & A. Ogus – Gersten’s conjecture and the homology of schemes, Ann. Sci. École Norm. Sup. 7 (1974), p. 181–201. [7] J.-L. Colliot-Thélène – Cycles algébriques de torsion et K-théorie algébrique, in Arithmetic algebraic geometry (Trento, 1991), Lecture Notes in Math., vol. 1553, Springer, 1993, p. 1–49. [8]

, L’arithmétique du groupe de Chow des zéro-cycles, J. Théor. Nombres Bordeaux 7 (1995), p. 51–73.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

28

[9]

[10] [11]

[12] [13] [14] [15]

[16] [17]

[18] [19] [20] [21]

[22] [23] [24]

J.-L. COLLIOT-THÉLÈNE

, Zéro-cycles sur les surfaces sur un corps p-adique : quelques résultats obtenus au moyen de la K-théorie algébrique. Notes distribuées à l’école d’été de Sestri Levante, juillet 2004. Disponible sur la page http://www.math.u-psud. fr/~colliot/liste-cours-exposes.html. J.-L. Colliot-Thélène & W. Raskind – K2 -cohomology and the second Chow group, Math. Ann. 270 (1985), p. 165–199. , Groupe de Chow de codimension deux des variétés définies sur un corps de nombres : un théorème de finitude pour la torsion, Invent. math. 105 (1991), p. 221–245. J.-L. Colliot-Thélène & S. Saito – Zéro-cycles sur les variétés p-adiques et groupe de Brauer, Int. Math. Res. Not. 1996 (1996), p. 151–160. J.-L. Colliot-Thélène, J.-J. Sansuc & C. Soulé – Torsion dans le groupe de Chow de codimension deux, Duke Math. J. 50 (1983), p. 763–801. C. S. Dalawat – Le groupe de Chow d’une surface rationnelle sur un corps local, Compos. Math. 141 (2005), p. 344–358. K. Fujiwara – A proof of the absolute purity conjecture (after Gabber), in Algebraic geometry 2000, Azumino (Hotaka), Adv. Stud. Pure Math., vol. 36, Math. Soc. Japan, 2002, p. 153–183. W. Fulton – Intersection theory, Ergebnisse Math. Grenzg., vol. 2, Springer, 1984. A. Grothendieck – Le groupe de Brauer, III, in Dix exposés sur la cohomologie des schémas, Advanced Studies in Pure Mathematics, vol. 3, Masson et NorthHolland, 1968. L. Illusie – On Gabber’s refined uniformization, exposés à l’Université de Tokyo, http://www.math.u-psud.fr/~illusie/, 2008. U. Jannsen – Hasse principles for higher-dimensional fields, prépublication http://arxiv.org/abs/0910.2803. U. Jannsen & S. Saito – Kato homology of arithmetic schemes and higher class field theory over local fields, Doc. Math., extra vol. (2003), p. 479–538. , Bertini and Lefschetz theorems for schemes over discrete valuation rings, with applications to higher class field theory, prépublication, http://arxiv. org/abs/0911.1470. , Kato conjecture and motivic cohomology over finite fields, prépublication, http://arxiv.org/abs/0910.2815. U. Jannsen, S. Saito & K. Sato – Étale duality for constructible sheaves on arithmetic schemes, prépublication, http://arxiv.org/abs/0910.3759. A. J. de Jong – Smoothness, semi-stability and alterations, Publ. Math. I.H.É.S. 83 (1996), p. 51–93.

ASTÉRISQUE 339

(1012)

ZÉRO-CYCLES DES VARIÉTÉS p-ADIQUES

29

[25] B. Kahn – Résultats de “pureté” pour les variétés lisses sur un corps fini, in Algebraic K-theory and algebraic topology (Lake Louise, AB, 1991), NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., vol. 407, Kluwer Acad. Publ., 1993, p. 57–62. [26] K. Kato – A Hasse principle for two-dimensional global fields, J. reine angew. Math. 366 (1986), p. 142–183. [27] K. Kato & S. Saito – Unramified class field theory of arithmetical surfaces, Ann. of Math. 118 (1983), p. 241–275. [28] G. Laumon – Homologie étale, in Séminaire de géométrie analytique (École Norm. Sup., Paris, 1974-75), Astérisque, vol. 36–37, Soc. Math. France, 1976, p. 163–188. [29] S. Lichtenbaum – Duality theorems for curves over p-adic fields, Invent. math. 7 (1969), p. 120–136. [30] A. S. Merkur0 ev & A. A. Suslin – K-cohomologie des variétés de SeveriBauer et homomorphisme de norme résiduelle (en russe), Izv. Akad. Nauk SSSR Ser. Mat. 46 (1982), p. 1011–1046, 1135–1136. [31] R. Parimala & V. Suresh – Zero-cycles on quadric fibrations : finiteness theorems and the cycle map, Invent. math. 122 (1995), p. 83–117. [32] M. Rapoport & T. Zink – Über die lokale Zetafunktion von Shimuravarietäten. Monodromiefiltration und verschwindende Zyklen in ungleicher Charakteristik, Invent. math. 68 (1982), p. 21–101. [33] W. Raskind – Torsion algebraic cycles on varieties over local fields, in Algebraic K-theory : connections with geometry and topology (Lake Louise, AB, 1987), NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., vol. 279, Kluwer Acad. Publ., 1989, p. 343–388. [34] S. Saito – On the cycle map for torsion algebraic cycles of codimension two, Invent. math. 106 (1991), p. 443–460. [35] S. Saito & K. Sato – A finiteness theorem for zero-cycles over p-adic fields. With an appendix by Uwe Jannsen, Ann. of Math. 172 (2010), p. 1593–1639. [36] , A p-adic regulator map and finiteness results for arithmetic schemes, Documenta Mathematica (2010), p. 525–594, extra volume : Andrei A. Suslin’s sixtieth birthday. , Zero-cycles on varieties over p-adic fields and Brauer groups, prépubli[37] cation http://arxiv.org/abs/0906.2273. [38] P. Salberger – Torsion cycles of codimension 2 and l-adic realizations of motivic cohomology, in Séminaire de Théorie des Nombres, Paris, 1991–92, Progr. Math., vol. 116, Birkhäuser, 1993, p. 247–277. [39] K. Sato – Injectivity of the torsion cycle map of codimension two of varieties over p-adic fields with semi-stable reduction, J. reine angew. Math. 501 (1998), p. 221–235.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

30

[40]

J.-L. COLLIOT-THÉLÈNE

, p-adic étale Tate twists and arithmetic duality, Ann. Sci. École Norm. Sup. 40 (2007), p. 519–588.

[41] M. Spiess – On indecomposable elements of K1 of a product of elliptic curves, K-Theory 17 (1999), p. 363–383. [42] T. Szamuely – Corps de classes des schémas arithmétiques, Astérisque 332 (2010), p. 257–286, Séminaire Bourbaki, volume 2008/2009, exposé no 1006. [43] J. Tate – W C-groups over p-adic fields, Secrétariat mathématique, Soc. Math. France, 1958, Séminaire Bourbaki, volume 1957/1958, exposé no 156. Reproduit dans Séminaire Bourbaki, Vol. 4, Exp. No. 156, 265–277, Soc. Math. France, Paris, 1995.

Jean-Louis COLLIOT-THÉLÈNE CNRS Université Paris XI UMR 8628 du CNRS Département de Mathématiques Bâtiment 425 F–91405 ORSAY Cédex E-mail : [email protected]

ASTÉRISQUE 339

339

ASTÉRISQUE 2011

SÉMINAIRE BOURBAKI VOLUME 2009/2010 EXPOSÉS 1012-1026 (1013) p-adic families of modular forms Matthew EMERTON

SOCIÉTÉ MATHÉMATIQUE DE FRANCE Publié avec le concours du CENTRE NATIONAL DE LA RECHERCHE SCIENTIFIQUE

Séminaire BOURBAKI 62e année, 2009-2010, no 1013, p. 31 à 61

Novembre 2009

p-ADIC FAMILIES OF MODULAR FORMS [after Hida, Coleman, and Mazur] by Matthew EMERTON

INTRODUCTION The theory of p-adic families of modular forms grew out of two highly related traditions in the arithmetic theory of modular forms: the theory of congruences of modular forms (which dates back to work of Ramanujan) and the (more recent) theory of Galois representations attached to modular forms. The first example of a p-adic family of modular forms was the Eisenstein family, considered by Serre in [37]. This is a family of q-expansions, parametrized by the weight k, whose coefficients are p-adically continuous functions of k. Serre’s immediate goal in studying this family was to obtain an understanding of the possible congruences between the q-expansion coefficients of modular forms in different weights, especially of the constant terms, since such congruences lead to congruences between special values of ζ-functions. The papers [23, 22] led to a decisive shift in the theory, placing it at the centre of the arithmetic theory of modular forms. In these papers, Hida constructed p-adic families of cuspforms, varying continuously with the weight k, which were also simultaneous eigenforms for the Hecke operators. Thus, in light of the known construction of Galois representations attached to Hecke eigenforms, one found that associated to these p-adic families of cuspidal eigenforms there were corresponding p-adic families of p-adic Galois representations. The existence of such families led Mazur to develop his general theory of deformations of Galois representations [31], which in turn inspired further developments [45, 43]. Hida’s constructions had a certain limitation: if f is a Hecke eigenform of weight k ≥ 1 and level N prime to p, then f appears in a Hida family if and only if (at least) one of the roots of the pth Hecke polynomial of f is of slope zero (i.e. a p-adic unit). This restriction was removed by the work of Coleman and Mazur [10], who constructed p-adic analytic (more precisely, rigid analytic) curves of eigenforms containing any

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

32

M. EMERTON

such form f , whether or not its pth Hecke polynomial admits a unit root; these are the so-called eigencurves. The eigencurves are fundamentally analytic objects. One can also ask whether there is an algebraic family (or more precisely, a scheme) that parametrizes all the f as above, regardless of the slopes of the roots of the pth Hecke polynomial. Indeed, there is such an object; all the eigenforms f (of arbitrary weight but some fixed level N ) are parametrized by the Zp points of Spec T(N ), where T(N ) is the p-adic Hecke algebra of level N . These points are no longer parametrized by weight; indeed, Spec T(N ) is (at least conjecturally) of relative dimension three over Spec Zp . It is conjectured that every continuous, two-dimensional, semi-simple odd p-adic Galois representation of GQ that is unramified outside finitely many primes corresponds to a point of Spec T(N ) for some appropriate value of N . This is one of the main motivations for the study of the families Spec T(N ), and the related p-adic families of eigenforms constructed by Hida and Coleman–Mazur. In Section 1 of this exposé we recall the basic theory of modular forms, Hecke operators, and the Galois representations associated to Hecke eigenforms. In Section 2, we outline the definitions and basic results and conjectures regarding the p-adic Hecke algebras T(N ), and the families of Hida and Coleman–Mazur. We focus more on systems of Hecke eigenvalues attached to eigenforms, rather than on the eigenforms themselves. This is in keeping with our focus on the relationship with Galois representations (although it takes us somewhat far in spirit from the concrete viewpoint of [37]). Acknowledgments I would like to thank J-P. Serre for his helpful comments on an earlier version of this article. 0.1. Notation As usual Q, R, and C denote the fields of rational, real, and complex numbers, and Z denotes the ring of integers. For any prime p, we let Zp denote the ring of p-adic integers, and Qp denote the field of p-adic numbers. We let Q denote the algebraic closure of Q in C, and let Z denote the integral closure of Z in Q. For each prime p, we fix an algebraic closure Qp of Qp , and let Zp denote the integral closure of Zp in Qp . We also fix an embedding ıp : Q ,→ Qp . This restricts to an embedding Z ,→ Zp . We write Fp to denote the residue field of Zp . It is an algebraic closure of the field Fp of p elements. We let ordp : Qp → Z ∪ {∞} denote the p-adic valuation, normalized so that ordp (p) = 1. If x ∈ Qp , then ordp (x) is also called the slope of x. (Thus x has finite slope if and only if x 6= 0, while x has slope × zero if and only if x ∈ Zp .)

ASTÉRISQUE 339

(1013)

p-ADIC FAMILIES OF MODULAR FORMS

33

1. MODULAR FORMS, HECKE ALGEBRAS, AND GALOIS REPRESENTATIONS 1.1. Modular forms Let 

H = τ ∈ C | =(τ ) > 0



denote the complex upper half-plane. The group SL2 (Z) acts on H in the usual way: ! a b aτ + b τ= . cτ + d c d Let O( H ) denote the space of holomorphic functions on H . If k is an integer, then we define the weight k-action of SL2 (Z) on O( H ) as follows: (f |k γ)(τ ) := (cτ + d)−k f (γτ ),  for f ∈ O( H ) and γ = ac db ∈ SL2 (Z); as the notation indicates, this is a right action. If N ≥ 1, define  Γ1 (N ) := γ ∈ SL2 (Z) | γ ≡ ( 10 ∗1 ) mod N . Definition 1.1. — A modular form (resp. cuspform) of weight k and level N is a holomorphic function f ∈ O( H ) that is invariant under the weight k-action of Γ1 (N ), and for which (1)

lim (f |k γ)(iy)

y→∞

exists and is finite (resp. vanishes) for each γ ∈ SL2 (Z). We let M k (N ) (resp. S k (N )) denote the space of modular forms (resp. cuspforms)of weight k and level N . Remark 1.2. — If f ∈ O(H) is invariant under the weight k-action of Γ1 (N ), then, in order to check if f is a modular form or a cuspform, it suffices to study the limit (1) for finitely many γ ∈ SL2 (Z) (namely, for a set of coset representatives for Γ1 (N )\SL2 (Z)). Remark 1.3. — If f is a modular form of weight k and level N , then, applying the invariance property of f to the matrix ( 10 11 ) ∈ Γ1 (N ), one finds that f (τ +1) = f (τ ). We may thus expand the function f (τ ) as a Fourier series f (τ ) :=

∞ X

cn (f )q n ,

n=−∞

where q := exp(2πiτ ). Condition (1), with γ = 1, then shows that cn (f ) = 0 for n < 0 (resp. for n ≤ 0 if f is a cuspform). We refer to this Fourier series as the q-expansion of f .

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

34

M. EMERTON

Clearly M k (N ) and S k (N ) are vector subspaces of O( H ). In fact they are also finite dimensional. (See [39] for a discussion of this and other basic facts concerning modular forms.) Example 1.4. — If k < 0, then M k (N ) = 0. When k = 0, the space M 0 (N ) consists simply of the constant functions on H (and so S 0 (N ) = 0). To avoid these trivial cases, we will typically assume that k ≥ 1 in all that follows. As k increases, the dimensions of both M k (N ) and S k (N ) grow essentially linearly in k (with the exception that M k (N ) = 0 if N = 1 or 2 and k is odd). Example 1.5. — The simplest examples of modular forms of positive weight are the Eisenstein series Ek ∈ M k (1). These are defined for even k ≥ 4. (It is easily shown that M k 1) vanishes if k is odd or 0 < k < 4.) The q-expansion of Ek is given by the following formula: ã ∞ Å −Bk X X k−1 n Ek (τ ) = + d q , 2k n=1 d|n

where Bk is the kth Bernoulli number. There is a direct sum decomposition

M k (1) = CEk ⊕ S k (1). More generally, for any N , we may decompose M k (N ) into the direct sum of a space of Eisenstein series (typically of dimension greater than one when N > 1) and the space of cuspforms. (See Example 1.18 below.) 1.2. Hecke operators Fix integers k ≥ 1 and N ≥ 1. Write  Γ0 (N ) := γ ∈ SL2 (Z) | γ ≡ ( ∗0 ∗∗ ) mod N . Note that Γ0 (N ) contains Γ1 (N ) as a normal subgroup, and that the map ! a b 7→ d mod N c d induces an isomorphism (2)



Γ0 (N )/Γ1 (N ) −→ (Z/N Z)× .

A simple computation, using the normality of Γ1 (N ) in Γ0 (N ), shows that the weight k-action of Γ0 (N ) preserves M k (N ) and S k (N ). When restricted to these spaces, this action obviously factors through the quotient Γ0 (N )/Γ1 (N ), and hence, via the isomorphism (2), we obtain an action of the group (Z/N Z)× on M k (N ) and S k (N ). If d ∈ (Z/N Z)× , then we denote the corresponding automorphism of Mk (N ) by hdi. (These operators are sometimes referred to as the diamond operators.)

ASTÉRISQUE 339

(1013)

p-ADIC FAMILIES OF MODULAR FORMS

35

Remark 1.6. — We note a simple but important identity for the action of the diamond operator h−1i, namely h−1if = (−1)k f,

(3)

for any f ∈ M k (N ). This is easily verified by considering the weight k-action of the  0 matrix −1 0 −1 ∈ Γ0 (N ) on f . Definition 1.7. — If ` is a prime not dividing N , then we define the automorphism S` of M k (N ) via the formula S` = h`i`k−2 . Since the diamond operators leave S k (N ) invariant, so do the operators S` . In fact, although it is traditional to single out the operators S` as defined above, it is the operators `S` = h`i`k−1 that will be more important for us, as we see already in the next definition. Definition 1.8. — If ` is a prime not dividing N , then we define the endomorphism T` of M k (N ) via the formula (4)

(T` f )(τ ) =

∞ X n=0

cn` (f )q n +

∞ X

`cn (S` f )q n` .

n=0

Remark 1.9. — It is not immediately obvious that T` , which we have defined simply by its effect on q-expansions, actually preserves the space M k (N ). In fact T` can be thought of as a certain double coset operator, corresponding to the double coset GL2 (Z` ) ( 0` 01 ) GL2 (Z` ) (see e.g. [39, Ch. 3]). From this point of view, it is easy to verify that it preserves the space M k (N ), as well as the subspace S k (N ) of cuspforms. The operator S` also has a double coset interpretation; it corresponds to the double  coset GL2 (Z` ) 0` 0` GL2 (Z` ). This is one reason to consider S` as a primary object, rather than the diamond operator h`i. Definition 1.10. — We let Tk (N ), or simply Tk when the level N is understood,  denote the Z-subalgebra of End M k (N ) generated by the operators `S` and T` as ` ranges over the primes not dividing N . The algebra Tk (N ) is called the Hecke algebra (for the given weight k and level N ). Remark 1.11. — Following [39, Ch. 3], one can extend Definition 1.8 and define Hecke operators Tm acting on M k (N ) for any positive integer m prime to N . The algebra Tk (N ) defined above then coincides with the Z-algebra of endomorphisms of M k (N ) generated by the collection of these operators Tm . The following result encapsulates the basic properties of the algebra Tk , and of its action on M k (N ).

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

36

M. EMERTON

Proposition 1.12. — The algebra Tk is commutative, reduced, and free of finite rank over Z. Furthermore, the tensor product C ⊗Z Tk acts faithfully on M k (N ). Remark 1.13. — The commutativity part of the statement is not difficult to verify; for example, it is easily checked using the description of the Hecke operators in terms of double cosets. The additional properties of Tk are then equivalent to the following statements about the eigenspaces and eigenvalues of the Hecke operators: 1. Every eigenvalue of any of the Hecke operators is an algebraic integer. (Here one sees the importance, when k = 1, of taking `S` rather than S` in the definition of Tk , so as to avoid introducing denominators.) 2. The systems of simultaneous eigenvalues for the action of the Hecke operators on M k (N ) (which are collections of algebraic integers, by 1) are closed under the action of Gal(Q/Q). 3. The space M k (N ) decomposes as a direct sum of simultaneous eigenspaces for the Hecke operators. Definition 1.14. — We say that f ∈ M k (N ) is a Hecke eigenform if it is a simultaneous eigenvector for the Hecke operators `S` and T` (where ` ranges over all primes not dividing N ), or equivalently, if there is a ring homomorphism λ : Tk → C such that T f = λ(T )f for all T ∈ Tk . We refer to a homomorphism λ : Tk → C as a system of Hecke eigenvalues. (Any such λ is the system of Hecke eigenvalues attached to some Hecke eigenform. Also, according to the preceding remark, any such λ factors through the ring of algebraic integers Z in Q.) If λ is a system of Hecke eigenvalues, then we write M k (N )[λ] to denote the corresponding subspace of Hecke eigenforms. As already noted in the preceding remark, the space M k (N ) admits the direct sum decomposition M M k (N ) = Mk (N )[λ], λ

where the direct sum is taken over all systems of Hecke eigenvalues. Remark 1.15. — The formula (4) shows that if λ is a system of Hecke eigenvalues, then the q-expansion of a Hecke eigenform f ∈ M k (N )[λ], and hence the eigenform f itself, is to a large extent determined by the system of Hecke eigenvalues λ. For example, if N = 1, then the group of diamond operators is trivial, and so `S` f = `k−1 f . Formula (4) then shows that λ(T` )cn (f ) = cn` (f ) + cn/` (f )`k−1 for every prime number ` (where we set cn/` = 0 if ` - n). Thus the Fourier coefficients cn (f ) (n ≥ 1) are determined recursively by the single coefficient c1 (f ), and so f is

ASTÉRISQUE 339

(1013)

p-ADIC FAMILIES OF MODULAR FORMS

37

determined up to a scalar by its associated system of Hecke eigenvalues. In particular, the λ-eigenspace in M k (1) is one-dimensional.(1) If N > 1, then we find that f is determined by λ, together with the Fourier coefficients cm (f ), for those positive integers m divisible only by primes dividing N . Thus f need not be uniquely determined (up to a scalar) by λ, and the λ-eigenspace in M k (N ) can be of dimension greater than one. However, the structure of this eigenspace is well-understood, either using the theory of so-called oldforms and newforms as in [2], or in terms of the action of GL2 (A) on the space of modular forms of weight k [25, 6]. We do not recall the details here, since they will not be important for us. As we will explain in the following subsection, our attention will be focussed on the systems of eigenvalues λ themselves, rather than on the associated Hecke eigenforms. Remark 1.16. — Given a system of Hecke eigenvalues λ appearing in M k (N ), it fol× lows from the definition of the operators S` that there is a Q -valued character ε of (Z/N Z)× such that λ(`S` ) = ε(`)`k−1 . Thus we may recover the value of the weight k from the system of eigenvalues λ. Indeed, if ` is any prime not dividing N , then k = (log` |λ(`S` )|) + 1. Example 1.17. — If k ≥ 4 is even, then the Eisenstein series Ek ∈ M k (1) is a Hecke eigenform. The corresponding system of Hecke eigenvalues λ is given by λ(`S` ) = `k−1 ,

λ(T` ) = 1 + `k−1 .

(Here ` is an arbitrary prime, since we are in the case N = 1.) Example 1.18. — Let ψ1 : (Z/M1 Z)× → C× and ψ2 : (Z/M2 Z)× → C× be characters, and let k ≥ 1 (unless M1 = M2 = 1, in which case we require that k ≥ 4) be chosen so that ψ1 (−1)ψ2 (−1)(−1)k = 1. Then the following system of Hecke eigenvalues, which we denote by λψ1 ,ψ2 ,k , appears in M k (M1 M2 ): λψ1 ,ψ2 ,k (`S` ) = ψ1 (`)ψ2 (`)`k−1 ,

λψ1 ,ψ2 ,k (T` ) = ψ1 (`) + ψ2 (`)`k−1 .

In the case when M1 = M2 = 1, we obtain the systems of Hecke eigenvalues associated to the Eisenstein series Ek , as considered in the preceding example. In general, we refer to such a system of Hecke eigenvalues as an Eisenstein system of eigenvalues. If we write M Ek (N ) = M k (N )[λ], λ Eisenstein

(1)

A slight amount of caution is required here, because c0 (f ) is not directly determined by the cn (f ) for n ≥ 1. However, since k ≥ 1, then in fact c0 (f ) is so determined, as one easily sees, since a constant function cannot be modular of weight k > 0. As Serre notes [37, Rem. 2), p. 221], one can directly determine c0 (f ) from the cn (f ) for n ≥ 1 as follows: −c0 (f ) is the P value at s = 0 of the ∞ meromorphic function defined by analytic continuation of the Dirichlet series c (f )n−s . n=1 n

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

38

M. EMERTON

where the sum ranges over all Eisenstein systems of Hecke eigenvalues for which M1 M2 = N, then we refer to modular forms f ∈ Ek (N ) as Eisenstein series. There is a direct sum decomposition

M k (N ) = Ek (N ) ⊕ S k (N ). Unlike the Eisenstein systems of eigenvalues considered in Example 1.18, the systems of eigenvalues appearing in the spaces of cuspforms do not admit an elementary description. As we will see in the following subsection, they correspond to certain Galois representations. Example 1.19. — We close this subsection with a careful presentation of the preceding concepts in the case N = 1 and k = 12. In this case

M 12 (1) = E12 (1) ⊕ S 12 (1), where E12 (1) is one-dimensional, spanned by E12

∞ X X 691 691 = + ( d11 )q n = + q + 2049q 2 + 177148q 3 + · · · , 32760 n=1 32760 d|n

and S 12 (1) is also one-dimensional, spanned by Ramanujan’s famous cuspform ∆(τ ) = q

∞ Y

(1 − q n )24 =

n=1

∞ X

τ (n)q n = q − 24q 2 + 252q 3 + · · · .

n=1

(Here τ (n) = cn (∆) is by definition the nth Fourier coefficient of ∆.) Each of these modular forms is a Hecke eigenform, and correspondingly T12 admits two systems of Hecke eigenvalues. If we write λ1 (resp. λ2 ) to denote the system of Hecke eigenvalues attached to E12 (resp. ∆), then (5)

λ1 × λ2 : T12 ,→ Z × Z.

Note that since each of these eigenforms has been normalized so that c1 = 1, we may read off the corresponding systems of Hecke eigenvalues from the Fourier coefficients, as in Remark 1.15. The product λ1 ×λ2 provides an embedding λ1 ×λ2 : T12 ,→ Z×Z. It was first observed by Ramanujan that this embedding is not an isomorphism. Indeed, Ramanujan showed that X τ (n) ≡ d11 mod 691 d|n

for every natural number n, or equivalently, λ1 (T` ) ≡ λ2 (T` ) mod 691

ASTÉRISQUE 339

(1013)

p-ADIC FAMILIES OF MODULAR FORMS

39

for each prime `. On the other hand, it is easily verified (just by considering the cases when ` = 2 and 3) that no such congruence holds modulo any higher power of 691, nor modulo any other prime. Thus (5) induces an isomorphism ∼

T12 −→ {(u, v) ∈ Z × Z | u ≡ v mod 691}. ∼

If p is a prime and p 6= 691, then Zp ⊗Z T12 −→ Zp × Zp ; this reflects the fact that the distinct systems of eigenvalues λ1 and λ2 remain distinct when reduced modulo p. On the other hand, the tensor product Z691 ⊗Z T12 does not factor as a product in any non-trivial way; rather, it is a local ring, reflecting the congruence of λ1 and λ2 modulo 691. 1.3. Galois representations As in the preceding section, fix integers k ≥ 1 and N ≥ 1. From a certain point of view, it is the systems of Hecke eigenvalues appearing in M k (N ) that are of the greatest interest, rather than the modular forms, or even the Hecke eigenforms, themselves. This is because they give rise to Galois representations, as we now recall. Choose a prime number p. If λ is a system of Hecke eigenvalues appearing in

M k (N ), then since λ takes values in the ring Z of algebraic integers, we may compose it with our chosen embedding ıp : Q ,→ Qp , and so regard λ as taking values in Zp . For the remainder of this subsection, we regard all systems of Hecke eigenvalues as being Zp -valued. If λ : Tk → Zp is a system of Hecke eigenvalues, then we let λ : Tk → Fp be the homomorphism obtained by composing λ with the map Zp → Fp given by reducing modulo the maximal ideal of Zp . Let Σ denote the (finite) set of primes dividing N p, let QΣ denote the maximal algebraic extension of Q in Q that is unramified outside of the primes in Σ, and write GQ,Σ := Gal(QΣ /Q). Recall that if ` is a prime not in Σ, then attached to ` is a Frobenius element Frob` ∈ GQ,Σ , well-defined up to conjugacy, with the property that there is a prime ideal l lying over ` in the ring of algebraic integers in QΣ that is preserved by Frob` , such that for any algebraic integer x ∈ QΣ , Frob` (x) ≡ x` mod l. The Čebotarev density theorem furthermore implies that the union of these conjugacy classes is dense in GQ,Σ . If M is any integer divisible only by primes dividing N p, and if ζM denotes a primitive M th root of unity, then ζM ∈ QΣ , and so there is a group homomorphism χM : GQ,Σ → (Z/M Z)× describing the action of the elements of GQ,Σ on ζM , namely, for any σ ∈ GQ,Σ , we have χ

σ(ζM ) = ζMM

(σ)

.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

40

M. EMERTON

We refer to χM as the mod M cyclotomic character. It can also be characterized by the formula χM (Frob` ) ≡ ` mod M, for any prime ` not dividing N p. Also, if c ∈ GQ,Σ denotes complex conjugation, then χM (c) = −1. We also define the p-adic cyclotomic character χ : GQ,Σ → Z× p to be the projective n limit over n of the mod p -cyclotomic characters χpn . Again, the character χ is characterized by the formula χ(Frob` ) = ` for any ` not dividing N p, and we also have that χ(c) = −1. The various cyclotomic characters give the basic examples of characters (i.e. onedimensional representations) of the group GQ,Σ . The following theorem shows that Hecke eigenforms are a source of two-dimensional representations of this group. Theorem 1.20. — If λ : Tk (N ) → Zp is a system of Hecke eigenvalues appearing in M k (N ), then there is a continuous, semi-simple representation ρλ : GQ,Σ → GL2 (Qp ), uniquely determined (up to equivalence) by the condition that for each prime ` - N p, the matrix ρλ (Frob` ) has characteristic polynomial equal to X 2 − λ(T` )X + λ(`S` ). Remarks on the proof. — The uniqueness statement of the theorem is easily proved. Indeed, if ρ1 and ρ2 are two representations both satisfying the conditions of the theorem, then by assumption their characteristic polynomials agree on the set of elements Frob` , which by Čebotarev density are dense in GQ,Σ . Since they are continuous, their characteristic polynomials then agree on all elements of GQ,Σ . It follows that ρ1 and ρ2 are equivalent, as claimed, since a semi-simple finite-dimensional representation of a group is uniquely determined, up to equivalence, by its characteristic polynomials. In the case when λ is an Eisenstein system of Hecke eigenvalues, the existence of ρλ is also easily proved; see Example 1.24 below. On the other hand, if λ is a system of eigenvalues attached to a cuspform, then the construction of ρλ is much less trivial. Its construction is due to Eichler, Shimura, and Igusa [13, 38, 24] (in the case k = 2), to Deligne [11] (for k > 2), and to Deligne and Serre [12] (for k = 1). It is useful to give a name to the characteristic polynomials appearing in Theorem 1.20. Definition 1.21. — If λ : Tk (N ) → C is a system of Hecke eigenvalues, then for each prime ` - N , we define the `th Hecke polynomial of λ to be the polynomial X 2 − λ(T` )X + λ(`S` ).

ASTÉRISQUE 339

(1013)

p-ADIC FAMILIES OF MODULAR FORMS

41

Remark 1.22. — Since GQ,Σ is profinite, the representation ρλ may be conjugated so as to take values in GL2 (Zp ), and we let ρ◦λ denote such a GL2 (Zp )-valued representation underlying ρλ . The GL2 (Zp )-valued representation ρ◦λ is not always uniquely determined up to equivalence by λ. However, if we let ρλ denote the semi-simplification of the representation GQ,Σ → GL2 (Fp ) obtained by reducing ρ◦λ modulo the maximal ideal of Zp , then ρλ is uniquely determined, up to equivalence, by λ, and in fact, even by λ (as the notation suggests). Indeed, ρλ is uniquely characterized, up to equivalence, by the condition that for each prime ` - N p, the matrix ρλ (Frob` ) has characteristic polynomial equal to X 2 −λ(T` )X +λ(`S` ). (The proof of the uniqueness is identical to that given in the proof of Theorem 1.20.) ×

Remark 1.23. — As in Remark 1.16, write λ(`S` ) = ε(`)`k−1 for some Q -valued character ε of (Z/N Z)× . We may compose ε with the mod N cyclotomic character × × χN to obtain a Q -valued character of GQ,Σ , which we regard as being Qp -valued via our chosen embedding ıp : Q ,→ Qp . It then follows from the condition on the determinant of ρ(Frob` ) in the statement of Theorem 1.20, together with Čebotarev density and the given relationship between λ(`S` ) and ε(`), that det ρλ := (ε ◦ χN )χk−1 , where as above χ denotes the p-adic cyclotomic character. In particular, if c ∈ GQ,Σ denotes complex conjugation, then one computes that det ρλ (c) = ε(−1)(−1)k−1 = −1 (the last equality following from (3)). One says that ρλ is odd. Similarly, the representation ρλ is odd. Example 1.24. — If λψ1 ,ψ2 ,k is an Eisenstein system of Hecke eigenvalues attached to characters ψi : (Z/Mi Z)× → C× and the weight k, as in Example 1.18, then it is easy to write down a corresponding Galois representation ρλψ1 ,ψ2 ,k satisfying the conditions of Theorem 1.20; namely, we can take ρλψ1 ,ψ2 ,k = (ψ1 ◦ χM1 ) ⊕ (ψ2 ◦ χM2 )χk−1 . On the other hand, if λ arises from a cuspform, then ρλ does not admit a description in terms of characters. Indeed, one has the following result [35, Thm. 2.3]. Proposition 1.25. — If the system of Hecke eigenvalues λ is attached to a cuspform, then the representation ρλ associated to λ by Theorem 1.20 is irreducible. (p)

For any prime p, write Tk to denote the subalgebra of Tk generated by the elements `S` and T` for ` not dividing N p (i.e. we omit the Hecke operators at p). If λ : Tk → Zp is a system of Hecke eigenvalues, we write λ(p) to denote the restriction of (p) λ to Tk , and refer to λ(p) as the p-deprived system of Hecke eigenvalues associated to

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

42

M. EMERTON

(p)

(p)

λ. Similarly, we let λ : Tk → Fp denote the restriction of λ. The conditions on the Galois representation ρλ given in Theorem 1.20 evidently depend only on λ(p) , and in fact we can use the existence of the Galois representations attached to λ to show that λ(p) already determines λ. Indeed, we have the following more general result. Proposition 1.26. — If λ1 and λ2 are two systems of Hecke eigenvalues such that λ1 (T` ) = λ2 (T` ) for all but finitely many primes ` not dividing N, then λ1 and λ2 coincide. Proof. — This is proved by the same argument used to establish the uniqueness claim of Theorem 1.20. Let q be some fixed prime not dividing N , and choose p to be distinct from q. Let ρλ1 and ρλ2 denote the Galois representations associated to λ1 and λ2 as in Theorem 1.20, regarded as representations over Qp . Then ρλ1 and ρλ2 have the same traces on the elements Frob` , for all but finitely many `. Čebotarev density implies that the set of elements Frob` (where ` ranges over all but finitely many primes not dividing N p) is dense in GQ,Σ , and so, since ρλ1 and ρλ2 are continuous, we see that their traces coincide. Thus they have isomorphic semi-simplifications (since we are working over the field Qp of characteristic zero), and so their characteristic polynomials coincide on any element of GQ,Σ . Applying this to Frobq , we find that λ1 (Sq ) = λ2 (Sq ) and that λ1 (Tq ) = λ2 (Tq ). Since q was an arbitrary prime not dividing N , the proposition follows. The preceding proposition has the following technical corollary. (p)

Corollary 1.27. — The ring Tk

has finite index in Tk . (p)



Proof. — Since Tk is finite over Z, it suffices to show that C ⊗Z Tk −→ C ⊗Z Tk . Equivalently, we must show that distinct systems of Hecke eigenvalues remain distinct after omitting the eigenvalues corresponding to the Hecke operators at p. This follows from the proposition. Remark 1.28. — The finite index of Corollary 1.27 can be greater than 1. For exam(2) ple, if N = 23, k = 2, and p = 2, then the index of T2 in T2 is equal to 2. (More √ √ (2) precisely, T2 ∼ = Z[ 5].) = Z[(1 + 5)/2], while T2 ∼ Remark 1.29. — Corollary 1.27 (or better, its proof) shows that λ 7→ λ(p) induces a bijection between the set of homomorphisms Tk → C and the set of homomorphisms (p) Tk → C, and hence between the set of homomorphisms Tk → Zp and the set of (p) homomorphisms Tk → Zp .

ASTÉRISQUE 339

(1013)

p-ADIC FAMILIES OF MODULAR FORMS

43

2. p-ADIC FAMILIES OF SYSTEMS OF HECKE EIGENVALUES 2.1. The p-adic Hecke algebra Let N be a positive integer, and fix a prime p not dividing N . If k ≥ 1 is a positive integer, then for each prime ` not dividing N we define the operators S` and T` on L the direct sum ki=1 M i (N ) in the obvious way: S` and T` act on each summand via the Hecke operator with the same name. (p)

(p)

Definition 2.1. — We let T≤k (N ), or simply T≤k if the level N is understood, Lk denote the Z-algebra of endomorphisms of i=1 M i (N ) generated by the operators `S` and T` , as ` ranges over all primes not dividing N p. Since each operator S` and T` is determined by its action on each of the direct summands, there is a natural injection (p)

T≤k ,→

(6)

k Y

Ti .

i=1

Remark 2.2. — We could consider the analogous algebra in which we included the operators pSp and Tp . However, for our later purposes, it is important to omit these operators from the algebra under consideration. Proposition 2.3. — The image of (6) has finite index in

Qk

i=0

Ti .

Proof. — Given that the source and target of (6) are both finite Z-algebras, it suffices to show that (6) becomes an isomorphism after tensoring with C over Z. This follows from the fact that the p-deprived systems of eigenvalues appearing in M k (N ) are distinct for different values of k, by Remarks 1.16 and 1.29. Example 2.4. — Take N = 1, p = 2, and k = 6. The spaces Mi (1) for 1 ≤ i ≤ 6 vanish unless i = 4 or 6, in which case they are one-dimensional, spanned by E4 and ∼ ∼ E6 respectively. Thus T4 −→ Z and T6 −→ Z, and so (6) becomes in this case an embedding (2)

T≤6 ,→ Z × Z.

(7) Now E4 =

ã ∞ ÅX X 1 + d3 q n , 240 n=1 d|n

while E6 =

ã ∞ Å −1 X X 5 n + d q . 504 n=1 d|n

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

44

M. EMERTON

One immediately checks that 1 + `3 ≡ 1 + `5 mod 12, for all ` 6= 2. Furthermore, no analogous congruence holds modulo any larger modulus, and thus the embedding (7) induces an isomorphism (2)



T≤6 −→ {(u, v) ∈ Z × Z | u ≡ v mod 12}. A similar calculation shows that (3)



T≤6 −→ {(u, v) ∈ Z × Z | u ≡ v mod 6}. These examples exhibit congruences similar to those discussed in Example 1.19, but involving congruences between systems of Hecke eigenvalues in different weights. If k 0 ≥ k, then k M

0

M i (N ) ⊂

i=0

k M

M i (N ),

i=0

and so restriction induces a surjection (p)

(p)

T≤k0 → T≤k . Tensoring this with Zp over Z, we obtain a surjection (8)

(p)

(p)

Zp ⊗Z T≤k0 → Zp ⊗Z T≤k .

Definition 2.5. — The p-adic Hecke algebra T(N ), or simply T if the level N is understood, is defined to be the projective limit (9)

(p)

Zp ⊗Z T≤k , T := lim ←− k

where the transition maps are the maps (8). Remark 2.6. — Note that since any prime ` 6= p is invertible in Zp , the operator (p) S` = `−1 (`S` ) lies in each of the algebras Zp ⊗Z Tk , for each ` - N p, and so we may regard each of these algebras as being generated by the elements S` and T` (` - N p), just as well as by `S` . Also, since the transition maps (8) take the elements S` and T` in the source to the elements S` and T` in the target, these elements give rise to well-defined elements S` and T` in the projective limit T, for any prime ` - N p. From the various embeddings (6), we obtain an embedding Y T ,→ Zp ⊗Z Tk . k≥1

The target of this embedding is a countable product of non-zero rings; in particular, it is not Noetherian. On the other hand, we have the following result regarding the source.

ASTÉRISQUE 339

(1013)

p-ADIC FAMILIES OF MODULAR FORMS

45

Theorem 2.7. — The ring T is a p-adically complete, Noetherian Zp -algebra, and is in fact the product of finitely many complete Noetherian local Zp -algebras. (p)

Remarks on the proof. — Each Zp ⊗Z Tk is a finite Zp -algebra, and so is a product of finitely many complete local finite Zp -algebras. It follows that T is p-adically complete, and is a product of a countable collection of complete local Zp -algebras. The fact that this product involves only finitely many local algebras is not formal; it is equivalent to a statement about Fp -valued systems of Hecke eigenvalues that is the subject of Proposition 2.8 below. The fact that these local factors are Noetherian is also not formal; it is proved via a consideration of relation between the ring T and Galois representations, as discussed in the following subsection. (More precisely, each local component of T is canonically the quotient of a certain Galois pseudodeformation ring, and hence is Noetherian; see [30, §1.4] for a discussion of the latter, and in particular Lemma 1.4.2 for a proof of the Noetherianness of pseudo-deformation rings.(2)) Informally speaking, this theorem can be thought of as showing that the phenomenon exhibited in Example 2.4 is typical: as k grows, the power of p dividing the index of the image of (6) in its target grows progressively larger, reflecting the existence of many congruences modulo powers of p between systems of eigenvalues appearing in various weights. We present one concrete manifestation of this abundance of congruences in the following proposition (due to Jochnowitz [26], generalizing an argument of Serre in the case N = 1), which is an important ingredient in the proof of Theorem 2.7. Indeed, its statement is a straightforward reformulation of the claim that T has only finitely many maximal ideals. In order to state the proposition, we introduce additional notation. (p) Suppose given a p-deprived system of Hecke eigenvalues λ(p) : Tk → Zp . We then write λ

(p)

(p)

: Tk → Fp to denote the reduction of λ(p) modulo the maximal ideal of Zp .

Proposition 2.8. — As λ(p) ranges over all p-deprived systems of eigenvalues of all weights k ≥ 0, there are only finitely many possibilities for the collection of eigenvalues  (p) (p) λ (`S` ), λ (T` ) `-N p . One has the following precise conjecture regarding the Krull dimension of the ring T. Conjecture 2.9. — The ring T is equidimensional of Krull dimension 4, i.e. each irreducible component of Spec T is of dimension 4. (2)

Technically, the results of [30, §1.4] only apply when p is odd; however, with the appropriate modifications, they should also apply in the case when p = 2.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

46

M. EMERTON

Since T is a torsion free and p-adically complete Zp -algebra, this is equivalent to Spec T having relative dimension 3 over Spec Zp . This conjecture is motivated by the known and conjectured relations between the ring T and Galois representations. (See Remark 2.14 below.) We will prove in Corollary 2.28 below that each irreducible component of Spec T has Krull dimension at least 4. 2.2. Galois representations again As in the preceding subsections, we regard all systems of Hecke eigenvalues as taking values in Zp . Definition 2.10. — A p-adic system of Hecke eigenvalues is a homomorphism of Zp -algebras ξ : T → Zp . (p)

Suppose that λ(p) : Tk → Zp is a p-deprived system of Hecke eigenvalues. Since the target of λ(p) is a Zp -algebra, this homomorphism extends to a homomorphism (p) λ(p) : Zp ⊗Z Tk → Zp . Composing this homomorphism with the natural surjection (p) T → Tk , we obtain a homomorphism ξ : T → Zp . We refer to p-adic systems of Hecke eigenvalues arising in this way as classical. Theorem 2.11. — If ξ : T → Zp is any p-adic system of Hecke eigenvalues, then there is a continuous, semi-simple representation ρξ : GQ,Σ → GL2 (Qp ), uniquely determined (up to equivalence) by the condition that for each prime ` - N p, the matrix ρξ (Frob` ) has characteristic polynomial equal to X 2 − ξ(T` )X + ξ(`S` ). Sketch of proof. — The uniqueness proof is identical to that given in the proof of Theorem 1.20. As for existence, if ξ is a classical system, arising from the p-deprived system of eigenvalues λ(p) , then we can clearly set ρξ := ρλ . To construct ρξ in general, one uses the fact that the classical ξ are dense in the set of all ξ (in a suitable sense), and then constructs ρξ by an interpolation argument. Just as in the case of Theorem 1.20, one shows that if ξ is a p-adic system of Hecke eigenvalues, then ρξ is odd, and so we see that T parametrizes a family of odd two-dimensional p-adic Galois representations. Furthermore, one has the following fundamental conjecture to the effect that all odd two-dimensional Galois representations should be of this form. (See e.g. the conjecture on p. 108 of [20].) Conjecture 2.12. — If Σ is any finite set of primes containing p, and if ρ : GQ,Σ → GL2 (Qp ) is continuous, semi-simple, and odd, then ρ = ρξ for some p-adic system of Hecke eigenvalues of some level N divisible only by primes in Σ distinct from p.

ASTÉRISQUE 339

(1013)

p-ADIC FAMILIES OF MODULAR FORMS

47

Remark 2.13. — In fact, one expects to be able to take N to be the tame (i.e. primeto-p) Artin conductor of ρ. Remark 2.14. — One can use techniques from Galois cohomology to show that if ρ is an odd, irreducible, continuous two-dimensional p-adic representation of GQ,Σ (for some fixed finite set of primes Σ), then the expected dimension of a neighbourhood of ρ in the space of all such representations is three. (See Corollary 3 on p. 405 of [31]. This reference treats the case of mod p Galois representations, but is easily adapted to the context of p-adic Galois representations.) Taken together with Conjecture 2.12, this motivates Conjecture 2.9. Building on ideas of Gouvêa and Mazur [21] (in particular, the infinite fern, as considered in Subsection 2.5 below), together with the techniques of Wiles [45] and Taylor–Wiles [43], Böckle [3] has proved a strong result in the direction of Conjectures 2.12 and 2.9. Since the statement of his result is a little technical, we do not recall it here. However, by appealing to a result of Kisin, one can improve the part of Böckle’s theorem that pertains to Conjecture 2.9, as follows. Before stating it, we recall that Q(ζp3 ) contains a unique quadratic extension of Q when p is odd, and three such extensions when p = 2. For any such quadratic L ⊂ Q(ζp3 ), we write GL,Σ := Gal(QΣ /L). Theorem 2.15. — Let ξ : T → Zp be classical, and suppose that ρξ |GL,Σ is irreducible, for each quadratic extension L ⊂ Q(ζp3 ). Then Spec T has dimension 3 in a neighbourhood of ξ. Proof. — It follows from [29, Thm., p. 277] that this dimension is at most 3, while Corollary 2.28 below establishes the opposite inequality. This proves the result. 2.3. Families parametrized by weight: the Eisenstein family Since Spec T is (at least conjecturally) of relative dimension 3 over Spec Zp , one can think of the set of all p-adic systems of Hecke eigenvalues ξ as depending on three parameters. Unfortunately, even in those cases when Conjecture 2.9 is known, there is no particularly canonical choice of these three parameters. A little more formally, if Spec T has Krull dimension 4, then Noether normalization allows one to construct a finite map Spec T → Spec Zp [[T1 , T2 , T3 ]]. However, there is no canonical choice for such a map. On the other hand, there is a canonical map Spec T → Spec Zp [[T ]], as we now explain. Write q = p if p is odd, and q = 4 if p = 2, and set Γ = 1 + qZp . Let

L := {` prime | ` ≡ 1 mod N q}.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

48

M. EMERTON

We regard L as a subset of Γ. Dirichlet’s theorem on primes in arithmetic progression shows that L is in fact dense in Γ. Lemma 2.16. — The map L → T given by ` 7→ S` extends uniquely to a continuous homomorphism of groups Γ → T× . Proof. — If ` ∈ L , and if λ is any system of Hecke eigenvalues of weight k, then λ(S` ) = `k−2 . (Because ` ≡ 1 mod N , the diamond operator h`i is trivial.) The func(p) tion x 7→ xk−2 is continuous on Γ, and so the map ` 7→ S` from L → Zp ⊗Z Tk (p) × extends to a continuous homomorphism Γ → (Zp ⊗Z Tk ) , for any weight k. The lemma now follows by an easy passage to the limit. n

Write Zp [[Γ]] := lim Zp [Γ/Γp ]. This is the so-called completed group ring of Γ ←− n

over Zp ; there is an evident embedding of the usual group ring Zp [Γ] ,→ Zp [[Γ]]. If x ∈ Γ, we write [x] to denote the corresponding element of Zp [[Γ]] (so as to avoid confusion with the same element x regarded as belonging to the ring of coefficients ∼ Zp ). There is an isomorphism of Zp -algebras Zp [[T ]] −→ Zp [[Γ], determined by the condition T 7→ [1 + q] − 1. The continuous map Γ → T× of the preceding lemma extends uniquely to a homomorphism of Zp -algebras (10)

w : Zp [[Γ]] → T,

which we may equally well regard as a map Zp [[T ]] → T. Passing to Specs, we get the canonical map (11)



Spec T → Spec Zp [[Γ]] −→ Spec Zp [[T ]]

referred to above. What is the meaning of this map? Well, giving a Zp -valued point of Spec Zp [[Γ]] is the same as giving a continuous × character κ : Γ → Zp . Thus Spec Zp [[Γ]] is the space of characters of Γ. (The iso∼ morphism Spec Zp [[Γ]] −→ Spec Zp [[T ]] is then given by mapping a character κ to the value T = κ(1+q)−1; in this way, the space of continuous characters of Γ is identified with the maximal ideal of Zp , or, in more geometric terms, the open unit disk around × the origin of Qp .) If k is an integer, then we may define a character κk : Γ → Zp via the formula κk (x) = xk−2 . These points are Zariski dense in Spec Zp [[Γ]] (in fact, any infinite collection of them is Zariski dense), and so we regard Spec Zp [[Γ]] as a certain kind of interpolation of the set of integers, and refer to it as weight space. In particular, the Zp -valued point κk is said to be the point of weight k. Now suppose that ξ : T → Zp is classical, arising from the system of Hecke eigenvalues λ : Tk → Zp . One computes that the composite ξ ◦ w is equal to κk , the point of weight k. Thus we may think of the w as mapping a system of Hecke eigenvalues to

ASTÉRISQUE 339

(1013)

p-ADIC FAMILIES OF MODULAR FORMS

49

its corresponding weight (which explains our choice of notation). From this, we also see that w is injective (since there exist systems of Hecke eigenvalues of arbitrarily high weight), and hence that (11) is dominant. Now the weight is a very natural parameter to consider, and so it is reasonable to ask whether we can find families of systems of Hecke eigenvalues, and hence families of Galois representations, that are parametrized by the weight. Somewhat more precisely, we can ask whether we can find a closed subscheme Z ,→ Spec T such that the composite Z ,→ Spec T → Spec Zp [[T ]] is dominant with finite fibres; such a subscheme Z could then be thought of as a family of Galois representations, parametrized by the weight.(3) Of course, if we impose no other conditions on Z, then such subschemes Z exist for very general geometric reasons; on the other hand, a further natural condition to impose is that Z contain a Zariski dense set of points corresponding to classical systems of Hecke eigenvalues. The scheme Z could then be regarded as a one-dimensional family of systems of Hecke eigenvalues, parametrized by weight, and interpolating a collection of classical systems of Hecke eigenvalues. Example 2.17. — The most basic example of a one-dimensional family of systems of Hecke eigenvalues, parametrized by weight, is the Eisenstein family. This is the original p-adic family of modular forms, introduced by Serre in [37]. We describe it here, in the language of systems of Hecke eigenvalues that we have introduced. For simplicity we take N = 1, and we fix an even residue class i mod p − 1 if p (p) is odd. Consider the p-deprived systems of Hecke eigenvalues λk associated to the Eisenstein series Ek , for k ≥ 4, and congruent to i mod p − 1 if p is odd (resp. k ≥ 4 and even if p = 2). Recall that these are given by (p)

λk (`S` ) = `k−1 ,

(p)

λk (T` ) = 1 + `k−1 ,

where ` ranges over all primes distinct from p. We wish to rewrite these formulas slightly. Recall that Z× p = µp−1 × Γ (if p is odd) or µ2 × Γ (if p = 2). In either case, let µ denote the first factor, and write ω : Z× p →µ (p)

to denote the corresponding projection. Then we may rewrite the formulas for λk as k−2 k−2 (p) (p) λk (`S` ) = `ω(`)i−2 `ω(`)−1 , λk (T` ) = 1 + `ω(`)i−2 `ω(`)−1 , where we set i = 0 if p = 2. We may evidently interpolate these formulas into a Zp [[Γ]]-valued point of Spec T. Namely, there is a homomorphism E : T → Zp [[Γ]], defined by S` 7→ ω(`)i−2 [`ω(`)−1 ], T` 7→ 1 + `ω(`)i−2 [`ω(`)−1 ]. (3)

We are using the word “parametrized” in a somewhat liberal sense, in that we are allowing our family to be a multi-valued function of the weight, i.e. we are asking that Z → Spec Zp [[T ] have finite fibres, but not that it necessarily be injective.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

50

M. EMERTON

By construction, the composite κk ◦ E is equal to λk , for any k ≡ i mod p − 1 (or any even k, if p = 2). Again by construction, E ◦ w is the identity on Zp [[Γ]]. Thus, in more geometric terms, we have constructed a map Spec Zp [[Γ]] → Spec T which is a section to the weight map w : Spec T → Spec Zp [[Γ]], namely a family of Eisenstein systems of eigenvalues, parametrized by their weight. 2.4. Families parametrized by weight: Hida families and the eigencurve In our discussion of p-adic systems of Hecke eigenvalues, we have systematically ignored the Hecke operators Sp and Tp . This is important; for example, for the family λk of the Example 2.17, we have λk (Sp ) = pk−2 and λk (Tp ) = 1 + pk−1 . These functions do not interpolate well as p-adic functions of k. However, if we consider the pth Hecke polynomial X 2 − λk (Tp )X + pλk (Sp ), we see that it has the form X 2 − (1 + pk−1 )X + pk−1 = (X − 1)(X − pk−1 ). One of the two roots of this polynomial is in fact constant in the family, and so interpolates without difficulty in the family. It is the second root which does not interpolate well. This motivates the idea of changing our context slightly, and considering points not just in Spec T, but in Spec T × Gm (here the fibre product is with Spec Z[T, T −1 ] over Spec Z, or equivalently, with Spec Zp [T, T −1 ] over Spec Zp ). To any system of Hecke eigenvalues λ appearing in some M k (N ), we can plot a pair of Qp -valued points in this fibre product, whose first coordinate (for either point) is the associated classical p-adic system of eigenvalues ξ, and whose second coordinates are the roots of the pth Hecke polynomial of λ. Definition 2.18. — Let X denote the set of Qp -valued points of Spec T × Gm consisting of pairs (ξ, α), where ξ : Spec T → Zp is classical, attached to some system of Hecke eigenvalues λ : Tk → Zp with k ≥ 1, and α is a root of the pth Hecke polynomial X 2 − λ(Tp )X + pλ(Sp ). ×

Let X ord denote the subset of X consisting of pairs (ξ, α) for which α ∈ Zp . (The superscript ord is for ordinary.) Remark 2.19. — The reason for singling out the subset X ord of X is that (since any system of Hecke eigenvalues is a Zp -valued point of Spec T) these are precisely the points of X that consist of Zp -valued points of Spec T × Gm . The following theorem, due to Hida [23, 22], describes the interpolation of the points in X ord . (The map Spec T × Gm → Spec Zp [[Γ]] appearing in the statement of the theorem is the one obtained by first projecting onto the factor Spec T, and then applying the map w.)

ASTÉRISQUE 339

(1013)

p-ADIC FAMILIES OF MODULAR FORMS

51

Theorem 2.20. — The Zariski closure C ord of X ord in Spec T × Gm is onedimensional. The composite C ord → Spec T × Gm → Spec Zp [[Γ]] is finite, and is furthermore étale in the neighbourhood of those points of X ord that are attached to systems of Hecke eigenvalues appearing in weight k ≥ 2. Definition 2.21. — We refer to C ord as the Hida family, or ordinary family, of tame level N . Remark 2.22. — We will see in Subsection 2.5 below that it is necessary to restrict to weights k ≥ 2 in the final statement of the theorem. The curve C ord is (almost) precisely a family of the type we envisaged in the previous subsection. (We say “almost” because it lies in Spec T × Gm rather than in Spec T itself.) On the other hand, not every classical system of eigenvalues appears in C ord ; it is certainly possible that if λ : Tk → Zp , then both roots of the pth Hecke polynomial may have positive slope. We thus turn to our next result, due to Coleman and Mazur [10], which deals with the interpolation of the entire set X . In this case, taking the algebraic Zariski closure of these points in Spec T × Gm turns out to be too coarse of an operation, and we cannot hope to construct an algebraic family of the type envisaged in the previous subsection that contains all the points of X . Rather, we will construct a rigid analytic family, lying inside the associated rigid analytic space (Spec T × Gm )an .(4) Theorem 2.23. — The rigid analytic Zariski closure C of X in (Spec T × Gm )an is one-dimensional. More precisely, the composite

C ,→ (Spec T × Gm )an → (Spec Zp [[Γ]])an

(12)

is flat, and has discrete fibres. Furthermore, for any positive constant C, there are only finitely many points (ξ, α) in any given fibre satisfying ordp (α) ≤ C. (In other words, the slopes of the Gm -coordinates go to ∞ in each fibre.) Definition 2.24. — The curve C is called the eigencurve of tame level N . The analytification of C ord is called the slope zero part, or the ordinary part, of the eigencurve. It is a union of connected components of C . (4)

Concretely, if T =

Qm

i=1

Zp [[T1 , . . . , Tri ]]/(fi,1 , . . . , fi,si ), then

Spec T × Gm =

`m

i=1

Spec Zp [[T1 , . . . , Tri ]][T, T −1 ]/(fi,1 , . . . , fi,si ),

and (Spec T × Gm )an is the rigid analytic space

`m

i=1 { (T1 , . . . , Tri , T )

Also, (Spec Zp

[[Γ]])an

|

|T1 |, . . . , |Tri | < 1, T 6= 0, fi,1 (T1 , . . . , Tri ) = · · · = fi,si (T1 , . . . , Tri ) = 0 }.

∼ = (Spec Zp [[T ]])an = {T | |T | < 1}, i.e. the open unit disk in Qp .

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

52

M. EMERTON

Remark 2.25. — The map (12) is in fact étale in the neighbourhood of a point (ξ, α) ∈ X , unless α is a repeated root of the pth Hecke polynomial of the system of eigenvalues λ giving rise to ξ. (Compare the discussion of [10, p. 5].) It is conjectured that such repeated roots cannot occur when the weight k ≥ 2 [9]. Remark 2.26. — By construction, each of C ord and C contains a(n algebraic or rigid analytic, as the case may be) Zariski dense set of points (ξ, α) for which ξ is a classical system of eigenvalues. It is natural to ask whether the converse holds, namely, if (ξ, α) × is any Qp -valued point of C ord of C lying over the weight κk : Γ → Zp , for some positive integer k, then is ξ classical? The answer is no in general, for trivial reasons. One already sees this with the Eisenstein family of Example 2.17. Indeed, in the notation of that example (and assuming p is odd for simplicity), if k 6≡ i mod p − 1 then the associated system of eigenvalues is not associated to a modular form of level 1; rather, its values on `S` (p) and T` (for ` - N p) coincide with those of the system of eigenvalues λ1,ωi−k ,k (in the notation of Example 1.18), corresponding to an Eisenstein series of level p. Hida showed in general that if (ξ, α) is a Zp -valued (or equivalently, Qp -valued) point of C ord lying over the character κk for k ≥ 2 (or, more generally, a character of the form ψκk , where ψ has finite order and k ≥ 2), then there is a system of eigenvalues λ : Tk (N p) → Zp such that ξ(`S` ) = λ(`S` ) and ξ(T` ) = λ(T` ) for all ` - N p. In the non-ordinary case, the situation is more complicated. The fibre of C over any κk (or over any character ψκk , where ψ is of finite order) is typically infinite, and all but finitely many of the points do not arise from a classical eigenform (of any level). However, Coleman showed [7, 8] that if (ξ, α) is such a point, and if the slope of α is less than k − 1, then just as in the ordinary case, there is a system of eigenvalues λ : Tk (N p) → Zp such that ξ(`S` ) = λ(`S` ) and ξ(T` ) = λ(T` ) for all ` - N p. (One can show, e.g. using Theorem 2.33 below, that, conversely, if such a λ exists, then the slope of α is at most k − 1. Of course, all but finitely many of the points lying over κk have slope > k − 1.) Idea of proofs. — The first step in the proof of Theorems 2.20 and 2.23 is to define a space of p-adic modular forms on which the p-adic Hecke algebra T acts. In fact one can literally work with such a space, namely the space of generalized p-adic modular functions of Katz (as defined in [27], see also [19] and [23]) — this is the approach taken in [23] for the ordinary case and in [10] for the general case — or with a surrogate, constructed from the group cohomology of Γ1 (N ) and certain of its subgroups. The cohomological approach to the ordinary case is developed in [22], and for the general case is developed in [42, 1]. There is another approach, via the p-adically completed cohomology of modular curves [15, §4], which is somewhat different,

ASTÉRISQUE 339

(1013)

p-ADIC FAMILIES OF MODULAR FORMS

53

and which we will say a little about below. To simplify the exposition, from now on we will speak simply of “the space of p-adic modular forms”, meaning either the space of generalized p-adic modular functions, or one of the cohomological surrogates of [1, 22, 42]. The next step is to introduce an additional Hecke operator on this space, the socalled Up -operator. In the context of p-adic modular forms, this operator has the following effect on q-expansions: Up f =

∞ X

anp (f )q n .

n=0 ∗

We let T denote the quotient of T[Up ] that acts faithfully on the space of p-adic modular forms. Evidently, Spec T∗ ,→ Spec T × A1 . Suppose for a moment that f is a modular form of weight k and level N, with p - N , and let α and β be the roots of the pth Hecke polynomial. Then f (τ ) − βf (pτ ) is a modular form of level N p, which is a Up -eigenform with eigenvalue α. (This can be checked directly on the level of q-expansions.) Thus Spec T∗ contains the set X , and hence also the Zariski closure of this set. The technical difficulty that arises in establishing the theorems is that Spec T∗ is much bigger than either C ord or C , roughly speaking because Up has a huge kernel on the space of the p-adic modular forms, while we are trying to construct curves lying in Spec T × Gm , i.e. systems of eigenvalues of T∗ for which the associated Up -eigenvalue is non-zero. It is at this point that the proofs of the two theorems diverge somewhat, with the proof of Theorem 2.20 being technically simpler than that of Theorem 2.23. The points of X ord correspond to eigenforms whose Up -eigenvalue is ordinary. If f is any eigenform for T∗ whose Up -eigenvalue α is of positive slope, then limn Upn f = αn f → 0 as n → 0. Thus by iterating Up on the space of p-adic modular forms and passing to a limit, we can cut out the ordinary part of the space of p-adic modular forms, on which Up acts with only ordinary eigenvalues. (This process can be summarized by using the limits of powers of Un to construct the so-called ordinary projector, which projects to the ordinary part.) The quotient of T∗ acting faithfully on this ordinary part is denoted by Tord , and C ord = Spec Tord . The key fact, underlying the proof of Theorem 2.20, is that Tord is finite over Zp [[Γ]]. This can be proved in various ways; either using the theory of mod p modular forms, if one is working with generalized p-adic modular functions (this is the approach taken in [23]), or by arguments with group cohomology (this is the approach of [22]). As already indicated, the proof of Theorem 2.23 is more technical. The reason is as follows: if m∗ is a maximal ideal of T∗ lying over a maximal ideal m of T, and if m∗ is not ordinary (i.e. if Up ∈ m∗ ), then it follows from [19, Prop. II.3.14] that T∗m∗ ∼ = Tm [[Up ]].

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

54

M. EMERTON

(Here T∗m∗ and Tm denote completions, and Tm [[Up ]] is the formal power series ring in Tm with variable Up .) Thus if ξ : Tm → Zp is a system of eigenvalues, we can extend it to a system of eigenvalues of T∗m by assigning any positive-slope value of Up that we choose; even if ξ is classical, attached to some system of eigenvalues λ : Tk → Zp , the algebra T∗ has no way of distinguishing the positive slope roots of the pth Hecke polynomial of λ from any other positive slope elements of Zp . Thus one cannot reasonably interpolate the points X by algebra alone; it is necessary to use some analysis. In the generalized p-adic modular functions setting, the key step is to replace this space by a certain subspace of so-called overconvergent modular forms. (This is the approach of [10].) In the cohomological framework, this step can be taken at the beginning, by working with rigid analytic (rather than merely continuous) modular symbols (as is done in [42, 1]). In these settings, the operator Up is a compact operator, and so has a reasonable spectral theory. One can then analyze, and obtain finiteness results for, all of its non-zero eigenspaces, rather than just the ordinary eigenspaces. The analysis of these eigenspaces is at the heart of the proof of Theorem 2.23. As mentioned above, there is another approach to the proof of Theorem 2.23, via p-adically completed cohomology [15]. In this setting, one does not directly have an action of the Up -operator, but rather has an action of the entire group GL2 (Qp ), and the introduction of the Up -operator and the passage to its non-zero eigenspaces is effected in a single step, by applying the locally analytic Jacquet module functor of [14]. We make some further technical remarks. In the paper [10], the authors prove Theorem 2.23 only in the case when N = 1. The generalization to arbitrary N can be found in [4, Part II], or [15, §4]. Also, in most of the papers cited, the authors work with Hecke algebras that contain the Hecke operators U` for `|N, ` 6= p, as well as the operators S` and T` that we have considered. We have avoided any consideration of these additional operators, since they are not essential for the consideration of Galois representations. It is not difficult to deduce the results in the form that we have stated them from the corresponding results cited, which perhaps involve these additional operators. 2.5. The infinite fern The composition of the closed embedding C ord ,→ Spec T × Gm with the projection onto the first factor gives a map (13)

C ord → Spec T,

which is very close to being injective on Zp -points. Indeed, if (ξ1 , α) and (ξ2 , β) are two such points mapping to the same point of Spec T, then ξ1 = ξ2 = ξ (say), and

ASTÉRISQUE 339

(1013)

p-ADIC FAMILIES OF MODULAR FORMS

55

we see (by Theorem 2.29 below) that ρξ | GQp admits unramified quotients on which Frobp acts by α and β respectively. Thus if α 6= β, we see that ρξ | GQp is unramified. It is then conjectured (as a special case of [18, Conj. 3c]), and is proved in most cases [5], that ξ is a classical system of Hecke eigenvalues, arising from a weight 1 Hecke eigenform of level N . Hence (13) has (or at least, is expected to have) at most finitely many double points, arising from classical systems of Hecke eigenvalues in weight one. On the other hand, if we consider the analogous map (14)

C → (Spec T)an ,

then every system of eigenvalues λ : Tk → Zp gives rise to a pair of points (ξ, α) and (ξ, β), where ξ is a p-adic system of Hecke eigenvalues associated to λ, and X 2 − λ(T` )X + λ(`S` ) = (X − α)(X − β). Unless α = β (which, as we already noted, is expected to be impossible unless k = 1), we see that the image of (14) admits a double point at ξ. Thus the image of (14) is a very complicated curve, with an infinite number of double points. It is known as the infinite fern [32, 21]. The following theorem, due to Gouvêa and Mazur [21], shows that it is a kind of “space-filling curve” in (Spec T)an . Theorem 2.27. — Each component of the Zariski closure of the infinite fern in (Spec T)an is at least two-dimensional. Sketch of proof. — Since C is defined to be the Zariski closure of X , we see that the Zariski closure of the image of (14) is equal to the Zariski closure of the set ξ of classical p-adic systems of Hecke eigenvalues. Suppose that this Zariski closure contains a component Z that is one-dimensional. Since the singular locus of Z is a Zariski closed proper subset of Z , we may find a classical ξ lying in the smooth locus of Z . Since Z is one-dimensional, the map C → Z must be surjective in a neighbourhood of ξ. Let λ : Tk → Zp , for some k ≥ 1, be the system of eigenvalues giving rise to ξ, and let α and β be the two roots of the pth Hecke polynomial of λ. Then (unless α = β), the image of C is branched at ξ, contradicting the fact that Z is smooth at ξ. If α = β, then by appealing to the result of Coleman mentioned in Remark 2.26, it is easy to see that we may find arbitrarily small perturbations ξ 0 of ξ, for which ξ 0 is classical and such that the corresponding roots α0 and β 0 of the pth Hecke polynomial are distinct. Then we may apply the above argument to ξ 0 instead, and again derive a contradiction. The previous result has the following corollary on the Krull dimension of Spec T, which is again due to Gouvêa and Mazur [21].

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

56

M. EMERTON

Corollary 2.28. — The Krull dimension of each component of Spec T is at least 4. Sketch of proof. — It is equivalent to show that each component of the associated rigid analytic space (Spec T)an is at least three-dimensional. Any such component Y contains a component Z of the image of (14), which is two-dimensional. Twisting by characters of p-power conductor then provides a one-dimensional deformation of Z inside Y , showing that Y is at least three-dimensional. 2.6. Galois representations over Hida families and the eigencurve The following result, due to Mazur–Wiles [34] and Wiles [44], gives a Galoistheoretic interpretation of the points of C ord , and in particular, of the Gm -coordinate. Before stating it, we note that the chosen embedding ıp : Q ,→ Qp induces a map GQp → GQ → GQ,Σ (where we have written GQp and GQ to denote Gal(Qp /Qp ) and Gal(Q/Q) respectively, and where the second arrow is the natural surjection). For any representation ρ of GQ,Σ , we write ρ| GQp to denote the restriction of ρ to a representation of GQp via this map. Recall that GQp contains a normal subgroup Ip ∼ (the inertia subgroup), such that GQp /Ip −→ GFp , the absolute Galois group of Fp . This latter group is topologically generated by the Frobenius automorphism Frobp . We say that a representation of GQp is unramified if it is trivial when restricted to Ip ; any such representation is then endowed with an action of Frobp . Theorem 2.29. — If (ξ, α) is Zp -valued point of C ord , then ρξ | GQp admits a onedimensional unramified quotient on which Frobp acts with eigenvalue α. One has the following conjecture, which is an analogue for C ord of Conjecture 2.12. It was first made by Mazur and Tilouine [33]. Conjecture 2.30. — If Σ is any finite set of primes containing p, and if ρ : GQ,Σ → GL2 (Qp ) is a continuous, semi-simple, and odd representation whose restriction to GQp admits a one-dimensional unramified quotient on which Frobp acts through the eigenvalue α, then there is a Zp -valued point (ξ, α) in the Hida family for some level N divisible only by primes in Σ distinct from p such that ρ = ρξ . In their papers [40, 41], Skinner and Wiles have established this conjecture in a large number of cases. Theorem 2.31. — Let ρ and α be as in the statement of Conjecture 2.30, and let ρ : GQ,Σ → GL2 (Fp ) be the representation obtained by descending ρ to Zp and then reducing modulo the maximal ideal of Zp . If det ρ|Ip = ψχk−1 for some finite order character ψ and some integer k ≥ 2 (recall that χ denotes the p-adic cyclotomic

ASTÉRISQUE 339

(1013)

p-ADIC FAMILIES OF MODULAR FORMS

57

character), and if the semi-simplification of ρ | GQp (which is necessarily the direct sum of two characters, and which is well-defined independently of the choice of Zp -model of ρ giving rise to ρ) is the direct sum of distinct characters, then Conjecture 2.30 holds for ρ. Remark 2.32. — Suppose that ρ and α are as in the preceding result, and let (ξ, α) be the point on the Hida family (of an appropriately chosen level), whose existence is given by the theorem, for which ρ = ρξ . The assumption on det ρ in the theorem (and in particular, the assumption that k ≥ 2) implies, by the result of Hida recalled above, that ξ is obtained from a system of eigenvalues λ : Tk (N p) → Zp . Thus, in the context of this result, one concludes that ρ actually arises from the system of Hecke eigenvalues attached to a classical modular form (of level possibly divisible by p). The following result gives a Galois-theoretic interpretation of the points on C , analogous to Theorem 2.29. It is due to Kisin [28]. The statement requires the language + of Fontaine’s theory [17]. Recall that Fontaine has defined a ring Bcris , equipped with commuting actions of the group GQp and of a “Frobenius” operator ϕ. If V is any rep+ + resentation of GQp over Qp , then Dcris (V ) := (Bcris ⊗Qp V )GQp is a Qp -vector space of dimension at most that of V , equipped with an operator ϕ induced by the operator ϕ + on Bcris . Theorem 2.33. — If (ξ, α) is a Qp -valued point of C , and if ρ∨ ξ | GQ

p

the contragredient representation to ρξ | GQp , then dimensional subspace on which ϕ acts via α.

+ Dcris (ρ∨ ξ | GQp )

denotes

contains a one-

Sketch of proof. — If the p-adic system of Hecke eigenvalues ξ is classical, arising from a system of Hecke eigenvalues λ : Tk → Zp attached to some modular form of weight k ≥ 1, then the representation ρ∨ ξ | GQ is in fact crystalline, with Hodge–Tate p

+ weights equal to 0 and 1 − k, and so Dcris (ρ∨ ξ | GQp ) is two-dimensional over Qp . In this context, it is known that the characteristic polynomial of ϕ is equal to the pth Hecke polynomial of λ [36]. Thus, if α and β are the two roots of this polynomial, then we + + ϕ=α ϕ=β see that Dcris (ρ∨ and Dcris (ρ∨ are both non-zero. The theorem is ξ | GQp ) ξ | GQp ) then proved by showing that these non-zero spaces interpolate over the curve C . ×

+ ϕ=α Remark 2.34. — In the context of Theorem 2.33, if α ∈ Zp , then Dcris (ρ∨ ξ | GQp ) is non-zero if and only if ρξ | GQp contains an unramified quotient on which Frobp acts via α. Thus Theorem 2.29 is a consequence of Theorem 2.33.

The following conjecture is analogous to Conjecture 2.30 in the ordinary case. (See the hope expressed in Remark (2) of [28, p. 450].)

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

58

M. EMERTON

Conjecture 2.35. — If Σ is any finite set of primes containing p, and if ρ : GQ,Σ → GL2 (Qp ) + ϕ=α is a continuous, semi-simple, and odd representation such that Dcris (ρ∨ is ξ | GQ ) p

non-zero, then there is a Qp -valued point (ξ, α) in C for some level N divisible only by primes in Σ distinct from p such that ρ = ρξ . There has been recent progress on this conjecture (see the corollary on p. 3 of [30] as well as the forthcoming paper [16]).

REFERENCES [1] A. Ash & G. Stevens – “p-adic deformations of cohomology on GL(n): the non-ordinary case”, preprint, 2000. [2] A. O. L. Atkin & J. Lehner – “Hecke operators on Γ0 (m)”, Math. Ann. 185 (1970), p. 134–160. [3] G. Böckle – “On the density of modular points in universal deformation spaces”, Amer. J. Math. 123 (2001), p. 985–1007. [4] K. Buzzard – “Eigenvarieties”, in L-functions and Galois representations, London Math. Soc. Lecture Note Ser., vol. 320, Cambridge Univ. Press, 2007, p. 59– 120. [5] K. Buzzard & R. Taylor – “Companion forms and weight one forms”, Ann. of Math. 149 (1999), p. 905–919. [6] W. Casselman – “On some results of Atkin and Lehner”, Math. Ann. 201 (1973), p. 301–314. [7] R. F. Coleman – “Classical and overconvergent modular forms”, Invent. Math. 124 (1996), p. 215–241. [8]

, “Classical and overconvergent modular forms of higher level”, J. Théor. Nombres Bordeaux 9 (1997), p. 395–403.

[9] R. F. Coleman & B. Edixhoven – “On the semi-simplicity of the Up -operator on modular forms”, Math. Ann. 310 (1998), p. 119–127. [10] R. F. Coleman & B. Mazur – “The eigencurve”, in Galois representations in arithmetic algebraic geometry (Durham, 1996), London Math. Soc. Lecture Note Ser., vol. 254, Cambridge Univ. Press, 1998, p. 1–113. [11] P. Deligne – “Formes modulaires et représentations `-adiques”, Séminaire Bourbaki, volume 1968/1969, exposé no 355, Lecture Notes in Math., vol. 179, 1971, p. 139–172.

ASTÉRISQUE 339

(1013)

p-ADIC FAMILIES OF MODULAR FORMS

59

[12] P. Deligne & J-P. Serre – “Formes modulaires de poids 1”, Ann. Sci. École Norm. Sup. 7 (1974), p. 507–530. [13] M. Eichler – “Quaternäre quadratische Formen und die Riemannsche Vermutung für die Kongruenzzetafunktion”, Arch. Math. 5 (1954), p. 355–366. [14] M. Emerton – “Jacquet modules of locally analytic representations of p-adic reductive groups. I. Construction and first properties”, Ann. Sci. École Norm. Sup. 39 (2006), p. 775–839. [15]

, “On the interpolation of systems of eigenvalues attached to automorphic Hecke eigenforms”, Invent. Math. 164 (2006), p. 1–84.

[16]

, “Local-global compatibility in the p-adic Langlands programme for GL2 (Qp )”, preprint, 2010.

[17] J.-M. Fontaine – “Le corps des périodes p-adiques”, Astérisque 223 (1994), p. 59–111. [18] J.-M. Fontaine & B. Mazur – “Geometric Galois representations”, in Elliptic curves, modular forms, & Fermat’s last theorem (Hong Kong, 1993), Ser. Number Theory, I, Int. Press, Cambridge, MA, 1995, p. 41–78. [19] F. Q. Gouvêa – Arithmetic of p-adic modular forms, Lecture Notes in Math., vol. 1304, Springer, 1988. [20]

, “Deforming Galois representations: controlling the conductor”, J. Number Theory 34 (1990), p. 95–113.

[21] F. Q. Gouvêa & B. Mazur – “On the density of modular representations”, in Computational perspectives on number theory (Chicago, IL, 1995), AMS/IP Stud. Adv. Math., vol. 7, Amer. Math. Soc., 1998, p. 127–142. [22] H. Hida – “Galois representations into GL2 (Zp [[X]]) attached to ordinary cusp forms”, Invent. Math. 85 (1986), p. 545–613. [23]

, “Iwasawa modules attached to congruences of cusp forms”, Ann. Sci. École Norm. Sup. 19 (1986), p. 231–273.

[24] J.-i. Igusa – “Kroneckerian model of fields of elliptic modular functions”, Amer. J. Math. 81 (1959), p. 561–577. [25] H. Jacquet & R. P. Langlands – Automorphic forms on GL(2), Lecture Notes in Math., vol. 114, Springer, 1970. [26] N. Jochnowitz – “Congruences between systems of eigenvalues of modular forms”, Trans. Amer. Math. Soc. 270 (1982), p. 269–285. [27] N. M. Katz – “p-adic interpolation of real analytic Eisenstein series”, Ann. of Math. 104 (1976), p. 459–571. [28] M. Kisin – “Overconvergent modular forms and the Fontaine-Mazur conjecture”, Invent. Math. 153 (2003), p. 373–454.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

60

M. EMERTON

[29]

, “Geometric deformations of modular Galois representations”, Invent. Math. 157 (2004), p. 275–328.

[30]

, “The Fontaine-Mazur conjecture for GL2 ”, J. Amer. Math. Soc. 22 (2009), p. 641–690.

[31] B. Mazur – “Deforming Galois representations”, in Galois groups over Q (Berkeley, CA, 1987), Math. Sci. Res. Inst. Publ., vol. 16, Springer, 1989, p. 385–437. [32]

, “An “infinite fern” in the universal deformation space of Galois representations”, Collect. Math. 48 (1997), p. 155–193.

[33] B. Mazur & J. Tilouine – “Représentations galoisiennes, différentielles de Kähler et “conjectures principales””, Publ. Math. I.H.É.S. 71 (1990), p. 65–103. [34] B. Mazur & A. J. Wiles – “On p-adic analytic families of Galois representations”, Compositio Math. 59 (1986), p. 231–264. [35] K. A. Ribet – “Galois representations attached to eigenforms with Nebentypus”, in Modular functions of one variable, V (Proc. Second Internat. Conf., Univ. Bonn, Bonn, 1976), Lecture Notes in Math., Vol. 601, Springer, 1977, p. 17–51. [36] T. Saito – “Modular forms and p-adic Hodge theory”, Invent. Math. 129 (1997), p. 607–620. [37] J-P. Serre – “Formes modulaires et fonctions zêta p-adiques”, in Modular functions of one variable, III (Proc. Internat. Summer School, Univ. Antwerp, 1972), Lecture Notes in Math., vol. 350, Springer, 1973, p. 191–268. [38] G. Shimura – “Correspondances modulaires et les fonctions ζ de courbes algébriques”, J. Math. Soc. Japan 10 (1958), p. 1–28. [39]

, Introduction to the arithmetic theory of automorphic functions, Publications of the Mathematical Society of Japan, No. 11. Iwanami Shoten, Publishers, Tokyo, 1971.

[40] C. M. Skinner & A. J. Wiles – “Residually reducible representations and modular forms”, Publ. Math. I.H.É.S. 89 (1999), p. 5–126. [41]

, “Nearly ordinary deformations of irreducible residual representations”, Ann. Fac. Sci. Toulouse Math. 10 (2001), p. 185–215.

[42] G. Stevens – “Rigid analytic modular symbols”, preprint, 1994. [43] R. Taylor & A. J. Wiles – “Ring-theoretic properties of certain Hecke algebras”, Ann. of Math. 141 (1995), p. 553–572. [44] A. J. Wiles – “On ordinary λ-adic representations associated to modular forms”, Invent. Math. 94 (1988), p. 529–573. [45]

, “Modular elliptic curves and Fermat’s last theorem”, Ann. of Math. 141 (1995), p. 443–551.

ASTÉRISQUE 339

(1013)

p-ADIC FAMILIES OF MODULAR FORMS

61

Matthew EMERTON Department of Mathematics Northwestern University 2033 Sheridan Rd EVANSTON, IL 60208-2730 – USA E-mail : [email protected]

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

339

ASTÉRISQUE 2011

SÉMINAIRE BOURBAKI VOLUME 2009/2010 EXPOSÉS 1012-1026 (1014) Algèbres amassées et applications Bernhard KELLER

SOCIÉTÉ MATHÉMATIQUE DE FRANCE Publié avec le concours du CENTRE NATIONAL DE LA RECHERCHE SCIENTIFIQUE

Séminaire BOURBAKI 62e année, 2009-2010, no 1014, p. 63 à 90

Novembre 2009

ALGÈBRES AMASSÉES ET APPLICATIONS [d’après Fomin-Zelevinsky, ...] par Bernhard KELLER

INTRODUCTION

Les algèbres amassées (cluster algebras), inventées [38] par Sergey Fomin et Andrei Zelevinsky au début des années 2000, sont des algèbres commutatives, dont les générateurs et les relations sont construits de façon récursive. Parmi ces algèbres se trouvent les algèbres de coordonnées homogènes sur les grassmanniennes, les variétés de drapeaux et beaucoup d’autres variétés qui jouent un rôle important en géométrie et théorie des représentations. La motivation principale de Fomin et Zelevinsky était de trouver un cadre combinatoire pour l’étude des bases canoniques dont on dispose [68] [81] dans ces algèbres et qui sont étroitement liées à la notion de positivité totale [82] dans les variétés associées. Il s’est avéré rapidement que la combinatoire des algèbres amassées intervenait également dans de nombreux autres sujets, par exemple dans – – – –

la géométrie de Poisson [52] [53] [54] [5] ... ; les systèmes dynamiques discrets [41] [69] [24] [63] ... ; les espaces de Teichmüller supérieurs [32] [33] [30] [34] ... ; la combinatoire et en particulier l’étude de polyèdres tels les associaèdres de Stasheff [19] [18] [60] [76] [36] [37] [85] [86] ... ; – la géométrie algébrique (commutative ou non commutative) et en particulier l’étude des conditions de stabilité de Bridgeland [6], les algèbres Calabi-Yau [64] [55], les invariants de Donaldson-Thomas [66] [75] [91] [46] ... ; – et la théorie des représentations des carquois et des algèbres de dimension finie, voir par exemple les articles de synthèse [8] [92] [93] [50] [71].

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

64

B. KELLER

Nous renvoyons aux articles d’initiation [40] [107] [104] [105] [106] et au portail des algèbres amassées [35] pour plus d’informations sur les algèbres amassées et leurs liens avec d’autres sujets mathématiques (et physiques). Dans cet exposé, nous donnons une introduction concise aux algèbres amassées (section 1) et présentons deux applications : – la démonstration de la périodicité de certains systèmes dynamiques discrets, d’après Fomin-Zelevinsky [41] et l’auteur [71] [72] (section 2.3) ; – la construction de bases duales semi-canoniques, d’après Geiss-Leclerc-Schröer [48] (section 3.4). Ces applications sont fondées sur la catégorification additive des algèbres amassées à l’aide de catégories de représentations de carquois (avec relations). Nous en décrivons les idées principales à la section 4. Nous y esquissons également des développements récents importants liés à la catégorification monoïdale d’algèbres amassées [58] [88] et à leur étude via les carquois à potentiel [22] [23].

1. DESCRIPTION ET PREMIERS EXEMPLES 1.1. Description Une algèbre amassée est une Q-algèbre commutative munie d’un ensemble de générateurs distingués (les variables d’amas) regroupés dans des parties (les amas) de cardinal constant (le rang) qui sont construites récursivement par mutation à partir d’un amas initial. L’ensemble des variables d’amas peut être fini ou infini. Théorème 1.1 ([39]). — Les algèbres amassées n’ayant qu’un nombre fini de variables d’amas sont paramétrées par les systèmes de racines finis. La classification est donc analogue à celle des algèbres de Lie semi-simples complexes. Nous allons préciser le théorème (dans le cas simplement lacé) à la section 2. 1.2. Premier exemple Pour illustrer la description et le théorème, présentons [106] l’algèbre amassée AA2 associée au système de racines A2 . Par définition, elle est engendrée sur Q par les variables d’amas xm , m ∈ Z, soumises aux relations d’échange xm−1 xm+1 = 1 + xm , m ∈ Z. Ses amas sont par définition les paires de variables consécutives {xm , xm+1 }, m ∈ Z. L’amas initial est {x1 , x2 } et deux amas sont reliés par une mutation si et seulement si ils ont exactement une variable d’amas en commun.

ASTÉRISQUE 339

(1014)

ALGÈBRES AMASSÉES ET APPLICATIONS

65

Les relations d’échange permettent d’exprimer toute variable d’amas comme fonction rationnelle des variables initiales x1 , x2 et donc d’identifier l’algèbre AA2 à une sous-algèbre du corps Q(x1 , x2 ). Afin d’expliciter cette sous-algèbre, calculons les xm pour m ≥ 3. Nous avons : (1.2.1) (1.2.2) (1.2.3)

1 + x2 x1 1 + x3 x1 + 1 + x2 x4 = = x2 x1 x2 1 + x4 x1 x2 + x1 + 1 + x2 1 + x2 1 + x1 x5 = = ÷ = . x3 x1 x2 x1 x2 x3 =

Notons que, contrairement à ce qu’on pourrait attendre, le dénominateur dans 1.2.3 reste un monôme ! En fait, toute variable d’amas dans une algèbre amassée quelconque est un polynôme de Laurent, voir le théorème 2.1. Continuons le calcul : (1.2.4) (1.2.5)

1 + x5 x2 + 1 + x1 x1 + 1 + x2 = ÷ = x1 x4 x2 x1 x2 1 + x1 = x2 . x7 = (1 + x1 ) ÷ x2 x6 =

Il est alors clair que la suite des xm , m ∈ Z, est 5-périodique et que le nombre de variables d’amas est effectivement fini et égal à cinq. Outre les deux variables initiales x1 et x2 nous avons trois variables non initiales x3 , x4 et x5 . En examinant leurs dénominateurs, nous voyons qu’elles sont en bijection naturelle avec les racines positives α1 , α1 + α2 , α2 du système de racines de type A2 . Ceci se généralise à tout diagramme de Dynkin, voir le théorème 2.1. 1.3. Algèbres amassées de rang 2 À tout couple d’entiers positifs (b, c) est associée une algèbre amassée A(b,c) . On la définit de la même manière que AA2 , mais en remplaçant les relations d’échange par ( xbm + 1 si m est impair, xm−1 xm+1 = xcm + 1 si m est pair. L’algèbre A(b,c) n’a qu’un nombre fini de variables d’amas si et seulement si bc ≤ 3, autrement dit si la matrice " # 2 −b −c

2

est la matrice de Cartan d’un système de racines Φ de rang 2. Le lecteur pourra s’amuser à vérifier que, dans ce cas, les variables d’amas non initiales sont toujours paramétrées par les racines positives de Φ.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

66

B. KELLER

2. ALGÈBRES AMASSÉES ASSOCIÉES AUX CARQUOIS 2.1. Mutation des carquois Un carquois est un graphe orienté, c’est-à-dire un quadruplet Q = (Q0 , Q1 , s, t) formé d’un ensemble de sommets Q0 , d’un ensemble de flèches Q1 et de deux applications s et t de Q1 dans Q0 qui, à une flèche α, associent respectivement sa source et son but. En pratique, on représente un carquois par un dessin comme dans l’exemple qui suit : % // α Q: 3^ 5 /6 µ

λ

1

ν

/2o

β

/

4.

γ

Une flèche α dont la source et le but coïncident est une boucle ; un 2-cycle est un couple de flèches distinctes β et γ telles que s(β) = t(γ) et t(β) = s(γ). De même, on définit les n-cycles pour tout entier positif n. Un sommet i d’un carquois est une source (respectivement un puits) s’il n’existe aucune flèche de but i (respectivement de source i). Appelons bon carquois un carquois fini sans boucles ni 2-cycles dont l’ensemble des sommets est l’ensemble des entiers 1 . . . n pour un entier positif n. À un isomorphisme fixant les sommets près, un tel carquois Q est donné par la matrice antisymétrique B = BQ dont le coefficient bij est la différence entre le nombre de flèches de i à j et le nombre de flèches de j à i pour tous 1 ≤ i, j ≤ n. Réciproquement, toute matrice antisymétrique B à coefficients entiers provient d’un bon carquois QB . Soient Q un bon carquois et k un sommet de Q. La carquois muté µk (Q) est le carquois obtenu à partir de Q comme suit : β / k α / j , on rajoute une nouvelle flèche (1) pour tout sous-carquois i [αβ] : i → j ; (2) on renverse toutes les flèches de source ou de but k ; (3) on supprime les flèches d’un ensemble maximal de 2-cycles disjoints deux à deux.

Si B est la matrice antisymétrique associée à Q et B 0 celle associée à µk (Q), on a ( −bij si i = k ou j = k ; 0 bij = bij + sgn(bik ) max(0, bik bkj ) sinon. C’est la règle de mutation des matrices antisymétriques (plus généralement : antisymétrisables) introduite par Fomin-Zelevinsky dans [38], voir aussi [42].

ASTÉRISQUE 339

(1014)

67

ALGÈBRES AMASSÉES ET APPLICATIONS

On vérifie sans peine que µk est une involution. Par exemple, les carquois (2.1.1)

E1

1Y et 

2o

3

2



3

sont reliés par la mutation par rapport au sommet 1. Notons que, du point de la théorie des représentations, ces carquois sont très différents. Deux carquois sont équivalents par mutation s’ils sont reliés par une suite finie de mutations. On vérifie facilement, par exemple à l’aide de [73], que les trois carquois suivants sont équivalents par mutation (2.1.2)

E1 

o F2 F4 7o

F3 

o

E5 

8o

1





o 

F6  10

9o

10

-5  7m 8 (

4 6Z 9  2

7 5 4X

|

 8 3



6

~

5j

3Q

*

10 "

1

La classe de mutation commune de ces carquois comporte 5739 carquois (à isomorphisme près). La classe de mutation de la « plupart » des carquois est infinie. La classification des carquois ayant une classe de mutation finie est un problème difficile, résolu récemment par Felikson-Shapiro-Tumarkin [28] : outre les carquois à deux sommets et les carquois associés à des surfaces à bord marquées [37], la liste contient 11 carquois exceptionnels, dont le plus grand est dans la classe de mutation des carquois 2.1.2. 2.2. Mutation des graines, algèbres amassées Soient n ≥ 1 un entier et F le corps Q(x1 , . . . , xn ) engendré par n indéterminées x1 , . . . , xn . Une graine (appelée aussi X-graine) est un couple (R, u), où R est un bon carquois et u une suite u1 , . . . , un qui engendre librement le corps F . Si (R, u) est une graine et k un sommet de R, la mutation µk (R, u) est la graine (R0 , u0 ), où R0 = µk (R) et u0 est obtenu à partir de u en remplaçant l’élément uk par l’élément u0k défini par la relation d’échange Y Y (2.2.1) u0k uk = ut(α) + us(α) . s(α)=k

t(α)=k

On vérifie que µ2k (R, u) = (R, u). Par exemple, les mutations de la graine 1

/2

/ 3 , {x1 , x2 , x3 }

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

68

B. KELLER

par rapport aux sommets 1 et 2 sont les graines (2.2.2) ß ™ / 3 , 1 + x2 , x2 , x3 1o 2 et 1o x1

&

2o

3,

ß ™ x1 + x3 x1 , , x3 . x2

Fixons maintenant un bon carquois Q. La graine initiale est (Q, {x1 , . . . , xn }). Un amas associé à Q est une suite u qui apparaît dans une graine (R, u) obtenue à partir de la graine initiale par mutation itérée. Les variables d’amas sont les éléments des amas. L’algèbre amassée AQ est la sous-algèbre de F engendrée par les variables d’amas. Clairement, si (R, u) est une graine associée à Q, l’isomorphisme naturel ∼ Q(u1 , . . . , un ) → Q(x1 , . . . , xn )

induit un isomorphisme de AR sur AQ qui préserve les variables d’amas et les amas. L’algèbre amassée AQ est donc un invariant de la classe de mutation de Q. Il est utile d’introduire un objet combinatoire qui code la construction récursive des graines : le graphe d’échange. Par définition, ses sommets sont les classes d’isomorphisme de graines (les isomorphismes renumérotent les sommets et les variables) et ses arêtes correspondent aux mutations. Par exemple, le graphe d’échange obtenu à partir du /2 / 3 est le 1-squelette de l’associaèdre de Stasheff [97] carquois Q : 1 [19] : 2 ◦





◦ ◦ ◦



1 ◦

3





◦ Le sommet 1 correspond à la graine initiale et les sommets 2 et 3 aux graines 2.2.2. Fixons un bon carquois connexe Q. Si son graphe sous-jacent est un diagramme de Dynkin simplement lacé de type ∆, nous disons que Q est un carquois de Dynkin de type ∆. Théorème 2.1 ([39]). — (a) Toute variable d’amas de AQ est un polynôme de Laurent à coefficients entiers [38].

ASTÉRISQUE 339

(1014)

ALGÈBRES AMASSÉES ET APPLICATIONS

69

(b) L’algèbre amassée AQ n’a qu’un nombre fini de variables d’amas si et seulement si Q est équivalent par mutation à un carquois de Dynkin. Dans ce cas, le graphe ∆ sous-jacent à ce carquois est unique et s’appelle le type amassé de Q. (c) Si Q est un carquois de Dynkin de type ∆, alors les variables d’amas non initiales de AQ sont en bijection avec les racines positives du système de racines Φ de ∆ ; plus précisément, si α1 , . . . , αn sont les racines simples, alors pour toute racine positive α = d1 α1 + · · · + dn αn , il existe une unique variable d’amas non initiale Xα de dénominateur xd11 · · · xdnn . Un monôme d’amas est un produit de puissances positives de variables d’amas qui appartiennent toutes au même amas. La construction d’une « base canonique » de l’algèbre amassée AQ est un problème important et encore très largement ouvert, voir par exemple [96] [26] [17]. On s’attend à ce qu’une telle base contienne tous les monômes d’amas, d’où la conjecture : Conjecture 2.2 ([39]). — Les monômes d’amas sont linéairement indépendants sur le corps Q. Si Q est un carquois de Dynkin, on sait [15] que les monômes d’amas forment une base de AQ . Si Q est acyclique, c’est-à-dire n’admet aucun cycle orienté, la conjecture résulte d’un théorème de Geiss-Leclerc-Schröer [51], qui montrent l’existence d’une « base générique » contenant les monômes d’amas. La conjecture a aussi été démontrée pour des classes d’algèbres amassées à coefficients (voir la section 3), par exemple dans les travaux [44] [51] [21]. Conjecture 2.3 ([39]). — Les variables d’amas s’écrivent comme des polynômes de Laurent à coefficients entiers positifs en les variables de tout amas. La catégorification monoïdale développée par Leclerc [80] et Hernandez-Leclerc [58] (voir la section 4.3) a permis récemment de montrer cette conjecture d’abord pour les carquois de type An et D4 , voir [58], puis pour tout carquois admettant une orientation bipartite [88], c’est-à-dire une orientation où tout sommet est une source ou un puits. Elle est démontrée de façon combinatoire par Musiker-Schiffler-Williams [86] pour tous les carquois associés à des surfaces à bord marquées [37] et par Di Francesco-Kedem [25] pour les carquois associés au T -système de type A. Nous renvoyons à [40] et [42] pour de nombreuses autres conjectures sur les algèbres amassées et à [23] pour la solution d’une bonne partie de ces conjectures grâce à la catégorification (voir la section 4.4).

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

70

B. KELLER

2.3. Y -graines, application à la conjecture de périodicité Soient n ≥ 1 un entier et G le corps Q(y1 , . . . , yn ) engendré par des indéterminées yi . Une Y -graine est un couple (R, v), où R est un bon carquois et v une suite v1 , . . . , vn qui engendre librement le corps G (nous nous écartons quelque peu de la définition dans [42]). Si (R, v) est une Y -graine et k un sommet de R, la mutation µk (R, v) est la Y -graine (R0 , v 0 ), où R0 = µk (R) et   vi−1 si i = k,      v (1 + v )m si le nombre de flèches i → k est m ≥ 1, i k vi0 = −1 −m  vi (1 + vk ) si le nombre de flèches k → i est m ≥ 1,      v sinon. i Par exemple, les Y -graines obtenues à partir de y1 → y2 par mutation sont, en écrivant les variables vi à la place des sommets i : ã Å ã Å ã Å y2 y1 y2 1 + y1 µ1 1 + y1 + y1 y2 1 1 µ2 ← 7→ → 7→ ← y1 1 + y1 1 + y1 + y1 y2 y2 y2 y1 (1 + y2 ) Å ã 1 µ2 µ1 7→ → y1 (1 + y2 ) 7→ (y2 ← y1 ) . y2 Les Y -graines jouent un rôle important dans la théorie de Teichmüller supérieure de Fock-Goncharov [32] [31] et dans l’étude par Kontsevich-Soibelman [75] des invariants de Donaldson-Thomas des carquois à potentiel [22]. Elles sont liées aux X-graines par des conjectures de dualité [32] étudiées systématiquement par Fomin-Zelevinsky dans [42]. En particulier, dans [42], les auteurs montrent que les deux types de graines se déterminent mutuellement si, en même temps que AQ , on considère aussi AQ e, où ‹ est l’extension principale de Q obtenue à partir de Q en rajoutant de nouveaux Q sommets n + 1, . . . , 2n et une nouvelle flèche (n + i) → i pour tout 1 ≤ i ≤ n. Ces liens combinés avec la catégorification additive (voir section 4.1) ont permis récemment une application des algèbres amassées à l’étude de systèmes dynamiques discrets issus de la physique mathématique. Soient ∆ et ∆0 deux diagrammes de Dynkin simplement lacés. Notons 1, . . . , n et 1, . . . , n0 leurs sommets et A et A0 leurs matrices d’incidence, le coefficient en position (i, j) valant 1 s’il existe une arête entre i et j et 0 sinon. Notons h et h0 les nombres de Coxeter de ∆ et ∆0 . Le Y -système associé à ∆ et ∆0 est un système infini d’équations de récurrence en des variables Yi,j,t associées aux sommets (i, j) du produit ∆ × ∆0 et dépendant d’un paramètre de temps discret t ∈ Z. Les équations du Y -système sont Qn aij j=1 (1 + Yj,i0 ,t ) (2.3.1) Yi,i0 ,t−1 Yi,i0 ,t+1 = Q 0 , 0 n −1 ai0 j 0 j 0 =1 (1 + Yi,j 0 ,t ) pour tout sommet (i, i0 ) du produit et tout entier t.

ASTÉRISQUE 339

(1014)

ALGÈBRES AMASSÉES ET APPLICATIONS

71

Théorème 2.4. — Toutes les solutions du Y -système sont périodiques par rapport au paramètre t de période divisant 2(h + h0 ). Ce théorème vient confirmer la « conjecture de périodicité » formulée par Al. B. Zamolodchikov [103, (12)] pour ∆0 = A1 , par Kuniba-Nakanishi [77, (2a)] pour ∆0 = Am et par Ravanini-Valleriani-Tateo [90, (6.2)] dans le cas général. Le théorème a été démontré – pour (An , A1 ) par Frenkel-Szenes [43] (qui donnent des solutions explicites) et par Gliozzi-Tateo [56] (à l’aide de calculs de volumes de 3-variétés) ; – par Fomin-Zelevinsky [41] pour (∆, A1 ), où ∆ n’est pas nécessairement simplement lacé (ils utilisent les méthodes de leur théorie des algèbres amassées et un calcul sur ordinateur pour les types exceptionnels ; ce calcul peut être évité maintenant grâce à [102]) ; – pour (An , Am ) par Volkov [100], qui construit des solutions explicites grâce à des considérations de géométrie projective élémentaire, et par Szenes [98], qui interprète le système comme un système de connexions plates sur un graphe ; la démonstration d’un énoncé équivalent est due à Henriques [57] ; – pour (∆, ∆0 ) quelconques dans [71] [72] à l’aide de la catégorification additive, voir la section 4.1. Pour des diagrammes non simplement lacés ∆ et ∆0 , deux variantes généralisées de la conjecture existent : la première se ramène au théorème 2.4 par la technique du « pliage », voir [41] ; la deuxième, formulée par Kuniba-Nakanishi [77, (2a)] et Kuniba-Nakanishi-Suzuki [78, B.6], fait intervenir le double de la somme des nombres de Coxeter duaux (voir par exemple le chapitre 6 de [67]) ; elle a été démontrée dans [61] [62].

3. ALGÈBRES AMASSÉES À COEFFICIENTS Nous allons généraliser légèrement la définition donnée à la section 2 pour obtenir la classe des « algèbres amassées antisymétriques de type géométrique ». Cette classe contient de nombreuses algèbres d’origine géométrique munies de « bases duales semi-canoniques ». La construction d’une grande partie d’une telle base est l’une des applications les plus remarquables des algèbres amassées. Nous renvoyons à [42] pour la définition des « algèbres amassées antisymétrisables à coefficients dans un semi-corps », qui constituent la classe la plus générale considérée jusqu’à maintenant.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

72

B. KELLER

3.1. Définition ‹ un carquois glacé de type (n, m), c’est-à-dire Soient 1 ≤ n ≤ m des entiers. Soit Q un bon carquois à m sommets et qui ne comporte aucune flèche entre sommets i, j ‹ est le soustous les deux strictement plus grands que n. La partie principale de Q carquois plein Q dont les sommets sont 1, . . . , n (un sous-carquois est plein si, avec deux sommets, il contient toutes les flèches qui les relient). Les sommets n + 1, . . . , m ‹ sont les sommets gelés. L’algèbre amassée associée au carquois glacé Q AQ e ⊂ Q(x1 , . . . , xm ) est définie de la même façon que l’algèbre amassée associée à un carquois (section 2) sauf que – seules les mutations par rapport à des sommets non gelés sont admises et aucune flèche entre sommets gelés n’est introduite lors des mutations ; – les variables xn+1 , . . . , xm , qui font partie de tous les amas, sont appelées coefficients plutôt que variables d’amas ; – le type amassé du carquois glacé est celui de sa partie principale (s’il est défini). Souvent, on considère des localisations de AQ e obtenues en inversant certains des coefficients. Si K est une extension de Q et A une K-algèbre (associative avec 1), ‹ sur A est la donnée d’un une structure d’algèbre amassée à coefficients de type Q isomorphisme ϕ de AQ e ⊗Q K sur A. Un tel isomorphisme est déterminé par les images des coefficients et des variables de la graine initiale ϕ(xi ), 1 ≤ i ≤ m. Nous appellerons ‹ et des ϕ(xi ). graine initiale de A la donnée du carquois Q 3.2. Exemple : le cône sur la grassmannienne des plans d’un espace vectoriel Soit n ≥ 1 un entier. Soit A l’algèbre des fonctions polynomiales sur le cône audessus de la grassmannienne des plans de Cn+3 . Cette algèbre est engendrée par les coordonnées de Plücker xij , 1 ≤ i < j ≤ n + 3, assujetties aux relations de Plücker : pour tout quadruplet d’entiers i < j < k < l compris entre 1 et n + 3, nous avons (3.2.1)

xik xjl = xij xkl + xjk xil .

Notons que les monômes dans cette relation sont naturellement associés aux diagonales et aux côtés du carré i j

l k L’idée est d’interpréter cette relation comme une relation d’échange dans une algèbre amassée (à coefficients). Pour décrire cette algèbre, considérons, dans le plan affine

ASTÉRISQUE 339

(1014)

73

ALGÈBRES AMASSÉES ET APPLICATIONS

euclidien, un polygone régulier P dont les sommets sont numérotés de 1 à n + 3. Considérons la variable xij comme associée au segment [ij] joignant les sommets i et j. Proposition 3.1 ([39, Example 12.6]). — L’algèbre A a une structure d’algèbre amassée à coefficients telle que – les coefficients soient les variables xij associées aux côtés de P ; – les variables d’amas soient les variables xij associées aux diagonales de P ; – les amas soient les n-uplets de variables d’amas correspondant à des diagonales qui forment une triangulation de P . En outre, les relations d’échange sont exactement les relations de Plücker et le type amassé est An . Une triangulation de P détermine une graine initiale pour l’algèbre amassée et les relations d’échange vérifiées par les variables d’amas initiales déterminent le carquois ‹ Par exemple, on vérifie que, dans le dessin suivant, la triangulation et le glacé Q. carquois glacé (dont les sommets gelés sont entourés de boîtes) se correspondent 0 5

05 U 1 45

4

01

/ 04 g ? 03

 34

2 3

02 O w

/ 12

 23

De nombreuses autres algèbres de coordonnées (homogènes) de variétés algébriques classiques admettent également des structures d’algèbres amassées (supérieures, voir la section 3.3), notamment les grassmanniennes [95] et les doubles cellules de Bruhat [3]. Certaines de ces algèbres n’ont qu’un nombre fini de variables d’amas et donc un type amassé bien défini. Voici quelques exemples extraits de [40], où N est un sous-groupe unipotent maximal : Gr2,n+3

Gr3,6

Gr3,7

Gr3,8

SL3 /N

SL4 /N

SL5 /N

Sp4 /N

SL2

SL3

An

D4

E6

E8

A1

A3

D6

B2

A1

D4

Un analogue de la proposition 3.1 pour les doubles cellules de Bruhat réduites [4] est dû à Yang et Zelevinsky [102]. Ils obtiennent ainsi une algèbre amassée (à coefficients principaux) avec une description explicite des variables d’amas pour tout diagramme de Dynkin.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

74

B. KELLER

3.3. Exemple : le sous-groupe unipotent maximal de SL(n + 1, C) Soient n un entier positif et N le sous-groupe de SL(n + 1, C) formé des matrices triangulaires supérieures dont les coefficients diagonaux sont tous égaux à 1. Pour 1 ≤ i, j ≤ n + 1 et g ∈ N , soit Fij (g) la sous-matrice carrée de taille maximale de g qui comporte le coefficient gij dans son coin inférieur gauche. Soit fij (g) le déterminant de Fij (g). Nous considérons les fonctions polynomiales fij : N → C pour 1 ≤ i ≤ n ‹à m et i + j ≤ n + 2. L’algèbre amassée supérieure associée à un carquois glacé Q sommets est la sous-algèbre de Q(x1 , . . . , xm ) formée des éléments qui s’expriment ‹ comme des polynômes de Laurent en les variables de tout amas associé à Q. Théorème 3.1 ([3]). — L’algèbre des fonctions polynomiales C[N ] a une structure d’algèbre amassée supérieure dont la graine initiale est donnée par f12 O

f22

~

.. .O

/ f13 O

/ f14

/ f23

/ ···

/ ··· O

/ f1,n+1 |

f2,n

/ ··· ~

fn,2

Il n’est pas difficile de vérifier que cette structure est de type amassé A3 pour n = 3, D6 pour n = 4 et qu’elle a une infinité de variables d’amas pour n ≥ 5. Un théorème de Fekete [27], généralisé dans [2], affirme qu’une matrice carrée à n+1 lignes et à coefficients réels est totalement positive (i.e. tous ses mineurs sont > 0) si les (n + 1)2 mineurs suivants sont strictement positifs : tous les mineurs formés des k premières lignes et de k colonnes consécutives pour 1 ≤ k ≤ n+1. Une matrice g ∈ N à coefficients réels est totalement positive si tous les mineurs non identiquement nuls sur N sont strictement positifs en g. Cette condition est équivalente à ce que l’on ait fij (g) > 0 pour les fij de la graine initiale du théorème. Comme les relations d’échange ne font pas intervenir de soustraction, tout amas non initial C donne également un critère de positivité : la matrice g ∈ N est totalement positive si et seulement si l’on a uij (g) > 0 pour toute variable d’amas uij dans C.

ASTÉRISQUE 339

(1014)

ALGÈBRES AMASSÉES ET APPLICATIONS

75

3.4. Exemple d’application aux bases duales semi-canoniques L’exemple 3.3 se généralise. Soient, en effet, G un groupe algébrique semi-simple complexe et N un sous-groupe unipotent maximal de G. Alors d’après [3], l’algèbre C[N ] est munie d’une structure canonique d’algèbre amassée supérieure. Soit n l’algèbre de Lie du groupe algébrique N . Dans [83], Lusztig construit une base distinguée, la base semi-canonique, de (l’espace vectoriel complexe sous-jacent à) l’algèbre enveloppante U (n). Le dual restreint de la cogèbre U (n) est canoniquement isomorphe à C[N ], qui est donc muni de la base duale de celle construite par Lusztig, appelée base duale semi-canonique. Théorème 3.2 (Geiss-Leclerc-Schröer [48]). — Tout monôme d’amas de C[N ] (munie de la structure d’algèbre amassée du théorème 3.1) fait partie de la base duale semi-canonique. Notons que ce théorème implique la conjecture 2.2 sur l’indépendance des monômes d’amas pour cette classe d’algèbres amassées. L’algèbre U (n) est également munie de la base canonique obtenue par spécialisation à partir de la base canonique [68] [81] du groupe quantique Uq (n). Théorème 3.3 (Geiss-Leclerc-Schröer [47]). — Pour G = SL(n + 1, C), la base canonique coïncide avec la base semi-canonique de U (n) si et seulement si n ≤ 4. Néanmoins, Geiss-Leclerc-Schröer conjecturent qu’au moins les « parties rigides » des bases duales canonique et semi-canonique coïncident. Plus précisément, un cas particulier de la conjecture 23.2 de [51] nous donne la conjecture qui suit. Conjecture 3.4 (Geiss-Leclerc-Schröer [51]). — Tout monôme d’amas de C[N ] appartient aussi à la base duale canonique. Dans un travail de longue haleine qui a abouti à [51], Geiss-Leclerc-Schröer ont généralisé les théorèmes 3.1 et 3.2 de l’algèbre C[N ] aux algèbres de coordonnées de cellules unipotentes de groupes de Kac-Moody simplement lacés. Nous renvoyons à [50] pour une introduction et une synthèse des résultats dans le cas fini, et à [21] pour une extension (partielle) au cas non simplement lacé.

4. CATÉGORIFICATIONS Les démonstrations du théorème de périodicité (Théorème 2.4) et du théorème sur la base duale semi-canonique (Théorème 3.2) s’appuient sur la « catégorification additive » des algèbres amassées ; celle des cas connus de la conjecture de positivité 2.3 sur la « catégorification monoïdale ». Nous allons esquisser les idées principales de ces

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

76

B. KELLER

méthodes. La section 4.4 est consacrée à la méthode de la catégorification à l’aide des « carquois à potentiel ». À ce jour, c’est la seule méthode qui permette de traiter des algèbres amassées associées à des carquois finis arbitraires (sans boucles ni 2-cycles). 4.1. Catégorification additive : la catégorie amassée Soit Q un carquois fini d’ensemble de sommets {1, . . . , n}. Un chemin de Q est une composition formelle (j|αs | · · · |α1 |i) d’un nombre s positif ou nul de flèches αi telle que s(αi ) = t(αi−1 ) pour 1 ≤ i ≤ s. En particulier, pour tout sommet i, nous avons le chemin paresseux ei = (i|i) de longueur nulle, neutre pour la composition naturelle des chemins. Une représentation (complexe) de Q est la donnée V d’espaces vectoriels complexes de dimension finie Vi , i ∈ Q0 , et d’applications linéaires Vα : Vi → Vj pour toute flèche α : i → j de Q. Une représentation est donc un diagramme d’espaces vectoriels de la forme donnée par Q. Un morphisme de représentations est un morphisme de diagrammes. On obtient ainsi la catégorie rep(Q) des représentations de Q. C’est une catégorie abélienne équivalente à la catégorie des modules de C-dimension finie sur une algèbre, à savoir l’algèbre des chemins CQ (une base de cette algèbre est formée des chemins de Q ; le produit de deux chemins composables est leur composition, le produit de chemins non composables est nul). En particulier, nous avons des notions naturelles de sous-représentation, de représentation simple, de somme directe et de représentation indécomposable (= représentation non nulle qui n’est pas somme directe de deux sous-représentations non nulles). Supposons que Q est un carquois de Dynkin de type ∆. Alors d’après le théorème de Gabriel [45], on a une bijection de l’ensemble des classes d’isomorphisme de représentations indécomposables de Q sur l’ensemble des racines positives de ∆ ; à une P représentation indécomposable V , cette bijection associe la racine ni=1 (dim Vi )αi , où les αi sont les racines simples. En composant cette bijection avec celle de la partie c) du théorème 2.1 de Fomin-Zelevinsky, nous obtenons une bijection de l’ensemble des classes d’isomorphisme de représentations indécomposables sur l’ensemble des variables d’amas non initiales de l’algèbre amassée AQ : à une représentation indécomposable V , cette bijection associe l’unique variable d’amas non initiale XV dont le dénominateur est xd11 · · · xdnn , où di = dim Vi . Il est remarquable que le numérateur de XV admette aussi une interprétation naturelle en termes de la représentation V . Pour expliciter cette interprétation, nous avons besoin de quelques notations supplémentaires : soient V une représentation quelconque de Q et di = dim Vi , i ∈ Q0 . Pour un élément e ∈ Nn , notons Gre (V ) l’ensemble des sous-représentations U de V telles que dim Ui = ei . La donnée d’un point de Gre (V ) est donc la donnée d’une famille de sous-espaces vectoriels Ui ⊂ Vi telle que Ui soit de dimension ei et que Vα (Ui ) ⊂ Uj pour toute flèche α : i → j de Q. Cette description montre que Gre (V ) est une sousvariété fermée du produit des grassmanniennes Grei (Vi ). En particulier, c’est une

ASTÉRISQUE 339

(1014)

ALGÈBRES AMASSÉES ET APPLICATIONS

77

variété projective (singulière en général). Elle est appelée grassmannienne des sousreprésentations (quiver Grassmannian) et étudiée dans [16], par exemple. On note χ(Gre (Vi )) la caractéristique d’Euler-Poincaré de son espace topologique sous-jacent. Posons P P n X Y e + (dj −ej ) 1 j→i j i→j CC(V ) = d1 , χ(Gre (V )) xi dn x1 · · · xn e i=1 où les sommes dans l’exposant portent sur les flèches de but i respectivement de source i. Théorème 4.1 (Caldero-Chapoton [12]). — Si V est indécomposable, nous avons XV = CC(V ). Notons que la formule pour CC(V ) a un sens pour toute représentation de tout carquois fini Q. Supposons maintenant que Q est un carquois sans cycles orientés quelconque. Une représentation V de Q est rigide si son groupe d’auto-extensions Ext1 (V, V ) dans la catégorie rep(Q) s’annule. La partie c) du théorème 2.1 ne s’applique plus, mais nous avons néanmoins une paramétrisation des variables d’amas non initiales en termes de représentations de Q. Théorème 4.2 ([14]). — L’application V 7→ CC(V ) induit une bijection de l’ensemble des classes d’isomorphisme de représentations rigides indécomposables de Q sur l’ensemble des variables d’amas non initiales de AQ . Ce théorème fournit une interprétation catégorique de la quasi-totalité des variables d’amas de l’algèbre amassée. Il se pose la question d’étendre cette interprétation aux variables initiales, aux relations d’échange et aux amas. Pour ce faire, on agrandit la catégorie des représentations : soit DQ la catégorie dérivée bornée de la catégorie abélienne rep(Q). Les objets de DQ sont donc les complexes bornés de représentations et ses morphismes sont obtenus à partir des morphismes de complexes en inversant formellement les quasi-isomorphismes. La catégorie DQ est une catégorie triangulée ; on note Σ son foncteur suspension (qui n’est autre que le foncteur de décalage des complexes X 7→ X[1]). Les ensembles de morphismes de DQ sont des espaces vectoriels de dimension finie et DQ admet un foncteur de Serre, c’est-à-dire une auto-équivalence S : DQ → DQ telle qu’on ait des isomorphismes bifonctoriels D Hom(X, Y ) = Hom(Y, SX) , où D = HomC (?, C) est la dualité des espaces vectoriels complexes. La catégorie amassée est la catégorie d’orbites CQ = DQ /(S −1 ◦ Σ2 )Z

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

78

B. KELLER

de DQ sous l’action du groupe cyclique engendré par l’automorphisme S −1 ◦ Σ2 . Elle est due à Buan-Marsh-Reineke-Reiten-Todorov [9] et, de façon indépendante et sous une forme très différente, à Caldero-Chapoton-Schiffler [13] pour les carquois de Dynkin de type A. La catégorie CQ est canoniquement triangulée [70]. Ses espaces de morphismes sont de dimension finie et son foncteur de Serre (induit par S) est isomorphe au carré de son foncteur suspension (induit par Σ). Cela signifie que CQ est Calabi-Yau de dimension 2. Notons π : DQ → CQ le foncteur de projection canonique et Ext1 (L, M ) = Hom(L, ΣM ) pour des objets L et M de CQ . Grâce à la propriété de Calabi-Yau, on a D Ext1 (L, M ) = Ext1 (M, L). On appelle rigide un objet L tel que Ext1 (L, L) s’annule. Notons Pi la représentation qui correspond au CQ-module CQei , i ∈ Q0 . On peut montrer [9] que tout objet L de CQ se décompose de façon unique (à isomorphisme près) sous la forme M L = π(M ) ⊕ Σπ(Pi )mi , i∈Q0

pour une représentation M et des multiplicités mi . On pose Y i CC(L) = CC(M ) · xm i . i∈Q0

Théorème 4.3 ([14]). — a) On a CC(L ⊕ M ) = CC(L) · CC(M ) pour tous L et M dans CQ , b) si L et M sont des objets de CQ tels que Ext1 (L, M ) soit de dimension 1 et L → E → M → ΣL et M → E 0 → L → ΣM soient deux triangles non scindés, on a (4.1.1)

CC(L) · CC(M ) = CC(E) + CC(E 0 ).

c) L’application CC induit une bijection de l’ensemble des objets rigides de CQ sur l’ensemble des monômes d’amas de AQ . d) Par cette bijection, les objets rigides indécomposables correspondent aux variables d’amas, et un ensemble d’indécomposables rigides T1 , . . . , Tn correspond à un amas si et seulement si Ext1 (Ti , Tj ) = 0 pour tous i, j. Les propriétés a) et b) fournissent une interprétation des relations d’échange. La démonstration du théorème est fondée sur le travail de plusieurs groupes d’auteurs : Buan-Marsh-Reiten-Todorov [11], Buan-Marsh-Reiten [10], Buan-Marsh-ReinekeReiten-Todorov [9], Marsh-Reineke-Zelevinsky [84], ... et surtout Caldero-Chapoton [12]. Une autre démonstration de la formule de multiplication 4.1.1 est due à Hubery

ASTÉRISQUE 339

(1014)

ALGÈBRES AMASSÉES ET APPLICATIONS

79

[59] pour des carquois dont le graphe sous-jacent est un diagramme de Dynkin étendu, et à Xiao-Xu [101] dans le cas général. La construction de la catégorie amassée a été généralisée des algèbres CQ à une classe d’algèbres de dimension globale 2 par Amiot [1]. Une version généralisée de l’application de Caldero-Chapoton et de la formule de multiplication 4.1.1 est due à Palu [89]. L’extension des résultats de Palu au cas de certaines algèbres amassées à coefficients est obtenue dans [44]. La mutation dans une catégorie 2-Calabi-Yau générale est construite par Iyama-Yoshino [65]. La démonstration du théorème de périodicité 2.4 dans [71] [72] est fondée sur ces travaux. 4.2. Catégorification additive : modules sur les algèbres préprojectives Nous allons décrire l’idée de base de la démonstration du théorème 3.2. Soient ∆ un diagramme de Dynkin, g l’algèbre de Lie simple complexe qui lui correspond et n une sous-algèbre nilpotente maximale de g. Soit N le groupe algébrique unipotent associé à n. L’algèbre de coordonnées C[N ] est le dual restreint de la cogèbre U (n). La base duale semi-canonique de l’algèbre C[N ] est duale de la base semi-canonique de U (n) construite par Lusztig [83]. Dans un premier temps, nous allons décrire (suivant [48]) la base duale semi-canonique en termes de modules sur l’algèbre préprojective : soit Q un carquois de Dynkin de type ∆. Soit Q le double carquois, obtenu à partir de Q en rajoutant une flèche α∗ : j → i pour chaque flèche α : i → j. Soit Λ l’algèbre préprojective de Q, c’est-à-dire le quotient de l’algèbre des chemins CQ (voir P la section 4.1) par l’idéal bilatère engendré par la somme [α, α∗ ] prise sur l’ensemble des flèches de Q. C’est une algèbre de dimension finie sur C qui est auto-injective (c’està-dire injective comme module sur elle-même). Appelons Λ-module un Λ-module à gauche de dimension finie sur C. Pour un tel module M , son vecteur dimension est la suite des entiers dim ei M , i ∈ Q0 . Soit d une famille d’entiers positifs indexés par Q0 . On note rep(Λ, d) la variété formée des familles de matrices Mα : Cds(α) → Cdt(α) , α ∈ Q1 , qui vérifient les relations de Λ, c’est-à-dire définissent une structure de Λ-module sur la somme directe des Cdi , i ∈ Q0 . La variété rep(Λ, d) porte une action naturelle par Q « changement de base » du groupe Gd = GL(di , C) et les orbites de cette action sont en bijection avec les classes d’isomorphisme de Λ-modules de vecteur dimension d. Notons Md l’espace vectoriel des fonctions constructibles et Gd -invariantes sur la variété rep(Λ, d). Notons U (n)d la composante graduée de U (n) associée au vecteur P di αi , où les αi sont les racines simples. Lusztig [83] a défini une injection linéaire λd de U (n)d dans Md . Chaque Λ-module M définit une forme linéaire sur Md , à savoir la forme qui, à une fonction f , associe sa valeur f (M ) en l’orbite déterminée par M .

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

80

B. KELLER

Par composition, le module M nous donne une forme linéaire λ

δM : U (n)d →d Md → C. Or nous avons l’isomorphisme canonique ι : U (n)∗d → C[N ]d . Comme dans [48], nous posons ϕM = ι(δM ). Nous obtenons ainsi une application M 7→ ϕM de la classe des Λ-modules dans l’algèbre C[N ]. On peut expliciter cette application en termes de caractéristiques d’Euler-Poincaré de variétés de drapeaux de sous-représentations de M , voir [48]. Cette description montre que l’application M 7→ ϕM est constructible sur la variété algébrique rep(Λ, d). Donc chaque composante irréductible de cette variété contient un ouvert dense où la fonction M 7→ ϕM est constante. On appelle génériques les modules M appartenant à de tels ouverts. Alors la base duale semicanonique n’est autre que {ϕM | M est générique}. Un module M est rigide si l’espace Ext1 (M, M ) s’annule. De façon équivalente [48], l’orbite de M dans rep(Λ, d), où d est le vecteur dimension de M , est ouverte. En particulier, si M est rigide, alors il est générique et la fonction ϕM appartient à la base duale semi-canonique. Pour montrer le théorème 3.2, il suffit donc de montrer que chaque monôme d’amas est de la forme ϕM pour un module rigide M . Pour cela, on procède par récurrence : on montre [49] que les éléments de la graine initiale sont des (0) (0) (0) images de modules rigides indécomposables canoniques T1 , T2 , . . . , Tm . Puis on relève l’opération de mutation des amas à une classe convenable de suites T1 , . . . , Tm de modules rigides indécomposables. Appelons accessibles les modules rigides dont les facteurs directs indécomposables sont obtenus par mutation itérée à partir de la suite (0) (0) (0) T1 , T2 , . . . , Tm . Le théorème suivant est l’analogue précis du théorème 4.3. Ses parties a) et b) permettent de relier la mutation des amas à la mutation des modules rigides et donc d’effectuer la récurrence qui termine la démonstration du théorème 3.2. Théorème 4.4 (Geiss-Leclerc-Schröer [48]). — On a a) ϕL⊕M = ϕL ϕM , b) Si Ext1 (L, M ) est de dimension 1 et que l’on a les suites exactes non scindées 0 → L → E → M → 0 et 0 → M → E 0 → L → 0 , alors on a ϕL ϕM = ϕE + ϕE 0 . c) L’application M 7→ ϕM induit une bijection de l’ensemble des classes d’isomorphisme de modules rigides accessibles sur l’ensemble des monômes d’amas. d) Les rigides indécomposables accessibles correspondent aux variables d’amas et une suite T1 , . . . , Tm de tels modules correspond à un amas si et seulement si l’on a Ext1 (Ti , Tj ) = 0 pour tous i et j.

ASTÉRISQUE 339

(1014)

81

ALGÈBRES AMASSÉES ET APPLICATIONS

4.3. Catégorification monoïdale Les catégorifications additives décrites ci-dessus se sont avérées très utiles et on sait les construire pour de grandes classes d’algèbres amassées. De l’autre côté, elles semblent difficiles à exploiter pour démontrer la conjecture de positivité 2.3, et la notion même de catégorification additive semble peu naturelle. La catégorification monoïdale, introduite par Leclerc [80] et Hernandez-Leclerc [58], consiste à réaliser une algèbre amassée comme l’anneau de Grothendieck d’une catégorie abélienne monoïdale. Elle est donc très naturelle. En outre, comme nous allons le voir, son existence donne immédiatement la conjecture de positivité 2.3 (et la conjecture d’indépendance 2.2). De l’autre côté, les catégorifications monoïdales semblent très difficiles à construire. Les notions suivantes [80] sont fondamentales pour la suite : un objet simple S d’une catégorie abélienne monoïdale est premier s’il n’admet pas de factorisation tensorielle non triviale ; il est réel si son carré tensoriel est encore simple. Soient A une algèbre amassée à coefficients et AZ son sous-anneau engendré par les variables d’amas et les coefficients. Suivant [58], une catégorification monoïdale de A est la donnée d’une catégorie abélienne monoïdale M et d’un isomorphisme d’anneaux ∼ ϕ : AZ → K0 (M ) tel que ϕ induise a) une bijection de l’ensemble des monômes d’amas sur l’ensemble des classes d’objets simples réels de M et b) une bijection de l’ensemble des variables d’amas et des coefficients sur l’ensemble des classes d’objets simples, réels et premiers de M . Le tableau suivant, extrait de [80], résume les correspondances entre les structures associées à une algèbre amassée et leurs relèvements dans une catégorification additive respectivement monoïdale. algèbre amassée A

catégorification additive C

catégorification monoïdale M

+

?



×





monôme d’amas

objet rigide

objet simple réel

variable d’amas

indécomposable rigide

simple premier réel

L’existence d’une catégorification monoïdale M d’une algèbre amassée A a des conséquences très fortes pour A : en effet, l’algèbre A est alors munie d’une « base canonique », à savoir la base fournie par les objets simples de M et cette base contient les monômes d’amas car ceux-ci correspondent bijectivement aux classes dans K0 (M )

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

82

B. KELLER

de certains objets simples. En particulier, la conjecture d’indépendance est vérifiée pour A . De même, la conjecture de positivité est vérifiée pour A . En effet, si on exprime une variable d’amas x comme polynôme de Laurent P (u1 , . . . , um ) ud11 · · · udmm

x=

en les variables d’un amas u1 , . . . , um , alors les coefficients de P sont les multiplicités de certains objets simples dans la classe du produit tensoriel ϕ(xud11 · · · udmm ) et sont donc des entiers positifs. L’existence d’une catégorification monoïdale ϕ : AZ → K0 (M ) donne également des renseignements précieux sur la structure monoïdale de M : en effet, elle montre que le comportement des objets simples réels de M est gouverné par la combinatoire des amas de A . Dans [58], Hernandez-Leclerc exhibent des catégorifications monoïdales conjectu‹ rales Ml pour les algèbres amassées Al associées à certains carquois glacés Q(∆, l), où ∆ est un diagramme de Dynkin simplement lacé et l ∈ N un « niveau ». Voici ‹ 5 , 3), où les sommets gelés sont marqués par des •. l’exemple du carquois Q(D /◦o /• ◦J ◦J /

◦Y

◦o 

◦O

◦o 

◦O ◦o

/

◦Y

/

/

◦o 

◦o 





/

◦O ◦o

/

• • •

Hernandez-Leclerc construisent les catégories Ml comme des sous-catégories monoïdales de la catégorie des représentations de dimension finie de l’algèbre affine quantique Uq (b g) associée à ∆. Ils construisent un morphisme d’anneaux ϕ : (Al )Z → K0 (Ml ) qui envoie les variables de l’amas initial sur les classes de certains modules de Kirillov-Reshetikhin et conjecturent que ϕ est une catégorification monoïdale de Al (Conjecture 13.2 de [58]). Ils démontrent leur conjecture pour l ≤ 1 et ∆ de type An , n ≥ 1, ou D4 ainsi que pour ∆ de type A2 et l = 2 (et observent que pour ∆ = A1 et tout l ∈ N, la conjecture résulte du travail de Chari-Pressley [20]). Dans [88], Nakajima construit des catégorifications monoïdales conjecturales Nl ‹ pour les carquois Q(R, l) associés à un carquois R bipartite et un niveau l ∈ N. Les catégories Nl sont réalisées comme des catégories de faisceaux pervers sur des variétés de carquois gradués [87] munies du produit tensoriel construit géométriquement

ASTÉRISQUE 339

(1014)

ALGÈBRES AMASSÉES ET APPLICATIONS

83

dans [99]. Si R est un carquois de Dynkin et l = 1, la catégorie Nl est équivalente à Ml et Nakajima démontre qu’elle est une catégorification monoïdale de l’algèbre Al confirmant ainsi la conjecture de Hernandez-Leclerc. Pour un carquois bipartite R quelconque et l = 1, il montre que l’anneau (Al )Z se plonge dans K0 (Nl ) de telle façon que les monômes d’amas sont envoyés sur des objets simples. Ceci entraîne la conjecture de positivité pour Al . 4.4. Catégorification via les carquois à potentiel Inspirés par des travaux de physiciens (voir par exemple la section 6 dans [29]) Derksen-Weyman-Zelevinsky ont étendu [22] l’opération de mutation des carquois aux carquois à potentiel et leurs représentations décorées. Décrivons brièvement ces d l’algèbre des notions en suivant [22]. Soit en effet Q un carquois fini. Notons CQ chemins complétée, c’est-à-dire la complétion de CQ par rapport à l’idéal bilatère engendré par les flèches de Q. L’espace CQ admet donc une base topologique formée de d est le complété tous les chemins de Q. L’homologie de Hochschild continue HH0 (CQ) de l’espace quotient de CQ par le sous-espace [CQ, CQ] engendré par tous les commutateurs. Il admet une base topologique formée de tous les cycles de Q, c’est-à-dire les orbites sous l’action du groupe cyclique Z/tZ de chemins cycliques de longueur t ≥ 0. Pour chaque flèche α de Q, la dérivée cyclique [94] est l’unique application linéaire continue d → CQ d ∂α : HH0 (CQ) P qui envoie la classe d’un chemin p sur la somme vu prise sur toutes les décompositions p = uαv en des chemins u et v de longueur supérieure ou égale à zéro. Soit W un d L’algèbre de Jacobi P(Q, W ) potentiel sur Q, c’est-à-dire un élément de HH0 (CQ). d par l’idéal bilatère engendré par les dérivées est la complétion du quotient de CQ cycliques ∂α W , où α parcourt les flèches de Q. Une représentation décorée (M, V ) de (Q, W ) est formée d’un module M sur P(Q, W ) et d’une famille d’espaces vectoriels Vi , i ∈ Q0 , où M et les Vi sont supposés de dimension finie sur C. Par exemple, si Q n’a pas de cycles orientés (et donc W = 0 et P(Q, W ) = CQ), toute représentation décorée (M, V ) fournit un objet M π(M ) ⊕ Σπ(Vi ⊗C Pi ) i∈Q0

de la catégorie amassée (voir la section 4.1). Dans [22] et [23], Derksen-Weyman-Zelevinsky construisent et étudient l’opération de mutation pour les carquois à potentiel et leurs représentations décorées. Des difficultés techniques nombreuses et subtiles sont dues au fait que cette opération n’est ni fonctorielle ni définie partout. Ceci est aussi la raison pour laquelle les représentations décorées ne forment pas, en général, une catégorie. Néanmoins, la théorie

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

84

B. KELLER

développée par Derksen-Weyman-Zelevinsky est assez proche de la catégorification additive. Elle en diffère par le fait que les objets combinatoires centraux ne sont plus les variables d’amas mais les F -polynômes et g-vecteurs introduits dans [42] et qui sont peut-être encore plus fondamentaux que les variables d’amas. Dans [23], Derksen-Weyman-Zelevinsky appliquent leur théorie en démontrant de nombreuses conjectures formulées dans [42]. Ils y parviennent sous la seule hypothèse que les algèbres amassées considérées proviennent de bons carquois (non glacés), c’est-à-dire de matrices antisymétriques quelconques, ce qui représente un progrès remarquable par rapport aux approches précédentes. La construction de bases et la conjecture de positivité restent néanmoins des problèmes complètement ouverts dans cette généralité. Les idées de Derksen-Weyman-Zelevinsky ont été liées à la catégorification additive au sens des sections 4.2 et 4.1 dans [7] [74] [1]. Un lien important avec les surfaces à bord marquées est établi dans [79]. Remerciements Je remercie Caroline Gruson, Bernard Leclerc et Rached Mneimné pour leurs conseils avisés sur une version antérieure de ce texte.

REFERENCES [1] C. Amiot – Cluster categories for algebras of global dimension 2 and quivers with potential, Ann. Inst. Fourier (Grenoble) 59 (2009), p. 2525–2590. [2] A. Berenstein, S. Fomin & A. Zelevinsky – Parametrizations of canonical bases and totally positive matrices, Adv. Math. 122 (1996), p. 49–149. [3] , Cluster algebras. III. Upper bounds and double Bruhat cells, Duke Math. J. 126 (2005), p. 1–52. [4] A. Berenstein & A. Zelevinsky – Tensor product multiplicities, canonical bases and totally positive varieties, Invent. Math. 143 (2001), p. 77–128. [5] , Quantum cluster algebras, Adv. Math. 195 (2005), p. 405–455. [6] T. Bridgeland – Stability conditions on triangulated categories, Ann. of Math. 166 (2007), p. 317–345. [7] A. B. Buan, O. Iyama, I. Reiten & D. Smith – Mutation of cluster-tilting objects and potentials, à paraître dans Am. J. Math. [8] A. B. Buan & R. J. Marsh – Cluster-tilting theory, in Trends in representation theory of algebras and related topics, Contemp. Math., vol. 406, Amer. Math. Soc., 2006, p. 1–30. [9] A. B. Buan, R. J. Marsh, M. Reineke, I. Reiten & G. Todorov – Tilting theory and cluster combinatorics, Adv. Math. 204 (2006), p. 572–618.

ASTÉRISQUE 339

(1014)

ALGÈBRES AMASSÉES ET APPLICATIONS

85

[10] A. B. Buan, R. J. Marsh & I. Reiten – Cluster mutation via quiver representations, Comment. Math. Helv. 83 (2008), p. 143–177. [11] A. B. Buan, R. J. Marsh, I. Reiten & G. Todorov – Clusters and seeds in acyclic cluster algebras, Proc. Amer. Math. Soc. 135 (2007), p. 3049–3060, with an appendix coauthored in addition by P. Caldero and B. Keller. [12] P. Caldero & F. Chapoton – Cluster algebras as Hall algebras of quiver representations, Comment. Math. Helv. 81 (2006), p. 595–616. [13] P. Caldero, F. Chapoton & R. Schiffler – Quivers with relations arising from clusters (An case), Trans. Amer. Math. Soc. 358 (2006), p. 1347–1364. [14] P. Caldero & B. Keller – From triangulated categories to cluster algebras. II, Ann. Sci. École Norm. Sup. 39 (2006), p. 983–1009. [15] , From triangulated categories to cluster algebras, Invent. Math. 172 (2008), p. 169–211. [16] P. Caldero & M. Reineke – On the quiver Grassmannian in the acyclic case, J. Pure Appl. Algebra 212 (2008), p. 2369–2380. (1) [17] G. Cerulli Irelli – Canonically positive basis of cluster algebras of type A2 , prépublication arXiv:0904.2543. [18] F. Chapoton – Enumerative properties of generalized associahedra, Sém. Lothar. Combin. 51 (2004/05), Art. B51b. [19] F. Chapoton, S. Fomin & A. Zelevinsky – Polytopal realizations of generalized associahedra, Canad. Math. Bull. 45 (2002), p. 537–566. [20] V. Chari & A. Pressley – Quantum affine algebras, Comm. Math. Phys. 142 (1991), p. 261–283. [21] L. Demonet – Categorification of skew-symmetrizable cluster algebras, prépublication arXiv:0909.1633. [22] H. Derksen, J. Weyman & A. Zelevinsky – Quivers with potentials and their representations I: Mutations, Selecta Mathematica 14 (2008), p. 59–119. [23] , Quivers with potentials and their representations II: Applications to cluster algebras, prépublication arXiv:0904.0676. [24] P. Di Francesco & R. Kedem – Q-systems as cluster algebras. II. Cartan matrix of finite type and the polynomial property, Lett. Math. Phys. 89 (2009), p. 183–216. [25] , Positivity of the T -system cluster algebra, prépublication arXiv:0908.3122. [26] G. Dupont – Generic variables in acyclic cluster algebras, prépublication arXiv:0811.2909. [27] M. Fekete – Über ein Problem von Laguerre, Rend. Circ. Mat. Palermo 34 (1912), p. 89–100, 110–120.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

86

B. KELLER

[28] A. Felikson, M. Shapiro & P. Tumarkin – Skew-symmetric cluster algebras of finite mutation type, prépublication arXiv:0811.1703. [29] B. Feng, A. Hanany, Y.-H. He & A. M. Uranga – Toric duality as Seiberg duality and brane diamonds, J. High Energy Phys. 12 (2001), Paper 35, 29. [30] V. V. Fock & A. B. Goncharov – Cluster X -varieties, amalgamation, and Poisson-Lie groups, in Algebraic geometry and number theory, Progr. Math., vol. 253, Birkhäuser, 2006, p. 27–68. [31] , Moduli spaces of local systems and higher Teichmüller theory, Publ. Math. Inst. Hautes Études Sci. 103 (2006), p. 1–211. [32] , Cluster ensembles, quantization and the dilogarithm, Ann. Sci. Éc. Norm. Supér. 42 (2009), p. 865–930. [33] , Cluster ensembles, quantization and the dilogarithm. II. The intertwiner, in Algebra, arithmetic, and geometry: in honor of Yu. I. Manin. Vol. I, Progr. Math., vol. 269, Birkhäuser, 2009, p. 655–673. , The quantum dilogarithm and representations of quantum cluster va[34] rieties, Invent. Math. 175 (2009), p. 223–286. [35] S. Fomin – Cluster algebras portal, http://www.math.lsa.umich.edu/ ~fomin/cluster.html. [36] S. Fomin & N. Reading – Generalized cluster complexes and Coxeter combinatorics, Int. Math. Res. Not. 2005 (2005), p. 2709–2757. [37] S. Fomin, M. Shapiro & D. Thurston – Cluster algebras and triangulated surfaces. I. Cluster complexes, Acta Math. 201 (2008), p. 83–146. [38] S. Fomin & A. Zelevinsky – Cluster algebras. I. Foundations, J. Amer. Math. Soc. 15 (2002), p. 497–529. , Cluster algebras. II. Finite type classification, Invent. Math. 154 [39] (2003), p. 63–121. [40] , Cluster algebras: notes for the CDM-03 conference, in Current developments in mathematics, 2003, Int. Press, Somerville, MA, 2003, p. 1–34. , Y -systems and generalized associahedra, Ann. of Math. 158 (2003), [41] p. 977–1018. [42] , Cluster algebras. IV. Coefficients, Compos. Math. 143 (2007), p. 112– 164. [43] E. Frenkel & A. Szenes – Thermodynamic Bethe ansatz and dilogarithm identities. I, Math. Res. Lett. 2 (1995), p. 677–693. [44] C. J. Fu & B. Keller – On cluster algebras with coefficients and 2-Calabi-Yau categories, Trans. Amer. Math. Soc. 362 (2010), p. 859–895. [45] P. Gabriel – Représentations indécomposables, Séminaire Bourbaki, vol. 1973/1974, exp. no 444, Lecture Notes in Math., vol. 431, Springer, 1975, p. 143– 169.

ASTÉRISQUE 339

(1014)

ALGÈBRES AMASSÉES ET APPLICATIONS

87

[46] D. Gaiotto, G. W. Moore & A. Neitzke – Wall-crossing, Hitchin systems and the WKB approximation, prépublication arXiv:0907.3987. [47] C. Geiss, B. Leclerc & J. Schröer – Semicanonical bases and preprojective algebras, Ann. Sci. École Norm. Sup. 38 (2005), p. 193–253. [48]

, Rigid modules over preprojective algebras, Invent. Math. 165 (2006), p. 589–632.

[49]

, Partial flag varieties and preprojective algebras, Ann. Inst. Fourier (Grenoble) 58 (2008), p. 825–876.

[50]

, Preprojective algebras and cluster algebras, in Trends in representation theory of algebras and related topics, EMS Ser. Congr. Rep., Eur. Math. Soc., Zürich, 2008, p. 253–283.

[51]

, Cluster algebra structures and semicanonical bases for unipotent groups, prépublication arXiv:math/0703039.

[52] M. Gekhtman, M. Shapiro & A. Vainshtein – Cluster algebras and Poisson geometry, Mosc. Math. J. 3 (2003), p. 899–934, 1199. [53]

, Cluster algebras and Weil-Petersson forms, Duke Math. J. 127 (2005), p. 291–311.

[54]

, On the properties of the exchange graph of a cluster algebra, Math. Res. Lett. 15 (2008), p. 321–330.

[55] V. Ginzburg – Calabi-Yau algebras, prépublication arXiv:math/0612139. [56] F. Gliozzi & R. Tateo – Thermodynamic Bethe ansatz and three-fold triangulations, Internat. J. Modern Phys. A 11 (1996), p. 4051–4064. [57] A. Henriques – A periodicity theorem for the octahedron recurrence, J. Algebraic Combin. 26 (2007), p. 1–26. [58] D. Hernandez & B. Leclerc – Cluster algebras and quantum affine algebras, Duke Math. J. 154 (2010), p. 265–341. [59] A. Hubery – Acyclic cluster algebras via Ringel-Hall algebras, prépublication http://www.maths.leeds.ac.uk/~ahubery/Cluster.pdf. [60] C. Ingalls & H. Thomas – Noncrossing partitions and representations of quivers, Compos. Math. 145 (2009), p. 1533–1562. [61] R. Inoue, O. Iyama, B. Keller, A. Kuniba & T. Nakanishi – Periodicities of T and Y -systems, dilogarithm identities, and cluster algebras I: type Br , prépublication arXiv:1001.1880. [62]

, Periodicities of T and Y -systems, dilogarithm identities, and cluster algebras II: types Cr , F4 , and G2 , prépublication arXiv:1001.1881.

[63] R. Inoue, O. Iyama, A. Kuniba, T. Nakanishi & J. Suzuki – Periodicities of T -systems and Y -systems, Nagoya Math. J. 197 (2010), p. 59–174.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

88

B. KELLER

[64] O. Iyama & I. Reiten – Fomin-Zelevinsky mutation and tilting modules over Calabi-Yau algebras, Amer. J. Math. 130 (2008), p. 1087–1149. [65] O. Iyama & Y. Yoshino – Mutation in triangulated categories and rigid Cohen-Macaulay modules, Invent. Math. 172 (2008), p. 117–168. [66] D. Joyce & Y. Song – A theory of generalized Donaldson-Thomas invariants. II. Multiplicative identities for Behrend functions, prépublication arXiv:0901.2872. [67] V. G. Kac – Infinite-dimensional Lie algebras, third ed., Cambridge Univ. Press, 1990. [68] M. Kashiwara – Bases cristallines, C. R. Acad. Sci. Paris Sér. I Math. 311 (1990), p. 277–280. [69] R. Kedem – Q-systems as cluster algebras, J. Phys. A 41 (2008), 194011, 14. [70] B. Keller – On triangulated orbit categories, Doc. Math. 10 (2005), p. 551– 581. [71]

, Cluster algebras, quiver representations and triangulated categories, in Triangulated categories, London Math. Soc. Lecture Note Ser., vol. 375, Cambridge Univ. Press, 2010, p. 76–160.

[72]

, The periodicity conjecture for pairs of Dynkin diagrams, prépublication arXiv:1001.1531.

[73]

, Quiver mutation in Java, applet Java http://www.math.jussieu.fr/ ~keller/quivermutation.

[74] B. Keller & D. Yang – Derived equivalences from mutations of quivers with potential, Advances in Math. 226 (2011), p. 2118–2168. [75] M. Kontsevich & Y. Soibelman – Stability structures, Donaldson-Thomas invariants and cluster transformations, prépublication arXiv:0811.2435. [76] C. Krattenthaler – The F -triangle of the generalised cluster complex, in Topics in discrete mathematics, Algorithms Combin., vol. 26, Springer, 2006, p. 93–126. [77] A. Kuniba & T. Nakanishi – Spectra in conformal field theories from the Rogers dilogarithm, Modern Phys. Lett. A 7 (1992), p. 3487–3494. [78] A. Kuniba, T. Nakanishi & J. Suzuki – Functional relations in solvable lattice models. I. Functional relations and representation theory, Internat. J. Modern Phys. A 9 (1994), p. 5215–5266. [79] D. Labardini-Fragoso – Quivers with potentials associated to triangulated surfaces, Proc. Lond. Math. Soc. 98 (2009), p. 797–839. [80] B. Leclerc – Algèbres affines quantiques et algèbres amassées, notes d’un exposé au séminaire d’algèbre à l’institut Henri Poincaré le 14 janvier 2008.

ASTÉRISQUE 339

(1014)

ALGÈBRES AMASSÉES ET APPLICATIONS

89

[81] G. Lusztig – Canonical bases arising from quantized enveloping algebras, J. Amer. Math. Soc. 3 (1990), p. 447–498. [82]

, Total positivity in reductive groups, in Lie theory and geometry, Progr. Math., vol. 123, Birkhäuser, 1994, p. 531–568.

[83]

, Semicanonical bases arising from enveloping algebras, Adv. Math. 151 (2000), p. 129–139.

[84] R. J. Marsh, M. Reineke & A. Zelevinsky – Generalized associahedra via quiver representations, Trans. Amer. Math. Soc. 355 (2003), p. 4171–4186. [85] G. Musiker – A graph theoretic expansion formula for cluster algebras of type Bn and Dn , prépublication arXiv:0710.3574. [86] G. Musiker, R. Schiffler & L. Williams – Positivity for cluster algebras from surfaces, prépublication arXiv:0906.0748. [87] H. Nakajima – Quiver varieties and finite-dimensional representations of quantum affine algebras, J. Amer. Math. Soc. 14 (2001), p. 145–238. [88]

, Quiver varieties and cluster algebras, prépublication arXiv:0905.0002.

[89] Y. Palu – Cluster characters for 2-Calabi-Yau triangulated categories, Ann. Inst. Fourier (Grenoble) 58 (2008), p. 2221–2248. [90] F. Ravanini, A. Valleriani & R. Tateo – Dynkin TBAs, Internat. J. Modern Phys. A 8 (1993), p. 1707–1727. [91] M. Reineke – Cohomology of quiver moduli, functional equations & integrality of Donaldson-Thomas type invariants, prépublication arXiv:0903.0261. [92] I. Reiten – Tilting theory and cluster algebras, prépublication http://www. institut.math.jussieu.fr/~keller/ictp2006/lecturenotes/reiten.pdf. [93] C. M. Ringel – Some remarks concerning tilting modules and tilted algebras. Origin. Relevance. Future, in Handbook of tilting theory, London Math. Soc. Lecture Note Ser., vol. 332, Cambridge Univ. Press, 2007, p. 49–104. [94] G.-C. Rota, B. Sagan & P. R. Stein – A cyclic derivative in noncommutative algebra, J. Algebra 64 (1980), p. 54–75. [95] J. S. Scott – Grassmannians and cluster algebras, Proc. London Math. Soc. 92 (2006), p. 345–380. [96] P. Sherman & A. Zelevinsky – Positivity and canonical bases in rank 2 cluster algebras of finite and affine types, Mosc. Math. J. 4 (2004), p. 947–974, 982. [97] J. D. Stasheff – Homotopy associativity of H-spaces. I, II, Trans. Amer. Math. Soc. 108 (1963), p. 275–292; ibid. 108 (1963), p. 293–312. [98] A. Szenes – Periodicity of Y-systems and flat connections, Lett. Math. Phys. 89 (2009), p. 217–230.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

90

B. KELLER

[99] M. Varagnolo & E. Vasserot – Perverse sheaves and quantum Grothendieck rings, in Studies in memory of Issai Schur (Chevaleret/Rehovot, 2000), Progr. Math., vol. 210, Birkhäuser, 2003, p. 345–365. [100] A. Y. Volkov – On the periodicity conjecture for Y -systems, Comm. Math. Phys. 276 (2007), p. 509–517. [101] J. Xiao & F. Xu – Green’s formula with C∗ -action and Caldero-Keller’s formula for cluster algebras, prépublication arXiv:0707.1175. [102] S. W. Yang & A. Zelevinsky – Cluster algebras of finite type via Coxeter elements and principal minors, Transform. Groups 13 (2008), p. 855–955. [103] A. B. Zamolodchikov – On the thermodynamic Bethe ansatz equations for reflectionless ADE scattering theories, Phys. Lett. B 253 (1991), p. 391–394. [104] A. Zelevinsky – From Littlewood-Richardson coefficients to cluster algebras in three lectures, in Symmetric functions 2001: surveys of developments and perspectives, NATO Sci. Ser. II Math. Phys. Chem., vol. 74, Kluwer Acad. Publ., 2002, p. 253–273. [105]

, Cluster algebras: origins, results and conjectures, in Advances in algebra towards millennium problems, SAS Int. Publ., Delhi, 2005, p. 85–105.

[106]

, What is . . . a cluster algebra?, Notices Amer. Math. Soc. 54 (2007), p. 1494–1495.

[107]

, Cluster algebras, arXiv:math.RT/0407414.

notes

for

2004

IMCC,

prépublication

Bernhard KELLER Université Paris Diderot – Paris 7 Institut de Mathématiques de Jussieu U.M.R. 7586 du CNRS Case 7012 Bâtiment Chevaleret F–75205 Paris Cedex 13 E-mail : [email protected]

ASTÉRISQUE 339

339

ASTÉRISQUE 2011

SÉMINAIRE BOURBAKI VOLUME 2009/2010 EXPOSÉS 1012-1026 (1015) Linear stability of black holes Sergiu KLAINERMAN

SOCIÉTÉ MATHÉMATIQUE DE FRANCE Publié avec le concours du CENTRE NATIONAL DE LA RECHERCHE SCIENTIFIQUE

Séminaire BOURBAKI 62e année, 2009-2010, no 1015, p. 91 à 139

Novembre 2009

LINEAR STABILITY OF BLACK HOLES [d’après M. Dafermos et I. Rodnianski] by Sergiu KLAINERMAN

The treatment of perturbations of Kerr spacetime has been prolixious in its complexity. Perhaps at a later time the complexity will be unravelled by deeper insights. But meantime the analysis has led into a realm of the rococo, splendorous, joyful and immensely ornate. S. Chandrasekhar, The mathematical theory of black holes.

1. INTRODUCTION While the splendorous remains, a layer of complexity has now been unravelled. I report on the recent, remarkable, ongoing progress made on the linear stability of black holes, more precisely on the boundedness and decay properties of solutions to linear equations in a Kerr spacetime. The Kerr spacetimes K (m, a) are explicit solutions of the Einstein vacuum equations (discovered by R. Kerr in 1963) depending on two parameters 0 ≤ a ≤ m, corresponding physically to black holes of mass m and angular momentum am. The case a = m = 0 corresponds to the Minkowski space while a = 0, m > 0, corresponds to the much older Schwarzschild solution (K. Schwarzschild 1915). The problem of linear stability of the Kerr family is an old problem which has received a lot of attention in the Physics literature immediately after the discovery of these fascinating solutions of the Einstein equations in vacuum, which, embedded in the larger 3-parameter family of the so-called Kerr-Newman spacetimes, form the basis of our understanding of black holes. The obvious question raised by the discovery of any interesting, explicit solution of a complex, non-linear system, such as the Einstein equations, is that of their stability under small perturbations. Roughly the problem here is to show that all spacetime developments of initial data sets, sufficiently close to the initial data set of a Kerr spacetime, behaves in the large like (possible another) Kerr solution. This is not only a deep mathematical question but one with serious astrophysical implications. Indeed if the Kerr family would be unstable under perturbations, black holes would be nothing more than mathematical

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

92

S. KLAINERMAN

artifacts. The Einstein equations are, of course, nonlinear and hyperbolic, thus the issue of stability is an extremely difficult and a dicey one. Given the geometric, covariant structure of the equations, with no universal notions of space and time variables, it is not even a-priori clear what that means. Linear stability, though still tricky, is somewhat easier to define. It is clear, for example, that any first order approximation of the equations, at the level of the space-time metric, in any reasonable coordinate system, will generate some system of wave equations in the Kerr background we want to perturb. Thus it is natural to ask, and this must certainly be relevant to the full nonlinear problem, whether solutions to linear wave equations in a fixed Kerr background are well behaved. If it turns out that solutions of these linear equations are amplified, due to the non-trivial features of the background geometry, then there is a reasonable chance that the background itself might be unstable. It is not enough, however, to establish that solutions are not amplified; to have a chance to prove non-linear stability we also need to show that solutions decay at a sufficient rate. There is a lot of confusion in this regard among some physicists who seem to believe that somehow the lack of linear instability is a strong indication of nonlinear stability. This, of course, is not true even near solutions of minimal energy of simple nonlinear PDE’s, as the case of the Burger equation ∂t u + u∂x u = 0 easily demonstrates. The solution u = 0 is a global minimum for the energy integral R E(t) = |u(t, x)|2 dx, yet any compactly supported, smooth, small perturbation of the zero initial leads to blow up in finite time. To be useful, a result on linear stability has to establish, quantitatively, not just a lack of amplification but also a realistic decay. In fact all known stability results, for strongly nonlinear wave equations (Einstein equations are quasilinear), depend on precise decay information for the linearized solutions. The methods by which one establishes these decay estimates are also a very important issue. Thus, in the Minkowski space R1+3 , it is easy to derive decay estimates for solution to the standard wave equation φ = 0 using explicit representation formulas in the physical or Fourier variables. These formulas, however, depend heavily on the specific features of the Minkowski space and do not survive under small perturbations of the Minkowski metric. In other words, such methods are intrinsically not robust. A far more useful method for deriving decay estimates for the wave equation, and more generally for linear field equations, is that of invariant vector fields, see [28], [29]. That method, first introduced to prove stability results for quasilinear wave equations, plays a fundamental role in all known proofs of the stability of the Minkowski space, see [12], [30], [32], [5]. In the case of the Kerr metric, or rather the more accessible case of the Schwarzschild metric, one can use the specific symmetries of the space to separate variables and then concentrate on the pointwise properties of the corresponding

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

93

eigenvalue problem. This method is not only not robust but, to our knowledge, was not even satisfactory to derive unconditional decay results for general solutions of the wave equation. In the physics literature, where the problem of linear stability of Schwarzschild and Kerr spacetimes has received a tremendous amount of attention (see e.g. [40], [45], [39], [38], [41], the monograph [9] and the references therein), this method of mode decomposition led to nothing more, in the words of Press and Teukolsky (see [38]), than “an unsuccessful search for instabilities”. On the other hand mathematical rigorous efforts based on this approach can only lead to statements of decay without a rate or precise rates of decay of specific modes, both of which, in principle, compatible with the scenario in which a general solution of the corresponding linear problem is not even uniformly bounded. For the results in this direction, see [34], [31], [21] in Schwarzschild and an attempt [22] in Kerr. Moreover, even if ultimately successful, such methods would leave us with a heavy machinery to prove some form of linear stability without any clue on how to approach to the non-linear problem. A simple version of the vector field method was first used by Kay and Wald, see [27], to prove the boundedness of solutions of the wave equation in a Schwarzschild spacetime. The first attempt to use the vector field method, to prove integrated local energy decay in Schwarzschild is due to Blue and Soffer [6]. Their work however had serious flaws. The first complete results on pointwise decay for solutions of the wave equation on the Schwarzschild background have been obtained, independently, by Blue-Sterbenz [7], and Dafermos-Rodnianski [16]. In [16] Dafermos and Rodnianski also introduced the crucial red shift vector field, which led to stronger decay rates along the event horizon in Schwarzschild and, more importantly, played a central role in extending the boundedness and decay results to Kerr space-times, see [20], [17]. Other important contributions were made by S. Alinhac in [3], Dafermos-Rodnianski in [19], Marzuola-Metcalfe-Tataru-Tohaneanu in [35] and Luk in [33] for the problem in Schwarzschild, and by Tataru-Tohaneanu in [43] and Andersson-Blue in [4], for Kerr spacetimes. I will review these results following, mainly the works of Dafermos-Rodnianski, in particular their general exposition in [17] and the recent paper [18].

2. INITIAL VALUE PROBLEM We recall that an initial data set consists of a 3-dimensional manifold Σ, a complete Riemannian metric g(0) , a symmetric 2-tensor k(0) , and a well specified set of initial conditions corresponding to the matterfields under consideration. These have

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

94

S. KLAINERMAN

to be restricted to a well known set of constraint equations. We restrict the discussion to asymptotically flat initial data sets, i.e. outside a sufficiently large compact set K, Σ(0) \ K is diffeomorphic to the complement of the unit ball in R3 and admits a system of coordinates in which g(0) is asymptotically euclidean and k(0) vanishes, at appropriate order. A Cauchy development of an initial data set is a globally hyperbolic spacetime ( M , g), verifying the Einstein field equations, in the presence of a matterfield with energy momentum Q, (2.1)

1 Rαβ − Rgαβ = Qαβ , 2

and an embedding i : Σ −→ M such that i∗ (g(0) ), i∗ (k(0) ) are the first and second fundamental forms of i(Σ(0) ) in M . In what follows I will restrict the discussion to the Einstein vacuum equations, i.e. the case when the energy momentum tensor vanishes identically and the equations take the purely geometric form (2.2)

Rαβ = 0.

smf_bbk_1015_blackhole5.jpg

Figure 1.

The most primitive question asked about the initial value problem, solved in a satisfactory way, for very large classes of evolution equations, is that of local existence and uniqueness of solutions. For the Einstein equations this type of result was first

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

95

established by Y. F.-Bruhat [23] with the help of wave coordinates(1). According to this result any smooth initial data set admits a unique, smooth, local (up to an isometry) globally hyperbolic (2) Cauchy development. In the case of nonlinear systems of differential equations the local existence and uniqueness result leads, through a straightforward extension argument, to a global result concerning the maximal time interval of existence. If this interval is bounded the solution must become infinite at its upper boundary. The formulation of the same type of result for the Einstein equations is a little more subtle; something similar was achieved in [10]. Theorem 1 (Bruhat-Geroch). — For each smooth initial data set there exists a unique, smooth, maximal, future, globally hyperbolic development (MFGHD). Thus any construction, obtained by an evolutionary approach from a specific initial data set, must be necessarily contained in its maximal development MFGHD. This may be said to solve the problem of global(3) existence and uniqueness in General Relativity; all further questions, one could say, concern the qualitative properties of these maximal developments. The central issue becomes that of existence and character of singularities 2.1. Special solutions We recall that EVE admits a remarkable family of explicit, stationary solutions given by the two parameter family of Kerr solutions among which one distinguishes the Schwarzschild family of solutions, of mass m > 0, Å ã Å ã 2m 2m −1 2 (2.3) gS = − 1 − dt2 + 1 − dr + r2 dσS2 . r r Though the metric seems singular at r = 2m it turns out that one can glue together two regions r > 2m and two regions r < 2m of the Schwarzschild metric to obtain a metric which is smooth along H = {r = 2m}, see [24], called the Schwarzschild horizon. The portion of r < 2m to the future of the hypersurface t = 0 is a black hole whose future boundary r = 0 is singular. The similar region to the past of t = 0 is called a white hole. The region r > 2m, called the domain of outer communication, is free of singularities. To see how to explicitly extend the metric, introduce the tortoise coordinate r∗ = r + 2m ln(r/2m − 1) and the Kruskal null coordinates, U = e−(t−r∗ )/4m , 3 −r/2M V = e(t+r∗ )/4m , relative to which the metric takes the form ds2 = − 32m er dU dV + (1)

These allow one to cast the Einstein vacuum equations in the form of a system of nonlinear wave equations. (2) Any past directed, in-extendable causal curve of the development intersects Σ0 . (3) A proper definition of global solutions in GR requires a special discussion concerning the proper time of causal geodesics.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

96

S. KLAINERMAN

Figure 2. Kruskal’s maximally extended Schwarzschild space-time. Note the two disconnected external regions, r > 2m, the black and white holes and the curvature singularity at r = 0. Note the behavior of light cones at the event horizon, r = 2m.

r2 dσ 2 . Observe now that r = 2M corresponds precisely to U · V = 0. Indeed r is r r an implicit function of U · V through the relation ( 2m − 1)e 2m = −U V . In the new coordinates, after a simple conformal compactification, the completed space-time has the form given in Figure 3A. + − Here the boundaries I and I , called future and past null infinities, are idealized boundaries of the space-time corresponding to end points, of future directed, respectively past directed, null geodesics. The points i+ and i− correspond to end points of future and past time-like geodesics while i0 corresponds to space-like infinity. The Schwarzschild family is included in a larger two parameter family of solutions K (a, m) discovered by Kerr. A given Kerr space-time, with 0 ≤ a < m, has a well defined domain of outer communication r > r+ := m + (m2 − a2 )1/2 . In BoyerLindquist coordinates, well adapted to r > r+ , the Kerr metric has the form ∆ − a2 sin2 θ 2 2a sin2 θ(r2 + a2 − ∆) dt − dtdφ Σ Σ (r2 + a2 )2 − ∆a2 sin2 θ Σ + sin2 dφ2 + dr2 + Σdθ2 Σ ∆ with Σ = r2 + a2 cos2 θ, ∆ = r2 + a2 − 2mr. As in the Schwarzschild case, the exterior Kerr metric extends smoothly across the Kerr event horizon, H = {r = r+ }. It can be shown that the future and past sets of any point in the domain of outer communication intersect any time-like curve, passing through points of arbitrary large values of r, in gK = −

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

97

finite time as measured relative to proper time along the curve. This fact is violated by points in the region r ≤ r+ , which consists of the union between a black hole region, extended towards the future, and a white hole region to the past. Thus physical signals (i.e. future time-like or null geodesics) which initiate at points in r ≤ r+ cannot be registered by far away observers(4). The extended Kerr is singular only at r = 0. Thus the singularities in Kerr cannot have any effect on the domain of outer communication which is, in fact, entirely smooth even analytic. The boundary of the domain of outer communication {r = r+ } is called the event horizon. In the non-degenerate case, a < m, the event horizon consists of two null hypersurfaces intersecting transversally on a compact 2 sphere. The exterior Kerr metrics are stationary, which means, roughly, that the coefficients of the metric are independent of the time variable t. One can reformulate this by saying that the vector field T = ∂t is Killing(5) (everywhere in the domain of outer communication) and time-like at points with r large, i.e. the so-called asymptotic region (where the space-time is close to flat). One can also easily check that T is tangent to the horizon H = N ∪ N , which is itself a null hypersurface, i.e. the restriction of the metric to the tangent space to H is degenerate (see Figure 4A). In addition to being stationary the coefficients of the Kerr metric are independent of the circular variable φ. Thus Kerr is stationary and axially symmetric. The Schwarzschild metrics corresponding to a = 0 are not just axially symmetric but spherically symmetric, which means that the metric is left invariant by the whole rotation group of the standard sphere S2 . A well known theorem of Birkhoff shows that they are the only such solutions of the vacuum Einstein equations. Another peculiarity of a Schwarzschild metric, not true in the case of Kerr, is that the stationary Killing vector field T = ∂t is orthogonal to the hypersurface t = 0. A stationary space-time which has this property is called static. This is also equivalent to the fact that the Schwrazschild metric is invariant with respect to the reflection t → −t. Moreover T is time-like for all r > 2m and null along the Schwarzschild horizon H = {r = 2m}. This is not the case for Kerr solutions in which case T = ∂t is only time-like for r > m + (m2 − a2 cos2 θ)1/2 , null for r = m + (m2 − a2 cos2 θ)1/2 and space-like in the region between r+ and r = m + (m2 − a2 cos2 θ)1/2 , called the ergosphere. Finally we remark that the Kerr family has unacceptable features for a > m. To summarize: 1. The Kerr family K (a, m), 0 ≤ a ≤ m provides a two parameter family of asymptotically flat solutions of the Einstein vacuum equations exhibiting a smooth (4)

They must end in the singularity at r = 0, in Schwarzschild space-time. Their behavior in Kerr is more complicated. (5) A vector field X is said to be Killing if its associated 1 parameter flow consists of isometries of g, i.e. the Lie derivative of the metric g with respect to X vanishes, L X g = 0.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

98

S. KLAINERMAN

domain of outer communication and its complement, separated by the event horizon {r = r+ }. For a < m the event horizon consists of two null hypersurfaces intersecting transversally on a compact 2 sphere. 2. All solutions are stationary, i.e they admit a Killing vector field T which is timelike in the asymptotic region. The Schwarzschild space-time (i.e. a = 0) is also static. Moreover the Kerr family is axially symmetric, i.e. it admits another, circular, Killing vector field Z which vanishes on the axis of symmetry. The Schwarzschild space-time is spherically symmetric. 3. The stationary vector field T is tangent along the horizon and space-like for all a > 0. It remains space-like in a small region of DOC called ergo-region. In the case a = 0, T is null along the horizon and time-like everywhere in DOC. 4. In all cases 0 ≤ a < m, DOC contains trapped null geodesics, i.e. null geodesics which are entirely contained in a region of DOC with a bounded value of r. In the case a = 0, all trapped null geodesics are either tangent to the time-like surface {r = 3m} or asymptotic to it. 2.2. Stationary space-times We formalize below the notion of an asymptotically flat stationary, vacuum, spacetime. Assume that ( M , g) is a smooth vacuum Einstein space-time of dimension 3 + 1 and T is a smooth Killing vector field on M . Assume given a space-like hypersurface Σ0 ⊆ M such that outside a sufficiently large compact set K of Σ, every orbit of T intersects Σ0 at only one point. Moreover we assume the existence of a coordinate system (x0 , x1 , x2 , x3 ) in M (end) = T(Σ0 \ K) (i.e. the union of orbits of T which p intersect Σ0 \ K) such that T = ∂t and, with r = (x1 )2 + (x2 )2 + (x3 )2 , the com−2 ponents of the space-time metric verify(6), for some k ≥ 0, g00 = −1 + 2m ), r + Ok (r j k 2J x −1 −3 1 2 3 gij = δij +Ok (r ), g0i = −ijk r3 +Ok (r ), for some m > 0, J = (J , J , J ) ∈ R such that, |J|2 < m2 . We can then define the exterior region, or domain of outer communication, by E = I − ( M (end) ) ∩ I + ( M (end) ), −

+

where I ( M (end) ), I ( M (end) ) denote, respectively, the past and future sets of M (end) . One further assumes that E is globally hyperbolic, i.e. any inextensible time-like or null curve in E must intersect Σ0 . Finally we define B ∪ W , the union of the black hole and white hole regions, as the complement of E in M and the event horizon H as the boundary of E. One can show that H is achronal (i.e. no points in H can be connected by time-like curves) and that T must be tangent to H . One can also show, using the theorem of Hawking, that H is non-expanding (see appendix). We denote by Ok (ra ) any smooth function in M (end) which verifies |∂ i f | = O(ra−i ) for any 0 ≤ i ≤ k.

(6)

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

99

One can easily check that the event horizon H of any of the Kerr family K (a, m), 0 ≤ a < m, verifies the following properties: 1. H is spanned by two smooth null hypersurfaces N and N which intersect transversally along a 2 sphere S. Moreover N (resp. N ) is spanned by the union of the future (past), in-extendible, complete, null geodesics orthogonal to S. 2. Both N and N have vanishing null second fundamental forms (see appendix). The second condition is in fact an easy consequence of the non-expanding nature of H and the Einstein vacuum equations. A fundamental conjecture in General Relativity is to prove that the converse is true, i.e. any, regular, stationary solution of the Einstein vacuum equations verifying the above properties must be isometric to K (a, m), 0 ≤ a < m. The simple motivation behind this conjecture is that one expects, due to gravitational radiation, that general, dynamic, solutions of the Einstein field equations settle down, asymptotically, into a stationary regime. Thus the conjecture, if true, would give a description of all the asymptotic states of the Einstein vacuum equations. The conjecture is, essentially, solved in the analytic case (see [14] for an up to date account) and only partially solved in the category of smooth space-times, see [26] and [1]. In the next section we attempt to give a somewhat precise formulation of the problem of stability of Kerr.

3. STABILITY OF KERR 3.1. Stability of the Minkowski space The Minkowski space R3+1 is, of course, the simplest solution of the Einstein vacuum equations. Is it stable? Among all Kerr solutions, the Minkowski space is also the only one free of singularities, or geodesically complete. Roughly speaking this means that any freely moving observer in M can be extended indefinitely, as measured relative to its proper time. Such a space-time is said to have a regular MFGHD. Does this property persist under small perturbations? The result stated below is a rough version of the global stability of Minkowski, the complete result also provides very precise information about the decay of the curvature tensor along null and time-like directions as well as many other geometric information concerning the causal structure of the corresponding space-time, see [12], [30], [32] and [5]. Of particular interest are peeling properties i.e. the precise decay rates of various components of the curvature tensor along future null geodesics. Theorem 2 (Global Stability of Minkowski). — Any asymptotically flat initial data set which is sufficiently close to the trivial one has a regular MFGHD.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

100

S. KLAINERMAN

3.2. Cosmic censorship In general, however, we expect maximal developments to be incomplete, with singular boundaries. An important result in this direction is the recent formation of trapped surfaces result of D. Christodoulou [11]. Together with the well known singularity theorem of R. Penrose, his result shows that there exists a large class of regular initial data whose MFGHD is incomplete. The unavoidable presence of singularities, for sufficiently large initial data sets, as well as the analysis of explicit examples (such as Schwarzschild and Kerr) have led Penrose to formulate two fundamental conjectures, concerning the character of general solutions to the Einstein equations. Here I restrict my discussion only to the so called weak cosmic censorship conjecture (WCC), which is the only one relevant to the issue of stability of Kerr. To understand the statement of (WCC), consider the different behavior of null rays in Schwarzschild and Minkowski space-times. In Minkowski space light originating at any point p = (t0 , x0 ) propagates, towards future, along the null rays of the null cone t − t0 = |x − x0 |. Any free observer in R1+3 , following a straight time-like line, will necessarily meet this light cone in finite time, thus experiencing the event p. On the other hand, any point p in the trapped region r < 2m of the Schwarzschild space, is such that all null rays initiating at p remain trapped in the region r < 2m. In particular events causally connected to the singularity at r = 0 cannot influence events in the domain of outer communication r > 2m, which is thus entirely free of singularities. The same holds true in any Kerr solution with 0 ≤ a < m. WCC is an optimistic extension of this fact to the future developments of general, asymptotically flat initial data. The desired conclusion of the conjecture is that any such development, with the possible exception of a non-generic set of initial conditions, has the property that any sufficiently distant observer will never encounter singularities or any other effects propagating from them. To make this more precise one needs define what a sufficiently distant observer means. This is typically done by + introducing the notion of future null infinity I which, roughly speaking, provides end points for the null geodesics which propagate to asymptotically large distances. The future null infinity is constructed by conformally embedding the physical space¯ ,g ¯ ), g ¯ = Ω2 g in M , with time ( M , g) under consideration to a larger space-time(7) ( M + a null boundary I (where Ω = 0, dΩ 6= 0). +

Definition 1. — The future null infinity I is said to be complete if any future null geodesic along it can be indefinitely extended relative to an affine parameter.

(7)

Note however that the boundary of this extended space-time is not smooth, generically.

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

101

Weak Cosmic Censorship Generic asymptotically flat initial data have maximal future developments possessing a complete future null infinity. Using the language introduced above, we are finally ready to state the following. Conjecture 3 (Global stability of Kerr). — Any small perturbation of the initial data set of a Kerr space-time has a global future development with a complete future null infinity which, within its domain of outer communication(8), behaves asymptotically like a (another) Kerr solution.

4. STABILITY OF MINKOWSKI SPACE To understand what would be needed in a proof of stability of Kerr it pays to review some of the main ideas in the proof of stability of the Minkowski space. For lack of space and time I will be very schematic. Also, for brevity, I will be discussing only the proof in [12] and [30]. I will just note that the proof in [32] is based also on a variation of the vector field method discussed below, even though the geometric set-up is different. 4.1. Vector field method The centerpiece, keystone, of the entire proof is a geometric method to derive decay estimates for components of the curvature tensor based on a generalization of the energy method for wave equations. The method has two distinct parts, a geometric version of the multiplier method and the method of commuting vector fields. 1. Multiplier method. — One starts with the Bianchi identities which, due to the vanishing of their trace, take the form of a Maxwell type system(9): (4.1)

D[ Rαβ]γδ = 0,

Dδ Rαβγδ = 0.

A remarkable feature of this system is the existence of a fully symmetric, traceless, covariant four tensor Qαβγδ , depending quadratically on R, which verifies the divergence condition Dδ Qαβγδ = 0, (8) (9)

That means, roughly, outside the black hole region. The two equations in (4.1) are in fact equivalent.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

102

S. KLAINERMAN

and such that Q(X1 , X2 , X3 , X4 ) is positive for any future, causal vector fields X1 , X2 , X3 , X4 . Thus, for any three such vector fields X, Y, Z we find with Qµ (X, Y, Z) the one form obtained by contraction with X, Y, Z, Dµ Qµ (X, Y, Z)

(4.2) Err(X, Y, Z)

=

=

Err(X, Y, Z)

1 Q( (X) π, Y, Z) + Q(X, 2

(Y )

π, Z) + Q(X, Y,

(Z)

π)



where (X) π = L X g is the deformation tensor of X. We integrate the above identity on past domains of dependence(10) D(0, 1), sandwiched between two space-like hypersurfaces Σ0 and Σ1 with future unit normals denoted by T . Let N denote the null boundary of D(0, 1) and L the geodesic null generator (i.e. DL L = 0 and g(L, L) = 0), of N , normalized by the condition g(L, T ) = −1 on Σ0 ∩ N . Then, with Q = Q[R], Err = Err[R] as above, (4.3) Z Z Z ZZ Q(X, Y, Z, L) + N

Q(X, Y, Z, T ) −

Q(X, Y, Z, T ) = Σ1

Σ0

Err(X, Y, Z). D(0,1)

Clearly, if X, Y, Z are Killing we deduce, Err[R](X, Y, Z) = 0, and thus derive a conservation law. In the particular case when the vector fields X, Y, Z are also causal we derive a very useful coercive estimate for the left-hand side of (4.3) in terms of the integral on Σ0 , which may be interpreted as initial condition. In view of the fact that the energy-momentum Q is traceless with respect to any pair of indices, the same remains true if we replace Killing vector fields by conformal Killing ones, i.e. such that (X) π is proportional to the metric g or, in other words, the traceless part (X) π ˆ vanishes identically. 2. Commuting vector field method. — In addition to the procedure outlined above, the generalized energy method allows us to make use of commutation with selected vector fields. In fact, for any vector field X one can show that a suitable modified Lie derivative of R, denoted by d L X R, verifies the following version of (4.1) (4.4)

Dδ ( d L X R)αβγδ = Jαβγ ( (X) πˆ , R).

We can thus replace Q = Q[R] with Q[ d L X R] and repeat the procedure above to derive integral inequalities for suitable directional derivatives of R. The procedure outlined above, based on Killing and conformal Killing vector fields, seems to require a space-time with a lot of symmetries, such as the Minkowski space. It pays at this point to consider how the method works in that case. (10)

D is such that the causal past set of any point in D, in the slab between Σ0 , Σ1 , is included in D.

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

103

4.2. Minkowski space Rn+1 The Minkowski space Rn+1 comes equipped with two important geometric structures: I. Family of Killing and conformal Killing vector fields – – – –

Generators of translations in the xµ directions: Tµ = ∂x∂ µ . Generators of rotations in the (µ, ν) plane: Lµν = xµ ∂ν − xν ∂µ . Generator of scaling: S = xµ ∂µ . Generators of inverted translations(11): Kµ = 2xµ xρ ∂x∂ ρ − (xρ xρ ) ∂x∂ µ .

Of particular importance for us are the causal vector fields T0 = ∂t and K0 = (t2 + r2 )∂t + 2txi ∂i , which can be used to derive coercive energy identities. Here r2 = |x|2 = (x1 )2 + · · · (xn )2 . II. Canonical double null foliation. — This is given by the level surfaces of two optical functions u = t − r and u = t + r, i.e. solutions of the Eikonal equation mαβ ∂α u∂β u = mαβ ∂α u∂β u = 0. With respect to u, u the vector fields T0 , S and K0 take the form (4.5)

T0 =

1 (L + L), 2

S=

1 (uL + uL), 2

K0 =

1 2 (u L + u2 L) 2

where L = −mαβ ∂β u∂α = ∂t + ∂r and L = −mαβ ∂β u∂α = ∂t − ∂r are the null generators of the corresponding null hypersurfaces. Observe also that the rotation vector fields Lij = xi ∂j − xj ∂i (denoted also by Oij ) are tangent to the leaves of both foliations. To see how these vector fields can be used consider solutions of the standard wave equation φ = 0, with compactly supported data. Let Q = Q[φ] be the associated energy momentum tensor (see (6.2) in Section 6), i.e. Dβ Qαβ = 0. The standard energy identity, associated to the time translation T0 = ∂t allows us to derive the standard energy conservation identity Z Z 2 |∂φ| = |∂φ|2 ≤ I0 Σt

Σ0

P with I0 a constant depending only on the initial data of φ and |∂φ|2 = nα=0 |∂α φ|2 . Using the causal conformal Killing vector field K0 (see details in Section 6 for dimensions n ≥ 3), we can also estimate Z Z 2 |φ| . (1 + r2 )|∂φ|2 ≤ I0 . Σt

Σ0

(11)

Observe that the vector fields Kµ can be obtained applying the standard inversion to the vector fields Tµ .

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

104

S. KLAINERMAN

The Killing vector fields Tµ and Lµν commute with , while S preserves the space of solutions to φ = 0 (since [, S] = 2). This leads us to introduce the generalized Sobolev norms k X X (4.6) Q k [φ](t) = k L Xi1 L Xi2 ... L Xij φkL2 (Σt ) j=0 Xi1 ,..,Xij

with the sum taken over all Killing vector fields T, L and scaling vector field S. The crucial point of this method is that these generalized energy type norms are bounded by initial data, i.e.,

Q k [φ](y) . Q k [φ](0) . I0 . Proposition 1 (Global Sobolev inequalities). — Let φ be an arbitrary function in Rn+1 such that Q k [φ] is finite for some k > n2 . Then for t > 0, we have with u = t−|x| and u = t + |x| 1 (4.7) |φ(t, x)| ≤ c Q k [φ]. n−1 1 2 (1 + u) (1 + |u|) 2 Since Q k [φ] is bounded, for solutions of φ = 0, depending only on initial data at t = 0, we deduce a strong, realistic, uniform decay estimate. A similar analysis can be done for solutions of the Maxwell equations or the linearized Bianchi equations in Minkowski space. It is also important to realize that one can be more economical with the vector fields we use. Thus, for example, one can derive the same information using only the vector fields T0 , S, K0 and rotations Oij = Lij , i, j = 1, . . . , n. The upshot of the vector field method is that it allows us to derive realistic decay estimates by a flexible procedure which can be easily generalized to perturbations of the Minkowski space. 4.3. Deformation method Since a general perturbation of Minkowski space cannot preserve any symmetries the best we can hope for is to substitute them by approximate symmetries. We are thus looking to replace some of the conformal Killing vector fields of Minkowski with almost conformal Killing, i.e. vector fields whose deformation tensors are sufficiently small so that we can still derive useful estimates for the curvature tensor. The idea is to define these vector fields starting from two special functions whose role is to replace the optical functions u, u of the Minkowski space. In the original proof of [12] this is done by choosing a suitable defined optical function u and a suitable time function t. The function u is then defined to be u = t − 2u. In [30] one picks instead two exact optical functions u and u. One can then define vector fields T0 , S, K0 by mimicking the formulas (4.5) (with L = −gαβ ∂α u∂β , L = 2T0 − L and T0 the unit future normal to the maximal foliation Σt ) and rotation vector fields by a geometric method tied

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

105

u, t or u, u. To make the method work we need to make sure that the errors generated in the energy inequalities derived above are sufficiently small. To see, very roughly, what this entails consider (in the case of the (t, u) foliations of [12]) a quantity of the form: Z Q[ ” L O R](K0 , K0 , T0 , T0 ). Q (t) = Σt

Based on the vector field method outlined above one can show that the time dependent quantity Q (t) verifies, schematically, an identity of the form

Q (t) = Q (0) + E(t),

E(t) = E1 (t) + E2 (t),

with Z

E1 (t) =

t

Z ds

0

Q[ ” L O R](K0 , K0 , T0 )

Σs

and E2 (t) the additional error term generated by the right-hand side of (4.4). Here Q( ” L O R, K0 , K0 , T0 ) is an expression quadratic in ” L O R and linear in the deformation tensors of K0 and T0 . To make this work, i.e. obtain a global bound for Q (t), by a Gronwall inequality, we see that we need appropriate (and compatible!) decay estimates for both R and the traceless parts of the deformation tensors of K0 , T0 and O. We summarize the above considerations as follows: 1. The proof of stability of Minkowski space in [12] and [30] requires precise decay information for the curvature tensor R. 2. In a first approximation one may assume that R verifies a linear field equation(12) in Minkowski space (linearized Bianchi). The vector field method allows one to derive realistic decay estimates for components of R. 3. One can derive, essentially, the same decay estimates for the true curvature tensor of a perturbed solution of the Einstein equations, by a deformation method in which one deforms part of the geometric structure of the Minkowski space ((u, t) or (u, u)) and an appropriate number of conformal Killing vector fields (i.e. T0 , K0 , S0 and rotations O). The key here is to derive, simultaneously, suitable decay estimates for R and the traceless parts of the deformation tensors of these vector fields. These estimates have to be strong enough to be able to control the error terms generated in the energy estimates.

(12)

Note also that the result of Lindblad-Rodnianski [32] is based on a linearization at the level of the metric, which brings in the standard wave equation.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

106

S. KLAINERMAN

4.4. Non-linear stability of Kerr In view of the above discussion a proof of the non-linear stability of the Kerr family requires: 1. A robust method to derive decay estimates for linear field equations in a fixed Kerr background. Such a method has to take into account the geometric features of the Kerr metric, such as the event horizon, ergo-region and trapped null geodesics. It cannot rely only on the continuous symmetries of the Kerr metric, i.e. its Killing vector fields, which are both too limited and have serious degeneracies. 2. Find an effective linearization procedure, such as the linearized Bianchi equations(13) in the stability of Minkowski space, to which the methods sketched above apply. 3. Find a way to deform the geometry of the Kerr solution, taking into account that any small perturbation of a Kerr metric may lead, asymptotically, to a different Kerr metric.

5. LINEAR STABILITY OF THE KERR FAMILY As discussed above a first, essential step, in the proof of stability of the Kerr solution is to establish its linear stability, which amounts to prove appropriate decay estimates for solutions to the specific linear field equations in a fixed Kerr background which arise by a suitable linearization. In a somewhat simplified version of linear stability, one would like to show, by robust methods, that all solutions of the covariant wave equation (5.1)

g φ = 0,

0 ≤ a < m,

in K (a, m), 0 ≤ a < m (or more generally a fixed stationary, axially symmetric space-time with a non-degenerate horizon) with reasonable initial data on a spacelike hypersurface Σ0 , as in the figure below, are well behaved (14) in the future of Σ0 (see figure below). A more elementary task, and yet very difficult in the rotating case(15), a > 0, is to show that solutions remain bounded in the entire exterior region of the space-time. (13)

Note however that the exact analog of the Bianchi equations in a Kerr background are ill posed. Decay at rates comparable to those in the flat case. (15) The much simpler non-rotating case a = 0, corresponding to the Schwarzschild space-time, was solved previously in work by Kay and Wald. (14)

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

107

5.1. Difficulties The following are the main difficulties one has to overcome to prove linear stability, in the sense discussed above. – K (a, m) has only two linearly independent Killing vector fields, the stationary one T and the axially symmetric one Z. In the Schwarzschild case we have, of course, an action of the full rotation group SO(3) and thus two linearly independent rotation vector fields. – The stationary Killing field T degenerates in the ergo-region of K (a, m), i.e. it becomes space-like. The presence of an ergo-region is connected, physically, with the so called Penrose process according to which energy can be extracted from a rotating black hole and thus contribute to linear instability. This phenomenon is also known in the Physics literature as super-radiance. Even in Schwarzschild, T loses its time-like character on the horizon. Thus the basic energy identity provided by T loses information near the horizon, for a = 0, and loses coercivity, thus seemingly useless, for a > 0. – K (a, m) possesses a family of trapped null geodesics, i.e. future null geodesics + which neither go to I nor penetrate the black hole region. Though, fortunately, these are unstable they provide however very serious technical difficulties to derive decay information. In the case a = 0 the situation is somewhat simpler as one can show that all trapped geodesics are restricted, or asymptotic, to the surface r = 3m. 5.2. Main new ideas I try to summarize below some of the main new ideas which have crystallized in the wake of the pioneering works of Blue-Soffer, Blue-Sterbenz, Dafermos-Rodnianski, Tataru-Tohaneanu, Andersson-Blue, mentioned in the introduction. – The introduction (by Dafermos-Rodnianski) of a new vector field defined in a neighborhood of the horizon (called the red shift vector field), which I will denote by H, with coercive properties in a small neighborhood of the horizon, which compensates for the degeneracy of the stationary vector field T. – A robust mechanism, due to Dafermos-Rodnianski, for proving boundedness of solutions for Kerr space-times with a  m, despite the notorious problem of super-radiance. This is based on a decomposition, invariant relative to the actions of T and Z, into super-radiant and sub-radiant modes and the properties of the red shift vector field H. – Discovery on an effective treatment of the trapped region, based on the fact that all trapped null geodesics are unstable. In Schwarzschild this can be achieved by a suitable modification of the so called Morawetz vector field, which I will denote

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

108

S. KLAINERMAN

by M. In K (a, m), for a small enough, there are three competing methods [17], [43], [4] which deal effectively with the trapped region. They all depend, in one form or another, on the integrability properties of the geodesic flow, remarkable fact due to Carter [8]. – Decay in both Schwarzschild and Kerr is due to a third vector field, which is a suitable modification of K0 from Minkowski space(16). Recently, in [18], Dafermos and Rodnianski gave a new, more flexible, treatment of how to generate decay from null infinity without using K0 . – Traditionally energy estimates require integration, using appropriate vector fields, on large causal domains. Thus one was restricted to look for vector fields which are coercive in such regions and, unfortunately, there are not enough of those. The new methods, especially those of Dafermos-Rodnianski, point the way to a more flexible use of vector fields by concentrating on specific geometric regions where degeneracies occur (such as the event horizon) and finding new non-causal vector fields (such as the red shift H), which provides an effective cure for the missing information. The lack of causality of H can then be compensated by patching it with other vector fields, such as T or M. A similar patching procedure can be implemented in a neighborhood of null infinity, see [18]. 5.3. Main results The first result, on boundedness of solutions to the wave equation (5.1), applies to the exterior region of a fixed stationary, axially symmetric space-time M , sufficiently close to Schwarzschild, see [20]. Theorem 3 (Boundedness). — Any solution (5.1) with reasonable initial data on a space-like hypersurface Σ0 , is globally bounded(17) in the future of Σ0 . The result applies in particular to Kerr space-times K (a, m) with a  m. The same method can also be applied to derive boundedness of axially symmetric solutions of (5.1) for the whole range 0 ≤ a < m. The next result concerns decay of solutions in the Schwarzschild case a = 0. The result is expressed relative to the pair of optical functions u = t − r∗ and u = t + r∗ where r∗ = r + 2m ln(r − 2m). Observe that along the horizon, to the future of Σ0 , we have u = −∞ but, for the region we are interested in, we have u finite.

(16) (17)

Such a vector field is also used in the stability of the Minkowski space. It also has bounded, non-degenerate, total energy.

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

109

Theorem 4. — Let Σ0 as in Figure 8A above, in the exterior of the Schwarzschild space-time K (0, m). Any solution to the wave equation (5.1), with reasonable initial data on Σ0 , admits the following estimates: 1. There exists a constant C such that, uniformly(18) on all points to the future of Σ0 C |φ| ≤ . u 2. For any R > 2m, we have, with a constant CR , for all r ≥ R, |rφ| ≤

CR . |u|1/2

A similar theorem can be stated and proved for K (a, m) with a > 0 sufficiently small. In this case however the functions u = t − r∗ , u = t + r∗ where √ r∗ = r + r+ ln(r − r+ ), r+ = m + m2 − a2 are not optical functions. To avoid this problem one can measure decay in a different way. The idea is to start with hypersurface Σ0 , as in Figure 8B, and translate it using the flow φτ associated to the stationary Killing vector field T = ∂t . This defines a foliation Στ = φτ (Σ0 ). Theorem 5. — Let Σ0 and foliation Στ defined as above (see Figure 8B), in the exterior of the Kerr space-time K (a, m), with a sufficiently small. Any solution to the wave equation (5.1), with reasonable initial data on Σ0 , admits the following estimates: 1. There exists a constant C such that, uniformly(19), |r1/2 φ| ≤ Cτ −1+δ . 2. Also, uniformly, |rφ| ≤ Cτ −

1−δ 2

.

6. VECTOR FIELD METHOD FOR THE WAVE EQUATION We discuss here modifications of the vector field method for the wave equation in a globally hyperbolic Lorentzian space-time ( M , g), (6.1)

g φ =

0.

Multiplier method. We start with the energy momentum tensor, (6.2)

1 Qαβ = Qαβ [φ] = Dα φDβ φ − gαβ gµν Dµ φDν φ). 2

(18)

The result has been recently improved by J. Luk, see [33] using geometric methods. A similar result was also announced by Tataru in [42] using Fourier methods. (19) Note that the loss of δ was recently removed.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

110

S. KLAINERMAN

One can easily check that Qµν is symmetric and verifies the local conservation laws Dν Qµν = 0 as well as the positive energy conditions Q(X, Y ) ≥ 0, for all causal, future oriented vector fields X, Y. Unlike the Bel-Robinson tensor encountered above, the energy-momentum tensor of the wave equations is not traceless, in the interesting µν physical dimension n = 3. Indeed gµν Qµν = − n−1 2 g Dµ φDν φ. Given a vector field X with deformation tensor (X) π = L X g, i.e., (X) παβ = Da Xβ + Db Xα , we have 1 µν (X) Q πµν . 2 We integrate (6.3) on a past domain of dependence(20) sandwiched between two spacelike hypersurfaces Σ0 and Σ1 with future unit normal T . Let N denote the null boundary of the future set of D(0, 1) and L the geodesic null generator (i.e. DL L = 0 and g(L, L) = 0), of N , normalized by the condition g(L, T ) = −1 on Σ0 ∩ N . Integrating (6.3) in D(0, 1) we derive the formula Z Z Z ZZ 1 Q · (X) π. Q(X, L) + Q(X, T ) = Q(X, T ) − N Σ1 Σ0 D(0,1) 2 Dµ (Qµν X ν ) =

(6.3)

This formula is particularly useful if X is Killing and time-like in which case π = 0 and the two boundary integrands on the left are positive. In the particular case when g is the Minkowski metric and X = T0 = ∂t is the time derivative with respect to the standard coordinate t, we derive the standard law of conservation of energy. The method turns out to be useful, even if X is not Killing, by adding a lower order correction to the pointwise identity (6.3). More precisely we modify the energy momentum Q as follows,

(X)

1 1 Q(X, Y ) + w φ · Y (φ) − Y (w) φ2 , 2 4 with w a scalar function to be chosen appropriately. Q(w) (X, Y )

=

Proposition 2. — The following integral identity holds true in a past domain of dependence as above, Z Z Z Z Q(w) (X, L) + Q(w) (X, T ) = Q(w) (X, T ) − Err(φ; w, X) N

Σ1

Σ1

D(0,1)

with integrand Err = Err(φ; w, X) given by (6.4) (20)

Err =

1 Q· 2

(X)

 1 π + w · g(dφ, dφ) − (w)φ2 . 4

D(0, 1) is such that the causal past set of any point in D, in the slab between Σ0 , Σ1 , is included

in D.

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

111

Proof. — Consider Pµ = Qµν X ν + 12 wφDµ φ − 41 Dµ wφ2 and calculate its divergence, Dµ Pµ = 12 π µν Qµν + 12 wDµ φDµ φ − 14 (w)φ2 which we then integrate on our causal domain D. (X) Remark. — Typically we want to choose w = n−1 π to cancel the lagrangian 2 tr n−1 µν µν µ term in π Qµν = π ˆ Qµν − 2 trπD φDµ φ. In some situations, as in Examples 2, 3 below, it pays to choose instead w = 12 tr (X) π.

Below are two important examples (both due originally to C. Morawetz) in Minkowski space, in a domain D = {t, x)/t0 ≤ t ≤ t1 , |x| ≤ t − r0 } ⊂ Rn+1 sandwiched between Σ0 = {t = t0 } and Σ1 = {t = t1 }. Thus, L = ∂t + ∂r and T = T = ∂t . Example 1. — Let X be the conformal Killing vector field K0 = (t2 + |x|2 )∂t + 2txi ∂i with deformation tensor (K0 ) π = −4tm. Thus tr( (K0 ) π) = −4(n + 1)t. Since (K0 ) tr(Q) = − n−1 = n−1 π) to make the term 2 g(dφ, dφ), we choose w 2 (tr Q · (X) π + wg(dφ, dφ) vanish identically. We derive the conservation law, with (K0 ) w = n−1 π): 2 (tr Z Z Z (6.5) Q(w) (K0 , L) + Q(w) (K0 , T ) = Q(w) (K0 , T ). N

Σ1

Σ0

R

R

We can easily check that both Σ1 Q(w) (K0 , T ) and N Q(w) (K0 , L) are positive. In fact one can show, for n ≥ 3 (see [29]), for a small constant c > 0, with L = ∂t + ∂r , L = ∂t − ∂r  Q(w) (K0 , T ) ≥ c (t + r)2 |Lφ|2 + (t − r)2 |Lφ|2 + r2 |∇φ|2 + |φ|2 . Example 2. — Start with X = ∂r . We have Å ã 2 xi xj (X) π00 = (X) π0i = 0, (X) πij = δij − , r |x| |x|

i, j = 1, . . . , n.

Hence, tr (X) π = 2(n−1) . Thus, choosing w = 21 tr (X) π, we have, with ∇ / denoting the r induced covariant differentiation on the spheres S(t, r) of constant t and r Q·

(X)

2 / φ|2 π + w g(dφ, dφ) = Dα Dβ (X) π αβ = |∇ r Å ã 1 n−1 1 g (w)φ2 = ∆ φ2 . 4 4 r

In the particular case when n = 3, since ∆( 1r ) = −4πδ0 , we deduce Err(φ; w =

1 (X) 1 tr π, X = ∂r ) = |∇ / φ|2 + 2πδ0 φ2 . 2 r

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

112

S. KLAINERMAN

Therefore, with w = 12 tr (X) π = 2r Z t1 Z Z Z Z 1 2π |φ(t, 0)|2 dt+ Q(w) (∂r , ∂t )− Q(w) (∂r , ∂t )− Q(w) (∂r , L). |∇ / φ|2 = t0 D r Σ1 N Σ0 One can easily bound the surface integrals on the right-hand side by energy estimates (using vector field X = T = ∂t ) and thus derive a very useful space-time inequality for the left-hand side. Example 3. — Take as vector field X 0 = f (r)X = f (r)∂r . We have, in general, (f X) παβ = f (X) παβ + Dα f Xβ + Dβ Xα , tr((f X) π) = f tr (X) π + 2X(f ). Hence, for X = ∂r , we deduce Q·

(f X)

π

π + 2f 0 (r)|∂r φ|2 − f 0 (r)g(dφ, dφ)  1 = f (X) π(dφ, dφ) + 2f 0 |∂r φ|2 − f tr (X) π + f 0 g(dφ, dφ). 2 =

fQ ·

(X)

Therefore, for w = 12 tr (f X) , since Q·

(f X)

(X)

π(dφ, dφ) = |∇ / φ|2 and tr (X) π =

π + w g(dφ, dφ) = f (r)|∇ / φ|2 + 2f 0 (r)|∂r φ|2  1 (n − 1)f (r) 1 w = ∆ + f 0 (r) . 4 4 r

For n = 3, with F a primitive of f , i.e. f (r) = F 0 (r), hence, with w = w(f X) = 12 tr((f X) π), Err(φ; w, f X)

2(n−1) , r

1 1 (f X) π) 4  2 tr(

= f (r) (X) π(dφ, dφ) + 2f 0 (r)|∂r φ|2 −

=

1 2 4∆ F

and,

1 2 2 φ ∆ F (r). 4

To obtain a coercive estimate we thus need f, f 0 ≥ 0 and ∆2 F ≤ 0. One can easily rλ check that f (r) = 1+r λ , 0 ≤ λ ≤ 1 verifies these requirements. In the particular case λ = 1 we derive. Proposition 3. — The following estimate holds true for arbitrary solutions of φ = 0 in R3+1 , for an arbitrary R > 0, Z Z ∞Z  |Dφ|2 + |φ|2 . |Dφ|2 . 0

|x|≤R

Σ0

To summarize: The multiplier method consists in finding vector fields X and scalars w = w(X) such that at least one of the following statements holds true in a past causal domain D: – The vector field X is coercive, i.e. we have both Err(φ; w(X), X) ≥ 0 and Q(w) (X, L), Q(w) (X, T ) are positive at the future boundary of D. – The vector field X is positive, i.e. Err(φ; w(X), X) ≥ 0, and we have a way to estimate the boundary terms along N and Σ1 .

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

113

In practice it is very hard to find good vector fields X which achieve either of the two conditions. As we have seen, in the stability of the Minkowski space, one defines vector fields X, analogous to T0 , K0 on Minkowski space, such that the integrand Err is sufficiently small so that the corresponding space-time integral can be controlled. Also, as we shall see in the next section, it is very difficult to find globally defined vector fields X and scalars w = w(X) for which Err(φ; w(X), X) has definite sign. The new idea, pursued by Dafermos-Rodnianski, is to concentrate in regions of spacetimes, not necessarily causal domains (such as a small neighborhood of the event horizon in Schwarzschild or the entire ergo-region in Kerr), where the natural Killing vector fields of the space-time are degenerate and look for new vector fields for which Err(φ; X.λ) has a sign in the restricted region. Once we control these degenerate regions we can hope to get a global coercive vector field by a patching procedure. 6.1. Commuting vector field method As in the stability of the Minkowski space it is not enough to derive estimates by the multiplier method. One needs in addition to commute the equation with enough suitable vector fields. In the case of the wave equation this is provided by the following. Lemma 1. — For an arbitrary vector field X we have, g (Xφ) = X(g φ) −

(X) αβ

π

 Dα Dβ φ − 2Dβ (X) παβ − Dα (tr (X) π) Dα φ.

In particular, if X is Killing and g φ = 0 we infer that g (Xφ) = 0 and therefore we can apply to X(φ) the same multiplies method estimates as for φ. There are cases, however, where the error terms obtained by commutation are not small but contain instead terms which lead, by integration, to positive bulk integrals. This, as we shall see, is the case of the red shift vector field discussed below.

7. RED SHIFT In [17] Dafermos and Rodnianski prove a general result concerning the existence of a red shift vector field in a neighborhood of a non-degenerate Killing horizon. This is a null hypersurface N with a null generator L (see appendix for definitions) which is the restriction to N of a Killing vector field N, with complete orbits and flow (φτ )τ ≥0 , and such that ω = g(DL L, L) < 0, for an adapted null companion(21) L. It is easily seen that the future horizon of any K (a, m) with 0 ≤ a < m verifies these assumptions. The result below, however, is a lot more general. (21)

In fact ω can be made constant, related to the surface gravity of the Killing horizon.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

114

S. KLAINERMAN

Proposition 4 (Dafermos-Rodnianski). — Given such a null hypersurface, there exist a neighborhood U of N and a strictly time-like, smooth vector field H on U , both invariant(22), with respect to the N-flow φτ , τ ≥ 0, such that in U , for a constant c>0 (H)

(7.1)

π · Q ≥ c Q(H, H).

Moreover, given any Λ > 0, we can choose H such that, all along N  (H) π · Q ≥ c e3 (φ)2 + Λ (e4 (φ)2 + |∇ / φ|2 . The proof of the proposition is based on the following lemma. Lemma 2. — Assume given a small portion of null hypersurface N , in a neighborhood of a compact cross section S, with an adapted null pair (e3 , e4 ) (see appendix) such that ω = g(D4 e4 , e3 ) < 0. Extend e3 in a small, space-time, neighborhood of S by solving the differential equation DX X

= −A X + N ),

X| N = e3 ,

where N is an arbitrary smooth extension of e4 and A a sufficiently large positive constant, whose size depends on Λ > 0 below. Then, in a full neighborhood of S, along N , we have  1 (7.2) Q · (X) π ≥ ce3 (φ)2 + Λ (e4 φ)2 + |∇φ|2 . 2 Proof. — See appendix. The proof of the proposition follows easily by applying the lemma to the case when L is the restriction of the Killing vector field N (recall that N is a Killing horizon) with complete orbits. In that case it suffices to construct X in a small neighborhood of S (restricted, say, to a space-like hypersurface Σ passing through S) and then extend it by using the flow (φτ )τ ≥0 of N (or T in a stationary space-time such as Kerr), in a whole neighborhood of the horizon of the form U = ∪τ ≥0 φτ (U ) where U is a neighborhood of S in Σ. Since L N (X) π = L N L X g = L X L N g = 0 the positivity property of Q · (X) π on the neighborhood of S in Σ is preserved all through the neighborhood U of N . Moreover, the same is true for the deformation tensor of the vector field H = N + X. It remains, however, to check whether it is realistic to expect that L = N is both Killing and verifies the condition −ω > 0. This property defines in fact non-degenerate Killing horizons. In the particular case of the Schwarzschild space-time one can check directly that the stationary Killing field T verifies both properties along the event (22)

Or T-invariant in a stationary metric such as Kerr.

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

115

horizon. The same is true for all Kerr solutions with a < m, but in that case the vector field N differs from T, which is space-like on the horizon. In fact the following general result holds true, see [2]. Proposition 5. — Any non-expanding, bifurcate, null hypersurface ( N , N , S) admits a future directed Killing vector field N, defined in a neighborhood of S, which is tangent to the null generators of the horizon. Moreover, given an arbitrary null geodesic vector field L on N with affine parameter u, N must be of the form N = κ u L for some constant κ > 0. 7.1. The red shift vector field as commutator The red shift vector field provides useful estimates near horizon even when used as commutator. In view of Lemma 1 we have, with π the deformation tensor of vector field H, g (Hφ) = −π αβ Dα Dβ φ + · · · , where we ignore the terms linear in the first derivatives of φ, which may be assumed as having been already estimated. One can easily check that π3a = 0, see Appendix 10.2. Thus, π αβ Dα Dβ φ does not contain the derivatives D4 Da φ. Hence (see Appendix 10.2), since π44 = −2ω, π34 = ω, −π αβ Dα Dβ φ

=

2ωD23 φ − 2ωD3 D4 φ + π4a D3 Da + πab Da Db + · · · .

One can also eliminate the term D3 D4 φ using the equation φ = 0 since the principal terms of , expressed relative to our null frame, are of the form −2ωD3 D4 +δ ab Da Db . We deduce that (7.3)

g (Hφ) = 2ωD23 φ + Aa D3 Da φ + Bab Da Db φ + · · ·

with bounded A, B. Now, when applying the multiplier method to (7.3), i.e. replacing φ with H(φ) in the previous step, we can take advantage of the negative sign of 2ωD23 φ and absorb all other second derivatives choosing the constant Λ > 0 in (7.2) sufficiently large. 7.2. Modified Morawetz vector field in Schwarzschild To take care of the trapped region r = 3m in Schwarzschild one needs to construct a vector field of the form h(r)∂r similar to the one of Morawetz in Example 3 above. In fact it is better to work with the modified Regge-Wheeler coordinate r∗ = r + 2m log(r − 2m) − 3m − 2m log m, such that r∗ = 0 for r = 3m. In these  coordinates the Schwarzschild metric takes the form µ − dt2 + (dr∗ )2 + r2 dσ 2 with dr ∗ m −1 r∗ r∗ µ = (1 − 2m and ∂r∗ µ = 2m r ). Observe that dr = µ r 2 µ. Hence Γr ∗ r ∗ = r 2 = Γtt and ∗ ∗ Γtr∗ r∗ = Γrr∗ t = Γrtt = Γttt = 0. Also, for an arbitrary orthonormal frame e1 , e2 on the spheres of constants r and t, Da eb = ∇a eb − δrab ∂r∗ . We look for a vector field X = f ∂r∗ and scalar w = w(f ) such that Err(φ; w, X) ≥ 0 for an open neighborhood, in r∗ of r∗ = 0. To motivate the calculations consider

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

116

S. KLAINERMAN

first X = ∂r∗ for which we can easily calculate the only non-xero components, with respect to the frame ∂t , ∂r∗ , e1 , e2 of its deformation tensor, i.e. (X) πtt = −2 rm2 µ, 4µ (X) ∗ ∗ (X) πr r = 2 rm2 µ, πab = 2µ π = 4m r δab . Thus, tr r 2 + r . Also, since, g(dφ, dφ) = −µ−1 (∂t φ)2 + µ−1 (∂r∗ φ)2 + |∇ / φ|2 we derive Qαβ (X) παβ

1 π(dφ, dφ) − tr (X) πg(dφ, dφ) 2ã Å 2m 1 (X) r − 3m − tr |∇φ|2 = π g(dφ, dφ) + 2 r2 2 r2 r − 3m 2µ = 2 |∇ / φ|2 − g(dφ, dφ). 2 r r (X)

=

Thus, 1 αβ (X) r − 3m µ |∇ / φ|2 − g(dφ, dφ). Q παβ = 2 r2 r To eliminate the lagrangian term we are led to choose w = µr for which 1 r − 3m Q(w) · (X) π = |∇ / φ|2 2 r2 which, unlike the case of Minkowski space, does not have a definite sign. We look for a modification of X of the form f X = f (r∗ )∂r∗ . As in Example 3 above we find Q·

(f X)

(7.4)

π

= f Q · (X) π + 2Q(df, X) = f Q · (X) π + 2f 0 µ−1 Q(X, X) 2f µ r − 3m |∇ / φ|2 − g(dφ, dφ) + 2f 0 µ−1 Qr∗ r∗ = 2f 2 r r Å ã r − 3m 2f µ 1 2 0 −1 2 ∗ = 2f |∇ / φ| − g(dφ, dφ) + 2f µ (∂ φ) − µg(dφ, dφ) r r2 r 2 Å ã 2f µ r − 3m |∇ / φ|2 + 2f 0 µ−1 (∂r∗ φ)2 − f 0 + g(dφ, dφ). = 2f r2 r

Recalling Formula (6.4), with w = f 0 +

2µ r ,

and setting W = − 41 ∆(w), we derive

r − 3m |∇ / φ|2 + f 0 µ−1 (∂r∗ φ)2 + W φ2 . r2 To obtain a coercive estimate we need to choose a function f = f (r∗ ) such that f 0 ≥ 0, f r−3m ≥ 0 and W ≥ 0. This cannot be done, but one can find an f which verifies r2 the first two properties and such that W > 0 in a small neighborhood of r = 3m. Therefore, if the function φ is given by its decomposition into spherical harmonics X φ= φ` , Err(φ, w, f ∂r∗ ) = f

`≥0

then for the part φL =

P

φ` , with L sufficiently large, for which Z L(L + 1) 2 |∇ / φL | ≥ |φL |2 , r2 S2 S2

`≥L

Z

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

117

we can find an appropriate function fL , bounded, increasing and vanishing at r = 3m, and a scalar wL such that ErrL = Err(φL ; wL , fL (r∗ )∂r∗ ) has the lower bound Z Z Å ã r − 3m 2 2 0 −1 2 2 ∗ φL ) + F |φL | (|∇ / φ | + |∂ φ | ) + 2f µ (∂ ErrL ≥ c fL L t L r L r2 S2 S2 for some positive function F . For the remaining first L harmonics, one can find functions f` and scalars w` such that, for Errl = Err(φ` ; w` , f` (r∗ )∂r∗ ) Z Z  F (|∇ / φ` |2 + |∂t φ` |2 ) + 2f 0 µ−1 (∂r∗ φ` )2 + F |φ` |2 . Errl ≥ c S2

S2

Combining we obtain Z Å Z ã r − 3m 2 2 0 −1 2 2 ∗ . f (|∇ / φ| + |∂ φ| ) + 2f µ (∂ φ) + F |φ| Err(φ) ≥ c t r r2 S2 S2 Two alternative approaches for obtaining a positive definite quantity, without a decomposition into spherical harmonics, have been advanced. One relies on combining (7.4) with an appropriate choice of a scalar w and the red shift vector field, see [35]. The other, [18], exploits a combination given by the expression Qw1 [φ] ·(f1 X) π + Qw2 [O(φ)] ·(f2 X) π with angular momentum vector fields O. In all of these approaches the generated expression degenerates, relative to the principle terms, at the photon-sphere r = 3m, thus necessitating a loss of regularity to obtain a non-degenerate estimate. The corresponding construction in Kerr with small angular momentum is much more subtle, as the trapped set is no longer confined to a co-dimension one manifold r = 3m in physical space. Its structure has to be now captured in the cotangent space, where it is governed by the geodesic flow. In Kerr, the geodesic flow is integrable, which equivalently can be expressed in terms of the separability of the wave equation— respecting the decomposition XZ X φ(t, r, ϕ, θ) = eiωt eimϕ Sλ,m (aω, θ)uω λ,m (r), m≥0

λ

where Sλ,m (aω, θ) are the oblate spheroidal harmonics and λ is the Carter constant— an additional, to ω and m, integral of motion—or existence of a Killing (Carter) tensor. In the Kerr case with a  m, the (degenerate) analog of the Morawetz estimate can be derived with the help of three different approaches. In the first, one replaces a vector field f (r∗ )∂r∗ by an appropriately constructed pseudo-differential operator, [43]. The second approach, [17], combines different X estimates with appropriately defined ω ω functions fm,λ and scalars wm,λ dependent on the geometric frequencies ω, m, λ. In the third approach, [4], one explores a combination of X type identities for φ and for

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

118

S. KLAINERMAN

the quantity obtained by fusing the Carter tensor and φ. All three approaches rely on the integrability of the geodesic flow and in particular imply the estimate Z Z  f1 (|∇ / φ|2 + |∂t φ|2 ) + f2 (∂r∗ φ)2 + F |φ|2 Err(φ) ≥ c Σt

Σt

for some nonnegative function f1 , vanishing in a neighborhood of r = 3m, and positive functions f2 and F .

8. BOUNDEDNESS RESULTS 8.1. Simplest case Consider first a static space-time ( M , g) which is the MFGHD of an initial data set Σ0 and such that the Killing vector field T is everywhere time-like and orthogonal to Σ0 (23). Let t be the time function associated to t, i.e T(t) = 1 and t = 0 on Σ0 . Starting with a local system of coordinates x = (x1 , . . . , xn ) on Σ0 and parametrizing points along the orbits γ of T by the parameter t and the x coordinates on γ ∩ Σ0 we easily see that M = Σ0 × R and the space-time g metric takes the form (8.1)

g = −n2 (x)dt2 + gij (x)dxi dxj ,

with x = (x1 , . . . , xn ) an arbitrary coordinate system on Σ0 and g a Riemannian metric. Our assumptions imply, for a sufficiently small constant λ0 , uniformly in M λ0 ≤ n ≤ λ−1 0 ,

2 λ0 |ξ|2 ≤ gij ξ i ξ j ≤ λ−1 0 |ξ| .

Also, relative to our system of coordinates, T = ∂t . We normalize T by introducing the vector field e(0) = n−1 T = n−1 ∂t , unit future normal to the space-like foliation Σt defined by the level surfaces of t. We decompose a space-time vector field X relative to the unit time-like e(0) , (8.2)

X = X 0 e(0) + X,

g(e(0) , X) = 0,

and define the positive definite Riemannian metric, (8.3)

h(X, Y ) = X 0 · Y 0 + g(X, Y ).

Given an arbitrary tensor-field π we denote by |π| its norm with respect to the metric h. (23)

This implies, in particular, that all orbits of T are complete, see [13], and must intersect Σ0 (orthogonally), see [15].

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

119

Proposition 6. — Any solution φ of the wave equation (8.4)

g φ = 0,

φ|t=0 = φ(0) ,

∂t φ|t=0 = φ(1)

with smooth, compactly supported, initial data φ(0) , φ(1) on Σ0 is globally bounded. Proof. — According to our general procedure we have, with Q the energy-momentum tensor Z Z Q(T, e(0) ) ≤ C. Q(T, e(0) ) = (8.5) Σ0

Σt

Hence, since Q(T, e(0) ) =

and λ0 ≤ n ≤ λ−1 0 we deduce Z |Dφ|2 ≤ λ−2 |Dφ|2 . C 0

1 2 2 n|Dφ| ,

Z (8.6) Σt

Σ0

with a constant C depending only on λ0 and the initial data. In view of our definition above, we have |Dφ|2 = (e(0) φ)2 + |∇φ|2 , where ∇ denotes the induced covariant derivative on Σt . We plan to bound the L∞ norm of φ in terms of the L2 norms of P its higher derivative, according to Sobolev inequality, kφ(t)kL∞ . si=1 k∇i φ(t)kL2 for s > 23 . To get the higher derivative we commute  with T. Since T is Killing we must have T(φ) = 0 and therefore, repeating the first step Z |D(Tφ)|2 . C Σt

. C. Now we can write(24)  = −n−2 ∂t2 +∆g from from which, in particular, which we infer that k∆g φkL2 (Σt ) is uniformly bounded. Using the Bochner identity for ∆g , the boundedness of the curvature tensor of g (and of derivatives of n) and the first derivative estimates already established we then deduce that k∇2 φkL2 (Σt ) is uniformly bounded in t. Using the vanishing of φ at infinity (on each Σt ) and elliptic estimates, we can also derive a bound for kφkL2 (Σt ) . We can repeat the procedure, by commuting g once more with T, to establish bounds for all higher derivatives k∇k φkL2 (Σt ) , k ≥ 0. Thus, by Sobolev, φ is uniformly bounded. R

|∂t2 φ|2 Σt

8.2. First degenerate case We assume next the more realistic hypothesis that T is not time-like everywhere but degenerates in fact along a horizon, i.e. a null hypersurface N along which T is tangent to its generators. This, of course, is the situation in Schwarzschild. Since we have to work with space-like hypersurfaces transversal to the horizon we will not make use(25) of the condition that T is hypersurface orthogonal. We choose an original (24)

Note that ∆g differs from ∆g by first order terms in ∇φ. It can be shown, however, that a stationary space-time with T tangent to the generators of the horizon must be in fact static. (25)

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

120

S. KLAINERMAN

space-like hypersurface Σ0 and translate it using the flow of T to obtain a space-like foliation Στ , as in the picture below. It is easy to show that away from the horizon we still have Q(T, e0 ) ≥ C|Dφ|2 . The constant C however degenerates as we approach the horizon. Yet some control remains. Thus, precisely on the horizon, we have using an adapted null frame, as in appendix, normalized such that e0 = 21 (e3 + e4 ), and such that T = −ωe4 with ω = g(D4 e4 , e3 ) < 0. Therefore, the energy density Q(T, e0 ) = − 12 ω Q(e4 , e4 ) +  Q(e3 , e4 ) = − 21 ω (e4 φ)2 + |∇ / φ|2 . In other words we are only missing the transversal derivative e3 (φ). Similarly, the flux density Q(T, e4 ) = −ωQ(e4 , e4 ) = −ω|e4 (φ)|2 , i.e we are missing the angular derivatives ∇ / φ. Through a clever argument Kay and Wald, see [27], were able to overcome these difficulties and still derive a boundedness result without using any new vector field. Proposition 7. — Any solution φ of the wave equation (8.7) (8.7)

g φ = 0,

φ|t=0 = φ(0) ,

∂t φ|t=0 = φ(1)

in Schwarzschild space-time with smooth initial data φ(0) , φ(1) on Σ0 , decaying sufficiently fast at infinity, is globally bounded in the domain of outer communication E. Proof. — The red shift vector field of Dafermos-Rodnianski provides a far more powerful and compelling proof, which holds in fact for any stationary space-time in which T is everywhere time-like in the complement of the event horizon. The idea is that, precisely in a neighborhood of the horizon N , where the energy identity due to T becomes degenerate, we gain the missing information from the red shift vector field H. Indeed, along the horizon H is future time-like. Hence the energy density and flux density associated to the red shift vector field H provide precisely the information we would get from the Killing field T if there was no degeneracy at N . So far this information is purely local. To obtain a useful estimate we need to also make use of the fact that H π ·Q ≥ c Q(H, H) in a space-time neighborhood U of N , as + in Proposition 4. We first extend H to our entire domain D = I (Σ0 ) ∩ E by making sure that it coincides with T away from a slightly larger, T-invariant neighborhood V . We can also arrange that the extended H is also T invariant and that, in V \ U , we have | Q · H π | . Q(T, T). Indeed this can be first arranged on Σ0 , by an extension of the form f H + (1 − f )T (with a smooth f such that f = 1 on U ∩ Σ0 and f = 0 in the complement of V ∩ Σ) and then extended to the entire domain D by using the pushforward with φτ . We then apply Proposition 2 for vector field X = H and w = 0, in the domain

D(0, τ ), the region of D between Σ0 and Στ . Since H π = 0 in the complement of V

ASTÉRISQUE 339

(1015)

we have Z Z Q(H, e4 )+ N (0,τ )

Z Q(H, e0 ) =

ZZ Q(H, e0 )−

ZZ H

U (0,τ )

Σ0 τ

F (τ 0 )dτ 0 +

. F (0) − 0

Z . F (0) −

Z

τ

0

Z

π· Q .

V (0,τ )\ U (0,τ )

Since H π · Q ≥ c|Dφ|2 in U , | H π · Q| . |Dφ|2 in V \ U and Q(H, e0 ) ≥ |Dφ|2 , we deduce(26) Z Z ZZ Z F (τ ) := |Dφ|2 . |Dφ|2 − |Dφ|2 + Z

H

π·Q−

U (0,τ )

Στ

Στ

Στ

121

LINEAR STABILITY OF BLACK HOLES

R

N (0,τ ) Q(H, e4 ) ≥ 0,

|Dφ|2

D(0,τ )\ U (0,τ )

|Dφ|2

Στ 0 \ U

τ

F (τ 0 )dτ 0 + Cτ.

0

Thus, by Gronwall we derive a global bound for F (τ ), i.e. a bound for the L2 norm of all first derivatives of φ. To estimate the higher derivative we commute the wave equation not only with T but also with the red shift vector field H. Indeed, commutation with T provides estimates for supτ ≥0 kDT(φ)kL2 (Στ ) , estimate which degenerates only near the horizon N . This degeneracy is more than compensated by commuting the wave operator with H. Thus, repeated commutations with T and H and elliptic theory, as in the simpler case explained below, provide bounds for all higher derivatives of φ. 8.3. The super-radiant regime The method of proof described above can be extended to the case when the vector field T becomes space-like in a neighborhood of the horizon, as is the case in Kerr. The major difficulty in this case is that the global energy density associated T is not positive definite in the ergo-region and therefore ceases to provide any useful information, at least in a first approximation. The effect of super-radiance is well described in the physics literature, starting with the pioneering work of Penrose [36] and Zel0 dovich [44], and provides an amplification mechanism for linear waves. Nevertheless, Dafermos-Rodnianski were able to extend their methods to cover the case of axially symmetric stationary space-times which are sufficiently close to Schwarzschild. Thus, in addition to T the space-time has a second Killing vector field Z, with circular orbits, tangent to the horizon N . One can show, in this case, for a constant γ > 0, and a suitably defined null generator L = e4 , T = L − γZ along the horizon N . In other words N is also a Killing horizon for a Kerr space-time. (26)

In the last line of the inequality below we make use of the fact that, away R from the neighborhood |Dφ|2 in terms of

U of the horizon, the energy identity provided by T gives us a bound for

Στ 0 \ U

initial conditions.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

122

S. KLAINERMAN

Thus, the flux density associated to T is Q(T, L) = Q(L − γZ, L) = |Lφ|2 − γ(Zφ)(Lφ) = (Tφ)2 + γ(Zφ)(Tφ). Therefore, if |Tφ| > γ|Zφ|, we must have Q(T, L) > 0. This suggests a decomposition of φ = φ] + φ[ such that Q[φ] ](T, L) ≥ 0. It can be made precise by decomposing φ with respect to Fourier frequencies ω ∈ R relative to T, and discrete frequencies m, relative to Z. Thus, by a simple cut-off, φ] will be restricted to the frequency range |ω| > |γ|m, called sub-radiant regime, while φ[ , the super-radiant part of φ, has frequencies in the range ω ≤ γm. We expect that the arguments used in the previous subsection would work to treat the non super-radiant part φ] , for which T continues to provide a coercive energy identity. The real new issue is φ[ . One can show, and this is the main new insight of Dafermos-Rodnianski [20], that in stationary axisymmetric space-times near Schwarzschild, in particular in K (a, m) with a  m, the super-radiant frequencies of g φ = 0 are not trapped. The quantitative manifestation of this fact is reflected in the existence of a “simple” vector field X = f (r∗ )∂r∗ and a scalar w = w(X) with the property that Q(w) [φ[ ] ·(X) π ≥ Cr1 ,r2 χr1 ,r2 (|Dφ[ |2 + |φ[ |2 ) with a characteristic function χr1 ,r2 equal to one in the region 2m < r1 ≤ r ≤ r2 < ∞. The relative ease of the choice of X hinges on the fact that for φ[ the lagrangian term g(dφ[ , dφ[ ) is positive in a neighborhood of r = 3m—the trapped set in Schwarzschild. This inequality leads to a non-degenerate version of the Morawetz estimate and together with the red shift estimate allows one to control φ[ . I should note that the actual analysis is complicated by coupling between φ] and φ[ introduced by cut-offs in the physical space which are, unfortunately, required to justify the time Fourier frequencies ω.

9. DECAY MECHANISM The proof of decay in both Theorems 4 and 5 hinges on two basic steps plus a final iteration procedure based on the pigeon hole principle. We consider below the simpler case of decay in Schwarzschild. We consider T-invariant regions obtained by intersecting space-time domains of the form 2m < R1 < R < R2 or 2m ≤ r < R with the future of Σ0 , in the exterior domain E. Also, in what follows, N is the portion of the horizon r = 2m to the future of Σ0 . Step I. — The goal of the first step is to derive an estimate of the form ZZ (9.1) |Dφ|2 ≤ C V I0 , V

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

123

where V is an arbitrary large neighborhood of N , containing the trapped region and I0 a constant depending only on the initial data of φ (it depends in fact on the L2 (Σ ∩ V ) of the first two derivatives of φ. The proof of such an estimate requires the following substeps: 1. Using the red shift vector field H one can control the non-degenerate energy in a small T-invariant neighborhood U 1 of the horizon (such as 2m ≤ r < r1  3m), at the expense of having to control the space-time integral of |Dφ|2 in the complement of U 1 in a somewhat larger neighborhood (such as r ≤ r1 + ). 2. Using the modified Morawetz vector field M one can control the space-time  integral of f (r) |Dφ|2 + |φ|2 , with f (r) vanishing of order 2 at r = 3m, in a sufficiently large T-invariant neighborhood U 2 of the trapped region r = 3m which intersects U 1 (such as r1 −  < r < R, for an arbitrary R > 3m). 3. Commuting the equation with T and using elliptic theory (or, alternatively, commuting also with the angular momentum vector fields) we derive a similar estimate for f (r)|D2 φ|2 + |Dφ|2 . Thus, by losing one derivative, we control the space-time integral of |Dφ|2 in U 2 . 4. Combining this last estimate with the previous estimate in U 1 we derive a space-time estimate for |Dφ|2 in V = U 1 ∪ U 2 . We also derive a non-degenerate estimate along the horizon. Step II. — The goal now is to derive a decay estimate by using the previous step together with asymptotic information from future null infinity. Originally this was done by using a natural adaptation of the vector field K0 of Minkowski space. Here I will sketch instead the new procedure of [18]. To simplify matters I will first present their argument in Minkowski space. It will be quite transparent from the proof how to adapt it to the Schwarzschild case. In fact, once the first step above has been accomplished (which is a lot more delicate in a Kerr background because of the extended trapped region) the same proof also applies to Kerr. The idea is to foliate Minkowski space by hypersurfaces Σ(τ ) = ΣL τ ∪ ΣR (τ ) divided by r = R, for a fixed value R. The left piece is a space-like hyperplane ΣL (τ ) = {(t = τ, x)/|x| ≤ R} while the right piece is the null hypersurface ΣR (τ ) = {(t, x)/t − |x| = τ − R, |x| ≥ R}. Let DL (τ1 , τ2 ) and DR (τ1 , τ2 ) be the regions to the left and right with for τ1 ≤ τ2 as in the figure below (the figure on the right is the same as that on the left, but viewed in the compactified Penrose diagram of the Minkowski space). We start with the following estimate: ZZ (9.2) (|Dφ|2 + |φ|2 ) . CR E(τ ) DL (τ,∞)

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

124

S. KLAINERMAN

where E(τ ) is the non-degenerate energy of the slice Σ(τ ), i.e. Z Q[φ](T, N ) (9.3) E(τ ) := Σ(τ )

with N normal to Σ(τ ), i.e. N = ∂t on ΣL and N = L = ∂t + ∂r on ΣR . Thus, Z Z  2 (Lφ)2 + |∇ / φ|2 . E(τ ) = |Dφ| + ΣL (τ )

ΣR (τ )

This is, essentially, the estimate obtained at the first step. It also follows by using a variation of the Morawetz vector field discussed before. Observe also that E(τ ) is monotonically decreasing(27), i.e. E(τ2 ) ≤ E(τ1 ), in Minkowski space. In the region DR we apply the energy estimate(28) of Proposition 2 with X = rp (∂t + ∂r ) = rp L, 0 ≤ p ≤ 2 and appropriate choice of w. We derive the identity ZZ Z  ˆ 2 + (2 − p)|∇ ˆ 2+ rp−1 (Lφ) / φ|2 rp (Lφ) (9.4) DR (τ1 ,τ2 )

ΣR (τ2 )

Z

rp (∇ / φ)2 =

+

Z

I + (τ1 ,τ2 )

ˆ 2+ rp (Lφ)

Z

ˆ 2 rp |∇ / φ|2 − |Lφ|



DL ∩ DR

ΣR (τ1 )

ˆ = 1 (∂t + ∂r )φ + 1 φ. Ignoring the boundary term at future null infinity we where Lφ 2 2r derive, for p = 2 Z ZZ Z ˆ 2+ ˆ 2. ˆ 2 + IR (τ1 , τ2 ) r2 (Lφ) r(Lφ) r2 (Lφ) DR (τ1 ,τ2 )

ΣR (τ2 )

ΣR (τ1 )

with Z

 ˆ 2 . R2 r2 |∇ / φ|2 − |Lφ|

IR (τ1 , τ2 ) = DL ∩ DR (τ1 ,τ2 )

Z

|Dφ|2 . DL ∩ DR (τ1 ,τ2 )

Averaging with respect to R (in a small interval near a fixed value) and using (9.2), we derive ZZ Z ˆ 2. ˆ 2 + CR R2 E(τ1 ). (9.5) r(Lφ) r2 (Lφ) DR (τ1 ,τ2 )

Hence, in fact Z ΣR (τ2 )

(27)

ˆ 2+ r2 (Lφ)

ΣR (τ1 )

ZZ

ˆ 2. r(Lφ) DR (τ1 ,τ2 )

Z

ˆ 2 + CR E(τ1 ). r2 (Lφ)

Σ(τ1 )

In Schwarzschild or Kerr we expect some bounded amplification. Alternatively one can proceed exactly as in [18] by multiplying directly the wave equation, in null coordinates. (28)

ASTÉRISQUE 339

(1015)

125

LINEAR STABILITY OF BLACK HOLES

Using the pigeonhole principle applied to (9.5) we infer that there exists a dyadic sequence σn → ∞ such that ÇZ å Z ˆ 2 . σn−1 ˆ 2 + CR E(τ1 ) . (9.6) r(Lφ) r2 (Lφ) ΣR (σn+1 )

ΣR (τ1 )

We now consider (9.4) with p = 1. After averaging in R exactly as before and applying once more (9.2) we deduce Z Z ZZ  2 2 2 ˆ ˆ 2 + CR E(σn−1 ). ˆ |Lφ| + |∇ r(Lφ) r(Lφ) + / φ| . DR (σn ,σn−1 )

ΣR (σn )

ΣR (σn−1 )

Using (9.6) we derive ÇZ ZZ  2 2 −1 ˆ |Lφ| + |∇ / φ| . σn DR (σn ,σn−1 )

å ˆ r (Lφ) + CR E(τ1 ) 2

2

+ CR E(σn−1 ).

Σ(τ1 )

ˆ 2 = (Lφ)2 + φ2 + 1 ∂r (φ2 ). Thus, after an integration by parts Observe that (Lφ) r ZZ ZZ Z ˆ 2= (Lφ) (Lφ)2 − φ2 . DR (σn ,σn−1 )

Hence, ZZ

DL ∩ DR (σn ,σn−1 )

DR (σn ,σn−1 )

2

2

|Lφ| + |∇ / φ|



.

σn−1

DR (σn ,σn−1 )

ÇZ

å 2

2

r (Lφ) + CR E(τ1 )

+ CR E(σn−1 ).

Σ(τ1 )

On the other hand, in view of (9.2) ZZ (|Dφ|2 + φ2 ) . CR E(σn−1 ). DL (σn ,σn−1 )

Adding the last two inequalities together, we derive ÇZ å Z σn −1 2 2 E(τ )dτ . Cσn r (Lφ) + CR E(τ1 ) . + CR E(σn−1 ). σn−1

Thus, with I1 = Σ(τ1 ) and R

Σ(τ1 )

R Σ(τ1 )

r2 (Lφ)2 + CR E(τ1 ) depending only on the initial norm on Z

σn

E(τ )dτ . I1 σn−1 + CR E(σn−1 ).

σn−1

Finally we deduce, by another simple application of the pigeonhole principle and the monotonicity of E(τ ), that for all τ ≥ τ1 , (9.7)

E(τ ) . τ −2 I1

as desired.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

126

S. KLAINERMAN

10. APPENDIX 10.1. Null hypersurfaces Consider a null hypersurface N embedded in M , with unit normal L (which itself is tangent to N ). Clearly N is generated by all null geodesics tangent to L orthogonal to a 2-surface S. In what follows we assume that S has the topology of a 2-sphere. Definition. — The null second fundamental form of a null hypersurface N is defined by (10.1)

χ(X, Y )

= g(DX L, Y ),

where L is a fixed null vector field tangent to the null generators of N and X, Y arbitrary vector fields tangent to N . Observe that the definition depends tensorially on the choice of L, i.e. if L0 = aL we have χ0 = aχ. The trace tr χ can be defined, relative to an arbitrary frame L, e1 , e2 , with g(ea , eb ) = δab , by tr χ = χ11 + χ22 . One can easily check that the definition is independent of the frame or the choice of null normal L. The hypersurface N is said to be non-expanding if the trace of χ vanishes identically. We can foliate N by the level surfaces of an affine parameter s of L, i.e. L(s) = 1, s = 0 on S. We can then define the null companion L of L, at any point p of N , to be the unique null normal orthogonal to the level surface passing through p such that g(L, L) = −1. 10.2. Red shift vector field Consider first an arbitrary null hypersurface N and a null pair (e4 = L, e3 = L), g(e3 , e4 ) = −1 with L null, tangent to N and L hypersurface orthogonal, i.e. orthogonal to a foliation of N by 2-surfaces (see appendix). We complete the null pair to a null frame (e1 , e2 , e3 , e4 ) with e1 , e2 an orthonormal frame tangent to the foliation. We easily check the following D4 e4 = −ωe4 , (10.2)

D4 e3 = ωe3 + η 1 e2 + η 2 e2 Da e3 = χab eb + ζa e3

where ω = g(D4 e4 , e3 ), η a = g(D4 e3 , ea ), ζa = g(Da e4 , e3 ), χab = g(Da e3 , eb ) depend only on the original choice of the null pair (e3 , e4 ) along N . We extend e3 in a small neighborhood of N by solving the equation (10.3)

ASTÉRISQUE 339

D3 e3 = −ωe3

(1015)

LINEAR STABILITY OF BLACK HOLES

127

with ω an arbitrary function on N which we hope to choose later. The deformation tensor of X = e3 can be easily calculated along N π44 = −2ω, π34 = ω, π33 = 0, π3a = 0, π4a = η a − ζa , πab = χab . Therefore, Q · π = Q33 π44 + 2Q34 π34 − 2Q3a π4a + Qab πab / φ|2 − 2∇3 φ ∇ / φ · (η − ζ) + χab ∇a φ∇b φ = −ω(e3 φ)2 + 2ω |∇  1 / φ|2 − trχ − 2e3 (φ) · e4 (φ) + |∇ 2 = −ω(e3 φ)2 + 2ω |∇ / φ|2 + χ ˆ ab ∇a φ∇b φ − 2e3 φ ∇ / φ · (η − ζ) + trχ(e3 φ)(e4 φ). By assuming −ω ≥ κ > 0 and ω sufficiently large positive, we deduce, for some  positive constant c, 21 Q · π + trχ(e3 φ)(e4 φ) ≥ c (e3 φ)2 + |∇ / φ|2 . To get rid of the term trχ(e3 φ)(e4 φ) we need to make a modification of Equation (10.3). We use instead D3 e3

(10.4)

= −ωe3 − Ae4 .

With this modification all components of π remain the same, except π33 = A. Thus, Q·π =

/ φ|2 + A(e4 φ)2 + χ ˆ ab ∇a φ∇b φ − ω(e3 φ)2 + 2ω |∇ − 2e3 φ ∇ / φ · (η − ζ) + trχ(e3 φ)(e4 φ).

Thus, choosing −ω ≥ κ > 0 and constants ω, A sufficiently large, we deduce, for some positive c > 0  1 Q · π ≥ c(e3 φ)2 + Λ |e4 φ|2 + |∇ / φ|2 2 with Λ > 0 arbitrarily large, provided that ω and A are sufficiently large.

REFERENCES [1] S. Alexakis, A. D. Ionescu & S. Klainerman – “Uniqueness of smooth stationary black holes in vacuum: small perturbations of the Kerr spaces”, Comm. Math. Phys. 299 (2010), p. 89–127. [2]

, “Hawking’s local rigidity theorem without analyticity”, preprint arXiv:0902.1173.

[3] S. Alinhac – “Energy multipliers for perturbations of the Schwarzschild metric”, Comm. Math. Phys. 288 (2009), p. 199–224. [4] L. Andersson & P. Blue – “Hidden symmetries and decay for the wave equation on the Kerr spacetime”, preprint arXiv:0908.2265. [5] L. Bieri – “An extension of the stability theorem of the Minkowski space in general relativity”, Ph.D. Thesis, ETH, Zürich, 2007.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

128

S. KLAINERMAN

[6] P. Blue & P. Soffer – “Semilinear wave equations on the schwarzschild manifold. I. Local decay estimates”, preprint arXiv:gr-qc/0310091. [7] P. Blue & J. Sterbenz – “Uniform decay of local energy and the semi-linear wave equation on Schwarzschild space”, Comm. Math. Phys. 268 (2006), p. 481– 504. [8] B. Carter – “Hamilton-Jacobi and Schrödinger separable solutions of Einstein’s equations”, Comm. Math. Phys. 10 (1968), p. 280–310. [9] S. Chandrasekhar – The mathematical theory of black holes, International Series of Monographs on Physics, vol. 69, The Clarendon Press, Oxford Univ. Press, 1983. [10] Y. Choquet-Bruhat & R. Geroch – “Global aspects of the Cauchy problem in general relativity”, Comm. Math. Phys. 14 (1969), p. 329–335. [11] D. Christodoulou – The formation of black holes in general relativity, EMS Monographs in Math., European Mathematical Society (EMS), Zürich, 2009. [12] D. Christodoulou & S. Klainerman – The global nonlinear stability of the Minkowski space, Princeton Math. Series, vol. 41, Princeton Univ. Press, 1993. [13] P. T. Chruściel – “On completeness of orbits of Killing vector fields”, Classical Quantum Gravity 10 (1993), p. 2091–2101. [14] P. T. Chruściel & J. L. Costa – “On uniqueness of stationary vacuum black holes”, Astérisque 321 (2008), p. 195–265. [15] P. T. Chruściel & R. M. Wald – “Maximal hypersurfaces in stationary asymptotically flat spacetimes”, Comm. Math. Phys. 163 (1994), p. 561–604. [16] M. Dafermos & I. Rodnianski – “The red-shift effect and radiation decay on black hole spacetimes”, Comm. Pure Appl. Math. 62 (2009), p. 859–919. [17]

, “Lectures on black holes and linear waves”, preprint arXiv:0811.0354.

[18]

, “A new physical space approach to decay for the wave equation with application to black hole spacetimes”, preprint arXiv:0910.4957.

[19]

, “A note on energy currents and decay for the wave equation on a Schwarzschild background”, preprint arXiv:0710.017.

[20]

, “A proof of uniform boundedness of solutions to wave equations on slowly rotating Kerr backgrounds”, preprint arXiv:0805.4309.

[21] R. Donninger, W. Schlag & A. Soffer – “A proof of Price’s law on Schwarzschild black hole manifolds for all angular momenta”, preprint arXiv:grqc/0908.4292. [22] F. Finster, N. Kamran, J. Smoller & S.-T. Yau – “Decay of solutions of the wave equation in the Kerr geometry”, Comm. Math. Phys. 264 (2006), p. 465–503.

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

129

[23] Y. Fourès-Bruhat – “Théorème d’existence pour certains systèmes d’équations aux dérivées partielles non linéaires”, Acta Math. 88 (1952), p. 141–225. [24] S. W. Hawking & G. F. R. Ellis – The large scale structure of space-time, Cambridge Monographs on Math. Physics, vol. 1, Cambridge Univ. Press, 1973. [25] T. J. R. Hughes, T. Kato & J. E. Marsden – “Well-posed quasi-linear second-order hyperbolic systems with applications to nonlinear elastodynamics and general relativity”, Arch. Rational Mech. Anal. 63 (1976), p. 273–294. [26] A. D. Ionescu & S. Klainerman – “On the uniqueness of smooth, stationary black holes in vacuum”, Invent. Math. 175 (2009), p. 35–102. [27] B. S. Kay & R. M. Wald – “Linear stability of Schwarzschild under perturbations which are nonvanishing on the bifurcation 2-sphere”, Classical Quantum Gravity 4 (1987), p. 893–898. [28] S. Klainerman – “Uniform decay estimates and the Lorentz invariance of the classical wave equation”, Comm. Pure Appl. Math. 38 (1985), p. 321–332. [29]

, “The null condition and global existence to nonlinear wave equations”, in Nonlinear systems of partial differential equations in applied mathematics, Part 1 (Santa Fe, N.M., 1984), Lectures in Appl. Math., vol. 23, Amer. Math. Soc., 1986, p. 293–326.

[30] S. Klainerman & F. Nicolò – The evolution problem in general relativity, Progress in Math. Physics, vol. 25, Birkhäuser, 2003. [31] J. Kronthaler – “The Cauchy problem for the wave equation in the Schwarzschild geometry”, J. Math. Phys. 47 (2006), 042501, 29. [32] H. Lindblad & I. Rodnianski – “Global existence for the Einstein vacuum equations in wave coordinates”, Comm. Math. Phys. 256 (2005), p. 43–110. [33] J. Luk – “Improved decay for solutions to the linear wave equation on a Schwarzschild black hole”, preprint arXiv:0906.5588. [34] M. Machedon & J. Stalker – “Decay of solutions to the wave equation on a spherically symmetric background”, preprint. [35] J. Marzuola, J. Metcalfe, D. Tataru & M. Tohaneanu – “Strichartz estimates on Schwarzschild black hole backgrounds”, Comm. Math. Phys. 293 (2010), p. 37–83. [36] R. Penrose – “Gravitational collapse: the role of general relativity”, Nuovo Cimento 1 (1969), p. 252–276. [37]

, Singularities and time asymmetry in general relativity—an Einstein centenary survey, Cambridge Univ. Press, 1979.

[38] W. Press & S. Teukolsky – “Perturbations of a rotating black hole. II. Dynamical stability of the Kerr metric”, Astrophys. J. 185 (1973), p. 649–673.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

130

S. KLAINERMAN

[39] R. H. Price – “Nonspherical perturbations of relativistic gravitational collapse. I. Scalar and gravitational perturbations”, Phys. Rev. D 5 (1972), p. 2419–2438. [40] T. Regge & J. A. Wheeler – “Stability of a Schwarzschild singularity”, Phys. Rev. 108 (1957), p. 1063–1069. [41] A. A. Starobinsky – “Amplification of waves during reflection from a rotating black hole”, Sov. Phys. JETP 37 (1973), p. 28–32. [42] D. Tataru – “Local decay of waves on asymptotically flat stationary spacetimes”, preprint arXiv:0910.5290. [43] D. Tataru & M. Tohaneanu – “Local energy estimate on Kerr black hole backgrounds”, preprint arXiv:0810.5766. [44] Y. B. Zel0 dovich – “Amplification of cylindrical electromagnetic waves from a rotating body”, Sov. Phys. JETP 35 (1972), p. 1085–1087. [45] F. J. Zerilli – “Gravitational field of a particle falling in a Schwarzschild geometry analyzed in tensor harmonics”, Phys. Rev. D 2 (1970), p. 2141–2160.

Sergiu KLAINERMAN Princeton University Mathematics Department 1108 Fine Hall PRINCETON, NJ 08544-5263 - U.S.A. E-mail : [email protected]

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

131

smf_bbk_1015_blackhole2_2a-3a.jpg

Figure 3a. Complete Penrose diagram of Schwarzschild. Note the black hole and white hall regions, singularity at r = 0, event horizon r = 2m and the boundaries at infinity.

smf_bbk_1015_blackhole2_2a-3b.jpg

Figure 3b. The right disconnected exterior region of Schwarzschild. Note that T = ∂t becomes null along the horizon r = 2m and vanishes on the bifurcate sphere where the two branches of the horizon meet.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

132

S. KLAINERMAN

smf_bbk_1015_blackhole2b.jpg

Figure 4a. Exterior domain of Kerr. Note that the stationary vector field T, which is time-like in the far away (asymptotic) region of space-time, becomes space-like inside the ergo-region, near the horizon H = N ∪N.

smf_bbk_1015_ergosphere.jpg

Figure 4b. Kerr solution, on a fixed slice, as a rotating black hole. Note the axis of symmetry and the presence of the ergosphere outside the event horizon.

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

133

smf_bbk_1015_blackhole3_6-5b.jpg smf_bbk_1015_blackhole3_6-5a.jpg

Figure 5a. Minkowski space in standard coordinates.

Figure 5b. Penrose diagram of the Minkowski space. Note that both the past + − of I and future of I exhaust the entire space.

smf_bbk_1015_paris_1109e.jpg

Figure 6. Behavior of null geodesics in the domain of outer communication by contrast to those in a black hole.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

134

S. KLAINERMAN

smf_bbk_1015_paris_1109d.jpg

Figure 7. Domain of integration for Equation (4.2) with null boundary N and two space-like pieces, Σ0 , Σ1 .

smf_bbk_1015_paris_1109g.jpg

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

smf_bbk_1015_paris_1109g2-3-8a.jpg

Figure 7a. Decay in Schwarzschild can be measured with respect to the double null foliation given by the level hypersurfaces of u = t − r∗ and v = u = t + r∗ .

135

smf_bbk_1015_paris_1109g2-3-8b.jpg

Figure 7b. Decay in Kerr can be measured with respect to a foliation Στ obtained from Σ0 , by using the T- flow. Note that Σ0 consists of two null portions and a space-like one in the middle.

smf_bbk_1015_paris_1109d.jpg

Figure 8c. Domain of integration for Equation (6.3) with null boundary N and two space-like pieces, Σ0 , Σ1 .

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

136

S. KLAINERMAN

smf_bbk_1015_paris_1109i.jpg

smf_bbk_1015_paris_1109f.jpg

ASTÉRISQUE 339

(1015)

LINEAR STABILITY OF BLACK HOLES

137

smf_bbk_1015_paris_1109U2a.jpg

smf_bbk_1015_paris_1109U2.jpg

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

138

smf_bbk_1015_paris_1109b.jpg

ASTÉRISQUE 339

S. KLAINERMAN

(1015)

LINEAR STABILITY OF BLACK HOLES

139

smf_bbk_1015_blackhole4a.jpg

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

339

ASTÉRISQUE 2011

SÉMINAIRE BOURBAKI VOLUME 2009/2010 EXPOSÉS 1012-1026 (1016) Ergodicity of two dimensional turbulence Antti KUPIAINEN

SOCIÉTÉ MATHÉMATIQUE DE FRANCE Publié avec le concours du CENTRE NATIONAL DE LA RECHERCHE SCIENTIFIQUE

Séminaire BOURBAKI 62e année, 2009-2010, no 1016, p. 137 à 156

Novembre 2009

ERGODICITY OF TWO DIMENSIONAL TURBULENCE [after Hairer and Mattingly] by Antti KUPIAINEN

INTRODUCTION The problem of turbulence has been described as the last great unsolved problem of classical physics. Understanding of the complicated motion of fluids in the presence of obstacles or stirring has been a challenge to mathematicians, physicists and engineers for quite a time now. The equations governing macroscopic fluid motion, the Navier Stokes equations, have been known for close to two centuries. For an incompressible fluid in units where the density equals one they read (1)

∂t u + u · ∇u = ν∆u − ∇p + f.

u(t, x) ∈ Rd is the velocity field at time t at x ∈ Λ, a domain in Rd subject to the incompressibility condition (2)

∇·u=0

and suitable boundary conditions on ∂Λ. ν is the viscosity coefficient of the fluid, p(t, x) the pressure and f (t, x) the external force that sustains the flow. Given f and u(0, ·) the task is to find u and p. It is fair to say that theoretical understanding of the consequences of these equations is still in its infancy. On the mathematical side, existence of smooth solutions for the three dimensional NS equations is wide open and has been chosen by some as one of the major problems of mathematics (http: //www.claymath.org/millennium/). On the physical side, experimental violations of the Kolmogorov scaling theory of turbulence [12] are still waiting for theoretical understanding. In two dimensions, i.e., for flows on the plane, there has been some progress during the last ten years. On the physical side, 2d turbulence has been the subject of accurate

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

138

A. KUPIAINEN

numerical and experimental studies [5], [25] and mathematically the ergodic theory of the NS flow has been under intensive study. It is important to realize that for the problem of turbulence one is interested in a very particular kind of force in (1), namely one that has a fixed length scale L built into it. Examples of this are flows past obstacles, with L the characteristic size of the obstacle. In such a setup the flow exhibits universal statistical properties as the viscosity parameter tends to zero (actually the control parameter is a dimensionless quantity, the Reynolds number given by Lv ν where v is a velocity scale related to the forcing). E.g. time averages of measurements of suitable functions of u seem to show statistical properties only depending on the Reynolds number. It is therefore of some interest to inquire about the foundations for such statistical studies, i.e., about the ergodic properties of the NS flow in the turbulent setup of a fixed scale high Reynolds number forcing. A convenient model for isotropic and homogeneous turbulence (i.e., in the limit of large Reynolds number and away from the boundary ∂Λ) is to consider (1) Equation (1) on the torus T2 = R2 /(2πZ)2 and take f random, a Fourier series with a finite number of terms and coefficients independent white noises (see below). Then the deterministic dynamics of (1) is replaced by a Markov process and one may pose questions on its ergodic properties: whether the process has a unique stationary state and whether this is reached and with what rate from arbitrary initial conditions. This Markov process is a diffusion process of a very degenerate type. While the phase space is infinite dimensional the noise is finite dimensional. There are two general mechanisms that can contribute to the ergodic and mixing properties of stochastic flows. One is dissipation, coming in our case from the Laplacian in (1). Dissipation contributes to ergodicity by exponential contraction of phase space under the flow. A second mechanism comes from the spreading of the noise from its finite dimensional subspace due to the nonlinear term in (1). In finite dimensional diffusion processes this leads to hypoellipticity if the noise spreads to the full phase space: the transition kernels are smooth (for equations with smooth coefficients). Combined with some irreducibility of the process ergodicity follows. In our infinite dimensional setup the dissipation due to the Laplacian leads to strong damping of large enough (depending on the Reynolds number) Fourier modes. If we keep noise on all the other, low, modes then one can reduce the problem to a low mode dynamics, albeit with some (exponentially decaying) memory due to the large modes. Proofs of ergodicity and mixing of the dynamics were given in this case in the works [6], [10] and [18]. However, it seemed far from trivial to extend the hypoellipticity ideas to the infinite dimensional setup to control also the case of very (1)

To get to the turbulent state one actually has to modify (1) a bit, see Section 8.

ASTÉRISQUE 339

(1016)

ERGODICITY OF TWO DIMENSIONAL TURBULENCE

139

degenerate forcing where the number of forced modes does not depend on the Reynolds number. This was accomplished by Hairer and Mattingly [13], [15] who gave sharp sufficient conditions for the noise to produce ergodic and mixing dynamics. In what follows I will present the main points of their approach focusing on the difference to finite dimensional hypoelliptic diffusions. The papers [13], [15] are very clearly written and they contain plenty of background material, especially [15] which builds a more general formalism applicable also to some reaction-diffusion equations. [15] also corrects a mistake in [13] so it should be consulted for a thorough study. In the final section I discuss more informally what we have learned about 2d turbulence and what issues might be accessible to a rigorous mathematical analysis. I would like to thank J. Bricmont, M. Hairer and J. Mattingly for comments on this exposition and the European Research Council and Academy of Finland for financial support.

1. 2D NS EQUATIONS The fundamental fact that is behind both the mathematical and physical understanding of 3d NS equations is energy conservation: in the absence of forces smooth inviscid flow preserves the L2 norm of u(t, ·). In two dimensions there is a second conserved quantity, the enstrophy, which is related to the H 1 norm and which leads to quite different physics and to much better regularity. Let us first define the vorticity ω = ∇ × u, which in d = 2 is a (pseudo)scalar: ω = ∂1 u2 − ∂2 u1 . The NS equation becomes in terms of ω a transport equation: (3)

ω˙ = ν∆ω − u · ∇ω + g,

R where g = ∂1 f2 − ∂2 f1 . We will assume the average force vanishes, i.e., f (t, x)dx = 0. R Then (1) preserves the condition u(t, x)dx = 0 which we will assume. The incompressibility condition (2) allows to write u = A ω where the linear operator A is given in terms of the Fourier transform by (4)

d A ω(k) = i(k2 , −k1 )k−2 ωˆ (k)

for k ∈ Z2 \ 0. The enstrophy E is defined to be (half of) the L2 -norm of ω: Z 1 E = 2 ω(t, x)2 dx := 21 kω(t)k2 .

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

140

A. KUPIAINEN

For a smooth u the condition ∇ · u = 0 leads to the absence of contribution from the nonlinear term to the evolution of the enstrophy: Z Z dE (5) = −ν (∇ω)2 dx + ωgdx, dt where the first term on the RHS can be interpreted as an enstrophy dissipation rate and the second one as an enstrophy injection rate. Using Poincaré inequality k∇ωk ≥ kωk and simple estimates one deduces (6)

kω(t)k2 ≤ e−νt kω(0)k2 + ν −2 sup kg(t)k2 . t 1

This a priori estimate for the H norm of u is the main ingredient in the proof of global regularity of the 2d NS flow. We wish now to discuss a version of (3) where the force g is random. We work in the subspace of real valued L2 (T2 ) functions with ω ˆ (0) = 0. It will be convenient to use the following basis for this space. Let Z + be the “upper half plane” in Z2 consisting of k = (k1 , k2 ) with k2 > 0 or k2 = 0 and k1 > 0. Hence Z2 \ 0 = Z + ∪ (−Z + ). Let ek = sin kx for k ∈ Z + and ek = cos kx for −k ∈ Z + . For each k ∈ Z2 pick independent Brownian motions βk (t) with unit speed, denoted collectively by β(t) and numbers γk ∈ R. Let X (7) Qβ(t) = γk βk (t)ek . k∈Z2

The stochastic version of Equation (3) reads dω = (ν∆ω − u · ∇ω)dt + Qdβ.

(8)

Regularity of the stochastic flow proceeds in parallel with the deterministic case as long as γk have enough decay at infinity. The analog of the enstrophy conservation Equation (5) is obtained by an application of the Ito formula (9)

1

d E = 2 dkωk2 = −νk∇ωk2 dt + (ω, Qdβ) + dt

P where  = 2π 2 k γk2 can be interpreted as the enstrophy injection rate. Taking averages we get a probabilistic analog of (5) and (6): (10)

d E E = −νEk∇ωk2 +  dt

and (11)

Ekω(t)k2 ≤ e−2νt kω(0)k2 + ν −1 .

Actually (9) can be used to control exponential moments of the enstrophy [6], [13] Lemma A.1: (12)

ASTÉRISQUE 339

E exp(ηkω(t)k2 ) ≤ 2 exp(ηe−νt kω(0)k2 )

(1016)

ERGODICITY OF TWO DIMENSIONAL TURBULENCE

141

for all η ≤ ν , i.e., probability for large L2 norm is exponentially small. (9) also allows to control the time integral of the H 1 norm: Z t (13) E exp(ην k∇ω(t)k2 ) ≤ 2 exp(ηt + ηkω(0)k2 ) 0

again for all η ≤ ν . Such a priori estimates allow one to prove the existence and pathwise uniqueness of strong solutions to Equation (8) under quite general conditions on the noise coefficients γk , see e.g. [11] and [22]. Of course the less the γk decay at infinity, the harder it is to establish the regularity of the PDE’s. As explained in the introduction, for the turbulence problem only a finite number of the γk are nonzero. Thus from the point of view of regularity the turbulent case is easy (this is not true in 3d!). However, the less noise there is, the harder it is to establish the ergodicity of the flow.

2. INVARIANT MEASURE Let us now specialize to the case where γk = 0 for k ∈ / K where K is a finite set. Thus the noise is finite dimensional: β = {βk }k∈K can be identified with the Wiener process in Ω = C([0, ∞), RD ) where D = |K|, equipped with the Wiener measure W (db). The solution of Equation (8) is a one parameter family of continuous maps Φt : Ω × L2 (T2 ) → L2 (T2 ) such that ω(t) = Φt (β, ω0 ) solves Equation (8) with initial condition ω0 and noise realization β. Actually, Φt is (Fréchet) differentiable in β and ω0 . ω(t) is a Markov process with state space H = L2 (T2 ). It gives rise to transition probabilities Pt (ω0 , A) which are probability measures on H, giving the probability of entering the set A ⊂ H at time t given that at time 0 we have ω(0) = ω0 : Pt (ω0 , A) = E1A (ω(t)). The transition probabilities generate a semigroup P t on bounded measurable functions on H by the same formula: Z (14) P t φ = Pt (·, dω)φ(ω) and the adjoint semigroup acting on bounded (Borel) measures: Z (15) P ∗t µ = µ(dω0 )Pt (ω0 , ·). We are interested in the invariant (or stationary) probability measures µ∗ satisfying the equation (16)

P ∗t µ∗ = µ∗ .

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

142

A. KUPIAINEN

Existence of an invariant measure is straightforward given the strong probabilistic Rt (ω ) control of the flow. One considers the family of time averages µt 0 = t−1 0 dsPs (ω0 , ·) and shows it is tight. Prohorov’s theorem then yields a limit point which is shown to be invariant. Uniqueness of the invariant measure is much more subtle. It implies ergodicity, i.e., in particular the equivalence of time averages and ensemble averages: (ω ) limt→∞ µt 0 (φ) = µ(φ) for all φ ∈ L2 (H, µ) and µ-a.s. in ω0 . In practice one would like to have more, i.e., the convergence in some sense of the measures Pt (ω0 , ·) to µ∗ as t → ∞. This leads to various mixing concepts.

3. DISSIPATION AND SMOOTHING For finite dimensional diffusion processes it is well known that the uniqueness of the invariant measure follows from recurrence and smoothing properties of the transition probabilities. Let us sketch a special version of this argument having the application to NS in mind. The semigroup P t is called strong Feller if the image is continuous for φ measurable. This has drastic consequences for the supports of invariant measures. Recall that x belongs to the support of a finite Borel measure µ on a Polish space (our setup) if µ(U ) > 0 for all open U containing x. Then the supports of two distinct ergodic invariant probability measures for a strong Feller semigroup are disjoint. To see this, suppose µ ⊥ ν and x ∈ supp µ ∩ supp ν. Pick A with µ(A) = 1 and ν(A) = 0. By strong Feller there exists a U containing x such that supy,z∈U |Pt (y, A)−Pt (z, A)| ≤ 21 . Moreover by assumption α := min{µ(U ), ν(U )} > 0. Write µ = (1−α)¯ µ +αµU with µ ¯ and µU probability measures with µU (U ) = 1 (i.e., µU = µ1U /µ(U )) and ν similarly. ∗ ∗ Then, by invariance |µ(A) − ν(A)| = | P t µ(A) − P t ν(A)| and thus ∗







1 = |µ(A) − ν(A)| ≤ (1 − α)| P t µ ¯(A) − P t ν¯(A)| + α| P t µU (A) − P t νU (A)| Z 1 ≤ (1 − α) + α | P t (y, A) − P t (z, A)|µU (dy)νU (dz) ≤ 1 − 2 α, U ×U

a contradiction. Suppose now that we knew that there exists an x that necessarily belongs to the support of every invariant measure of a strong Feller semigroup. We could then conclude uniqueness. This is a reasonable strategy for the NS equation. Indeed, ω = 0 is such a point. This follows since the NS equation is dissipative. Without forcing the fluid slows down, i.e., the L2 norm decays exponentially (see Equation (6)). There is a non zero probability for the force to stay small enough so that any neighborhood

ASTÉRISQUE 339

(1016)

ERGODICITY OF TWO DIMENSIONAL TURBULENCE

143

of 0 can be reached. More precisely, the Ito formula (10) combined with Poincaré inequality kωk ≤ k∇ωk yields Z µ(dω)kωk2 ≤ /ν for every invariant probability measure µ. Hence, there exists R < ∞ such that every such measure has at least half its mass in the ball BR of radius R centered at 0 in H. Thus one needs to show: for all r > 0, there exits Tr < ∞ such that Ir := inf PTr (ω0 , Br ) > 0 ω0 ∈BR ∗

(see [9], Lemma 3.1). Then µ(Br ) = P t µ(Br ) ≥ 21 Ir > 0 for all r > 0. This strategy does not quite work in our case since the strong Feller property is very hard to show for P t and might very well not be true. One of the main accomplishments of Hairer and Mattingly was to replace it with a condition that is more natural for NS and yet allows one to conclude that the supports of invariant measures are disjoint.

4. ASYMPTOTIC STRONG FELLER PROPERTY A strong Feller semigroup maps bounded functions to continuous ones. Often the easiest way to prove this is to show a bit more [8], Lemma 7.1.5: Proposition 4.1. — A semigroup on a Hilbert space H is strong Feller if for all φ : H → R with kφk∞ := supx∈H |φ(x)| and kDφk∞ finite one has (17)

kD P t φ(x)k ≤ C(kxk)kφk∞ ,

where C : R+ → R and D is the Fréchet derivative. We will now argue that the condition (17) is not very natural for the NS dynamics. As mentioned in the introduction there are (at least) two ways ergodicity can result. One is due to smoothing by the noise, the other is due to dissipation that erases memory of the initial conditions. The former effect leads to a condition like (17), the latter not. Let us next discuss the latter effect in our case. Let Js,t with s < t be the derivative of the NS flow (8) between times s and t, i.e., for every ξ ∈ H, Js,t ξ := ξ(t) is the solution of the linear equation (18)

∂t ξ(t) = ν∆ξ(t) + A ω(t) · ∇ξ(t) + A ξ(t) · ∇ω(t) := L ω(t) ξ(t)

for t > s and ξ(s) = ξ. This linear equation is readily controlled in terms of the H 1 norm of ω ([13], Lemma 4.10): Z t  (19) kξ(t)k ≤ exp C(δ, ν)(t − s) + δ k∇ω(r)k2 dr kξ(s)k s

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

144

A. KUPIAINEN

for any δ > 0. Combining with the a priori estimate (13) then (20)

EkJs,t kp ≤ 2p exp C(, ν, η, p)(t − s) + ηkω(s)k2



for all η > 0, all p < ∞. Equations (19) and (20) indicate a possible exponential separation of trajectories. However, since the Laplacian is the Fourier multiplier −k 2 it is not surprising that the high Fourier modes of ξ are strongly damped for a time that can be taken as large as we wish as N is increased. This is expressed by [13], Lemma 4.17: Lemma 4.2. — For every p ≥ 1, every T > 0, and every two constants γ, η > 0, there exists an orthogonal projector π` onto a finite number of Fourier modes such that 2

E(k(1 − π` )J0,T kp + kJ0,T (1 − π` )kp ) ≤ γeηkw0 k . For such contracting dynamics (17) is not a natural condition to try to prove. Indeed, let ξh = (1 − π` )ξ be the projection of ξ to the high modes and consider the toy problem where we apply (1 − π` ) to Equation (18) and drop altogether the ω-dependent terms: ∂t ξh (t) = ν∆ξh (t). Then for a function φ(ω) = ψ((1 − π` )ω) depending only on the high modes we have D P t φ(ω0 )ξ = EDφ(ω(t))ξh (t). Since in this toy case kξh (t)k ≤ e−At kξk for A > 0 we conclude kD P t φ(x)k ≤ e−At kDφk∞ . This toy model and Lemma 4.2 motivate the following definition by Hairer and Mattingly ([13], Proposition 3.12). Definition 4.3. — A semigroup P t on a Hilbert space H is asymptotically strong Feller if there exist two positive sequences tn and δn with {tn } nondecreasing and {δn } converging to zero such that for all φ : H → R with kφk∞ and kDφk∞ finite,  (21) |D P tn φ(x)| ≤ C(kxk) kφk∞ + δn kDφk∞ for all n, where C : R+ → R. (Hairer and Mattingly actually give a “topological” definition of the asymptotically strong Feller condition which is implied by the one above.) The main point is the following result whose proof is similar to the one given above in the strong Feller case ([13], Theorem 3.16): Proposition 4.4. — If the semigroup is asymptotically strong Feller at x then x belongs to the support of at most one ergodic invariant measure.

ASTÉRISQUE 339

(1016)

ERGODICITY OF TWO DIMENSIONAL TURBULENCE

145

We saw above that it is not unreasonable to expect that the high mode dynamics give rise to the second term in (21). Thus the question remains: why would the low mode dynamics be strong Feller? The answer to this question lies in the hypoellipticity of the low mode dynamics.

5. HYPOELLIPTICITY Let us first think about the low mode dynamics in the Galerkin approximation, i.e., by putting the high modes to zero. More formally, consider the equation dω = (ν∆ω − π` (u · ∇ω))dt + Qdβ,

(22)

where we assume the forcing is on low modes (1 − π` )Qβ = 0 and set (1 − π` )ω = 0. Equation (22) defines a diffusion process in a finite dimensional space which we may identify with RN , N = dim π` H. The diffusion process is thus degenerate with the dimension D of the noise (much) smaller than N . The strong Feller property follows for such diffusions provided the generator of the diffusion process is hypoelliptic. Let us discuss this next. Recall the Fourier basis {ek } for H. Let the range of π` be the span of {ek } with P |k| ≤ M . Write ω = k ωk ek . Then the equation (22) reads dωk = vk (ω)dt + γk dβk ,

(23) where vk is given by

vk (ω) = −ν|k|2 ωk −

 1 1 X 1  (j1 `2 − j2 `1 ) − 2 wj w` 2 2 8π |`| |j| j+`=k

and wk = 12 ω−k + (24)

1 2i ωk

for k ∈ Z + and w−k = w ¯k . The generator of this diffusion is X L = X0 + Xk2 k∈K

where we recall that γk = 0 for k ∈ / K. The vector fields Xα are given by X X0 = vk ∂ωk k

Xk

= γk ∂ωk .

An operator of the form (24) with smooth vector fields Xα is known to generate a semigroup P t with smooth kernel (hence it is strong Feller) provided the Hörmander bracket condition is satisfied (L is then hypoelliptic). The condition is that the span of the vector fields Xj , j 6= 0 and [Xi1 , [Xi2 , . . . [Xik−1 , Xik ]] . . . ] for k > 1 and ij ∈ {0} ∪ K at each ω ∈ RN equals RN .

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

146

A. KUPIAINEN

To check this condition in the NS case is a purely algebraic exercise and the result is the following [9, 13]: Proposition 5.1. — The following conditions for the set K ⊂ Z2 \{0} are sufficient for the Hörmander bracket condition to be satisfied: (a) K is invariant under the reflection k → −k (b) K contains at least two elements of unequal length (c) K spans Z2 under linear combinations with integer coefficients. An example of a very degenerate forcing that suffices is given by the set K = {(1, 0), (−1, 0), (1, 1), (−1, −1)}, i.e there is forcing only on two wave vectors and their reflections. Note that Proposition 5.1 is true for arbitrary (large enough) Galerkin cutoff N . Hence the full infinite dimensional generator formally satisfies the Hörmander condition and one might be tempted to try to use this to return to the attempt to prove the strong Feller property for P t . However, it is likely that, as we let N increase, the derivatives of the kernel of P t with respect to the high modes blow up since the smoothing is very weak for them. It is much more natural to try to use in that regime the dissipation as coded in the asymptotic strong Feller condition. Let us finally remark that if all the γk in (23) are nonzero the generator L is elliptic. If N is large enough (of the order /ν 3 ) then one may use the dissipativity of the high mode dynamics to solve for the high modes in terms of the (temporal history) of the low modes and use the ellipticity of the latter to prove ergodicity and mixing of the full dynamics [6, 10, 18].

6. MALLIAVIN MATRIX Why does elliptic diffusion produce smoothness in transition kernels? One way to think about this is to consider trajectories of the flow. Noise will make the trajectories non-unique: a change in the initial condition can be compensated by the noise. In elliptic diffusions noise spans the whole space and the compensation is immediate, in hypoelliptic diffusions the nonlinearity spreads the noise in all directions thanks to the bracket condition. Thus a derivative of the solution in the initial condition should equal its derivative in a particular direction in the (history of) noise space. Since we are integrating over the noise the latter derivative can be integrated by parts and hence an estimate like the strong Feller property can emerge. To be more explicit recall that we wrote the solution of the stochastic NS equation as ω(t) = Φt (β, ω0 ) with Φt smooth in the noise β ∈ C([0, ∞)), RD ) and the

ASTÉRISQUE 339

(1016)

ERGODICITY OF TWO DIMENSIONAL TURBULENCE

147

initial condition ω0 . Also we have denoted the derivative in the initial condition by hDω0 ω(t), ξi = J0,t ξ = ξ(t) for ξ ∈ H. Thus (25)

hD P t φ(ω0 ), ξi = Eh(Dφ)(ω(t)), ξ(t)i.

Consider next the infinitesimal change in the solution corresponding to the change of the noise β in the direction V ∈ C([0, ∞)), RD ): hDβ ω(t), V i := ζ(t). ζ(t) satisfies the same linearized NS equation (18) but with forcing QV : (26)

dζ(t) = L ω(t) ζ(t)dt + QdV (t)

and zero initial condition. The natural space to vary the noise is the Cameron-Martin Rt space, i.e., to take V of the form V (t) = 0 v(s) ds with v ∈ L2loc ([0, ∞], RD ). By variation of constants, ζ(t) is then given by Z t (27) Js,t Qv(s)ds := At v. ζ(t) = 0

Actually, the v one will eventually use is itself a function of the noise (see Equation (40)), but it will be a.s. in L2loc . The upshot is that At : L2 ([0, t], RD ) → H is an a.s. bounded random operator and so the Fréchet derivative can be written as XZ t k Ds ω(t)vk (s)ds hDβ ω(t), V i = k∈ K

0

where the operator Dks is called the Malliavin derivative and heuristically corresponds to an instantaneous kick at time s to the direction k in noise space. Explicitly (28)

Dks ω(t) = Js,t γk ek .

Suppose now we can find a v such that (29)

ξ(t) = ζ(t) , i.e., J0,t ξ = At v.

Inserting this to Equation (25) we get (30)

hD P t φ(ω0 ), ξi = Eh(Dφ)(ω(t)), ζ(t)i = EhDβ φ(ω(t)), V i.

The derivative Dβ in Equation (30) can be integrated by parts in the Gaussian Wiener measure to obtain (31)

EhDβ φ(ω(t)), V i = E(φ(ω(t))Dβ∗ V ).

In other words, the expression Dβ∗ is the adjoint of Dβ in L2 (Ω, W ). If the process v is P R adapted to the Brownian filtration its expression is simply Dβ∗ V = k vk (s)dβk (s), the Ito integral. Otherwise a derivative of v with respect to the noise also appears and Dβ∗ V is called the Skorokhod integral of v. Combining (31) with (30) the desired bound follows: (32)

|hD P t φ(ω0 ), ξi| ≤ kφk∞ E|Dβ∗ V |.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

148

A. KUPIAINEN

It remains to solve Equation (29) for V (i.e., for v). Let A∗t be the Hilbert space adjoint of At , i.e., explicitly ∗ (A∗t ξ)(s) = Q∗ Js,t ξ

(33)

for s ≤ t. Then the Malliavin matrix is defined by Z t ∗ M (t) := At A∗t = Js,t QQ∗ Js,t (34) ds. 0

Suppose M (t) is invertible. Then, clearly a solution to (29) is given by −1 v = A∗t M0,t J0,t ξ.

(35)

To sketch the rest of the story in the finite dimensional setup we need a bound for the Skorokhod integral appearing in Equation (32) [24]: Z t X Z k (36) E(Dβ∗ V )2 ≤ E |v(s)|2 ds + E Ds vl (r) Dlr vk (s))dsdr. 0

kl

The first term is the usual identity for the L2 norm of the Ito integral, the second term appears for a non-adapted v, as is the one given by (35). To compute the Malliavin derivative of v in (36) note that all we need is to compute Dr Js,t since At and M (t) are expressed in terms of Js,t . This in turn is obtained by differentiating the equation (18): η := Dkr ξ satisfies ∂t η(t) = ν∆ξ(t) + A ω(t) · ∇ξ(t) + A ξt · ∇ω(t) := L ω(t) η(t) + B(Jr,t γk ek , η(t)) where B is the bilinear form in appearing in L . By variation of constants an expression involving only J emerges. Thus in the finite dimensional setup (so e.g. for the Galerkin NS) the main work to be done is to show that M (t)−1 has good probabilistic bounds. Indeed it turns out kM (t)−1 k is in Lp (Ω) for all p < ∞. In the infinite dimensional case with degenerate noise it is unlikely that M (t) is a.s. invertible. QQ∗ is proportional to the projection in H to the subspace generated by the noise. In the expression for M (t) the dynamics spreads the range beyond this subspace, however we expect the projection of the result to the high modes to be very small. The key estimate on the Malliavin matrix Hairer and Mattingly prove is that M (t) is unlikely to be small on vectors that have large projection to low modes: Proposition 6.1. — For every α, η, p and every orthogonal projection π` on a finite number of Fourier modes, there exists C such that Å ã  (M φ, φ) (37) P inf <  ≤ Cp exp ηkω0 k2 2 kπ` φk≥αkφk kφk holds for every  ∈ (0, 1), and for every ω0 ∈ H.

ASTÉRISQUE 339

(1016)

ERGODICITY OF TWO DIMENSIONAL TURBULENCE

149

We will not discuss the details of the proof which is the technical core of the paper [15] (see also [21]). However, for experts we want to make the following comments. One major difficulty Hairer and Mattingly face is that the integrand in the expression for the Malliavin matrix is not adapted, i.e., depends on the future noise. The usual way out of this problem in the finite dimensional theory is to use the semigroup ˆ (t)J ∗ with property J0,t = Js,t J0,s to rewrite M (t) = J0,t M 0,t Z t −1 ∗ −1 ˆ (t) := J0,s QQ∗ J0,s M 0 ∗ −1 ˆ the reduced Malliavin matrix (and the control v(s) = Q∗ J0,s M (t)−1 ξ). In finite dimensions J0,t is invertible and now the integrand is adapted. The proof then uses Norris’ lemma [23] which states that if a semimartingale is small then both its bounded variation part and local martingale part are small. In the infinite dimensional case with degenerate noise, J0,t is not invertible due to dissipation of the high modes. Hence one needs to work with non-adapted processes. The way out for Hairer and Mattingly is the polynomial nature of the nonlinearity. In the iterative proof, to show that (φ, M (t)φ) is small implies that s → (Js,t P (u(s))φ, φ) is small for the various multiple commutators P ; the P will always be a polynomial. One then writes u(s) = v(s) + Qβ(s) where v is more regular and expands P (u(s)) in powers of Qβ(s), ending up with a polynomial in the Wiener process β(s) with coefficients that are nonadapted processes, but with higher regularity. The basic lemma one now needs is that such a Wiener polynomial can be small only if all the coefficients are small (up to events of small probability).

7. LOW MODE CONTROL The approach to prove smoothness sketched in the previous section is a form of stochastic control where the noise is used to force solution to a prescribed region in phase space (for results on stochastic control in our setup, see also [2], [1]). We saw that an exact compensation of the change of initial condition by a change in the noise seems impossible, but Proposition 6.1 gives reason to hope that partial compensation is possible for the low modes. Since by Lemma 4.2 the high modes are contracted the idea of Hairer and Mattingly is to do an approximate control such that instead of the full control (29) we have ξ(t)−ζ(t) → 0 as t → ∞. Thus, as before let v ∈ L2loc (R+ , RD ) be a shift in the noise and ζ(t) = A0,t v be the corresponding (infinitesimal) shift in the solution. Let (38)

ρ(t) = ξ(t) − ζ(t).

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

150

A. KUPIAINEN

Then, instead of the identities (30) and (31) we obtain hD P t φ(ω0 ), ξi = E(φ(ω(t))Dβ∗ V ) + EhDφ(ω(t)), ρ(t)i

(39)



kφk∞ E|Dβ∗ V | + kDφk∞ Ekρ(t)k.

The asymptotic strong Feller property will follow provided v can be chosen such that E|Dβ∗ V | stays bounded as t → ∞ and Ekρ(t)k tends to zero exponentially. To find v, Hairer and Mattingly use a construction where at successive time intervals two steps are alternated, one where high modes contract, the second where low modes are controlled by the noise. Suppose at some time t we knew ρ(t) is mostly in the high mode subspace, i.e., kπ` ρ(t)k  kρ(t)k. Then, at least for a short time it pays to set v = 0 since the linearized dynamics contracts such a ρ strongly. However, we cannot do this for too long since the low mode part of ρ will increase. Then provided we can find a v that will compensate the low mode part on a fixed time interval while leaving the high mode part approximately intact we can iterate the procedure. The low mode control is a simple modification of the full control explained in the previous section. Let us take the time intervals as [n, n + 1] with n an odd integer for the first step and an even integer for the second step. Thus we set v(t) = 0 for t ∈ [n, n + 1], n odd. Let An := An,n+1 , Mn := An A∗n and Jn := Jn,n+1 . For n even take vn := v|[n,n+1] = A∗n (Mn + β)−1 Jn ρ(n).

(40)

Note that except for the parameter β this agrees with the full control (35). While for β = 0 the inverse in (40) most likely does not exist, for β > 0 it does. The point now is that for small enough β (40) does a good job for the low mode control while the high modes remain approximately intact. To see this, compute (41)

ρ(n + 1)

=

ξ(n + 1) − ζ(n + 1)

=

Jn ξ(n) − (Jn ζ(n) + An A∗n (Mn + β)−1 Jn ρ(n))

=

β(Mn + β)−1 Jn ρ(n).

By Proposition 6.1, eigenvectors of Mn with small eigenvalues have small projections to the low modes. Hence one expects that for small β the operator β(Mn + β)−1 is small on vectors ψ with kπ` ψk ≥ αkψk whereas it is obviously bounded by one elsewhere. Combining the two steps we get the iteration (42)

ρ(n + 2) = Jn+1 β(Mn + β)−1 Jn ρ(n).

Combining Lemma 4.2, the bound (20) and Proposition 6.1, Hairer and Mattingly prove ([13], Lemma 4.16)

ASTÉRISQUE 339

(1016)

ERGODICITY OF TWO DIMENSIONAL TURBULENCE

151

Proposition 7.1. — For every two constants γ, η > 0 and every p ≥ 1, there exists a constant β0 > 0 such that for n even  2 E kρn+2 kp | F n ≤ γeηkωn k kρn kp holds almost surely whenever β ≤ β0 . Iterating Proposition 7.1 the exponential decay of Ekρ(t)k then follows ([13], Lemma 4.13). What remains is to bound the term E|Dβ∗ QV | in (39) uniformly in t, i.e., to bound the two integrals in (36). The crux of the matter here is that both terms can be written as a sum over n of factors proportional to ρ(n) which provides a convergence factor. For the first term this is obvious by (40). For the second one we need to go back to the integration by parts formula Equation (31). By construction Rt P V (t) = 0 v(s)ds = n Vn where Vn is F n+2 measurable. Thus since the integration P ∗ Vn and the second factor becomes by parts is local in time Dβ∗ V = n Dβ| [n,n+2] Z X (43) E tr( Ds v(r), Dr v(s))dsdr. n

[n,n+2]2

For details of how to finish the argument we refer the reader to Section 4.8 in [13].

8. TURBULENCE We have seen that the NS dynamics has a unique stationary state under very general forcing conditions. Moreover, it can be proven that the dynamics is mixing [14] and the stationary state is reached exponentially fast from arbitrary initial conditions and for arbitrary large Reynolds numbers R (for earlier proofs of mixing in the case where an R-dependent number of modes are forced, see [6], [20]). Does this mean we have reached some understanding on the properties of this state, in particular on the phenomenon of turbulence? The proof outlined in the previous sections uses properties of the system that have counterparts in the phenomenological theory of turbulence. These are the dissipation of the high Fourier modes and the transfer of the noise from the forced modes to the unforced ones due to nonlinearity. The latter point is significant because most results of the NS dynamics are based on the energy and enstrophy conservation laws alone, and those bounds would hold even if the nonlinearity was zero. Therefore, the properties of the latter are not used. This being said it must be stressed that we have gained very little understanding of the actual nature of the invariant state. Crucial part of the proof is irreducibility which is based on the fact that ω = 0 belongs to the support of every invariant measure. Recall that this holds, because there is a small probability that the random

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

152

A. KUPIAINEN

forces are close to zero for any given time interval so the fluid flow slows down due to viscosity. This is clearly not the true reason one sees fast approach to stationarity in physical experiments. The mixing times resulting from visits to the origin will be much larger than the ones observed. To understand the real mechanism for mixing one has to understand much better the transfer of energy and enstrophy from the forcing scale to other scales. It was Kraichnan’s observation [16] that we should expect this transfer to be in two dimensions quite different from the three dimensional case. In three dimensions, according to the Richardson-Kolmogorov picture the forcing in low modes injects into the system energy which is transported due to the nonlinearity in NS equation to the higher modes and eventually dissipated by the viscous term by large enough modes. This transport of energy through scales in wave number space (i.e., |k| := κ) is called the Richardson energy cascade. In fact the theory predicts a constant flux of energy from the injection scale (in our case 1) to the dissipation scale κν (these claims can be formulated in terms of various correlation functions in the putative stationary state, see e.g. the review [19]). Kraichnan noted that the existence in 2d of the second conserved quantity of the inviscid flow, the enstrophy, means that one has to pose the question at what scales (if any) energy and enstrophy are dissipated and if there exist separate fluxes for the two. His observation was that the fluxes of energy and enstrophy are to opposite directions, energy flows towards low modes and enstrophy towards high ones. Moreover, energy tends to be not dissipated at all whereas enstrophy is dissipated at high modes like energy in the 3d case. The presence of the two cascades, the direct cascade of enstrophy and the inverse cascade of energy is very well established both numerically [5] and experimentally [25]. In what follows we will point out a couple of mathematical questions regarding this picture which would be nice to understand. To state the Kraichnan picture more precisely it is convenient to work on a torus of size N , i.e., T2N := (R/(2πN Z)2 rather than N = 1 we had before. Of course by simple scaling we can get rid of the N at the expense of changing ν and the forcing scale, but since the theory involves large separations of the scales of dissipation, forcing and injection it is natural to take N large (eventually to infinity) we rather not do that. Consider now the NS dynamics on T2N with the random forcing on Fourier modes of size |k| ∼ κf  N −1 (observe that now k ∈ (N −1 Z)2 , i.e., |k| ≥ 1/N ). We shall add to the NS equation (3) an extra term that damps the low Fourier modes more strongly than the viscous term does (note that νk 2 can be as small as ν/N 2 ). This is the Ekman friction term −τ ω for τ > 0. Stationary states for this system exist for the same reasons as before and uniqueness should follow in the presence of the friction as without provided the conditions of Proposition 5.1 hold. The Kraichnan

ASTÉRISQUE 339

(1016)

ERGODICITY OF TWO DIMENSIONAL TURBULENCE

153

theory makes predictions on this stationary state, call it µν,τ,N in the various limits N → ∞, τ → 0 and ν → 0. The starting point is conservation laws of energy and enstrophy following from the enstrophy balance equation (10) and a corresponding one for energy and taking into account the extra friction term in the equation. Since the unique stationary state is translation invariant these become local identities, for enstrophy (44)

νEν,τ,N (∇ω(x))2 + τ Eν,τ,N (ω(x))2 = 

and analogously for energy (45)

νEν,τ,N (∇u(x))2 + τ Eν,τ,N (u(x))2 = 0

with 0 the energy injection rate (per unit volume) which is proportional to κ−2 f . Eν,τ,N denotes expectation in the measure µν,τ,N . The first question to pose is what happens to the viscous dissipation of energy and enstrophy as ν → 0. All the evidence points to vanishing of energy dissipation (46)

lim νEν,τ,N (∇u(x))2 = 0.

ν→0

Enstrophy dissipation is more subtle as we will see below, but again it is believed [3] that it vanishes: (47)

lim νEν,τ,N (∇ω(x))2 = 0.

ν→0

It would be interesting to prove these statements and also to understand whether a limiting measure limν→0 µν,τ,N exists and is supported on solutions of the damped randomly forced Euler equation. Indeed, some indications that this could be done come from the work [7] where time averages of solutions and statistical solutions are controlled in that limit. They are shown to be given in terms of solutions of the Euler equation and in particular [7] prove the relation (47) in that setup. The main predictions of the Kraichnan theory come from the limit N → ∞ and τ → 0. The limit N → ∞ means we are considering the NS dynamics in R2 . It is an interesting problem to try to prove that the (weak) limit limN →∞ µν,τ,N = µν,τ,∞ exists. Note that we do not expect this state to be supported on L2 but rather on polynomially bounded (and presumably smooth) functions. The reason the large volume limit might exist is the damping of the low modes by the friction term. It −1 produces an effective low wave number cutoff (which turns out to be ∼ τ 3/2 0 2 ). Granting this, what happens if we now take τ → 0? Is there also a measure µν,0,∞ ? The prediction of the Kraichnan theory is that the viscous energy dissipation (46) vanishes as ν → 0 uniformly in τ . Thus in that limit Eν,τ,∞ (u(x))2 = (0 − o(ν))/τ , i.e., the average energy density is not bounded in the putative limiting measure µν,0,∞ . However, it is believed that µν,0,∞ is supported on smooth ω and in particular

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

154

A. KUPIAINEN

limτ →0 τ Eν,τ,∞ (ω(x))2 = 0. Then (44) implies dissipative anomaly for enstrophy: enstrophy dissipation remains nonzero as ν → 0, i.e., limν→0 νEν,0,∞ (∇ω(x))2 =  > 0. The Kraichnan theory makes more quantitative predictions of the distribution of energy and enstrophy according to wave number. Define the energy spectrum for κ ∈ R+ e(κ) = 2πˆ g (κ)/κ

(48)

where gˆ(|k|) is the Fourier transform of the vorticity 2-point function g(x − y) = Eν,τ,∞ ω(x)ω(y). Then energy density is given by Eν,τ,∞ u(x)2 =

(49)



Z

e(κ)dκ 0

and enstrophy density by

R∞ 0

e(κ) ∼

(50) − 21

1 6

e(κ)κ2 dκ. Kraichnan theory predicts ( ε2/3 κ−3 , κf  κ  κν 0

2/3 −5/3

κ

,

κτ  κ  κf , −1

where κν ∼ ν  is the dissipation scale and κτ ∼ τ 3/2 0 2 the friction scale. The picture painted by the Kraichnan theory on 2d turbulence is thus quite complex. With well separated scales of viscous dissipation, injection and friction energy flows from the injection scale towards small wave numbers and is eventually dissipated by the friction. In the absence of friction and in infinite volume energy flows to ever smaller wave numbers and energy density is not defined in the stationary state. Enstrophy in turn flows to high wave numbers and is dissipated there by the viscosity. Only in the state µν,0,∞ as ν → 0 one expects to have constant fluxes of energy and enstrophy, for some exact calculations (subject to regularity assumptions), see [3]. One has to be careful with the order of limits as is seen from the behavior of enstrophy dissipation. Note in particular that the stationary state µν,0,N which we have been discussing in the previous sections does not exhibit turbulence in the sense of cascades of energy and enstrophy. Here energy will reside in low modes, indeed, in experiments one often sees the formation of a few large vortices in the flow. If ν is taken to zero in this state then both energy and enstrophy will blow up and indeed, no limit measure exists [17]. In [17] it is proven that only by taking the injection rate  (and thus also 0 ) proportional to ν a nontrivial limiting measure exists. Formally this limit still corresponds to diverging Reynolds number, but one does not expect it to be a turbulent state with near constant fluxes of energy and enstrophy. What makes the Kraichnan theory intriguing is that e.g. the spectrum (50) seems to be very well verified numerically and experimentally. Moreover, the invariant measure seems to possess strong scale invariance properties, at least in the inverse cascade

ASTÉRISQUE 339

(1016)

ERGODICITY OF TWO DIMENSIONAL TURBULENCE

155

regime. There are even indications of conformal invariance [4]. Thus it is not excluded that some of its properties could be mathematically accessible.

REFERENCES [1] A. A. Agrachev, S. B. Kuksin, A. V. Sarychev & A. Shirikyan – “On finite-dimensional projections of distributions for solutions of randomly forced 2D Navier-Stokes equations”, Ann. Inst. H. Poincaré Probab. Statist. 43 (2007), p. 399–415. [2] A. A. Agrachev & A. V. Sarychev – “Navier-Stokes equations: controllability by means of low modes forcing”, J. Math. Fluid Mech. 7 (2005), p. 108–152. [3] D. Bernard – “Influence of friction on the direct cascade of 2D forced turbulence”, Europhys. Lett. 50 (2000), p. 333–339. [4] D. Bernard, G. Boffetta, A. Celani & G. Falkovich – “Inverse turbulent cascades and conformally invariant curves”, Phys. Rev. Lett. 98 (2007), 024501. [5] G. Boffetta – “Energy and enstrophy fluxes in the double cascade of twodimensional turbulence”, Journal of Fluid Mechanics 589 (2007), p. 253–260. [6] J. Bricmont, A. Kupiainen & R. Lefevere – “Exponential mixing of the 2D stochastic Navier-Stokes dynamics”, Comm. Math. Phys. 230 (2002), p. 87–132. [7] P. Constantin & F. Ramos – “Inviscid limit for damped and driven incompressible Navier-Stokes equations in R2 ”, Comm. Math. Phys. 275 (2007), p. 529– 551. [8] G. Da Prato & J. Zabczyk – Ergodicity for infinite-dimensional systems, London Math. Soc. Lecture Note Series, vol. 229, Cambridge Univ. Press, 1996. [9] W. E & J. C. Mattingly – “Ergodicity for the Navier-Stokes equation with degenerate random forcing: finite-dimensional approximation”, Comm. Pure Appl. Math. 54 (2001), p. 1386–1402. [10] W. E, J. C. Mattingly & Y. Sinai – “Gibbsian dynamics and ergodicity for the stochastically forced Navier-Stokes equation”, Comm. Math. Phys. 224 (2001), p. 83–106. [11] F. Flandoli – “Dissipativity and invariant measures for stochastic Navier-Stokes equations”, NoDEA Nonlinear Differential Equations Appl. 1 (1994), p. 403–423. [12] U. Frisch – Turbulence, Cambridge Univ. Press, 1995. [13] M. Hairer & J. C. Mattingly – “Ergodicity of the 2D Navier-Stokes equations with degenerate stochastic forcing”, Ann. of Math. 164 (2006), p. 993–1032. [14]

, “Spectral gaps in Wasserstein distances and the 2D stochastic NavierStokes equations”, Ann. Probab. 36 (2008), p. 2050–2091.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

156

[15]

A. KUPIAINEN

, “A theory of hypoellipticity and unique ergodicity for semilinear stochastic PDEs”, preprint arXiv:0808.1361.

[16] R. H. Kraichnan – “Inertial ranges in two dimensional turbulence”, Phys. Fluids 10 (1967), p. 1417–1423. [17] S. B. Kuksin – Randomly forced nonlinear PDEs and statistical hydrodynamics in 2 space dimensions, Zürich Lectures in Adv. Math., European Math. Soc. (EMS), Zürich, 2006. [18] S. B. Kuksin & A. Shirikyan – “Ergodicity for the randomly forced 2D NavierStokes equations”, Math. Phys. Anal. Geom. 4 (2001), p. 147–195. [19] A. Kupiainen – “Statistical theories of turbulence”, in Random Media 2000 (J. Wehr, ed.), Wydawnictwa ICM, Warszawa, 2004. [20] J. C. Mattingly – “Exponential convergence for the stochastically forced Navier-Stokes equations and other partially dissipative dynamics”, Comm. Math. Phys. 230 (2002), p. 421–462. [21] J. C. Mattingly & É. Pardoux – “Malliavin calculus for the stochastic 2D Navier-Stokes equation”, Comm. Pure Appl. Math. 59 (2006), p. 1742–1790. [22] R. Mikulevicius & B. L. Rozovskii – “Global L2 -solutions of stochastic Navier-Stokes equations”, Ann. Probab. 33 (2005), p. 137–176. [23] J. Norris – “Simplified Malliavin calculus”, in Séminaire de Probabilités, XX, 1984/85, Lecture Notes in Math., vol. 1204, Springer, 1986, p. 101–130. [24] D. Nualart – The Malliavin calculus and related topics, Probability and its Applications, Springer, 1995. [25] M. K. Rivera, W. B. Daniel, S. Y. Chen & R. E. Ecke – “Energy and enstrophy transfer in decaying two-dimensional turbulence”, Phys. Rev. Lett. 90 (2003), 104502.

Antti KUPIAINEN Department of Mathematics University of Helsinki P.O. Box 68 (Gustaf Hällströmin katu 2b) FI-00014 Helsinki – Finland E-mail : [email protected]

ASTÉRISQUE 339

339

ASTÉRISQUE 2011

SÉMINAIRE BOURBAKI VOLUME 2009/2010 EXPOSÉS 1012-1026 (1017) La correspondance de Langlands locale p-adique pour GL2(Qp) Laurent BERGER

SOCIÉTÉ MATHÉMATIQUE DE FRANCE Publié avec le concours du CENTRE NATIONAL DE LA RECHERCHE SCIENTIFIQUE

Séminaire BOURBAKI 62e année, 2009-2010, no 1017, p. 157 à 180

Mars 2010

LA CORRESPONDANCE DE LANGLANDS LOCALE p-ADIQUE POUR GL2 (Qp ) par Laurent BERGER

INTRODUCTION La correspondance de Langlands locale p-adique pour GL2 (Qp ) est une bijection entre certaines représentations de dimension 2 de Gal(Qp /Qp ) et certaines représentations de GL2 (Qp ). Ces représentations sont à coefficients soit dans une extension finie de Fp (correspondance en caractéristique p), soit dans une extension finie de Qp (correspondance p-adique). Dans le cas de la correspondance en caractéristique p, on peut faire une liste des objets du côté Gal(Qp /Qp ) ainsi que du côté GL2 (Qp ) et cela permet de définir une bijection numérique dont la construction est donnée dans la section 1. Dans le cas de la correspondance p-adique, on commence par expliquer comment faire le tri dans les représentations de Gal(Qp /Qp ) (théorie de Fontaine) et dans celles de GL2 (Qp ) (représentations « admissibles » au sens de Schneider et Teitelbaum). Ensuite, on donne les premiers exemples de correspondance construits par Breuil, ce qui est l’objet de la section 2. C’est en étudiant ces exemples que Colmez a compris comment construire de manière fonctorielle cette correspondance, grâce à la théorie des (ϕ, Γ)-modules. Cette construction est donnée dans la section 3 pour les représentations « triangulines ». La section 4 contient la construction générale, ainsi que quelques propriétés de cette correspondance 1. la compatibilité à la réduction modulo p, 2. le lien avec la correspondance de Langlands locale « classique ». Dans la section 5, nous donnons quelques applications de la correspondance, dont 1. le calcul de la réduction modulo p des représentations cristallines, 2. la démonstration de (nombreux cas de) la conjecture de Fontaine-Mazur.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

158

L. BERGER

Ce texte ne dit pas grand chose sur la motivation de ces constructions. Outre la compatibilité avec la correspondance de Langlands locale classique, une propriété très importante de la correspondance p-adique est sa réalisation dans la cohomologie complétée des tours de courbes modulaires (travaux en cours de rédaction). Enfin, l’extension de ces constructions à d’autres groupes que GL2 (Qp ) est particulièrement délicate et fait l’objet de nombreux travaux en cours dont il serait prématuré de parler (voir [11]). Notations Dans tout ce texte, E désigne une extension finie de Qp dont on note OE l’anneau des entiers, mE l’idéal maximal de OE et kE son corps résiduel. Les corps E et kE sont les corps des coefficients des représentations que l’on considère. ab La théorie du corps de classes local fournit une application Q× p → Gal(Qp /Qp ) dont l’image est dense et que l’on normalise en décidant que l’image de p est le frobenius géométrique. Cette application nous permet de considérer les caractères de × × ab Q× p à valeurs dans E ou kE comme des caractères de Gal(Qp /Qp ) . On note µλ le × × caractère de Qp qui est non-ramifié (c’est-à-dire trivial sur Zp ) et qui envoie p sur λ et on note |·| le caractère x 7→ p−valp (x) où valp (p) = 1.

1. REPRÉSENTATIONS EN CARACTÉRISTIQUE p Dans cette section, nous donnons la classification des représentations de Gal(Qp /Qp ) et de GL2 (Qp ) en caractéristique p, afin de définir la correspondance dans ce cas. 1.1. Représentations de Gal(Qp /Qp ) nr Soit Qnr p l’extension maximale non-ramifiée de Qp de telle sorte que Gal(Qp /Qp ) 1/d est le sous-groupe d’inertie I Qp de Gal(Qp /Qp ). Si n > 1 et d = pn −1, alors Qnr ) p (p nr 1/d 1/d est une extension modérément ramifiée de Qp et l’application g 7→ g(p )/(p ) définit par réduction modulo p un caractère ωn : I Qp → F× pn (c’est le caractère « de niveau n » θpn −1 du §1.7 de [45]). Par exemple, ω1 est la restriction à I Qp de ω, la réduction modulo p du caractère cyclotomique.

Si h ∈ Z, alors il existe une unique représentation semi-simple notée ind(ωnh ), n−1 de déterminant ω h et de restriction à I Qp , isomorphe à ωnh ⊕ ωnph ⊕ · · · ⊕ ωnp h . × est un La représentation ind(ωnh ) est alors Fp -linéaire. Si χ : Gal(Qp /Qp ) → kE r+1 caractère, alors on note ρ(r, χ) la représentation ind(ω2 ) ⊗ χ qui est absolument irréductible si r ∈ {0, . . . , p − 1}.

ASTÉRISQUE 339

(1017)

CORRESPONDANCE DE LANGLANDS LOCALE p-ADIQUE

159

Théorème 1.1. — Toute représentation kE -linéaire absolument irréductible de dimension 2 de Gal(Qp /Qp ) est isomorphe à ρ(r, χ) pour un r ∈ {0, . . . , p − 1}. Toute représentation kE -linéaire semi-simple de dimension 2 de Gal(Qp /Qp ) est donc isomorphe (après extension éventuelle des scalaires) à ρ(r, χ) ou bien à ω r µλ ⊕ ω s µν . 1.2. Représentations de GL2 (Qp ) Nous donnons à présent la classification des représentations kE -linéaires lisses (c’est-à-dire localement constantes) et absolument irréductibles de GL2 (Qp ) qui admettent un caractère central. On identifie le centre de GL2 (Qp ) à Q× p . Si r > 0, alors r 2 Sym kE est une représentation de GL2 (Fp ) qui fournit par inflation une représentation de GL2 (Zp ), et on l’étend à GL2 (Zp )Q× p en faisant agir p par l’identité. La représentation GL (Q ) 2 indGL2 (Z p)Q× Symr kE 2

p

p

2 est l’ensemble des fonctions f : GL2 (Qp ) → Symr kE qui sont localement constantes, × à support compact modulo GL2 (Zp )Qp et telles que f (kg) = Symr (k)f (g) si k ∈ GL2 (Zp )Q× p et g ∈ GL2 (Qp ). Cet espace est muni de l’action de GL2 (Qp ) donnée par (gf )(h) = f (hg). L’algèbre de Hecke   GL (Q ) 2 EndkE [GL2 (Qp )] indGL2 (Z p)Q× Symr kE 2

p

p

× se calcule à partir de la décomposition de GL2 (Zp )Q× p \GL2 (Qp )/GL2 (Zp )Qp et on peut montrer qu’elle est isomorphe à kE [T ], où T correspond à la double classe  p0 GL2 (Zp )Q× p · 0 1 · GL2 (Zp ). × Si χ : Q× p → kE est un caractère lisse et si λ ∈ kE , alors on pose GL (Q )

π(r, λ, χ) =

2 indGL2 (Z p)Q× Symr kE 2

p

p

T −λ

⊗ (χ ◦ det).

C’est une représentation lisse de GL2 (Qp ), de caractère central ω r χ2 . Théorème 1.2. — Si r ∈ {0, . . . , p − 1} et si (r, λ) ∈ / {(0, ±1), (p − 1, ±1)}, alors la représentation π(r, λ, χ) est irréductible. Si λ = ±1, alors on a deux suites exactes : 0 → Sp ⊗ (χµλ ◦ det) → π(0, λ, χ) → χµλ ◦ det → 0, 0 → χµλ ◦ det → π(p − 1, λ, χ) → Sp ⊗ (χµλ ◦ det) → 0, où la représentation Sp (la spéciale) ainsi définie est irréductible. Ce théorème est la réunion des résultats de [2] et [1] qui traitent le cas λ 6= 0 et des résultats de [8] qui traite le cas λ = 0 (les représentations dites supersingulières).

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

160

L. BERGER

Théorème 1.3. — Les représentations kE -linéaires lisses absolument irréductibles de GL2 (Qp ) admettant un caractère central sont les suivantes : 1. χ ◦ det ; 2. Sp ⊗ (χ ◦ det) ; 3. π(r, λ, χ) où r ∈ {0, . . . , p − 1} et (r, λ) ∈ / {(0, ±1), (p − 1, ±1)}. Ce théorème est démontré dans [2], [1] et [8]. On constate alors que toutes les représentations lisses irréductibles de GL2 (Qp ) admettant un caractère central sont admissibles, c’est-à-dire que si K est un sous-groupe ouvert compact de GL2 (Qp ), alors l’espace des vecteurs fixes par K est de dimension finie. Il existe une autre manière de construire des représentations lisses de GL2 (Qp ), en utilisant l’induction parabolique. On note B2 (Qp ) l’ensemble des matrices triangulaires supérieures de GL2 (Qp ). Si δ1 et δ2 sont deux caractères lisses de Q× p , alors on  note δ1 ⊗ δ2 le caractère de B2 (Qp ) défini par a0 db 7→ δ1 (a)δ2 (d). La représentation GL (Q )

2 p B(δ1 , δ2 ) = indB2 (Q (δ1 ⊗ ω −1 δ2 ) p)

est l’ensemble des fonctions f : GL2 (Qp ) → kE localement constantes et telles que f (bg) = (δ1 ⊗ ω −1 δ2 )(b)f (g) si b ∈ B2 (Zp ) et g ∈ GL2 (Qp ). Cet espace est muni de l’action de GL2 (Qp ) donnée par (gf )(h) = f (hg). On n’obtient pas de nouvelles représentations, comme le précise le résultat suivant. × Théorème 1.4. — Si λ ∈ kE et r ∈ {0, . . . , p − 1}, alors les semi-simplifiées des r+1 représentations B(χµ1/λ , χω µλ ) et π(r, λ, χ) sont isomorphes.

Ces isomorphismes sont démontrés dans [2] et [1]. Notons que la représentation B(1, ω) se réalise (via l’identification entre P1 (Qp ) et GL2 (Qp )/B2 (Qp )) comme l’espace C 0 (P1 (Qp ), kE ) des fonctions localement constantes sur P1 (Qp ), muni de l’action naturelle de GL2 (Qp ). On trouve donc par le théorème 1.2 que la spéciale s’identifie à Sp ' C 0 (P1 (Qp ), kE )/{constantes}. 1.3. La correspondance semi-simple modulo p Toute représentation kE -linéaire semi-simple de dimension 2 de Gal(Qp /Qp ) est (après extension éventuelle des scalaires), soit absolument irréductible et donc de la forme ρ(r, χ) par le théorème 1.1, soit une somme de deux caractères et de la forme × (ω r+1 µλ ⊕ µ1/λ ) ⊗ χ avec λ ∈ kE et r ∈ {0, . . . , p − 2}. La correspondance entre représentations kE -linéaires semi-simples de dimension 2 de Gal(Qp /Qp ) et représentations kE -linéaires semi-simples lisses de GL2 (Qp ), est

ASTÉRISQUE 339

(1017)

CORRESPONDANCE DE LANGLANDS LOCALE p-ADIQUE

161

définie par Breuil dans [8] comme suit ρ(r, χ) ↔ π(r, 0, χ), (ω

r+1

µλ ⊕ µ1/λ ) ⊗ χ ↔ π(r, λ, χ)ss ⊕ π([p − 3 − r], 1/λ, ω r+1 χ)ss ,

où [p−3−r] est le représentant de p−3−r mod p − 1 dans {0, . . . , p−2} et « ss » veut dire semi-simplifiée. Du côté GL2 (Qp ), les objets peuvent être de longueur 1, 2, 3 ou 4. Théorème 1.5. — La correspondance ci-dessus est bien définie. Il s’agit de vérifier que des paramètres (r, λ, χ) différents qui donnent des représentations isomorphes d’un côté donnent des représentations isomorphes de l’autre côté. On sait par exemple que les seuls entrelacements entre les ρ(r, χ) sont ρ(r, χ) = ρ(r, χµ−1 ) = ρ(p − 1 − r, χω r ) = ρ(p − 1 − r, χω r µ−1 ), et que de même les seuls entrelacements entre les π(r, 0, χ) sont π(r, 0, χ) = π(r, 0, χµ−1 ) = π(p − 1 − r, 0, χω r ) = π(p − 1 − r, 0, χω r µ−1 ). La correspondance ci-dessus a été définie par Breuil, en se fondant sur les calculs de réduction modulo p rappelés au §2.3. En étant plus soigneux, on peut aussi définir une correspondance sans semi-simplifier.

2. REPRÉSENTATIONS EN CARACTÉRISTIQUE 0 Dans cette section, nous rappelons la théorie de Fontaine pour les représentations p-adiques de Gal(Qp /Qp ), et la théorie de Schneider et Teitelbaum pour les représentations de GL2 (Qp ). Ensuite, nous donnons les premiers exemples de la correspondance p-adique. 2.1. Théorie de Hodge p-adique Il est facile de faire la liste des représentations kE -linéaires de dimension 2 de Gal(Qp /Qp ), mais l’étude des représentations E-linéaires de ce groupe est plus compliquée car il y en a beaucoup. De manière plus précise, si l’on se donne une représentation kE -linéaire W de dimension 2 de Gal(Qp /Qp ), alors les résultats de [39] montrent que l’ensemble des représentations E-linéaires V qui admettent un OE -réseau T stable par Gal(Qp /Qp ) avec T /mE T = W est en général l’ensemble des OE -points d’un espace rigide de dimension 5. L’objet de la théorie de Hodge p-adique est de faire le tri dans ces représentations, et de décrire aussi explicitement que possible celles qui proviennent de la géométrie arithmétique. Pour cela, Fontaine a introduit dans [27] un certain nombre d’anneaux Bcris ⊂ Bst ⊂ BdR qui sont des Qp -algèbres topologiques munies d’une action de

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

162

L. BERGER

Gal(Qp /Qp ). L’anneau BdR est muni d’une filtration et l’anneau Bst est muni d’un frobenius ϕ et d’un opérateur de monodromie N tels que N ◦ ϕ = pϕ ◦ N ; en=0 fin, on a Bcris = BN . Si V est une représentation E-linéaire de Gal(Qp /Qp ) et si st ∗ ∈ {cris, st, dR}, alors on pose D∗ (V ) = (B∗ ⊗Qp V )Gal(Qp /Qp ) . On peut montrer que D∗ (V ) est un E-espace vectoriel de dimension 6 dimE (V ) et on dit que V est cristalline ou semi-stable ou de de Rham si l’on a égalité pour ∗ égal à cris ou st ou dR. Le E-espace vectoriel DdR (V ) est alors muni d’une filtration par des sousespaces E-linéaires, l’espace Dst (V ) ⊂ DdR (V ) est un (ϕ, N )-module filtré et Dcris (V ) = Dst (V )N =0 . Si D est un (ϕ, N )-module filtré, alors on définit tN (D) comme étant la valuation p-adique de ϕ sur det(D) et tH (D) comme étant l’unique entier h tel que Filh (det(D)) = det(D) et Filh+1 (det(D)) = {0}. On dit que D est admissible si tH (D) = tN (D) et si tH (D0 ) 6 tN (D0 ) pour tout sous-objet D0 de D. Théorème 2.1. — Si V est une représentation semi-stable de Gal(Qp /Qp ), alors Dst (V ) est un (ϕ, N )-module filtré admissible et le foncteur V 7→ Dst (V ) donne une équivalence de catégories de la catégorie des représentations semi-stables vers la catégorie des (ϕ, N )-modules filtrés admissibles. Le fait que Dst (V ) est admissible et que le foncteur V 7→ Dst (V ) est pleinement fidèle est démontré dans [29]. Le fait que tout module admissible provient d’une représentation semi-stable est le résultat principal de [22]. L’intérêt de ce théorème est que, pour se donner une représentation semi-stable, il suffit de se donner un (ϕ, N )-module filtré admissible, qui est un objet tout à fait explicite. 2.2. Représentations admissibles de GL2 (Qp ) Les représentations de GL2 (Qp ) qui nous intéressent sont les GL2 (Qp )-banach unitaires, c’est-à-dire les espaces de Banach B munis d’une action continue de GL2 (Qp ) et dont la topologie est définie par une norme k·k telle que kg(v)k = kvk quels que soient g ∈ GL2 (Qp ) et v ∈ B. Si B est un tel objet, alors OB = {b ∈ B tels que kbk 6 1} est stable par GL2 (Qp ) et donc B = OB /mE OB est une représentation kE -linéaire lisse de GL2 (Qp ). On dit comme dans [43] qu’un GL2 (Qp )-banach unitaire B est admissible si B est admissible au sens usuel, c’est-à-dire que si K est un sous-groupe ouvert compact de GL2 (Qp ), K

alors B est de dimension finie. Si B est de longueur finie, alors la semi-simplifiée de B ne dépend pas du choix de la norme par le principe de Brauer-Nesbitt, et on l’appelle par abus de langage la réduction modulo mE de B. Si l’on prend par exemple B = C 0 (P1 (Qp ), E), alors B est un GL2 (Qp )-banach unitaire admissible. L’espace B contient de manière évidente les constantes et la Steinberg

ASTÉRISQUE 339

(1017)

CORRESPONDANCE DE LANGLANDS LOCALE p-ADIQUE

163

est par définition St = C 0 (P1 (Qp ), E)/{constantes}. C’est une représentation irréductible admissible de GL2 (Qp ), dont la réduction modulo mE est la spéciale Sp. Si B est un GL2 (Qp )-banach unitaire et si v ∈ B, alors on a une fonction GL2 (Qp ) → B donnée par g 7→ g(v) et on dit que v est localement analytique si g 7→ g(v) l’est. On dit de même que v est localement algébrique si g 7→ g(v) est  localement un polynôme en a, b, c, d et (ad − bc)−1 où g = ac db , et que v est localement constante (c’est-à-dire lisse) si g 7→ g(v) l’est. On note Ban , Balg et Blisse les sous-espaces vectoriels correspondants de B. En général, on peut très bien avoir Balg = {0} et Blisse = {0}, mais on a le résultat suivant qui est démontré dans [44]. Théorème 2.2. — Si B est un GL2 (Qp )-banach unitaire admissible, alors Ban est un sous-espace dense dans B. Ce résultat est à comparer avec le théorème correspondant de Vignéras dans [48], concernant les représentations `-adiques de groupes p-adiques avec ` 6= p. Dans ce cas, les vecteurs lisses forment déjà un sous-espace dense. 2.3. Premiers exemples Avant d’expliquer dans la section 3 la construction générale de la correspondance de Langlands p-adique pour GL2 (Qp ), nous donnons ici les premiers exemples historiques, construits par Breuil dans [9] et [10]. Commençons par le cas cristallin. On se donne un entier k > 2 et un nombre ap ∈ mE . On associe à ces données le ϕ-module filtré Dk,ap = Ee1 ⊕ Ee2 où   !  Dk,ap si i 6 0, 0 −1 i Mat(ϕ) = et Fil Dk,ap = Ee1 si 1 6 i 6 k − 1,  pk−1 ap  {0} si i > k. Ce ϕ-module filtré est admissible, et il existe donc une représentation cristalline Vk,ap × ∗ telle que Dcris (Vk,a ) = Dk,ap . Si χ : Gal(Qp /Qp ) → OE est un caractère, alors on p note Vk,ap ,χ = Vk,ap ⊗ χ. Théorème 2.3. — Toute représentation cristalline absolument irréductible de dimension 2 de Gal(Qp /Qp ) est de la forme Vk,ap ,χ avec k > 2, ap ∈ mE et × χ : Gal(Qp /Qp ) → OE un caractère cristallin. De plus, les seuls isomorphismes entre ces représentations sont donnés par Vk,ap ,χ = Vk,−ap ,χµ−1 . Si k > 2, soit Symk−2 E 2 la représentation Symk−2 E 2 de GL2 (Zp ), étendue à 0 GL2 (Zp )Q× p en envoyant p sur l’identité. On pose alors GL (Q )

Πk,ap ,χ =

indGL2 (Z p)Q× Symk−2 E2 0 2

p

p

T − ap

⊗ (χ ◦ det),

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

164

L. BERGER

où T est un opérateur de Hecke défini comme en 1.2. Cette représentation est alors localement algébrique, irréductible si ap 6= ±(pk/2 +pk/2−1 ), et peut se réaliser comme le produit tensoriel d’une représentation algébrique par une représentation lisse GL (Q )

2 p Πk,ap ,χ = Symk−2 E 2 ⊗ indB2 (Q (µλ1 ⊗ µpλ−1 ) ⊗ (χ ◦ det), p) 2

où λ1 et λ2 sont les racines de X 2 − ap X + pk−1 = 0. Enfin, les seuls isomorphismes entre ces représentations sont donnés par Πk,ap ,χ = Πk,−ap ,χµ−1 . Théorème 2.4. — La représentation Πk,ap ,χ admet un réseau de type fini stable sous GL2 (Qp ), et son complété pour ce réseau est un GL2 (Qp )-banach unitaire admissible et topologiquement irréductible. Notons que deux réseaux de type fini sont commensurables et donnent donc des complétés isomorphes. Le complété de Πk,ap ,χ est topologiquement irréductible y compris quand Πk,ap ,χ est réductible (ce qui se produit pour ap = ±(pk/2 + pk/2−1 )). Le théorème 2.4 avait été conjecturé par Breuil, et démontré pour k 6 2p dans le §3.3 de [9]. La démonstration consistait à trouver explicitement un réseau et à calculer sa réduction modulo mE . Une démonstration générale est donnée dans [6] (le cas ap = ±2p(k−1)/2 pose un problème particulier, voir [40]). Théorème 2.5. — La semi-simplifiée V k,ap ,χ de la réduction modulo mE de Vk,ap ,χ correspond (via la correspondance du §1.3) à la semi-simplifiée Πk,ap ,χ de la réduction modulo mE de Πk,ap ,χ . Ce résulat suivait des calculs de Breuil dans tous les cas où V k,ap ,χ était connu à l’époque (soit par la théorie de Fontaine-Laffaille de [30], soit par des calculs informatiques explicites d’exemples comme dans [42]), et renforçait l’idée que les Πk,ap ,χ étaient les bons objets. Le théorème 2.5 a été démontré en toute généralité dans [5] en s’appuyant sur les constructions de [6] rappelées en 3.3. Breuil a par ailleurs donné dans [10] une construction similaire pour des représentations semi-stables non cristallines. Si k > 2 et L ∈ E, soit Dk, L le (ϕ, N )-module filtré Dk, L = Ee1 ⊕ Ee2 avec Mat(N ) = ( 01 00 ) et   ! si i 6 0,  Dk, L pk/2 0 i Mat(ϕ) = et Fil Dk, L = E(e1 + L e2 ) si 1 6 i 6 k − 1,  0 pk/2−1  {0} si i > k. Ce (ϕ, N )-module filtré est admissible et il existe donc une représentation semi-stable Vk, L telle que Dst (Vk,∗ L ) = Dk, L .

ASTÉRISQUE 339

(1017)

CORRESPONDANCE DE LANGLANDS LOCALE p-ADIQUE

165

Soit log L le logarithme p-adique normalisé par log L (p) = L et W ( L ) la E-représentation de dimension 2 de B2 (Qp ) donnée par ! ! 1 log L (a/d) a b 7→ , 0 1 0 d ce qui fait que W ( L ) est une extension non-scindée de E par E. Si δk : B2 (Qp ) → E ×  est le caractère qui à a0 db associe |ad|(k−2)/2 dk−2 , alors on pose W (k, L ) = W ( L )⊗δk et on a une suite exacte GL (Q )

GL (Q )

s

GL (Q )

2 p 2 p 2 p W (k, L ) − → indB2 (Q δk → 0. δk → indB2 (Q 0 → indB2 (Q p) p) p)

Par ailleurs, on peut montrer que la représentation Symk−2 E 2 ⊗ |det|(k−2)/2 est une GL2 (Qp ) sous-représentation de indB2 (Q δk et on définit p) Σ(k, L ) =

s−1 (Symk−2 E 2 ⊗ |det|(k−2)/2 ) . Symk−2 E 2 ⊗ |det|(k−2)/2

Les analogues des théorèmes 2.4 et 2.5 sont alors vrais : la représentation Σ(k, L ) admet un réseau stable sous GL2 (Qp ), et le complété B(k, L ) de Σ(k, L ) est un GL2 (Qp )-banach unitaire admissible topologiquement irréductible. De plus, la semisimplifiée de la réduction de Vk, L modulo mE correspond à la semi-simplifiée de celle de B(k, L ) via la correspondance du §1.3 (voir [14] pour des cas particuliers, qui ont eux aussi renforcé l’idée que les B(k, L ) étaient les bons objets ; ce résultat est démontré en toute généralité dans [5] en s’appuyant cette fois sur les constructions de [19] rappelées en 3.3).

3. LA SÉRIE PRINCIPALE UNITAIRE Dans cette section, nous rappelons la théorie des (ϕ, Γ)-modules de Fontaine, puis nous expliquons son application à la construction de modèles de la restriction à B2 (Qp ) des représentations de GL2 (Qp ) associées aux représentations p-adiques « triangulines ». 3.1. Les (ϕ, Γ)-modules La théorie des (ϕ, Γ)-modules de Fontaine, introduite dans [26], permet de décrire toutes les représentations p-adiques de Gal(Qp /Qp ) au moyen de modules sur des anneaux de séries munis de certains opérateurs. P Soit E† le corps des séries f (X) = n∈Z an X n où an ∈ E, la suite {an }n∈Z est bornée et il existe ρ(f ) < 1 tel que f (X) converge sur ρ(f ) 6 |X| < 1. Cet anneau peut être muni de deux topologies : d’une part la topologie p-adique (la norme de

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

166

L. BERGER

Gauss), et d’autre part la topologie LF consistant à mettre sur chaque E†,ρ (les f (X) qui convergent sur ρ 6 |X| < 1) la topologie de Fréchet de la convergence uniforme sur les couronnes du type ρ 6 |X| 6 σ pour σ < 1. Le complété de E† pour la topologie p-adique est le corps local E des séries P n f (X) = où an ∈ E, la suite {an }n∈Z est bornée et a−n → 0 quand n∈Z an X † n → +∞. On note O E et O E les anneaux des entiers de E† et de E pour la norme de Gauss. Le complété de E† pour la topologie LF est l’anneau de Robba R des séries P n f (X) = où an ∈ E et il existe ρ(f ) < 1 tel que f (X) converge sur n∈Z an X ρ(f ) 6 |X| < 1. Soit Γ un groupe isomorphe à Z× p , dont on note [a] l’élément correspondant à a ∈ Z× . Tous les anneaux ci-dessus sont munis d’un frobenius ϕ défini par p p ϕ(f )(X) = f ((1+X) −1) et d’une action de Γ donnée par ([a]f )(X) = f ((1+X)a −1). Définition 3.1. — Si A est l’un des anneaux E† ou E ou R , alors un (ϕ, Γ)-module sur A est un A-module libre de rang fini d, muni d’un frobenius semi-linéaire ϕ tel que Mat(ϕ) ∈ GLd (A) et d’une action semi-linéaire et continue de Γ qui commute à ϕ. On dit qu’un (ϕ, Γ)-module sur A est étale s’il en existe une base dans laquelle † Mat(ϕ) ∈ GLd ( O E ) (si A est E† ou R ) ou dans laquelle Mat(ϕ) ∈ GLd ( O E ) (si A = E). Si D† est un (ϕ, Γ)-module sur E† , alors D = E ⊗ E† D† est un (ϕ, Γ)-module sur E et Drig = R ⊗ E† D† est un (ϕ, Γ)-module sur R . De plus, si D† est étale, alors D et Drig le sont aussi. Théorème 3.2. — Les foncteurs D† 7→ D et D† 7→ Drig sont des équivalences de catégories, de la catégorie des (ϕ, Γ)-modules étales sur E† , vers la catégorie des (ϕ, Γ)-modules étales sur E et sur R . Le fait que D† 7→ D est une équivalence de catégories (la « surconvergence » des (ϕ, Γ)-modules sur E) est le résultat principal de [16]. Le fait que D† 7→ Drig est une équivalence de catégories est démontré dans [34]. Il existe un anneau Enr défini par exemple dans [26] (c’est le complété de l’extension maximale non-ramifiée de E), qui est muni d’un frobenius ϕ et d’une action de Gal(Qp /Qp ) et qui contient E, ce qui fait que si D est un (ϕ, Γ)-module sur E, alors V (D) = ( Enr ⊗ E D)ϕ=1 est un E-espace vectoriel, muni de l’action de Gal(Qp /Qp ) donnée par g(x ⊗ d) = g(x) ⊗ [χcycl (g)](d).

ASTÉRISQUE 339

(1017)

CORRESPONDANCE DE LANGLANDS LOCALE p-ADIQUE

167

Théorème 3.3. — Si D est un (ϕ, Γ)-module étale de dimension d sur E, alors V (D) est une représentation E-linéaire de dimension d de Gal(Qp /Qp ) et le foncteur qui en résulte, de la catégorie des (ϕ, Γ)-modules étales sur E vers la catégorie des représentations E-linéaires de Gal(Qp /Qp ), est une équivalence de catégories. Afin de faire le lien entre les (ϕ, Γ)-modules et les représentations de GL2 (Qp ), il faut construire un certain opérateur noté ψ. Si A est l’un des anneaux E† ou E ou R, alors A est un ϕ(A)-module libre de rang p engendré par {(1 + X)i }06i6p−1 . Si Pp−1 f ∈ A, on peut donc écrire f = i=0 ϕ(fi )(1 + X)i et on pose ψ(f ) = f0 . Si D est un (ϕ, Γ)-module sur A, alors il en existe une base de la forme {ϕ(ei )}16i6d et si y ∈ D, P P on peut donc écrire y = di=1 yi ϕ(ei ) et on pose ψ(y) = di=1 ψ(yi )ei . Proposition 3.4. — L’opérateur ψ ainsi défini ne dépend pas des choix, commute à l’action de Γ, et vérifie ψ(ϕ(f )y)) = f ψ(y) et ψ(f ϕ(y)) = ψ(f )y si f ∈ A. Le point de départ de la construction de la correspondance de Langlands p-adique pour GL2 (Qp ) en utilisant les (ϕ, Γ)-modules est le suivant. Si D est un (ϕ, Γ)-module × est un caractère continu, notons D δ Qp l’ensemble des sur E et si δ : Q× p → E (n) suites {x }n∈Z d’éléments de D telles que ψ(x(n+1) ) = x(n) pour tout n. On munit D δ Qp d’une action de B2 (Qp ) en décidant que, si x ∈ D δ Qp , alors 1. g(x)(n) = δ(a) · x(n) si g = ( a0 a0 ) avec a ∈ Q× p ; 2. g(x)(n) = [a](x(n) ) si g = ( a0 01 ) avec a ∈ Z× ; p Ä k ä 3. g(x)(n) = x(n+k) si g = p0 10 avec k ∈ Z ; n

4. g(x)(n) = (1 + X)cp · x(n) si g = ( 10 1c ) et cpn ∈ Zp . Cette définition est donnée dans [20]. L’idée est alors d’étendre cette action de B2 (Qp ) à une action de GL2 (Qp ). Cette stratégie marche particulièrement bien pour les représentations « triangulines » de dimension 2, que nous étudions dans le §3.3. 3.2. Pentes des ϕ-modules sur l’anneau de Robba Un ingrédient important de la construction de la « série principale unitaire » du §3.3 est la théorie des pentes de frobenius pour les ϕ-modules sur R , théorie due à Kedlaya et développée dans [34]. On dit qu’un ϕ-module sur R est pur de pente a/h † s’il en existe une base dans laquelle Mat(p−a ϕh ) ∈ GLd ( O E ) (par exemple, être étale est équivalent à être pur de pente nulle). Un module pur d’une certaine pente est dit isocline. Le résultat principal de la théorie des pentes est le théorème 6.10 de [34]. Théorème 3.5. — Si D est un ϕ-module sur R , alors il admet une unique filtration {0} = D0 ⊂ D1 ⊂ · · · ⊂ D` = D par des sous-ϕ-modules saturés, vérifiant 1. pour tout i > 1, le module Di /Di−1 est isocline ; 2. si si est la pente de Di /Di−1 , alors s1 < s2 < · · · < s` .

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

168

L. BERGER

Comme la filtration est unique, si D est en plus un (ϕ, Γ)-module, alors les Di sont eux aussi des (ϕ, Γ)-modules. Un point délicat mais crucial de la théorie des pentes est qu’un ϕ-module sur R qui est pur de pente s n’admet pas de sous-objet de pente < s par le théorème 3.5, mais peut très bien admettre des sous-objets saturés de pente > s. 3.3. La série principale unitaire Si V est une représentation p-adique, alors en combinant les théorèmes 3.3 et 3.2, on peut lui associer le (ϕ, Γ)-module étale Drig (V ) sur R , qui est un objet de la catégorie de tous les (ϕ, Γ)-modules sur R . Définition 3.6. — On dit qu’une représentation p-adique V est trianguline si Drig (V ) est une extension successive de (ϕ, Γ)-modules de rang 1 sur R . En utilisant les résultats de [4] qui font le lien entre la théorie de Hodge p-adique et la théorie des (ϕ, Γ)-modules, on peut montrer le résultat suivant. Théorème 3.7. — Les représentations semi-stables sont triangulines. × Si δ : Q× p → E est un caractère continu, alors w(δ) = logp δ(u)/ logp u ne dépend pas de u ∈ 1 + pZp et est appelé le poids de δ. La pente de δ est u(δ) = valp (δ(p)).

On définit R (δ) comme étant le (ϕ, Γ)-module de rang 1 engendré par eδ avec ϕ(eδ ) = δ(p)eδ et [a](eδ ) = δ(a)eδ . La pente de R (δ) au sens du §3.2 est alors bien u(δ). Théorème 3.8. — Tout (ϕ, Γ)-module de rang 1 sur R est isomorphe à R (δ) pour × un caractère δ : Q× p →E . Si V est une représentation trianguline de dimension 2, alors Drig (V ) est une extension de deux (ϕ, Γ)-modules de rang 1 et on a donc une suite exacte 0 → R (δ1 ) → Drig (V ) → R (δ2 ) → 0. Le fait que Drig (V ) est étale force les relations u(δ1 ) + u(δ2 ) = 0 et (à cause du théorème 3.5) u(δ1 ) > 0. Si u(δ1 ) = u(δ2 ) = 0, alors R (δ1 ) et R (δ2 ) sont étales et V elle-même est extension de deux représentations. × Théorème 3.9. — Si δ1 et δ2 : Q× sont deux caractères continus, alors p → E 1 Ext ( R (δ2 ), R (δ1 )) est un E-espace vectoriel de dimension 1, sauf si δ1 δ2−1 est de la forme x−i avec i > 0, ou de la forme |x|xi avec i > 1 ; dans ces deux cas, Ext1 ( R (δ2 ), R (δ1 )) est de dimension 2.

ASTÉRISQUE 339

(1017)

CORRESPONDANCE DE LANGLANDS LOCALE p-ADIQUE

169

Ce théorème est démontré dans [17]. On note alors S l’espace S = {(δ1 , δ2 , L )} où L = ∞ si δ1 δ2−1 n’est pas de la forme x−i avec i > 0, ni de la forme |x|xi avec i > 1, et L ∈ P1 (E) sinon. Une version constructive du théorème 3.9 ci-dessus permet d’associer à tout s ∈ S une extension non-triviale Drig (s) de R (δ2 ) par R (δ1 ), et réciproquement. Si s ∈ S , alors on pose w(s) = w(δ1 ) − w(δ2 ). On définit S ∗ comme l’ensemble des s ∈ S tels que u(δ1 ) + u(δ2 ) = 0 et u(δ1 ) > 0 et on pose alors u(s) = u(δ1 ) si s ∈ S ∗ . On définit les ensembles « cristallins », « semi-stables » et « non-géométriques » de paramètres. 1. 2. 3. 4.

S cris = {s ∈ S ∗ tels que w(s) > 1 et u(s) < w(s) et L = ∞} ; ∗ = {s ∈ S ∗ tels que w(s) > 1 et u(s) < w(s) et L 6= ∞} ; S st ∗ ng S ∗ = {s ∈ S ∗ tels que w(s) n’est pas un entier > 1} ; ng S irr = S cris t S st ∗ ∗ t S∗ .

Théorème 3.10. — Si s ∈ S irr , alors Drig (s) est étale. Si V (s) est la représentation trianguline associée, alors V (s) est irréductible et V (s) = V (s0 ) si et seulement si s ∈ S cris et s0 = (xw(s) δ2 , x−w(s) δ1 , ∞). ∗ Toutes les représentations triangulines absolument irréductibles s’obtiennent ainsi (quitte à étendre les scalaires) et V (s) devient cristalline (ou semi-stable) sur une extension abélienne de Qp après torsion éventuelle par un caractère si s ∈ S cris (ou ∗ st ng si s ∈ S ∗ ), tandis qu’elle n’est pas de de Rham si s ∈ S ∗ . Un ingrédient important de la démonstration du théorème 3.10 donnée dans [17] est la théorie des pentes des ϕ-modules sur R rappelée au §3.2. Nous expliquons à présent la construction des représentations Π(s) de GL2 (Qp ) associées à s ∈ S irr . On note log L le logarithme normalisé par log L (p) = L (si L = ∞, on pose log∞ = valp ) et si s ∈ S , on note δs le caractère (x|x|)−1 δ1 δ2−1 . Si s ∈ S irr alors on ne peut avoir L 6= ∞ que si δs est de la forme xi avec i > 0. On peut définir la notion de fonction de classe C u pour u ∈ R>0 (voir [18]), généralisant le cas u ∈ Z>0 . On note B(s) l’ensemble des fonctions f : Qp → E qui sont de classe C u(s) et telles que x 7→ δs (x)f (1/x) se prolonge en 0 en une fonction de classe C u(s) . L’espace B(s) est alors muni d’une action de GL2 (Qp ) donnée par la formule suivante " ! # Å ã a b ay + b · f (y) = (x|x|δ1−1 )(ad − bc) · δs (cy + d) · f . cy + d c d L’espace M (s) est défini par 1. si δs n’est pas de la forme xi avec i > 0, alors M (s) est l’espace engendré par 1 et par les y 7→ δs (y − a) avec a ∈ Qp ;

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

170

L. BERGER

2. si δs est de la forme xi avec i > 0, alors M (s) est l’intersection de B(s) et de l’espace engendré par y 7→ δs (y − a) et y 7→ δs (y − a) log L (y − a) avec a ∈ Qp . c(s) où M c(s) est l’adhérence de M (s) dans B(s). On pose enfin Π(s) = B(s)/M Théorème 3.11. — Si V (s) = Vk,ap ,χ , alors on a un morphisme GL2 (Qp )-équivariant Πk,ap ,χ → Π(s) dont l’image est dense, et si V (s) = Vk, L , alors on a un morphisme GL2 (Qp )-équivariant Σ(k, L ) → Π(s) dont l’image est dense. Ce théorème n’exclut pas a priori que Π(s) soit nul. On a cependant le résultat suivant, qui implique alors le théorème 2.4 et son analogue semi-stable. On note δ le caractère (x|x|)−1 δ1 δ2 . Théorème 3.12. — Si s ∈ S irr , alors on a un isomorphisme de représentations de B2 (Qp ) entre Π(s)∗ ⊗ δ et l’ensemble des suites bornées de D(V (s)) δ Qp . Ce théorème avait tout d’abord été montré par Colmez pour s ∈ S st ∗ puis par Breuil ng et moi-même pour s ∈ S cris et enfin par Colmez pour s ∈ S (voir [6] et [19]). Si ∗ ∗ cris 0 w(s) −w(s) 0 s ∈ S ∗ et s = (x δ2 , x δ1 , ∞), alors Π(s) = Π(s ) et on a donc deux manières de construire cet espace. L’entrelacement entre Π(s) et Π(s0 ) peut alors s’interpréter en termes de la filtration sur Dcris (V (s)). Corollaire 3.13. — Le GL2 (Qp )-banach unitaire Π(s) est non-nul, topologiquement irréductible et admissible. Le théorème 3.12 implique que Π(s) est non-nul car, si y ∈ D(V (s))ψ=1 , alors la suite constante de terme y appartient à D(V (s)) δ Qp et on peut montrer que Dψ=1 6= 0 pour tout (ϕ, Γ)-module étale D. Les deux autres propriétés se déduisent de même de la théorie des (ϕ, Γ)-modules. Remarque 3.14. — On peut montrer que, si D est un (ϕ, Γ)-module étale, alors il existe un OE [[X]]-module « petit » D] ⊂ D tel que, si x ∈ Dδ Qp est une suite bornée, alors x(n) ∈ D] pour tout n. Si D = E, alors on peut prendre D] = X −1 OE [[X]][1/p].

4. LA CORRESPONDANCE POUR GL2 (Qp ) Cette section contient la construction générale de la correspondance, ainsi que plusieurs de ses propriétés.

ASTÉRISQUE 339

(1017)

CORRESPONDANCE DE LANGLANDS LOCALE p-ADIQUE

171

4.1. Les foncteurs de Colmez Inspiré par ses constructions rappelées au §3.3, Colmez a construit dans [21] deux foncteurs D(·) et Π(·). Le foncteur D(·) associe à un GL2 (Qp )-banach unitaire admissible Π un (ϕ, Γ)-module étale D(Π) sur E. Le foncteur Π(·) associe à un (ϕ, Γ)-module étale D absolument irréductible et de dimension 2 sur E un GL2 (Qp )-banach unitaire absolument irréductible et admissible Π(D). Construction de Π 7→ D(Π) : si Π est un GL2 (Qp )-banach unitaire admissible, alors Äon définit äun certain sous-espace W de Π stable sous l’action du monoïde Zp P = Zp \{0} ce qui fait que, si DW est le dual de W , alors DW est une re0 1 présentation de P . On peut donc munir DW d’une structure de OE [[X]]-module par (1 + X)z · v = ( 10 z1 ) v, d’une action de Γ par [a](v) = ( a0 01 ) v et d’un frobenius ϕ par  ϕ(v) = p0 10 v. On pose alors D(Π) = E ⊗ OE [[X]] DW et on vérifie que D(Π) est un (ϕ, Γ)-module étale sur E. Construction de D 7→ Π(D) : si D est un (ϕ, Γ)-module étale sur E et si δ est un caractère, alors D δ Qp est l’espace dont on a rappelé la définition à la fin de 3.1. Colmez a généralisé cette construction en définissant un faisceau U 7→ D δ U , avec des applications ResU : D δ V → D δ U si U ⊂ V sont des ouverts de Qp . On a par exemple D δ Zp = D et ResZp est l’application x = {x(n) }n∈Z 7→ x(0) . On ψ=0 : D δ Zp → D δ Z× a aussi D δ Z× , l’application ResZ× p = D p étant donnée p par 1 − ϕψ. Colmez a alors défini par une formule explicite assez compliquée une × application wδ : D δ Z× p → D δ Zp qui correspond moralement à l’application × × x 7→ 1/x de Zp → Zp . Cette application lui permet de définir D δ P1 comme l’ensemble des (x1 , x2 ) avec x1 , x2 ∈ D δ Zp vérifiant ResZ× (x1 ) = wδ (ResZ× (x2 )). p p 1 On dispose alors d’une application ResQp : D δ P → D δ Qp et l’action de B2 (Qp ) sur D δ Qp s’étend assez naturellement en une action de GL2 (Qp ) sur D δ P1 (on a par exemple ( 01 10 ) (x1 , x2 ) = (x2 , x1 )). Ces constructions marchent quelle que soit la dimension de D, mais D δ P1 n’est pas le genre d’objet que l’on cherche à construire (il est trop gros). Notons (Dδ P1 )bor le sous-module des x ∈ D δ P1 tels que ResQp (x) est une suite bornée de D δ Qp . Théorème 4.1. — Si D est un (ϕ, Γ)-module étale absolument irréductible et de dimension 2 sur E et si δ est le caractère (x|x|)−1 det(D) alors 1. (D δ P1 )bor est stable sous l’action de GL2 (Qp ) ; 2. si Π(D) = Dδ P1 /(Dδ P1 )bor , alors (Dδ P1 )bor est isomorphe à Π(D)∗ ⊗δ ; 3. on a un isomorphisme D(Π(D)) = D ⊗ δ −1 . Si D n’est pas de dimension 2, alors (D δ P1 )bor n’est en général pas stable par GL2 (Qp ) et le théorème est donc spécifique à la dimension 2. De plus, la démonstration du (1) est assez détournée, puisqu’elle est fondée sur le fait que le (1) est vrai pour

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

172

L. BERGER

les représentations triangulines par le théorème 3.12, et que comme les représentations triangulines forment un sous-ensemble Zariski-dense de toutes les représentations p-adiques, le résultat s’étend par continuité (méthode suggérée par Kisin). En utilisant l’équivalence de catégories entre représentations p-adiques et (ϕ, Γ)-modules étales, on peut donc associer à toute représentation E-linéaire absolument irréductible de dimension 2 de Gal(Qp /Qp ) un GL2 (Qp )-banach unitaire absolument irréductible et admissible Π(V ). Le résultat suivant de [41] (valable si p > 5) nous dit quels GL2 (Qp )-banach on obtient de cette manière. Théorème 4.2. — Un GL2 (Qp )-banach unitaire absolument irréductible et admissible Π est de la forme Π(V ) avec V absolument irréductible de dimension 2 si et seulement si Π n’est pas un sous-quotient d’une induite d’un caractère unitaire de B2 (Qp ). Corollaire 4.3. — La correspondance de Langlands p-adique pour GL2 (Qp ) donne une bijection entre les deux ensembles suivants de E-représentations 1. les représentations absolument irréductibles de dimension 2 de Gal(Qp /Qp ) 2. les GL2 (Qp )-banach unitaires absolument irréductibles et admissibles qui ne sont pas un sous-quotient d’une induite d’un caractère unitaire de B2 (Qp ). Le résultat principal de [37] nous dit par ailleurs que la plupart des représentations E-linéaires réductibles de dimension 2 de Gal(Qp /Qp ) sont dans l’image du foncteur D(·) composé avec l’équivalence de Fontaine (l’idée est là aussi de se ramener au cas des triangulines par un argument de continuité). On peut alors étendre la correspondance aux représentations réductibles de dimension 2 de Gal(Qp /Qp ). 4.2. Propriétés de la correspondance La correspondance de Langlands p-adique pour GL2 (Qp ) jouit d’un certain nombre de propriétés, qui ont d’ailleurs guidé sa construction. Tout d’abord, elle est compatible à la correspondance en caractéristique p du §1.3, par réduction modulo mE . En effet, dans [5] il est démontré que si Π(W ) est la représentation de GL2 (Qp ) associée à une représentation kE -linéaire W de Gal(Qp /Qp ) par la correspondance du §1.3, alors la restriction à B2 (Qp ) de Π(W )∗ ⊗ δ est isomorphe à l’ensemble des suites bornées de D(W ) δ Qp après semi-simplification. En réduisant modulo mE l’isomorphisme entre Π(D)∗ ⊗ δ et (D δ P1 )bor du théorème 4.1, on obtient le résultat suivant. Théorème 4.4. — Si V est une représentation E-linéaire absolument irréductible de dimension 2 de Gal(Qp /Qp ), alors V correspond à Π(V ) par la correspondance du §1.3.

ASTÉRISQUE 339

(1017)

CORRESPONDANCE DE LANGLANDS LOCALE p-ADIQUE

173

Ensuite, la correspondance permet de retrouver la correspondance de Langlands locale « classique » démontrée par Harris et Taylor dans [32] et par Henniart dans [33] (mais construite bien avant par Tunnell dans [46] pour GL2 (Qp )). Cette correspondance est une bijection entre des représentations du groupe de Weil-Deligne de Qp et des représentations lisses admissibles de GL2 (Qp ). Si D est une représentation de Weil-Deligne de Qp , on note Lisse(D) la représentation de GL2 (Qp ) associée. Si D est de plus munie d’une filtration de poids a < b, alors on note Alg(D) la représentation algébrique Symb−a−1 E 2 ⊗ deta . Si V est une représentation de Gal(Qp /Qp ) qui est potentiellement semi-stable, alors une généralisation des constructions du §2.1 permet de lui associer un (ϕ, N, GQp )-module filtré Dpst (V ) et donc comme dans [28] une représentation de Weil-Deligne et un module filtré. On note Lisse(V ) et Alg(V ) les représentations de GL2 (Qp ) que l’on en déduit. Rappelons que, si Π est un GL2 (Qp )-banach unitaire, alors Πalg a été défini au §2.2.

Théorème 4.5. — Si V est une représentation E-linéaire absolument irréductible de dimension 2 de Gal(Qp /Qp ), alors Π(V )alg 6= {0} si et seulement si V est potentiellement semi-stable à poids distincts a < b. Dans ce cas, on a Π(V )alg = Alg(V ) ⊗ Lisse(V ).

Ce théorème est montré dans [21] (en utilisant les résultats de [24] si V n’est pas trianguline) et nous ramène aux premiers exemples de la correspondance puisque les espaces Πk,ap ,χ du §2.3 pouvaient aussi être définis (si ap 6= ±(pk/2 + pk/2−1 )) par Πk,ap ,χ = Alg(Vk,ap ,χ ) ⊗ Lisse(Vk,ap ,χ ). Le théorème 2.2 montre que Π(D)an est dense dans Π(D) et le résultat suivant, lui aussi montré dans [21], indique comment retrouver Π(D)an en termes de la définition de Π(D) donnée au (2) du théorème 4.1.

Théorème 4.6. — Si D est un (ϕ, Γ)-module étale sur E, absolument irréductible et de dimension 2, alors Π(D)an est l’image de D† δ P1 dans Π(D) = D δ P1 /(D δ P1 )bor .

Si V est une représentation E-linéaire absolument irréductible de dimension 2 de Gal(Qp /Qp ) qui devient cristalline sur une extension abélienne de Qp , alors Breuil avait conjecturé dans [6] une description explicite de Π(V )an (comme deux induites paraboliques localement analytiques, amalgamées au-dessus d’une sous-représentation localement algébrique commune). Cette conjecture est démontrée dans [38].

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

174

L. BERGER

5. APPLICATIONS Nous donnons ici quelques applications de la correspondance et des propriétés de compatibilité qu’elle vérifie. 5.1. Représentations triangulines Un sous-produit de la construction de la correspondance, et en particulier de la série principale unitaire, a été de dégager la notion de représentation p-adique trianguline de Gal(Qp /Qp ). Comme on l’a dit au §3.3, les représentations semi-stables sont triangulines, mais il en existe beaucoup d’autres. On a par exemple le résultat suivant de [35] (pour les formes modulaires p-adiques et les représentations qui leur sont associées, voir le rapport [25] d’Emerton). Théorème 5.1. — Les représentations associées aux formes modulaires paraboliques surconvergentes de pente finie sont triangulines. Tout comme la théorie de Hodge p-adique est un outil indispensable pour l’étude des formes modulaires, la théorie des représentations triangulines est au cœur de l’étude des formes modulaires et automorphes p-adiques ; c’est par exemple le thème de [3] et des nombreux travaux qui s’en inspirent. 5.2. Réduction des représentations cristallines Soient k > 2 et ap ∈ mE et Vk,ap la représentation définie au §2.1. Si on en choisit un OE -réseau stable par Gal(Qp /Qp ), que l’on réduit ce réseau modulo mE et qu’on semi-simplifie cette réduction, on obtient une représentation kE -linéaire semi-simple V k,ap qui ne dépend pas du choix du réseau par le principe de Brauer-Nesbitt. La question se pose alors de déterminer V k,ap . Si k 6 p, alors la réponse est donnée par la théorie de Fontaine-Laffaille (voir [30]) et on trouve que V k,ap = ind(ω2k−1 ). En utilisant la théorie des modules de Wach (une spécialisation de la théorie des (ϕ, Γ)-modules dans le cas cristallin), on peut calculer V k,ap pour k = p + 1 et pour k > p + 2 si valp (ap ) > b(k − 2)/(p + 1)c (voir [7] pour ces calculs, et [47] pour des améliorations ponctuelles de la borne sur valp (ap )). Hors du disque valp (ap ) > b(k − 2)/(p + 1)c, on n’a pas de formule générale et les calculs informatiques de [42], étendus ensuite par Buzzard, montrent que la situation se complique quand k augmente. En particulier, V k,ap dépend de ap d’une manière de plus en plus compliquée. Le théorème 4.4 montre que V k,ap est déterminée par Πk,ap , ce qui permet de calculer V k,ap dans les cas où on peut calculer Πk,ap . C’est ce qu’a fait Breuil pour k 6 2p dans [9] (et pour k = 2p + 1 dans un travail non publié) et qu’ont fait Buzzard

ASTÉRISQUE 339

(1017)

CORRESPONDANCE DE LANGLANDS LOCALE p-ADIQUE

175

et Gee dans [15] pour 0 < valp (ap ) < 1. Le théorème suivant rassemble les résultats que l’on connaît pour l’instant (mars 2010). Théorème 5.2. — La représentation V k,ap est connue dans les cas suivants 1. Si 2 6 k 6 p + 1, alors V k,ap = ind(ω2k−1 ). 2. Pour k = p + 2 (a) si 1 > valp (ap ) > 0, alors V k,ap = ind(ω22 ) (b) si valp (ap ) > 1, et si λ2 − ap /p · λ + 1 = 0, alors V k,ap = ωµλ ⊕ ωµλ−1 . 3. Pour 2p > k > p + 3 (a) si 1 > valp (ap ) > 0, alors V k,ap = ind(ω2k−p ) (b) si valp (ap ) = 1, et si λ = ap /p · (k − 1), alors V k,ap = ω k−2 µλ ⊕ ωµλ−1 (c) si valp (ap ) > 1, alors V k,ap = ind(ω2k−1 ). 4. Pour k = 2p + 1 (et p 6= 2) (a) si valp (a2p + p) < 3/2, alors V k,ap = ind(ω22 ) (b) si valp (a2p +p) > 3/2, alors V k,ap = ωµλ ⊕ωµλ−1 où λ2 −(a2p + p)/(2pap )· λ + 1 = 0. 5. Pour k > 2p + 2, les résultats ne sont que partiels (a) si valp (ap ) > b(k −2)/(p−1)c, alors V k,ap = ind(ω2k−1 ) (qui est réductible si p + 1 divise k − 1) (b) si 0 < valp (ap ) < 1 et t représente k − 1 mod p − 1 dans {1, . . . , p − 1}, alors (i) V k,ap = ind(ω2t ) si p − 1 ne divise pas k − 3 (ii) V k,ap ∈ {ind(ω2t ), ωµλ ⊕ ωµλ−1 } pour un certain λ si p − 1 divise k − 3. On ne dispose pas pour l’instant de formule générale, même conjecturale, pour V k,ap . Mentionnons tout de même la conjecture suivante de Buzzard (pour p 6= 2). Conjecture 5.3. — Si k est pair et si V k,ap est réductible, alors valp (ap ) est un entier. Le problème analogue dans le cas semi-stable du calcul des V k, L se pose et le théorème 4.4 s’applique là aussi. 5.3. Conjecture de Fontaine-Mazur P n Si f = est une forme modulaire parabolique propre de poids k, n>1 an q de niveau N et de caractère ε, et si E = Qp ({an }n>1 ), alors grâce à [23] on sait qu’il existe une représentation E-linéaire Vf de dimension 2 de Gal(Q/Q) telle que Vf est non-ramifiée en tout ` - pN et telle que, pour ces `, on a det(X − Fr` ) = X 2 − a` X + ε(`)`k−1 .

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

176

L. BERGER

On sait par ailleurs que la restriction de Vf à Gal(Qp /Qp ) est potentiellement semi-stable et Fontaine et Mazur ont formulé dans [31] la conjecture suivante. Conjecture 5.4. — Si V est une représentation E-linéaire irréductible de dimension 2 de Gal(Q/Q) qui est non-ramifiée en presque tout ` 6= p, et dont la restriction à Gal(Qp /Qp ) est potentiellement semi-stable à poids de Hodge-Tate distincts, alors il existe une forme modulaire parabolique propre telle que V est la tordue de Vf par un caractère. Notons V la réduction modulo mE de V . Théorème 5.5. — La conjecture de Fontaine-Mazur est vraie, si l’on suppose que V satisfait certaines hypothèses techniques. Ce théorème a été démontré indépendamment par Kisin (voir [36]) et par Emerton (voir [24]). Les « hypothèses techniques » de Kisin sont les suivantes. 1. p 6= 2 et V est impaire, 2. V |Gal(Q/Q(ζp )) est absolument irréductible, ωχ ∗  3. V |Gal(Q /Qp ) n’est pas de la forme 0 χ . p

Les « hypothèses techniques » d’Emerton sont (1) et (2) et χ ∗  30 . V |Gal(Q /Qp ) n’est pas de la forme 0 ωχ ni de la forme p

χ ∗ 0 χ .

La méthode d’Emerton donne alors un résultat supplémentaire : si on suppose que V |Gal(Q /Qp ) est trianguline (au lieu de potentiellement semi-stable), alors V provient p d’une forme modulaire parabolique surconvergente de pente finie. L’outil le plus puissant dont on dispose pour l’instant afin de démontrer la modularité de certaines représentations galoisiennes est l’étude de leurs espaces de déformations. C’est cette méthode qui a permis à Wiles de démontrer dans [49] la modularité des courbes elliptiques semi-stables. Dans leur article [13], Breuil et Mézard ont proposé une conjecture reliant certains anneaux paramétrant les déformations potentiellement semi-stables d’une représentation V , et certaines représentations de GL2 (Zp ). Plus précisément, pour un certain type de paramètres (k, τ, V ) de déformations, ils définissent une multiplicité galoisienne µGal (k, τ, V ) et une multiplicité automorphe µAut (k, τ, V ). La multiplicité galoisienne mesure la « taille » de l’anneau des déformations de type (k, τ, V ), tandis que la multiplicité automorphe dépend de V et de la réduction modulo p de certaines représentations de GL2 (Zp ) associées à k et τ par la correspondance locale de Langlands « classique ». La conjecture de Breuil-Mézard est alors que µGal (k, τ, V ) = µAut (k, τ, V ). Le foncteur de Colmez V 7→ Π(V ) étant défini de manière assez naturelle, il s’étend aux familles et définit par suite un foncteur de l’espace des déformations de V vers

ASTÉRISQUE 339

(1017)

CORRESPONDANCE DE LANGLANDS LOCALE p-ADIQUE

177

l’espace des déformations de Π(V ) (c’est d’ailleurs cet argument qui est utilisé par Kisin dans [37]). Cette construction, ainsi que le théorème 4.5 concernant les vecteurs localement algébriques, permettent à Kisin de faire le lien entre multiplicité galoisienne et multiplicité automorphe, et par suite de démontrer la conjecture de Breuil-Mézard. Ceci lui donne des renseignements précis sur les anneaux de déformations de V qui lui permettent alors d’appliquer les techniques de modularité et de démontrer la conjecture de Fontaine-Mazur. Remarquons pour terminer que la conjecture de Breuil-Mézard était inspirée des calculs de [12] dont le principal résultat était la modularité de toutes les courbes elliptiques définies sur Q et c’est la rédaction de [13] qui a contribué à donner à Breuil l’idée de la construction de la correspondance de Langlands p-adique. La boucle est donc bouclée, puisque cette correspondance de Langlands p-adique pour GL2 (Qp ) permet à présent de démontrer une vaste généralisation du résultat de [12].

RÉFÉRENCES [1] L. Barthel & R. Livné – Irreducible modular representations of GL2 of a local field, Duke Math. J. 75 (1994), p. 261–292. [2] , Modular representations of GL2 of a local field : the ordinary, unramified case, J. Number Theory 55 (1995), p. 1–27. [3] J. Bellaïche & G. Chenevier – Families of Galois representations and Selmer groups, Astérisque 324 (2009). [4] L. Berger – Représentations p-adiques et équations différentielles, Invent. Math. 148 (2002), p. 219–284. [5] , Représentations modulaires de GL2 (Qp ) et représentations galoisiennes de dimension 2, Astérisque 330 (2010), p. 263–279. [6] L. Berger & C. Breuil – Sur quelques représentations potentiellement cristallines de GL2 (Qp ), Astérisque 330 (2010), p. 155–211. [7] L. Berger, H. Li & H. J. Zhu – Construction of some families of 2-dimensional crystalline representations, Math. Ann. 329 (2004), p. 365–377. [8] C. Breuil – Sur quelques représentations modulaires et p-adiques de GL2 (Qp ). I, Compositio Math. 138 (2003), p. 165–188. , Sur quelques représentations modulaires et p-adiques de GL2 (Qp ). II, [9] J. Inst. Math. Jussieu 2 (2003), p. 23–58. [10] , Invariant L et série spéciale p-adique, Ann. Sci. École Norm. Sup. 37 (2004), p. 559–610.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

178

[11] [12]

[13]

[14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29]

L. BERGER

, The emerging p-adic Langlands programme, in ICM 2010, vol. II, 2010, p. 203–230. C. Breuil, B. Conrad, F. Diamond & R. Taylor – On the modularity of elliptic curves over Q : wild 3-adic exercises, J. Amer. Math. Soc. 14 (2001), p. 843–939. C. Breuil & A. Mézard – Multiplicités modulaires et représentations de GL2 (Zp ) et de Gal(Qp /Qp ) en l = p, Duke Math. J. 115 (2002), p. 205–310, with an appendix by Guy Henniart. , Représentations semi-stables de GL2 (Qp ), demi-plan p-adique et réduction modulo p, Astérisque 331 (2010), p. 117–178. K. Buzzard & T. Gee – Explicit reduction modulo p of certain two-dimensional crystalline representations, Int. Math. Res. Not. 2009 (2009), p. 2303–2317. F. Cherbonnier & P. Colmez – Représentations p-adiques surconvergentes, Invent. Math. 133 (1998), p. 581–611. P. Colmez – Représentations triangulines de dimension 2, Astérisque 319 (2008), p. 213–258. , Fonctions d’une variable p-adique, Astérisque 330 (2010), p. 13–59. , La série principale unitaire de GL2 (Qp ), Astérisque 330 (2010), p. 213– 262. , (ϕ, Γ)-modules et représentations du mirabolique de GL2 (Qp ), Astérisque 330 (2010), p. 61–153. , Représentations de GL2 (Qp ) et (ϕ, Γ)-modules, Astérisque 330 (2010), p. 281–509. P. Colmez & J.-M. Fontaine – Construction des représentations p-adiques semi-stables, Invent. Math. 140 (2000), p. 1–43. P. Deligne – Formes modulaires et représentations `-adiques, Séminaire Bourbaki, vol. 1968/1969, exp. no 355, Lecture Notes in Math. 179 (1969), p. 139–172. M. Emerton – Local-global compatibility in the p-adic Langlands programme for GL2/Q , prépublication, 2010. , p-adic families of modular forms, Séminaire Bourbaki, vol. 2009/2010, exp. no 1013, Astérisque 33 ? (2011), p. 31–62. J.-M. Fontaine – Représentations p-adiques des corps locaux. I, in The Grothendieck Festschrift, Vol. II, Progr. Math., vol. 87, Birkhäuser, 1990, p. 249–309. , Le corps des périodes p-adiques, Astérisque 223 (1994), p. 59–111. , Représentations l-adiques potentiellement semi-stables, Astérisque 223 (1994), p. 321–347. , Représentations p-adiques semi-stables, Astérisque 223 (1994), p. 113– 184.

ASTÉRISQUE 339

(1017)

CORRESPONDANCE DE LANGLANDS LOCALE p-ADIQUE

179

[30] J.-M. Fontaine & G. Laffaille – Construction de représentations p-adiques, Ann. Sci. École Norm. Sup. 15 (1982), p. 547–608. [31] J.-M. Fontaine & B. Mazur – Geometric Galois representations, in Elliptic curves, modular forms, & Fermat’s last theorem (Hong Kong, 1993), Ser. Number Theory, I, Int. Press, Cambridge, MA, 1995, p. 41–78. [32] M. Harris & R. Taylor – The geometry and cohomology of some simple Shimura varieties, Annals of Math. Studies, vol. 151, Princeton Univ. Press, 2001, with an appendix by Vladimir G. Berkovich. [33] G. Henniart – Une preuve simple des conjectures de Langlands pour GL(n) sur un corps p-adique, Invent. Math. 139 (2000), p. 439–455. [34] K. S. Kedlaya – A p-adic local monodromy theorem, Ann. of Math. 160 (2004), p. 93–184. [35] M. Kisin – Overconvergent modular forms and the Fontaine-Mazur conjecture, Invent. Math. 153 (2003), p. 373–454. [36]

, The Fontaine-Mazur conjecture for GL2 , J. Amer. Math. Soc. 22 (2009), p. 641–690.

[37]

, Deformations of GQp and GL2 (Qp ) representations, Astérisque 330 (2010), p. 511–528.

[38] R. Liu – Locally analytic vectors of some crystabelian representations of GL2 (Qp ), I, prépublication, 2009. [39] B. Mazur – Deforming Galois representations, in Galois groups over Q (Berkeley, CA, 1987), Math. Sci. Res. Inst. Publ., vol. 16, Springer, 1989, p. 385–437. ¯ nas – On some crystalline representations of GL2 (Qp ), Algebra Number [40] V. Pašku Theory 3 (2009), p. 411–421. [41]

, The image of Colmez’s Montréal functor, prépublication, 2010.

[42] D. Savitt & W. Stein – Some data for C. Breuil about the local representation attached to a new form of medium weight, rapport, 2002. [43] P. Schneider & J. Teitelbaum – Banach space representations and Iwasawa theory, Israel J. Math. 127 (2002), p. 359–380. [44]

, Algebras of p-adic distributions and admissible representations, Invent. Math. 153 (2003), p. 145–196.

[45] J-P. Serre – Propriétés galoisiennes des points d’ordre fini des courbes elliptiques, Invent. Math. 15 (1972), p. 259–331. [46] J. B. Tunnell – On the local Langlands conjecture for GL(2), Invent. Math. 46 (1978), p. 179–200. [47] M. Vienney – Réduction des représentations cristallines de dimension 2 en poids inférieur à p2 , manuscrit, 2010.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

180

L. BERGER

[48] M.-F. Vigneras – Banach l-adic representations of p-adic groups, Astérisque 330 (2010), p. 1–11. [49] A. Wiles – Modular elliptic curves and Fermat’s last theorem, Ann. of Math. 141 (1995), p. 443–551.

Laurent BERGER Université de Lyon UMPA, ENS de Lyon UMR 5669 du CNRS 46, allée d’Italie F–69007 LYON E-mail : [email protected]

ASTÉRISQUE 339

339

ASTÉRISQUE 2011

SÉMINAIRE BOURBAKI VOLUME 2009/2010 EXPOSÉS 1012-1026 (1018) Métriques kählériennes extrémales sur les surfaces toriques Olivier BIQUARD

SOCIÉTÉ MATHÉMATIQUE DE FRANCE Publié avec le concours du CENTRE NATIONAL DE LA RECHERCHE SCIENTIFIQUE

Séminaire BOURBAKI 62e année, 2009-2010, no 1018, p. 181 à 201

Mars 2010

MÉTRIQUES KÄHLÉRIENNES EXTRÉMALES SUR LES SURFACES TORIQUES [d’après S. Donaldson] par Olivier BIQUARD

INTRODUCTION Dans les années 1980, E. Calabi, étudiant les variétés complexes, a initié le programme consistant à trouver dans une classe de Kähler une métrique privilégiée, appelée extrémale [5, 6]. La recherche dans ce domaine est devenue particulièrement active depuis le programme de S. Donaldson dans les années 2000 [10, 11]. Donnonsen la conjecture centrale dans le cas des métriques à courbure scalaire constante (un cas particulier de métrique extrémale) : Conjecture 0.1 (Yau-Tian-Donaldson). — Soit X une variété compacte complexe, munie d’une classe de Kähler entière L. Alors l’existence d’une métrique kählérienne à courbure scalaire constante dans L est équivalente à une propriété algébrique de (X, L), appelée K-polystabilité. La condition algébrique de K-(poly)stabilité sera précisée plus loin : introduite par Tian [27] dans l’étude des métriques de Kähler-Einstein, elle a été généralisée par Donaldson [13]. La conjecture est un analogue sur les variétés de la correspondance fameuse de Hitchin-Kobayashi entre fibrés holomorphes polystables et métriques de HermiteEinstein. La partie facile de la correspondance de Hitchin-Kobayashi sur les fibrés est le fait que l’existence d’une métrique de Hermite-Einstein implique la stabilité. Sur les variétés, même cette direction, à savoir la nécessité de la condition de K-polystabilité, est difficile : après les travaux initiaux de Donaldson [12] (voir l’exposé [4]), cette partie est maintenant largement démontrée [25, 23, 24]. En revanche, construire une métrique à courbure scalaire constante sur une variété K-polystable est considérablement plus délicat. Le théorème suivant de Donaldson est donc important, en ce qu’il

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

182

O. BIQUARD

exhibe la première classe assez large de variétés kählériennes sur laquelle est vérifiée la conjecture : Théorème 0.2 (Donaldson). — La conjecture est vraie sur les surfaces complexes toriques. Ce théorème est le point culminant d’un travail de plusieurs années sur les variétés toriques [13, 14, 15, 17], dont cet exposé tente de rendre compte. Sur la géométrie kählérienne torique en général, et son extension aux variétés sans multiplicité, on pourra consulter utilement le survey [16]. L’article [26] étudie ce qu’on peut dire quand la variété torique n’est pas K-polystable. Remerciements. Je remercie en particulier Paul Gauduchon et Vincent Minerbe pour leur aide précieuse dans la lecture des travaux évoqués ici.

1. GÉOMÉTRIE KÄHLÉRIENNE TORIQUE Une variété torique est une variété kählérienne (X n , ω), munie de l’action hamiltonienne effective d’un tore compact T n = t/Λ. 1.1. Structure symplectique On dispose alors d’une application moment, T -invariante, µ : X → t∗ . En clair, si (K 1 , . . . , K n ) est une base de champs de vecteurs induits par l’action de T n , et (x1 , . . . , xn ) sont les coordonnées sur t∗ données par xi = hK i , ·i, alors on peut voir les xi comme des fonctions sur X via µ, et elles satisfont (1)

dxi = −ıK i ω.

∂ On peut choisir des coordonnées angulaires θi , de sorte que K i = ∂θ , et la forme de i (1) Kähler s’exprime alors, sur l’ouvert où l’action de T est libre, comme

(2)

ω = dxi ∧ dθi .

Les (xi , θi ) sont les coordonnées action-angle (le choix des θi n’est pas unique). L’image P = µ(X) est un polytope convexe de t∗ , enveloppe convexe de l’image des points fixes de l’action de T ; il satisfait les axiomes des polytopes de Delzant : 1. P est simple : chaque sommet est l’intersection d’exactement n faces (de codimension 1) ; (1)

Dans tout l’exposé, on utilisera la convention usuelle de sommation implicite : si un symbole apparaît à la fois comme indice et comme exposant dans P unei formule, alors il est automatiquement sommé ; par exemple, la formule (2) signifie en réalité dx ∧ dθi . i

ASTÉRISQUE 339

(1018)

MÉTRIQUES EXTRÉMALE TORIQUES

183

2. P est rationnel : chaque face a une normale rationnelle ν ∈ QΛ ⊂ t ; on choisira toujours la normale entière (ν ∈ Λ), primitive, et pointant vers l’intérieur de P ; 3. à chaque sommet, l’ensemble des normales des n faces s’intersectant en ce sommet est une base entière de Λ. Réciproquement, tout polytope de Delzant est l’image par l’application moment d’une variété torique [9]. La construction se fait de la manière suivante : soient ν1 , . . . , νH les normales (entières, primitives, entrantes) aux H faces du polytope P , de sorte que P = ∩i {hνi , xi + ci > 0}. On considère l’action du tore T H sur CH , dont l’application moment est ( 21 |z 1 |2 −c1 , . . . , 12 |z H |2 −cH ). L’application RH → t, donnée P H → T dont on appellera le noyau N . par (yk ) 7→ H 1 yk νk , induit un morphisme T Alors on récupère la variété torique X par le quotient kählérien X = CH //N , muni de l’action résiduelle de T = T H /N . L’application moment µN : CH → n∗ est la restriction à n de µ. L’application moment de l’action de T sur X apparaît alors comme la restriction de µ à µ−1 N (0), et on peut vérifier explicitement que son image est le polytope P . 1.2. Structure complexe : les coordonnées complexes L’action du tore T n s’étend en une action holomorphe du complexifié TCn = (C∗ )n . Choisissant un point base dans X, un ouvert dense de X est ainsi identifié à (C∗ )n , et donc muni de coordonnées complexes za = exp(ya + iθa ), où ya ∈ R et θa est une variable angulaire. Dans ces coordonnées, la forme de Kähler est donnée par un ¯ potentiel T -invariant ϕ, de sorte que ω = ddC ϕ (= 2i∂ ∂ϕ). On calcule alors la forme de Kähler ω et la métrique g : (3)

ω=

∂2ϕ dya ∧ dθb , dya dyb

g=

∂2ϕ (dya dyb + dθa dθb ). dya dyb

1.3. Structure complexe : les coordonnées symplectiques L’invariance sous T des xi et de la structure complexe J implique immédiatement (dJdxi )Kj ,Kk = 0. En outre, puisque J est une structure complexe intégrable, la 2-forme dJdxi = dC dxi est de type (1, 1). Comme les (1, 1)-formes sont engendrées par les formes (4)

dxj ∧ dxk + (Jdxj ) ∧ (Jdxk ) et (Jdxj ) ∧ dxk − dxj ∧ (Jdxk ),

ces deux faits ensemble impliquent que dJdxi se décompose uniquement sur les formes de type (Jdxj ) ∧ dxk − dxj ∧ (Jdxk ). Ainsi l’idéal engendré par les 1-formes Jdxi est préservé par d, donc la distribution D = ∩ ker(Jdxi ) est intégrable. On peut alors choisir les coordonnées θi de sorte que D = ∩ ker dθi . On écrit alors (5)

Jdxi = Gij dθj ,

Jdθi = −Gij dxj ,

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

184

O. BIQUARD

où (Gij ) = (Gij )−1 , et on déduit de (2) la forme de la métrique : g = Gij dxi dxj + Gij dθi dθj .

(6)

En particulier, la matrice (Gij ) est symétrique, définie positive. En outre, notant f,k ∂f la dérivée ∂x k (souvent on notera plus simplement fk quand le contexte est clair), la 2-forme dJdθi = −Gij,k dxk ∧ dxj est à nouveau de type (1, 1). Compte tenu de (4), il faut que dJdθi = 0, donc G satisfait la symétrie des dérivées : Gij,k = Gik,j . Il existe donc une fonction réelle u, strictement convexe, telle que (Gij ) soit le hessien de u : Gij = uij . En notant (uij ) la matrice inverse de (uij ), on obtient la forme définitive de la métrique : g = uij dxi dxj + uij dθi dθj .

(7)

Ainsi, dans les coordonnées action-angle, la structure complexe est-elle entièrement codée par la seule donnée de la fonction u, appelée potentiel symplectique. Il est important de noter que u est déterminé uniquement à addition près d’une fonction affine. 1.4. La transformation de Legendre On fait maintenant le lien entre les deux points de vue, complexe et symplectique. On a une base de (1, 0)-formes, a = dθa − iuab dxb = −id(ua + iθa ).

(8)

Les (1, 0)-formes a , fermées, sont holomorphes, donc les (ua +iθa ) peuvent s’identifier aux coordonnées complexes ya + iθa vues en 1.2. Posons ainsi ya = ua .

(9)

Identifions la forme de Kähler ω dans les deux formalismes. Par (3), ω=

∂2ϕ ∂2ϕ dya ∧ dθb = uac dxc ∧ dθb , ∂ya ∂yb ∂ya ∂yb

à identifier avec ω = dxb ∧ dθb . Il faut donc que (10)

(

∂2ϕ ) = (uab )−1 = (uab ). ∂ya ∂yb

Autrement dit, il faut que le hessien de ϕ dans les coordonnées y soit l’inverse du hessien de u dans les coordonnées x, ce qui est typique de la transformée de Legendre : la solution est que ϕ et u sont images l’une de l’autre par la transformée de Legendre : (11)

ASTÉRISQUE 339

ϕ = ub xb − u.

(1018)

185

MÉTRIQUES EXTRÉMALE TORIQUES

De manière plus symétrique, on pense à y comme variable dans t et x dans t∗ , alors u et ϕ sont liés par la relation u(x) + ϕ(y) = hx, yi, où y(x) = du(x).

(12)

Exemple 1.1. — Dans Cn , on a ω = 2i (dz 1 ∧d¯ z 1 +· · ·+dz n ∧d¯ zn) = i 2

Pn 1

i 2

d( |z2| )∧dθi .

Il en résulte xi = |z2| . Le polytope P est donné par les équations xi > 0, et on calcule facilement la formule pour le potentiel symplectique : n

(13)

u(x) =

1X i x log xi . 2 1

1.5. La formule de Guillemin La construction de Delzant, déjà évoquée au début de cette section, produit une variété torique XP à partir d’un polytope de Delzant P . Cette construction se faisant par quotient kählérien d’un espace plat CH de grande dimension par un sous-groupe de U (1)H , la variété XP est naturellement pourvue d’une métrique kählérienne lisse. Pour l’exprimer, notons λi (x) = hνi , xi + ci l’équation de la i-ième face de P , où νi est la normale entière primitive entrante. Donc P = ∩N 1 {λi > 0}. Lemme 1.2 (formule de Guillemin [19, 20]). — Le potentiel symplectique uG de XP P pour la métrique de Delzant est donné par uG = 21 N 1 λi log λi . L’analogie de cette formule avec (13) n’est pas un hasard : la proposition se démontre en constatant que l’image de l’application moment du quotient XP est naturellement plongée dans l’image de l’application moment de CH , et en vérifiant que la restriction du potentiel symplectique de CH à cette image est un potentiel symplectique de XP . Exemple 1.3. — L’espace projectif complexe CP n = Cn+1 //U (1) admet pour polytope P = ∩n1 {xi > 0} ∩ {x1 + · · · + xn 6 1}. La formule de Guillemin fournit donc le potentiel symplectique de la métrique de Fubini-Study : u(x) =

1 1 x log x1 + · · · + xn log xn + (1 − x1 − · · · − xn ) log(1 − x1 − · · · − xn ) . 2

Dans cet exemple, la métrique quotient est Kähler-Einstein. En général il est très rare que la métrique de Delzant soit Kähler-Einstein ou même à courbure scalaire constante.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

186

O. BIQUARD

1.6. Les conditions au bord Le potentiel symplectique est bien défini sur le polytope de Delzant, mais ses dérivées explosent au bord du polytope, comme on le voit par exemple sur Cn . Il est important de connaître le comportement asymptotique des fonctions strictement convexes sur P qui sont des potentiels symplectiques de métriques kählériennes lisses sur X entier. La réponse est fournie par : Lemme 1.4 (Conditions au bord de Guillemin). — La métrique de Kähler associée à un potentiel symplectique u s’étend de manière lisse sur X si et seulement si – u s’exprime par rapport au potentiel de Guillemin uG comme u = uG + v avec v ∈ C ∞ (P¯ ) ; – la restriction de u à chaque face (de toute codimension) est encore strictement convexe. La démonstration se fait en passant en coordonnées complexes exp(ya + iθa ) et en écrivant précisément le comportement de la métrique près des faces du polytope (c’est-à-dire près des diviseurs correspondants). 1.7. La formule d’Abreu Un des intérêts majeurs du passage au potentiel symplectique est la simplicité des formules donnant la courbure d’une métrique kählérienne torique. En effet, on calcule facilement que, pour une fonction T -invariante f , (14)

ddC f = (fi uij )k dxk ∧ dθl ,

ΛddC f = (fi uij )j .

De la formule (3) pour g il résulte que la forme de Ricci ρ est donnée par Å 2 ã 1 ∂ ϕ ρ = − ddC log det . 2 ∂ya ∂yb D’après (10), les hessiens de ϕ par rapport à y, ou de u par rapport à x, sont inverses l’un de l’autre, et on obtient (15)

ρ=

1 C dd log det(uij ). 2

À l’aide de (14), on obtient alors les formules suivantes pour ρ et la courbure scalaire Scal = 2Λρ, formule due à Abreu [1] : (16)

ASTÉRISQUE 339

1 ρ = − uij dxk ∧ dθj , 2 ik

Scal = −uij ij .

(1018)

MÉTRIQUES EXTRÉMALE TORIQUES

187

1.8. La courbure scalaire comme application moment On peut considérer l’espace J des structures presque-complexes T -invariantes sur X, et compatibles à ω. Le groupe G = HamT (X, ω)/T du groupe des difféomorphismes hamiltoniens T -invariants de X par T agit sur J . On va voir que J est muni naturellement d’une structure kählérienne de dimension infinie, pour laquelle l’action de G est hamiltonienne, d’application moment égale à Scal. L’espace tangent à J en J s’identifie aux endomorphismes J˙ du fibré tangent, ˙ = 0. La strucT -invariants, symétriques pour la métrique gJ , et satisfaisant J J˙ + JJ ˙ = J ◦ J, ˙ et la structure symplectique par ture complexe J de J se décrit par J(J) R 1 1 n ˙ ˙ ˙ ˙ Ω(J1 , J2 ) = 2 P Tr(J J1 J2 )dv, où dv = dx · · · dx . À une fonction f (x1 , . . . , xn ) est associé le champ de vecteurs hamiltonien Xf par ∂ ιxf ω = −df , ce qui donne Xf = fi ∂θ . Les fonctions affines correspondent donc i aux champs de vecteurs infinitésimaux de l’action de T , ainsi l’algèbre de Lie de G s’identifie-t-elle à C ∞ (P¯ )/{fonctions affines}. L’action infinitésimale de f sur J est donnée au point J par LXf J = [LXf , J] qui se calcule en (17)

LXf J = dC fi ⊗

∂ ∂ + dfi ⊗ J . ∂θi ∂θi

À présent, il est immédiat que ˙ =− Ω(LXf J, J)

Z P

∂ dfi (J˙ )dv. ∂θi

˙ = Une application moment µ : J → (Lie G )∗ doit satisfaire dhµ(J), f i(J) ˙ −Ω(LXf J, J) sur J , ce qui conduit au candidat évident Z Z ∂ (18) hµ(J), f i = dfi (J )dv = − fij uij dv. ∂θi P P On utilise alors la formule d’intégration par parties, valable pour toute fonction f ∈ C ∞ (P¯ ) : Z Z Z (19) fij uij dv = f uij dv + f dσ. ij P

P

∂P

Ici σ est une mesure sur ∂P définie en demandant que, sur la face de normale entière ν, on ait |dv| = |ν ∧ dσ 2 |. La formule (19) se vérifie par un calcul direct près des faces utilisant l’asymptotique du potentiel symplectique u. Le second terme de (19) ne dépend pas de la structure complexe J, donc est constant sur J . On peut donc choisir comme application moment : Z (20) hµ(J), f i = −f uij ij dv. P

On a finalement montré :

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

188

O. BIQUARD

Lemme 1.5. — L’action de G sur J est hamiltonienne, avec application moment µ(J) = −uij ij = Scal(J). Au passage, de (17) on déduit l’action infinitésimale complexifiée de G : JLXf J = JLXf J = dC fi ⊗ J

∂ ∂ − dfi ⊗ . ∂θi ∂θi

∂ ∂ En particulier JLXf J( ∂x j ) = −fij ∂θ . Cela signifie que la variation infinitésimale i J˙ = JLXf J préserve la distribution D = ∩ ker(Jdxi ) introduite en 1.3 et donc s’ex∂ ∂ prime par une variation du potentiel symplectique. Compte tenu de J ∂x j = ujk ∂θ , k il devient clair que :

Lemme 1.6. — L’action infinitésimale complexifiée de f ∈ Lie(G ), à savoir JLXf J, correspond à une modification infinitésimale du potentiel symplectique par −f . Modifier infinitésimalement le potentiel symplectique par f correspond donc à modifier J par −JLXf J = −LJXf J (car J est intégrable), donc à l’action infinitésimale du champ de vecteurs −JXf sur J. Ainsi toutes les structures complexes correspondant à différents potentiels symplectiques sont-elles les mêmes. Il est équivalent de fixer la structure complexe en variant le potentiel de Kähler, ou de fixer la structure symplectique en variant le potentiel symplectique, le lien entre les deux points de vue étant donné par la transformation de Legendre. 1.9. Énergie, obstructions Étant donnée une action hamiltonienne d’un groupe G sur une variété kählérienne, on peut appliquer la théorie de la réduction kählériennne en lien avec la théorie géométrique des invariants pour l’action complexifiée de GC . En dimension infinie, même si le groupe n’admet pas de complexification, certaines propriétés restent formellement vraies. Cette observation est à la base du programme proposé par Donaldson [10, 11]. On va énumérer ici les conséquences formelles dans le cas torique (leur démonstration directe dans ce cas étant généralement très simple). La première manifestation de la théorie est l’existence d’une fonctionnelle E (de Kempf-Ness) sur une GC -orbite, dont la différentielle est l’application moment. En fait, la fonctionnelle est définie sur l’espace symétrique GC /G. Dans le cas kählérien, même si la complexification du groupe des difféomorphismes hamiltoniens n’existe pas, Donaldson a montré que l’espace symétrique correspondant s’interprète comme l’espace des potentiels de Kähler. Dans notre cas, le lemme 1.6 implique directement que G C /G s’identifie à l’espace des potentiels symplectiques. Si on veut résoudre l’équation Scal = A, la fonctionnelle E(u) doit donc satisfaire, en utilisant (19), Z Z Z Z ij ij dE(u) ˙ = (Scal −A)udv ˙ = (−uij − A)udv ˙ = (−u u˙ ij − Au)dv ˙ + udσ ˙ P

ASTÉRISQUE 339

P

P

∂P

(1018)

qui s’intègre immédiatement en Z (21) E(u) = − log det(uij )dv + FA (u), P

189

MÉTRIQUES EXTRÉMALE TORIQUES

Z

Z f dσ −

où FA (f ) = ∂P

Af dv. P

En géométrie kählérienne, cette fonctionnelle E est la K-énergie de Mabuchi, et la partie linéaire FA est le caractère de Futaki. La théorie générale dit que E est convexe le long des géodésiques de GC /G. Les géodésiques étant données par l’exponentielle du groupe, le lemme 1.6 implique que les géodésiques sont des droites dans l’espace des potentiels : ut = u0 + tf . La convexité se vérifie en effet immédiatement par un calcul : Z d2 E(u0 + tf ) = fij uik fkl ujl dv > 0 dt2 t=0 P avec égalité si et seulement si fij = 0, c’est-à-dire f est affine. Cela implique : Lemme 1.7. — Il existe au plus une solution u de Scal(u) = A. Unique veut dire à fonction affine près, mais on se rappellera que le potentiel symplectique est justement défini à fonction affine près. Le fait que les géodésiques dans l’espace des potentiels de Kähler soient des droites fut observé par Guan [18]. En particulier elles sont lisses, ce qui n’est pas connu dans le cas général (voir [7] pour la construction de ces géodésiques). L’unicité des métriques kählériennes toriques à courbure scalaire constante en découle immédiatement, le cas général est beaucoup plus difficile [12, 22, 21, 8]. Remarque 1.8. — Dans le cas presque complexe, les raisonnements ci-avant restent valables et montrent que les géodésiques de l’espace symétrique GC /G restent données par des droites Gij − tfij pour des fonctions f ∈ C ∞ (P¯ ). On peut aussi montrer que la courbure scalaire de la métrique (6) reste donnée par la formule −Gij ij . Cela s’interprète dans le cadre du programme de Donaldson : en géométrie presque-complexe, il est naturel de chercher une solution de Scal(J) = cst en variant J dans une orbite complexifiée de G , voir [10]. En général cela reste formel car on ne sait pas résoudre l’équation de variation de J et donc décrire l’orbite complexifiée. Dans le cas torique en revanche, l’orbite se décrit très bien comme les variations de (Gij ) par le hessien d’une fonction f , et le même raisonnement que dans le cas kählérien prouve l’unicité d’une éventuelle structure J à courbure scalaire constante. La question de l’existence est un problème ouvert et intéressant. Une obstruction générale est l’annulation de l’invariant de Futaki dans la direction des symétries infinitésimales. Ici les champs de vecteurs Xf de l’action de T sont

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

190

O. BIQUARD

donnés par des potentiels f affines. Récrivons alors la formule (19) comme Z Z ij (22) fij u dv = f (A − Scal(u))dv + FA (f ). P

P

On voit immédiatement que, pour une fonction affine f , si Scal(u) = A, il faut que FA (f ) = 0. Pour f = 1, on obtient la contrainte : (23)

A=

Vol ∂P . Vol P

Pour f = xi , compte tenu de cette valeur de A, la condition FA (xi ) = 0 se traduit par Z Z 1 1 i x dv = xi dσ. (24) Vol(P ) P Vol(∂P ) ∂P Autrement dit : Lemme 1.9. — Pour que Scal(u) = A, il est nécessaire que A = barycentre de P soit égal au barycentre de ∂P .

Vol ∂P Vol P

et que le

Bien sûr, il s’agit d’une contrainte forte sur le polytope. Pour un polytope général, on s’attend à pouvoir résoudre seulement l’équation (25)

− uij ij = A

avec A = c0 + ci xi .

On peut choisir les constantes ci de manière unique de sorte que la contrainte FA (f ) = 0 pour f affine demeure respectée. Les solutions de l’équation (25) sont des métriques kählériennes extrémales, c’est-à-dire que le gradient de leur courbure scalaire est un champ de vecteurs holomorphe (condition trouvée par Calabi pour R minimiser Scal2 dans une classe de Kähler). La conjecture 0.1 a une version pour les métriques extrémales, mais le théorème de Donaldson 0.2 est restreint au cas des métriques à courbure scalaire constante, et nous nous restreindrons donc aussi le plus souvent à ce cas dans la suite. Un cas particulier est celui d’une variété torique de Fano (ce qui, sur le polytope, revient à dire qu’on peut prendre comme équations des faces hνi , xi + 1). Au vu de (15), l’équation de Kähler-Einstein s’écrit alors 1 log det(uij ) − ϕ = 0. 2 Si, dans le second membre, on met une fonction linéaire ci xi , on obtient un soliton de Kähler-Ricci. Wang et Zhu [29] ont montré que toute variété torique de Fano admet un soliton de Ricci. Il est Kähler-Einstein quand l’invariant de Futaki s’annule.

ASTÉRISQUE 339

(1018)

191

MÉTRIQUES EXTRÉMALE TORIQUES

1.10. Énoncé précis du théorème Enfin, il y a une dernière obstruction plus générale qui provient de la formule (22) : si on choisit une fonction f convexe, alors le premier membre est positif ; ainsi, si Scal(u) = A, il faut que FA (f ) > 0, avec égalité si et seulement si f est affine. On peut alors énoncer précisément la conjecture et le théorème : Conjecture 1.10 (Donaldson). — Soit A ∈ C ∞ (P¯ ). Il existe une solution de l’équation Scal(u) = A si et seulement si, pour toute fonction convexe f ∈ C ∞ (P¯ ), on a FA (f ) > 0 avec égalité si et seulement si f est affine. Théorème 1.11 (Donaldson). — En dimension 2, pour A = est vraie.

Vol ∂P Vol P

, la conjecture

Dans la fin de cette section, on va rapidement faire le lien entre la condition du théorème 1.11 et la notion algébrique de K-stabilité. 1.11. K-stabilité Soit (W, Λ) une variété algébrique polarisée, munie d’une action de C∗ . Alors C∗ agit encore sur H 0 (W, Λk ), dont on notera la dimension dk , et sur son déterminant Λdk H 0 (W, Λk ) avec poids wk . Alors wk et dk sont des polynômes à coefficients rationnels, et F2 wk F1 + 2 + ··· = F0 + kdk k k L’invariant de Futaki de (W, Λ) est alors défini comme le terme F1 de ce développement. Une configuration test pour une variété polarisée (X, L) est la donnée : – d’un schéma V , muni d’une action de C∗ , et d’un morphisme plat p : V → C commutant avec l’action standard de C∗ sur C ; – d’un fibré en droite C∗ -équivariant L sur V de sorte que la fibre générale de p soit (X, Lr ). Un cas particulier est celui d’une configuration test produit : si (X, L) est munie d’une action de C∗ , alors il s’agit de la configuration test (V, L ) = (X × C, L). Si on a une configuration test, la fibre au-dessus de l’origine, (p−1 (0), L |p−1 (0) ), est munie d’une action de C∗ et admet donc un invariant de Futaki F1 qui est par définition l’invariant de Futaki de la configuration test. Dans le cas d’une configuration test produit, il coïncide avec l’invariant de Futaki de (X, L). On dit alors qu’une variété polarisée (X, L) est K-polystable si toute configuration test a un invariant de Futaki F1 6 0, avec égalité si et seulement si la configuration test

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

192

O. BIQUARD

est produit. Cette définition doit être pensée comme un analogue du critère de HilbertMumford pour la stabilité en théorie géométrique des invariants, les configurations test jouant le rôle des sous-groupes à un paramètre. C’est avec cette forme de stabilité que Donaldson [13] a énoncé la conjecture 0.1, mais il faut noter qu’on sait déjà que la forme précise de K-stabilité nécessaire et suffisante pour l’existence de métriques kählériennes à courbure scalaire constante n’est pas exactement celle-là [2, Remark 9], et des variantes ont été proposées (K-polystabilité « uniforme »). 1.12. Configurations test toriques ∂P Revenons à une variété torique. Fixons A = Vol Vol P , et rappelons l’invariant de Futaki FA (f ) défini dans (21). Par approximation, il est évident que les deux conditions suivantes sont équivalentes : 1. FA (f ) > 0 pour toute fonction convexe f ∈ C ∞ (P¯ ) ;

2. FA (f ) > 0 pour toute fonction convexe f affine par morceaux. Soient f1 , . . . , fh des fonctions affines à coefficients rationnels ; alors f = max(f1 , . . . , fh ) est une fonction convexe affine par morceaux. On va voir que f détermine une configuration test torique, dont l’invariant de Futaki est lié à FA (f ). Cela fournit une interprétation algébrique de l’hypothèse du théorème 1.11. À partir de P ⊂ Rn , on obtient un nouveau polytope convexe Q ⊂ Rn × R en posant Q = {(x, t), x ∈ P, 0 < t < R − f (x)}, où R est une constante fixée assez grande. t R

f

Q x

x

On choisit k ∈ N minimal de sorte que kQ soit donné par des équations à coefficients entiers. Alors kQ est le polytope associé à une variété torique (V, L ) (singulière car les conditions de Delzant ne sont pas satisfaites). Le tore complexe TCn+1 = TC × C∗ agit sur V . Pour simplifier la suite de l’exposition, on fera comme si k = 1. La face {xn+1 = 0} ∩ Q de Q est P , donc on obtient ainsi un plongement TC -équivariant i : X ,→ V . D’un autre côté, l’application Q → [0, R] ⊂ R, (x, t) 7→ t, correspond à une application C∗ -équivariante V → CP 1 (l’intervalle étant en effet le polytope associé à CP 1 ).

ASTÉRISQUE 339

(1018)

MÉTRIQUES EXTRÉMALE TORIQUES

193

Il est évident que la fibre de cette application en 0 est i(X), et on en déduit assez facilement que la fibre générale de p est (X, L). En revanche la fibre à l’infini peut être singulière, et il en résulte que (V, L ) est une configuration test pour (X, L), en prenant la projection sur C donnée par 1/t. Le point crucial est alors : Lemme 1.12. — L’invariant de Futaki de cette configuration test est lié à l’invariant de Futaki FA (f ) de la fonction convexe f par 1 − F1 = FA (f ). Vol P La condition de positivité de FA (f ) se traduit donc en la négativité de F1 . Il en résulte que l’hypothèse du théorème 1.11 revient à la K-polystabilité de la variété torique par rapport aux configurations test toriques. Démonstration. — La preuve repose sur les deux faits classiques suivants : – pour une variété torique, la dimension de l’espace des sections de Lk est obtenue comme le nombre de points entiers dans le polytope k P¯ ; – l’asymptotique du nombre de points entiers est donné par 1 (26) #(k P¯ ∩ Λ) = k n Vol P + k n−1 Vol dσ ∂P + O(k n−2 ), 2 2 où le volume de ∂P est calculé par rapport à la mesure entière introduite en 1.8. La démonstration consiste alors, à l’aide de quelques suites exactes, à montrer que dk = h0 (X, Lk ) et wk = h0 (V, L k ) − h0 (X, Lk ), puis à utiliser la formule (26) pour wk obtenir l’asymptotique de kd et donc F1 . k

2. DÉMONSTRATION DU THÉORÈME La preuve du théorème est longue et technique, elle occupe les trois articles [14, 15, 17]. On ne pourra en donner ici qu’une idée très succincte. 2.1. La méthode de continuité Comme souvent dans les résolutions d’équations aux dérivées partielles géométriques de type elliptique, la méthode est une méthode de continuité : on veut résoudre un problème P1 , on dispose d’une solution à un autre problème P0 , et on invente une suite « continue » de problèmes (Pt )t∈[0,1] qui interpole entre les deux. Avec un bon choix des problèmes intermédiaires, on essaie de montrer que l’ensemble des t ∈ [0, 1] pour lesquels on peut résoudre Pt est ouvert et fermé. L’ouverture se traite par un théorème de fonctions implicites, et est généralement la partie facile ; difficile en revanche est le problème de fermeture, qui consiste en un théorème de compacité pour les solutions des problèmes Pt .

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

194

O. BIQUARD

Dans le cas qui nous occupe, on cherche un potentiel symplectique u sur un polytope P , satisfaisant l’équation −uij ij = A. Une méthode de continuité évidente consisterait à partir du potentiel de Guillemin uG , de courbure scalaire A0 = −(uG )ij ij , puis ¯ à interpoler par une suite bien choisie (At ) de fonctions sur P entre A0 et A, pour résoudre le problème −(ut )ij ij = At . Malheureusement, dans cette méthode, il semble difficile d’obtenir les estimations a priori menant au théorème de compacité nécessaire. L’argument de Donaldson utilise en effet très spécifiquement un contrôle a priori de V i = uij j qui n’est valable qu’avec un second membre de l’équation constant (c’est la raison fondamentale pour laquelle le théorème fournit des métriques à courbure scalaire constante, mais pas des métriques extrémales). Cela mène à inventer une méthode de continuité beaucoup plus sophistiquée : l’idée est de varier, non plus le second membre de l’équation, mais le polytope P lui-même. La donnée devient alors un couple (P, σ) d’un polytope convexe dans t∗ et d’une mesure sur ∂P , proportionnelle sur chaque face à la mesure de Lebesgue. À une face E du polytope est naturellement associée une normale rentrante νE ∈ t par la relation 1 2 |νE ∧ dσ| = |dv|. L’équation de la face E a alors la forme λE (x) = hνE , xi + cE , où la constante cE est ajustée de sorte que λE s’annule sur E. Définissons le potentiel P dσ ∂P λE log λE , et A = Vol de Guillemin uG = 12 Voldv P . Le problème à résoudre devient : Problème : Étant donné (P, σ), résoudre l’équation −uij ij = A avec conditions au bord de Guillemin. L’invariant de Futaki FA de la section 1.9 reste bien défini, et Donaldson démontre : Théorème 2.1. — En dimension 2, il existe une solution du problème ci-dessus si et seulement si la condition suivante de stabilité est satisfaite : pour toute fonction convexe f ∈ C ∞ (P¯ ), on a FA (f ) > 0 avec égalité si et seulement si f est affine. Comme précédemment, la condition du théorème implique que P et ∂P ont même barycentre. Remarque 2.2. — Si le polytope n’est pas de Delzant, le problème n’est pas géométrique car le polytope ne correspond pas à une variété. Si le polytope est de Delzant et la mesure dσ n’est pas induite par la structure entière, les solutions restent des métriques à courbure scalaire constante, avec des singularités coniques le long des diviseurs correspondant aux arêtes du polytope. La question de l’existence de métriques kählériennes à courbure scalaire constante et singularités coniques le long de diviseurs est donc également résolue, dans le cas torique, par le théorème de Donaldson. Revenons à la méthode de continuité : un polytope P étant fixé, il est facile de voir que l’ensemble CP des mesures σ sur ∂P qui satisfont la condition de stabilité

ASTÉRISQUE 339

(1018)

MÉTRIQUES EXTRÉMALE TORIQUES

195

est convexe ; en outre il y a une construction simple d’une mesure naturelle σP ∈ CP dépendant continûment de P . Il en résulte immédiatement que deux polytopes (P0 , σ0 ) et (P1 , σ1 ) satisfaisant la condition de stabilité peuvent être joints par un chemin (Pt , σt ) de polytopes satisfaisant aussi cette condition. En dimension 2, le seul invariant de déformation d’un polytope convexe est son nombre d’arêtes. Pour démarrer la méthode de continuité, on a donc besoin de : Lemme 2.3. — Pour chaque entier k, il existe un polytope à k arêtes muni d’une solution de l’équation −uij ij = A. Démonstration. — On commence par observer qu’on dispose de solutions évidentes pour k = 3 et 4. Pour k = 3, le polytope de CP 2 décrit dans l’exemple 1.3, muni du potentiel de Guillemin, donne une solution. Pour k = 4 on dispose aussi du polytope de CP 1 × CP 1 , qui est un carré. Il y a un moyen géométrique simple d’ajouter une arête à un polygone : découper un petit coin près d’un sommet correspond à éclater la variété torique au point fixe correspondant.

(CP 1 × CP 1 )#CP 2

CP 1 × CP 1 c

Le modèle est fourni par C2 éclaté à l’origine, dont le polytope de Delzant Qc est donné par les équations x1 > 0, x2 > 0 et x1 + x2 > c. Le réel c > 0 est le volume du diviseur exceptionnel, et quand c → 0 on retrouve le polytope de C2 . Or cet éclatement est muni d’une métrique à courbure scalaire nulle, asymptotique à la métrique plate de C2 , la métrique de Burns, de potentiel symplectique donné par la formule (on pose x = x1 + x2 ) : 1 1 uc = x log x1 + x2 log x2 + (x − c) log(x − c) − x log x . 2 Il est alors possible, à la manière d’Arezzo-Pacard (voir [3], mais leur théorème ne s’applique pas directement ici en raison de la présence de champs de vecteurs holomorphes), d’effectuer le recollement suivant : partant d’un polygone P avec une solution u de −uij ij = A, on fixe un sommet et, quitte à changer de coordonnées par une transformation entière, on peut supposer que près du sommet le polytope est donné par les équations x1 > 0, x2 > 0. Alors pour c petit, il est possible de recoller le potentiel u avec uc et de déformer le résultat en une solution de l’équation Vol ∂Pc −uij ij = Ac = Vol Pc sur le polygone Pc = P ∩ Qc . (En réalité après cette modification de P , le barycentre de ∂Pc ne coïncide pas forcément avec celui de Pc et il est nécessaire de modifier légèrement la position des autres sommets pour satisfaire cette

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

196

O. BIQUARD

condition ; de ce fait le résultat n’est pas forcément un polytope de Delzant.) Ainsi au bout d’un nombre convenable d’éclatements obtient-on une solution initiale pour tout k. Cette proposition permet donc de démarrer la méthode de continuité : étant donné un polytope (P, σ) à k arêtes, on choisit un chemin (Pt , σt ) de polytopes satisfaisant la condition de stabilité vers un polytope (P0 , σ0 ). Lemme 2.4. — L’ensemble des t pour lesquels on peut résoudre l’équation Vol ∂Pt −uij ij = Vol Pt avec conditions au bord de Guillemin est ouvert. Démonstration. — Le fait que l’ensemble des t pour lesquels on peut résoudre l’équation soit ouvert est standard dans ce contexte : pour des problèmes d’application moment, la convexité (stricte) de l’énergie conduit immédiatement à l’inversibilité de la linéarisation du problème. Les conditions au bord de P ne font que traduire que le problème vu sur la variété torique associée XP correspond à des potentiels lisses sur XP entier (en différenciant (12), on s’aperçoit que u(x) ˙ + ϕ(y) ˙ = 0, donc une variation infinitésimale du potentiel symplectique se traduit par une variation opposée du potentiel kählérien sur XP ). La linéarisation se ramène ainsi à un problème elliptique (d’ordre 4), inversible, sur la variété compacte XP et on peut appliquer le théorème des fonctions implicites. (Les choses se compliquent un peu quand P ne satisfait pas les conditions de Delzant, puisqu’on ne dispose plus de la variété XP ; néanmoins XP continue alors à exister localement et cela permet de définir dans des cartes locales les espaces fonctionnels dans lesquels la linéarisation peut être inversée.) 2.2. Le théorème de compacité On en arrive maintenant à la partie la plus délicate du théorème : la compacité. Il s’agit d’une démonstration incroyablement difficile et longue, malgré l’apparence que beaucoup d’outils relèvent de l’analyse convexe à deux variables. Aussi ne pourra-t-on faire ici que quelques commentaires superficiels. Les différentes étapes sont : 1. 2. 3. 4.

contrôle L1 de la solution ; contrôle L∞ ; convergence près des arêtes, mais loin des sommets ; convergence près des sommets.

On a donc une solution de −uij ij = A sur (P, σ) et on veut contrôler u et toutes ses dérivées, et aussi avoir une borne inférieure sur le hessien (uij ). Comme u est défini à fonction affine près, on peut fixer un point p ∈ P tel que u > u(0) = 0. On dira alors que u est normalisée. La seule étape facile est la première ; la K-stabilité est directement utilisée pour montrer :

ASTÉRISQUE 339

(1018)

MÉTRIQUES EXTRÉMALE TORIQUES

197

Lemme 2.5. — Si on a une solution normalisée u de −uij ij = A sur (P, σ), alors Z Z udσ 6 c(P, σ, A), udv 6 c0 (P, σ, A). ∂P

P

Ce contrôle peut sembler faible, mais parce que u est convexe positive, il implique déjà une borne uniforme C 1 sur u sur tout compact de l’intérieur de P , ainsi que sur la restriction u|∂P sur tout compact inclus dans l’intérieur d’une arête ; ainsi ce contrôle permet-il de contrôler C 0 la fonction u en dehors des sommets, et C 1 sur les compacts. Démonstration. — La condition de stabilité implique, pour toute fonction convexe f ∈ C 0 (P¯ ), l’inégalité Z Z (27) FA (f ) = f dσ − Af dv > 0, ∂P

P

avec égalité si et seulement si f est affine. On en déduit, si f est normalisée, Z (28) f dσ 6 C(P, σ, A)FA (f ). ∂P

R (Par l’absurde : sinon existe une suite fn avec ∂P fn dσ = 1 et FA (fn ) → 0. Par la compacité pour les fonctions convexes mentionnée ci-dessus, on extrait une limite R fn → f où f est continue en dehors des sommets, vérifie ∂P f dσ = 1 et FA (f ) = 0. L’inégalité (27) reste valable pour ce type de fonctions, et on aboutit à ce que f doit être affine, ce qui est impossible puisqu’elle est normalisée.) R Appliquant (28) à u, et observant que par (19) on a FA (u) = P uij uij dv = 2 Vol P , on déduit immédiatement le lemme. Partant de ce contrôle, qui, comme on l’a vu, implique un contrôle C 1 à l’intérieur, et en utilisant l’équation −uij ij = A, on peut montrer des estimations intérieures pour la solution : sur toute partie compacte de l’intérieur, (29)

K −1 6 (uij ) 6 K,

|∇` u| 6 c` .

Ces contrôles intérieurs ne sont pas faciles et s’appuient notamment sur des idées de Trudinger-Wang pour certaines équations d’ordre 4 (voir [28] et les références qui s’y trouvent). En tout cas, ces estimations disent que, sur tout compact intérieur, une suite de solutions converge C ∞ vers une limite. Le problème est donc de comprendre le comportement au bord, et c’est la difficulté majeure du théorème. La seconde étape est le contrôle L∞ : la fonction u étant convexe positive, il suffit de la borner aux sommets du polygone. L’argument est délicat et s’appuie notamR ment sur l’intégration par parties FA (f ) = P uij fij dv, appliquée à des fonctions f judicieusement choisies.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

198

O. BIQUARD

La troisième étape est importante, elle permet de comprendre le comportement au bord. Elle s’appuie sur une astuce liée à la dimension 2 et sur le fait que le second membre de l’équation soit constant : dans ce cas, on a une borne uniforme sur le champ de vecteurs V i = uij j : (30)

|V | 6 c(P, σ, A).

Cette borne s’établit en observant que l’équation −uij ij = A implique div W = 0, où A W (x) = V (x) − 2 x, donc W = J grad H (ici J et grad sont pour la métrique plate (dx1 )2 + (dx2 )2 ), pour une fonction H dont on montre qu’elle satisfait l’équation uij Hij = 0. Pour ce type d’équation, on sait borner |dH| en fonction de la donnée au bord, et cela donne le contrôle sur V . Cette estimation a priori va servir pour empêcher l’effondrement des solutions. C’est un problème général en géométrie riemannienne de dimension 4 : si on a une suite de métriques, solutions d’une équation aux dérivées partielles non linéaire géométrique, alors la norme L2 de la courbure est contrôlée par la topologie (formules de Chern-Weil) ; si en outre le rayon d’injectivité est borné inférieurement (la métrique ne s’effondre pas), alors on peut souvent montrer qu’il y a compacité pour le problème, c’est-à-dire qu’on peut toujours extraire une limite d’une suite de solutions. L’idée de Donaldson est de fournir pour les fonctions convexes un substitut au rayon d’injectivité, suffisant pour contrôler la convergence des solutions. Au vu de la forme (7) pour la métrique, on bornera inférieurement un rayon d’injectivité en bornant inférieurement la matrice (uij ), c’est-à-dire en bornant supérieurement le hessien (uij ). Près d’une arête, qu’on peut supposer donnée par l’équation (entière) x1 = 0, Donaldson borne plutôt la quantité suivante, qui est une sorte de dérivée seconde dans la direction x1 :  1 (31) D(x1 , x2 ) = 1 u(0, x2 ) − u(x1 , x2 ) − x1 u1 (x1 , x2 ) . x Par exemple, pour u = 21 x1 log x1 + (x2 − ax1 )2 , on obtient le champ de vecteurs V = (2, a) et la quantité D(x1 , x2 ) = 12 + a2 (x1 )2 . L’idée de la preuve, suggérée par cet exemple, est qu’une borne sur V implique une borne sur D, mais la réalisation est très difficile : elle passe par un argument d’éclatement et des manipulations qui utilisent la transformation de Legendre pour estimer certaines aires. C’est le cœur de la démonstration. La borne sur D permet le contrôle de la métrique g, mais ce n’est pas une étape facile. On dérive une borne sur la courbure, comme souvent en géométrie riemannienne, par un argument d’éclatement : si la courbure n’est pas bornée, on extrait aux points de concentration des limites dont on montre qu’elles ne peuvent pas exister. Un autre

ASTÉRISQUE 339

(1018)

MÉTRIQUES EXTRÉMALE TORIQUES

199

argument est nécessaire pour borner inférieurement la métrique et surtout son rayon d’injectivité. À partir de là, des arguments plus classiques de compacité permettent, à partir de l’équation sur la courbure scalaire, de borner toutes les dérivées de la métrique et donc de u. Enfin la dernière étape consiste en le contrôle près des sommets, où des arguments similaires (mais pas plus faciles) sont employés près des sommets.

RÉFÉRENCES [1] M. Abreu – Kähler geometry of toric varieties and extremal metrics, Internat. J. Math. 9 (1998), p. 641–651. [2] V. Apostolov, D. M. J. Calderbank, P. Gauduchon & C. W. TønnesenFriedman – Hamiltonian 2-forms in Kähler geometry. III. Extremal metrics and stability, Invent. Math. 173 (2008), p. 547–601. [3] C. Arezzo & F. Pacard – Blowing up and desingularizing constant scalar curvature Kähler manifolds, Acta Math. 196 (2006), p. 179–228. [4] O. Biquard – Métriques kählériennes à courbure scalaire constante : unicité, stabilité, Séminaire Bourbaki, vol. 2004/2005, exp. no 938, Astérisque 307 (2006), p. 1–31. [5] E. Calabi – Extremal Kähler metrics, in Seminar on Differential Geometry, Ann. of Math. Stud., vol. 102, Princeton Univ. Press, 1982, p. 259–290. [6]

, Extremal Kähler metrics. II, in Differential geometry and complex analysis, Springer, 1985, p. 95–114.

[7] X. Chen – The space of Kähler metrics, J. Differential Geom. 56 (2000), p. 189– 234. [8] X. Chen & G. Tian – Geometry of Kähler metrics and foliations by holomorphic discs, Publ. Math. Inst. Hautes Études Sci. 107 (2008), p. 1–107. [9] T. Delzant – Hamiltoniens périodiques et images convexes de l’application moment, Bull. Soc. Math. France 116 (1988), p. 315–339. [10] S. K. Donaldson – Remarks on gauge theory, complex geometry and 4-manifold topology, in Fields Medallists’ lectures, World Sci. Ser. 20th Century Math., vol. 5, World Sci. Publ., River Edge, NJ, 1997, p. 384–403.

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

200

O. BIQUARD

[11]

, Symmetric spaces, Kähler geometry and Hamiltonian dynamics, in Northern California Symplectic Geometry Seminar, Amer. Math. Soc. Transl. Ser. 2, vol. 196, Amer. Math. Soc., 1999, p. 13–33.

[12]

, Scalar curvature and projective embeddings. I, J. Differential Geom. 59 (2001), p. 479–522.

[13]

, Scalar curvature and stability of toric varieties, J. Differential Geom. 62 (2002), p. 289–349.

[14]

, Interior estimates for solutions of Abreu’s equation, Collect. Math. 56 (2005), p. 103–142.

[15]

, Extremal metrics on toric surfaces : a continuity method, J. Differential Geom. 79 (2008), p. 389–432.

[16]

, Kähler geometry on toric manifolds, and some other manifolds with large symmetry, in Handbook of geometric analysis. No. 1, Adv. Lect. Math. (ALM), vol. 7, Int. Press, Somerville, MA, 2008, p. 29–75.

[17]

, Constant scalar curvature metrics on toric surfaces, Geom. Funct. Anal. 19 (2009), p. 83–136.

[18] D. Guan – On modified Mabuchi functional and Mabuchi moduli space of Kähler metrics on toric bundles, Math. Res. Lett. 6 (1999), p. 547–555. [19] V. Guillemin – Kaehler structures on toric varieties, J. Differential Geom. 40 (1994), p. 285–309. [20]

, Moment maps and combinatorial invariants of Hamiltonian T n -spaces, Progress in Math., vol. 122, Birkhäuser, 1994.

[21] T. Mabuchi – Uniqueness of extremal Kähler metrics for an integral Kähler class, Internat. J. Math. 15 (2004), p. 531–546. [22]

, An energy-theoretic approach to the Hitchin-Kobayashi correspondence for manifolds. I, Invent. Math. 159 (2005), p. 225–243.

[23]

, K-stability of constant scalar curvature polarization, prépublication arXiv:0812.4093.

[24]

, A stronger concept of K-stability, prépublication arXiv:0910.4617.

[25] J. Stoppa – K-stability of constant scalar curvature Kähler manifolds, Adv. Math. 221 (2009), p. 1397–1408. [26] G. Székelyhidi – Optimal test-configurations for toric varieties, J. Differential Geom. 80 (2008), p. 501–523. [27] G. Tian – Kähler-Einstein metrics with positive scalar curvature, Invent. Math. 130 (1997), p. 1–37. [28] N. S. Trudinger & X.-J. Wang – The affine Plateau problem, J. Amer. Math. Soc. 18 (2005), p. 253–289.

ASTÉRISQUE 339

(1018)

MÉTRIQUES EXTRÉMALE TORIQUES

201

[29] X.-J. Wang & X. Zhu – Kähler-Ricci solitons on toric manifolds with positive first Chern class, Adv. Math. 188 (2004), p. 87–103. [30] B. Zhou & X. Zhu – K-stability on toric manifolds, Proc. Amer. Math. Soc. 136 (2008), p. 3301–3307.

Olivier BIQUARD UPMC – Université Paris VI UMR CNRS 7586 Institut de Mathématiques de Jussieu Case 247 4 place Jussieu F-75005 PARIS E-mail : [email protected]

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

339

ASTÉRISQUE 2011

SÉMINAIRE BOURBAKI VOLUME 2009/2010 EXPOSÉS 1012-1026 (1019) Grandes matrices aléatoires et théorèmes d’universalité Alice GUIONNET

SOCIÉTÉ MATHÉMATIQUE DE FRANCE Publié avec le concours du CENTRE NATIONAL DE LA RECHERCHE SCIENTIFIQUE

Séminaire BOURBAKI 62e année, 2009-2010, no 1019, p. 203 à 237

Mars 2010

GRANDES MATRICES ALÉATOIRES ET THÉORÈMES D’UNIVERSALITÉ [d’après Erdős, Schlein, Tao, Vu et Yau] par Alice GUIONNET

INTRODUCTION Les grandes matrices aléatoires sont tout d’abord apparues en statistique dans les travaux de Wishart [37] pour décrire des tableaux de données aléatoires, puis en physique dans les travaux de Wigner [36] et Dyson pour approximer un opérateur modélisant le Hamiltonien d’un noyau excité. Montgomery, soutenu par les simulations d’Odlyzko, a conjecturé que leurs valeurs propres étaient également reliées aux fameux zéros de la fonction de Riemann sur la droite critique, les espacements de ceux-ci loin de l’axe réel étant distribués de façon identique. Les matrices aléatoires sont depuis intervenues dans de nombreux contextes, et leur spectre a attiré un intérêt grandissant. On a notamment étudié sa convergence globale, la distribution des espacements des valeurs propres à l’intérieur du spectre et les fluctuations des valeurs propres extrêmes, quand la taille des matrices tend vers l’infini. Le cas de matrices aléatoires à coefficients gaussiens s’est révélé plus facile à appréhender, compte tenu de formules explicites pour la loi jointe des valeurs propres. Néanmoins, comme pour le théorème central limite, il est attendu que ces théorèmes limites sont universels et ne dépendent que très peu de la nature des coefficients. Le but de cet exposé est de montrer les récents efforts fournis pour étudier cette universalité. Nous considérerons les matrices dites de Wigner qui sont hermitiennes et avec des coefficients indépendants et équidistribués (modulo l’hypothèse de symétrie). Plus précisément, on notera XN une matrice hermitienne N × N dont les coefficients (Xij )1≤i 0 telles que, pour tout t > C Ä ä Ä ä 0 0 (2) ν |z| > tC ≤ e−t µ |z| > tC ≤ e−t . Si ν est une mesure sur C, nous supposerons que les lois de la partie imaginaire et de la partie réelle sont indépendantes, et, si ν est une mesure sur la droite réelle, R que x3 dν(x) s’annule ou que ν a au moins trois points dans son support. Nous distinguerons ces deux cas par un paramètre β qui sera égal à deux (resp. un) si le support est dans C (resp. R). Théorème 0.1. — Soit B un intervalle compact de R et x ∈ (−2, 2). La probabilité qu’aucune valeur propre de XN ne tombe dans un intervalle 1 1 N 2 x + N − 2 ρsc (x)−1 B converge quand la dimension N tend vers l’infini vers une limite non triviale qui ne dépend que de β et de B (voir la description de cette limite en (11) dans le cas où β = 2). La probabilité qu’aucune valeur propre de XN ne tombe dans l’ensemble 1 1 2N 2 + N − 6 B converge quand la dimension N tend vers l’infini vers une limite non triviale qui ne dépend que de β et de B (voir (12) dans le cas où β = 2). Le second résultat décrit les fluctuations des valeurs propres de YN au bord du 2 spectre (proche de 2), qui sont d’ordre N − 3 , alors que le premier concerne celles des valeurs propres à l’intérieur du spectre, qui sont beaucoup plus petites, d’ordre N −1 . Ce dernier point permet d’étudier la convergence de la loi des espacements des valeurs propres à l’intérieur du spectre ; on a par exemple le résultat suivant (voir [24, p.84] et [1, Theorem 4.2.49] avec [7] pour le cas gaussien et [10, Theorem 3] pour le cas général) dans le cas où ν est une mesure sur C. Théorème 0.2. — Pour tout x ∈ (−2, 2), toute fonction lN tendant vers l’infini plus lentement que N , le nombre moyen de valeurs propres de YN à distance inférieure à lN /N ρsc (x) de x et dont les espacements sont plus petits que s/ρsc (x)N est approximativement égal à PGaudin ([0, s])lN , où PGaudin est la distribution de Gaudin.

ASTÉRISQUE 339

(1019)

GRANDES MATRICES ALÉATOIRES ET UNIVERSALITÉ

205

Le second point du théorème 0.1 donne les fluctuations de la plus grande valeur propre. Théorème 0.3. — Pour tout t ∈ R, la plus grande valeur propre λ1 de XN est telle que Å ã 2 λ1 lim P N 3 ( √ − 2) ≤ t = Fβ (t) N →∞ N avec Fβ la fonction de répartition de la loi de Tracy-Widom. Ces théorèmes ont été démontrés pour des matrices gaussiennes depuis une quinzaine d’années, notamment après les travaux de M. Mehta [24], P. Forrester [17] et C. Tracy et H. Widom [33, 34], grâce à la formule explicite de la loi jointe des valeurs propres et sa propriété de loi déterminantale. Les premiers pas vers l’universalité ont été franchis par K. Johansson [21] qui a considéré des matrices obtenues comme somme d’une matrice gaussienne et d’une matrice de Wigner plus générale, de nouveau grâce à une structure déterminantale de la loi jointe des valeurs propres. Dans ce dernier cas, il est primordial de supposer que les coefficients sont complexes hors de la diagonale (cas hermitien). L’universalité des fluctuations au bord du spectre a été démontrée par des techniques de moments et de combinatoire par A. Soshnikov. Cette approche ne pouvant permettre d’analyser les espacements des valeurs propres à l’intérieur du spectre, il a fallu attendre l’été dernier pour que cette question soit enfin résolue par deux équipes indépendantes, d’une part L. Erdős, B. Schlein et H.T. Yau et leurs collaborateurs, et d’autre part T. Tao et V. Vu. Les approches de ces deux équipes, quoique différentes, reposent sur des raisonnements beaucoup plus probabilistes et montrent une sorte de régularité des espacements des valeurs propres en fonction des coefficients de la matrice. Les travaux presque simultanés de T. Tao et V. Vu [32] et de L. Erdős, B. Schlein, S. Péché, J. Ramirez et H.T. Yau [9] ont permis d’étendre ces résultats d’une part sous la condition que les quatre premiers moments de ν et µ sont les mêmes que ceux de la gaussienne (ou plus généralement que ceux d’une matrice étudiée par K. Johansson [21]), et d’autre part, sous des conditions de régularité de la densité de ν et µ. Alliant ces résultats, L. Erdős, J. Ramirez, B. Schlein, T. Tao, V. Vu et H.T. Yau [10] ont pu démontrer l’universalité dans le cas hermitien sous des conditions de queues exponentielles seulement. Afin de pouvoir considérer des ensembles plus généraux, et en particulier les matrices à coefficients réels (cas symétrique), L. Erdős, B. Schlein et H.T. Yau [14] ont introduit une approche dynamique qui permet de largement généraliser les résultats de K. Johansson [21] sans utiliser de formule explicite ou de structure déterminantale. Ce point de vue a été étendu dans les travaux récents de L. Erdős, H.T. Yau et Y. Yin [15, 16] pour étudier les matrices de Wishart et à bande, et réduire considérablement les hypothèses sur les troisième et quatrième moments. L’universalité est attendue à

SOCIÉTÉ MATHÉMATIQUE DE FRANCE 2013

206

A. GUIONNET

l’intérieur (resp. au bord) du spectre seulement sous une condition de second (resp. quatrième) moment fini, voir [22] pour ce résultat dans le cas de coefficients complexes ayant une composante gaussienne. Le but de ces notes est de présenter les idées des preuves de ces différents résultats, en précisant particulièrement celles des travaux de L. Erdős, B. Schlein et H.T. Yau et T. Tao et V. Vu. Nous évoquerons également leurs liens avec les propriétés de délocalisation des vecteurs propres de XN . Nous ne parlerons pas des diverses généralisations à d’autres ensembles de matrices tels que les matrices de Wishart [25], [31] et [15] ou les matrices à bande généralisées [16]. Par ailleurs, les résultats que nous développerons concernent la situation de coefficients indépendants, totalement différente par exemple de celle rencontrée pour les modèles de matrices avec un potentiel non quadratique pour laquelle nous renvoyons le lecteur à [6]. Un des problèmes qui semble encore ouvert pour des matrices à coefficients indépendants concerne le cas de matrices à bande dont les coefficients sont nuls en dehors d’une bande de largeur W bien plus petite que N . Il est attendu que pour W √ bien plus grand que N les vecteurs propres sont délocalisés comme pour les matrices √ de Wigner, voir le corollaire 4.4, alors que pour W  N , ils sont localisés. Cette transition serait un modèle simple de la transition d’Anderson.

1. MATRICES GAUSSIENNES ET LOIS DÉTERMINANTALES Nous considérons dans cette section le cas de matrices de Wigner GN dites du GUE (pour Gaussian Unitary Ensemble), c’est-à-dire les matrices XN telles que ν (resp. µ) est une mesure gaussienne standard sur C (resp. R) ; pour toute fonction test f (3) Z Z Z Z x2 dxe− 2 x + iy f (x)dµ(x) = f (x) √ et f (z)dν(z) = f ( √ )dµ(x)dµ(y) . 2π 2 R Comme nous allons le voir, la particularité de ces matrices est que la loi jointe de leurs valeurs propres est explicite et est une loi déterminantale. Cette structure très particulière, que nous allons bientôt décrire, permet une analyse relativement aisée des propriétés locales du spectre. Il n’est pas difficile de constater que la loi de la matrice GN est laissée invariante par la conjugaison U GN U ∗ par une matrice unitaire, et par conséquent que les valeurs propres DN = (λ1 , . . . , λN ) de GN sont indépendantes de ses vecteurs propres, dont la matrice V N suit la mesure uniforme sur le groupe unitaire. En calculant le Jacobien du changement de coordonnés qui à GN associe DN et une paramétrisation de V N ,

ASTÉRISQUE 339

(1019)

207

GRANDES MATRICES ALÉATOIRES ET UNIVERSALITÉ

on trouve que la loi des valeurs propres est décrite par la mesure de probabilité (4)

1 d P N (x1 , . . . , xN ) = ZN

Y

2

|xi − xj |

N Y

2

e−xi /2 dxi .

i=1

1≤i