PERSPECTIVES IN MATHEMATICS, Vol. 10 J. Coates and S. Helgason, editorsAutomorphic Forms, Shimura Varieties, and L-functions Volume I Proceedings of a Conference held at the University of Michigan, Ann Arbor, July 6-16, 1988 Edited by Laurent Clozel James S, Milne Département de Mathématiques | Mathematics Department Université de Paris-Sud University of Michigan Orsay, France Ann Arbor, Michigan ACADEMIC PRESS, INC. Harcourt Brace Jovanovich, Publishers Boston SanDiego New York Berkeley London Sydney Tokyo TorontoCopyright © 1990 by Academic Press, Inc. Allrights reserved. No part of this publication may be reproduced or transmitted in any form or by any means, electronic ‘or mechanical, including photocopy, recording, or ‘any information storage and retrieval system, without permission in writing from the publisher. ACADEMIC PRESS, INC. 1250 Sixth Avenue, San Diego, CA 92101 United Kingdom Edition published by ACADEMIC PRESS LIMITED 24-28 Oval Road, London NW17DX. Ubrary of Congress Cataleging-in- Publication Data Automorphic forms, Shimura varieties, and L-functions: proceedings of a conference held at the University of Michigan, Ann Arbor, July 6-16, 1988/edited by Laurent Clozel, James S. Milne. Pp. cm, — (Perspectives in mathematics: vol. 10-11) Includes bibliographical references. ISBN 0-12-176651-9 (x. 1: alk. paper). — ISBN 0-12-176652-1 (\. 2: alk. paper) 1. Automorphic forms — Congresses. 2. Varieties (Universal algebra) — Congresses. 3. L-functions — Congresses. I. Clozel, Laurent. II. Milne, 1S. IIL Series. S1L3°3 — de20 89-18212 cip Printed in the United States of America a9 909192 987654321FOREWORD In 1977, the AMS Summer Research Institute was devoted to “Au- tomorphic Forms, Representations, and L-Functions”. One of its cen- tral topics was the relation between automorphic forms (in their mod- ern guise as automorphic forms on adéle groups) and various objects arising from algebraic geometry, most notably the Hasse-Weil zeta functions of varieties, Galois representations, and Grothendieck’s mo- tives. These conjectural relations had been explored by Shimura and others, but Langlands had formulated a systematic program to study them for Shimura varieties. At the time of the conference, Deligne and Langlands stated several fundamental conjectures concerning Shimura varieties, Galois representations, and L-functions. The decade following the conference saw substantial progress on many of these problems, and the conference was organized in Ann Arbor in 1988 to review this progress and to explore new avenues of research and new questions. In the theory of automorphic forms, advances have been made in the study of the Arthur-Selberg trace formula, the analytic properties of automorphic L-functions including in some cases their analytic continuation, Langlands’s functoriality principle including its proof in some important instances, the struc- ture and properties of the discrete spectrum for classical groups, and the p-adic interpolation of certain L-functions. Moreover the baffling problems raised by L-indistinguishability are better understood. As regards Shimura varieties, the basic conjecture of Shimura and Deligne about their canonical models has been proved in the strenthened form conjectured by Langlands at the Corvallis conference, and a combi- natorial conjecture of Langlands, allowing one to express their local zeta functions in automorphic terms, proved in some cases. Certain questions that arise in the comparison of the ¢-adic representations associated with Shimura varieties and automorphic forms (Zucker’s conjecture, formulas for the traces of Hecke operators in L?- cohomol- ogy spaces) have been solved. Important arithmetic consequences of the theory of automorphic forms and the functoriality principle have Vvvi FOREWORD been obtained or scem more accessible, for example, the construction of the Galois representations associated with Maass forms and proofs of the Tate conjecture for certain arithmetic varieties. There has also been progress in the study of the local zeta function of a Shimura variety at a prime of bad reduction. Finally, starting with Drinfeld, the analogues of these problems have been studied for function fields. ‘The articles in these Proceedings, which are expansions of the lec- tures given at the conference, are intended to reflect these advances. They are divided, in a somewhat arbitrary manner, between two vol- umes, The first volume contains expository articles on the trace for- mula (Labesse) and on the progress since Corvallis in understanding the analytic properties of L-functions (Shahidi). The articles of Milne and Clozel develop two different aspects of Langlands’s paper at Cor- vallis: while Milne’s article explains results on Langlands’s conjecture on conjugates of Shimura varieties and how they should extend to holomorphic automorphic forms and mixed Shimura varicties, that of Clozel takes up the more speculative question of defining a category of automorphic representations that has the structure of a Tannakian category. The article of Laumon is concerned with finding a geometric interpretation for certain Eisenstein series in the function field case. The papers of Arthur and Kottwitz concern, inter alia, the conj tural Hecke-Galois relations for Shimura varieties in the most general case; Kottwitz’s paper also includes a conjectural description of the number of points on a Shimura variety over a finite field. The second volume contains papers on Galois representations asso- ciated with automorphic forms (Blasius, Carayol, Taylor); bad reduc- tion of Shimura varieties (Rapoport); higher L-functions (Jacquet- Shalika); coherent cchomology and automorphic forms (Harris); a Lefschetz trace formula, conjectured by Deligne, of importance for zeta functions of Shimura varicties (Zink); the conjectures of Tate and Beilinson in the context of Shimura varieties and a review of the progress that has been made on them and related questions (Ramakr- ishnan); the p-adic L-functions associated with Shimura convolutions (Hida); and finally, the proof of the Zucker conjecture (Zucker). The conference was supported by generous grants from the National Science Foundation through the Presidential Young Investigator and Special Projects programs, We are indebted to the Mathematics De- partment of the University of Michigan, and especially Don Lewis, forFOREWORD vii Administration for pro- its assistance, and to the School of Busine: viding us with an air-conditioned auditorium during one of the hottest spells of the century in Ann Arbor. The manuscripts not submitted in ‘TEX were TpX-ed by Steve Tinney and Chris Weider, and Steve Tin- ney had the difficult task of preparing a uniform manuscript for the publisher from files in submitted in every known dialect of TgX. Fi- nally, the nonmathematical organization of the conference would not have been possible without the exceptional efforts of Lee Zukowski.CONTRIBUTORS Numbers in parentheses refer to the pages on which the authors’ contribu- tions begin. J. Arthur (1-1), Mathematics Department, University of Toronto, Toronto MSS IAI, Ontario, Canada D. Blasius (1-11), Mathematics Department, University of California, Los Angeles, California 90024 H, Carayol (15-II), Departement de Mathématiques, Université de Stras- bourg, F67084 Strasbourg, France L. Clozel (77-1), Departement de Mathématiques, Bat 425, Université de Paris-Sud, 91405 Orsay CEDEX, France M. Harris (41-II), Mathematics Department, Brandeis University, P.O. Box 9110, Waltham, Massachusetts 02254-9110 H, Hida (93-1), Mathematics Department, University of California, Los Angeles, California 90024 H. Jacquet (143-II), Mathematics Department, Columbia University, New York, New York 10027 R, Kottwitz (161-1), Mathematics Department, University of Chicago, 5734 University Avenue, Chicago, Illinois 60637 J.-P. Labesse (211-1), Centre de Mathématiques, Ecole Normale Supérieure, 75005 Paris, France G. Laumon (227-1), Departement de Mathématiques, Bat 425, Université de Paris-Sud, 91405 Orsay, Francex CONTRIBUTORS J.S. Milne (283-1), Mathematics Department, University of Michigan, Ann Arbor, Michigan 48109 D. Ramakrishnan (227-II), Mathematics Department, California Institute of Technology, Pasadena, California 91125 M. Rapoport (253-II), Mathematisches Institut, Wegelerstrasse 10, 5300 Bonn 1, West Germany F, Shahidi (415-1), Mathematics Department, Purdue University, West Lafayette, Indiana 47907 J. Shalika (143-IIl), Mathematics Department, Johns Hopkins Unversity, Baltimore, Maryland 21218 R, Taylor (323-II), Mathematics Department, Cambridge University, Cambridge, United Kingdom T. Zink (337-II), Mathematisches Institut, Wegelerstrasse 10, 5300 Bonn 1, West Germany S. Zucker (377-ID, Mathematics Department, Johns Hopkins University, Baltimore, Maryland 21218CONTENTS OF VOLUME I Unipotent Automorphic Representations: Global Motivation By JAMES ARTHUR Motifs et Formes Automorphes: Fonctorialité 5 By LAURENT CLOZEL Shimura Varieties and \-adic Representations By Rosert E. Korrwirz The Present State of the Trace Formula By J.-P. LaBEssE Faisceaux Automorphes Liés aux Séries d’Eisenstein . By G. LaumMon Canonical Models of (Mixed) Shimura Varieties and Automor- phic Vector Bundles . By J. S. MILNE Automorphic L-Functions: A Survey By F. SHAHIDI Applications du Principe de i .77 161 211 227 283 415CONTENTS OF VOLUME II Automorphic Forms and Galois Representations: Some Exam- ples... : By Don Brastus Non-abelian Lubin-Tate Theory . By H. CaraycL Automorphic Forms and the Cohomology of Vector Bundles on Shimura Varieties . a By MICHAEL FARRIS p-adic L-functions for base change lifts of GL2 to Gls . By Haruzo HIDA Exterior Square L-functions By Hervé JACQUET AND JosEPH ‘SHALIKA. Problems Arising from the Tate and Beilinson Conjectures in the context of Shimura Varieties By DiInAKAR RAMAKRISHNAN On the Bad Reduction of Shimura Varieties By M. Raproporr Representations of Galois Groups Associated to Hilbert Mod- ular Forms . : . By RICHARD TaYLor The Lefschetz Trace Formula for an Open Algebraic Surface By THomas ZINK L?-cohomology of Shimura Varieties . By STEVEN ZUCKER xii - 41 . 93 143Unipotent Automorphic Representations: Global Motivation JAMES ARTHUR Contents §1. Introduction . 2. 2. ee ee §2. Endoscopic data. . . Le ee B §3. The discrete part of the trace forma ..........14 §4. The conjectural multiplicity formula... . 2... . . . 20 §5. The expansion of Itise(f) - 0 ee BT §6. The sign characters ey andry . 2. 87 §7. The expansion of Eaise,(f). - be 42 §8. A combinatorial formula for Weel groups... .. 2... 54 §9. Concluding remarks . . . bee OT §1. INTRODUCTION In the paper [3], we gave a conjectural description of the discrete spectrum attached to the automorphic forms on a general reductive group. The main qualitative feature of this description was a Jordan decomposition into semisimple and unipotent constituents. This is in keeping with the dual nature of conjugacy classes and characters, and in fact, with a general parallelism between geometric objects and spec- tral objects that is observed in many mathematical contexts. Such a decomposition for automorphic representations would of course be parallel to the Jordan decomposition for rational conjugacy classes. It would also be analogous to the Jordan decomposition that is an essential part of the representation theory of finite algebraic groups. ‘The decomposition should actually apply uniformly to the automor- phic representations in certain families. The families or “packets” are indexed by certain parameters which are the source of the decomposi- tion. The quantitative side of the conjectures in [3] is a formula for the multiplicity with which a representation in any packet occurs in the Supported in part by NSERC Operating Grant A3483. ‘Automorphie Forms, Shimura Copyright © 1990 by Academic Press, Inc. Varieties, and L-Functions All rights of reproduction in any form reserved. ISBN 0-12-176651-92 JAMES ARTHUR discrete spectrum. It is a generalization of the formula for tempered representations which is implicit in the examples in [15]. In terms of the Jordan decomposition, tempered automorphic representations are semisimple. The multiplicity formula for nontempered automorphic representations contains some new signs. These are constructed out of the root numbers of certain L-functions, attached to the semisimple part of the given automorphic representation. In this paper, we shall try to give some motivation for the con- jectures. Some version of the conjectures, at least for many classical groups, ought to follow from the stable trace formula. This is cer- tainly so in the few cases where the stable trace formula has been established [15], [22]. In general, one would need to combine the theory of endoscopy with the ordinary (or twisted) trace formula to obtain a stable trace formula. There are still a number of problems to be solved, but one can guess what the final answer will be. The pur- pose of this paper is to show that it is compatible with the conjectures of [3]. For purposes of introduction, let G be a connected, simply con- nected group over a number field F. We shall be interested in the spectral side of the trace formula. The tial ingredient we shall study is a certain distribution Taise,t(F) 5 fecx(GA)), which is discrete in the parameters which describe the representations of G(A). It is given by an explicit formula (3.1), one term of which involves the trace of f on the discrete spectrum. When the stable trace formula has been established, the payoff will be an identity (1.1) Taise(f) = SG, H) SHC") 5 H in which H ranges over elliptic endoscopic groups, «(G, H) is a certain constant, and f + Sf#_.(f#) is a pullback to G of a stable distribu- tion on H(A). (Reczll that the endoscopic groups are a natural family of quasi-split groups attached to G. Recall too that a stable distri- bution is a special case of an invariant distribution, which arises as a natural consequence of the difference between rational conjugacy and geometric conjugacy. We refer the reader to [3, §3] for a brief discus- sion of these notions and of the Langlands-Shelstad transfer mapping fil) Asa distribution on G(A), ST#,.(f") is not generally stable.UNIPOTENT AUTOMORPHIC REPRESENTATIONS 3 However, the trace of f on the discrete spectrum is also usually not stable. Endoscopic groups were actually invented by Langlands with the aim of measuring this lack of stability. The endoscopic groups on the right hand side of (1.1) should all contribute to the multiplicity formula for representations in the di: crete spectrum. However, the trace of f on the discrete spectrum only one of several terms in the explicit formula for Iaisct(f). The other terms are the surviving remnants of Eisenstein series, and are parametrized by (conjugacy classes of) proper Levi subgroups of G. Each such term is a linear combination of distributions, which are ob- tained by taking the trace of a product of two operators, one being the action of f on the induced discrete spectrum, and the other being an intertwining operator that comes from Eisenstein series. These addi- tional terms have one important function. They account for that part of the discrete spectrum of a given H which under functoriality may into the continuous spectrum of G. However, the additional terms also contribute irrelevant information, which complicates the study of (1.1). The attempt to separate the extrancous information from the contribution of the discrete spectrum leads to combinatorial diffi- culties. The main point of this paper is to solve these combinatorial problems. The results are given in §5-§8. In §5 we expand Ijise(f) into a linear combination of irreducible characters. This hinges on the con- jectures of [3]. However, we have only the modest goal of showing that the conjectures are compatible with (1.1), so we are free to assume them. Each coefficient in the expansion contains a certain quotient of L-functions, which comes from the global intertwining operators. If the irreducible character is tempered, this quotient should equal the parity of the pole of the L-function at s = 1. If the irreducible charac- ter is nontempered, however, it will have a unipotent part. When the corresponding unipotent element is not even, the quotient must als be expressed in terms of the order of the L-function at the center of the critical strip. The exact relation is given by Proposition 5.1, which we prove in §6. Together with Lemma 7.1, it provides the justifica- tion for the sign characters which appear in the general multiplicity formula. In §7 we establish a parallel expansion of the right hand side of (1.1) into irreducible characters. This requires various properties from en- doscopy, some known and others which are expected to hold, which we discuss in §2 and §3. The endoscopic groups H consist of the4 JAMES ARTHUR quasi-split form of G, together with groups of smaller dimension. By reasons of induction, then, the stable distributions Sf{f,, are uniquely determined by (1.1), However, we must derive the expansion in §7 without reference to the left hand side of (1.1). The coefficients in the expansion have to be given as certain undetermined constants, which can be regarded as “stable multiplicities”, and which only later are tied precisely to the sign characters discussed above. For a parameter which contributes to the tempered discrete spectrum, the correspond- ing coefficient will be familiar from (15, §6,7] and [12, §12]. It is then just equal to 1, divided by the order of a certain finite group. Our aim is to show that with the assumption of the conjectures of [3], the left and right hand sides of (1.1) are equal. We would thus like to establish a term by term identification of the two parallel ex- pansions. However, this is not immediately obvious. What remains to be proved at the end of §7 is a sort of analogue for Weyl groups of the endoscopy identity (1.1). The expansion of Iaisc,t(f) contains certain constants i(«), which are defined if x is any connected component of a complex reductive group. The expansion for the right hand side of (1.1) is identical, except that é(x) is replaced by another constant e(z). In the first case, i(z) is given by a finite sum over elements in the Weyl set of 2. It is the analogue for Weyl groups of the left hand side of (1.1). The second constant e(z) is the analogue of the right hand side of (1.1), and is given as a finite sum over the isolated conjugacy classes in a. In §8 we prove that i(2) equals e(«) for every component «. This establishes the term by term identification of the expansions of each side of (1.1). ‘At the end of §8 the reader might be wondering whether the paper has provided the global motivation claimed in the title. It is true that the identity (1.1) is weaker than the conjectural multiplicity formula (and the local conjectures on which it is based). However, the identity can still provide significant information about the discrete spectrum, for either G or its endoscopic groups. This is especially so if for one of the groups, the conjectures are known to hold. The group GL(n) is such an example, thanks to recent work of Moeglin and Waldspurger [21]. The twisted version of (1.1), applied to GL(n), will relate the discrete spectrum of many classical groups to that of GL(n). In particular, it should yield some version of the multiplicity formula for the quasi-split orthogonal and symplectic groups. We shall finish the paper in §9 with an informal discussion of these questions. Throughout the paper we shall adopt the following notational con-UNIPOTENT AUTOMORPHIC REPRESENTATIONS 5 ventions. Suppose that 5 is a set on which a group T’ acts. We shall denote the set of orbits of P on ¥ by either Orb(I,S) or E/P. In gencral, if A and B are subsets of a group I’, we shall write Cent(4,B) = {b€B: b'ab=a, for alla € A} for the pointwise centralizer of A in B, and Norm(A,B) = {b€ B: b-'Ab= A} for the normalizer of A in B. Next, suppose that C is a finite union of connected components in a (nonconnected) algebraic group. Then C+ denotes the algebraic group generated by C, and C° js the connected component of 1 in C+. If s is any element in C, we set Cy = Cent(s,C°) . Then C; is also an algebraic group, with identity component C2 = (C.)° = Cent(s,C%)® . (This differs from the notation of [2] and some other papers, in which the symbol C, was reserved for the identity component of the central- izer.) We shall also write ZC) = Cent(C,C°) . This group is the intersection of C° with the center of C+, and is contained in Z(C°). Finally, if X is any topological space, 1o(X) denotes the set of connected components of X. §2. ENDoscopic DATA Suppose that G is a connected component of a reductive algebraic group over a number field F. Then G+ stands for the group generated by G, and G° is the connected component of 1 in G+. We shall assume that G(F) is not empty. As in [3, §6], we shall also assume that G is an inner twist of a component in a quasi-split group. More precisely, we assume that there is a map ni @ 3G",6 JAMES ARTHUR where G* is a component such that (G*)° is quasi-split, and such that G*(F) contains an element which preserves some F-splitting of (G*)° under conjugation. It is required that 7 extend to an isomorphism of G* with (G*)* such that for any o € Gal(F'/F), the map no(n7!): Gt + G* is an inner automorphism by an element in (G*)°. The standard situation is when Gt = G°. By allowing G to be a more general component, we are providing for applications of the twisted trace formula [5]. Associated to the connected component G® we have the L-group Go = GaxWr. It is a semidirect product of a complex connected group G° with the Weil group Wr of F. (As in [3], we follow the notation of Kottwitz [12], so that G® stands for the identity coset of the L-group. The symbol “G° can then be reserved for the full L-group of G°.) We have not assumed that G+ is a semidirect product of G° with a finite cyclic group, but this does not seem to be a serious concern. In particular, it is reasonable to define the L-group “G+ of G+ simply as a semidirect product of §G® by the cyclic group (Gt) of connected components in G+. The action of 7o(G+) on G° is dual to its action by outer autmorphisms on G°. The action of 7(G*) on Wr could be defined by some map of m9(G*) into H'! (Wp, Z(G°)). However, for simplicity we shall assume that (G+) and Wp (as subgroups of +G+) commute. Associated to the component G we have an “L-coset” 'G = GxWr, in which G is a coset of G in a group Gt such that 1gt = Gt»Wrp. Notice that . . Z(G) = Cent(G,G°) is in general a proper subgroup of the center Z(G) = Cent(G°,G°)UNIPOTENT AUTOMORPHIC REPRESENTATIONS 7 of G°. We must always be careful to distinguish between these two groups. The Galois group [ =r of F over F acts on both Z(G) and Z(G°). The subgroups of [-invariant elements are given by Z(G)? = Cent(4G,G°) and f . Z(G)? = Cent(4G°,G) . These too are not generally equal. Observe that ayry0 Ag = (2(6)") is the maximal T-invariant torus in the center of G+. It is of course not the dual group of the maximal split torus Ag in the center of Gt. It is associated, rather, to the dual of the real vector space aq = Hom(X*(G)r,R) - (X*(G)p denotes the module of F-rational characters on G+.) More precisely, X"(G)r = X,(Ag) , so that the complex dual space af; ¢ = X*(G)r@C is the Lie algebra of Ag. We shall write ka = (Ago) = 2(G)" 0 (Z(G)? for the group of fixed points of G in Ag. It is a closed subgroup of Age whose identity component equals A,. The theory of endoscopy for nonconnected groups is the subject of work in progress by Kottwitz and Shelstad. As in [3, §6], we shall guess at the ultimate form of some of this theory by extrapolating from the connected case. Thus, an endoscopic datum (H,H,s,€) should consist of a connected quasi-split group H over F, an extension I— HGH = Wp 1, a semisimple coset s in G/Z(G°), and an L-embedding € of 1 into *G®. The definition is similar to the one given in [3, §3,86] except8 JAMES ARTHUR that s is now a coset of Z(G°) instead of a single element in G. It is required that &(H) = Cent(s,G°)° , the connected centralizer in G° of any element in the coset s, and that (2.1) s€(h)s~? = a(wy)E(h) , hen, where w, is the image of h in Wp and a(-) represents a locally trivial clement in H'(Wp, Z(G°)). In other words, a(-) belongs to ker' (Wp, Z(G°)), the kernel of the map H (Wp, Z(G°)) — Q@H' (Wr, 2(G)) , in which v runs over the valuations of F. It is further required that the two extensions 1 and “H define the same map of Wp into Out(#), the group of outer automorphisms of H. Recall that an endoscopic datum is said to be elliptic if the sct €(H)s is not contained in any proper parabolic subset of 'G. Equivalently, the datum is elliptic if and only if the group &(2H)")/(Z(H)") 9 Z(G)" is finite, or again, if and only if €(Aq) equals Ag. Finally, two elliptic endoscopic data (H,H,s,€) and (H’, 1’, s’,£’) are equivalent if there exist dual isomorphisms a: H — H’ and 3: H' —+H, together with an element g € G° such that 96(B(h'))g~* = Eh’) , Wen, and Suppose that (H,H, s,€) is an elliptic endoscopic datum. We shall write Aut(H) for the group of elements g in G° such that gsg7} = s, and gé(H)g7! = €(H). Then Aut(H) is a reductive subgroup of G°. Notice that €(H)Z(G°)! is a closed subgroup of Aut(H). We shall need to know later that it is of finite index. Equivalently, we must establishUNIPOTENT AUTOMORPHIC REPRESENTATIONS 9 LEMMA 2.1. The identity component of Aut(H) equals - en) &A)(Z(G))" = HA Go) « Let s; be a fixed element in the coset s, and write Cy = {9 EG": 1957197! € Z(G)} = {gE G?: gsg7 = s} and . Cy = {9 €G: sigss'g7! = 1}. Then g — sigsy'g7! is an injective map from C,/C, onto a closed subgroup 2(s) of Z(G). LEMMA 2.2. The subgroup 2s) = {syzsp!27!: 2 € 2(G%)} is of finite index in Z(s). PRooF: Suppose that g belongs to C,. We can write g 1 belongs to the derived subgroup G9.,,, of G® and z belon, Then 19-1 ssaqisvtg-! 1 $195) G9 = $1915; 9, * $125) In particular, both g;, and z belong to Cy. But the element gi" lies in G9,,. The lemma follows from the fact that G., has finite center. Qa PROOF OF LEMMA 2.1: According to the first condition in its defi- nition, Aut(H) is contained in C,. Let Aut'(H) be the subgroup of clements g € Aut(H) such that s:gsy'g-! belongs to Z'(s). ‘The last Jemma tells us that Aut'(H) is of finite index in Aut(H). Let g be an element in Aut’(H). Then we can write g= m4, MmEC., 1 € Z(G). Suppose also that h is an element in €(H). The second condition in the definition of Aut(H) implies that ghg~! equals hy'h, for some element hy € €(H). We can write this as hyhta = (hgh) hig: .10 JAMES ARTHUR Both hy and g; commute with 1. It follows easily from (2.1) that hgh} also commutes with s;. Therefore hzyh~!z,* commutes with 81, and belongs to the subgroup Z(G) of Z(G°). Now hzyh72z" = o(21) where @ is the projection of h onto T = Gal(F/F). The action of T on Z(G) factors through a finite quotient Gal(E/F), and this action preserves the subgroup Z(G). We obtain a homomorphism g — o(a)zz? from Aut’ (H) to the finite group H' (Gal(E/F), Z(G)). Suppose that g lies in the kernel of this map. Then a(a)ez' = ofz)e*, oer, for some element z € Z(G). In other words, there is a decomposition zy = 224, for elements z in Z(G) and 24 in Z(G°)". We can therefore write g = 9/2, where the element g{ = giz lies in the centralizer Cs. In other words, g belongs to the subgroup C,2Z(G°)'. It remains only to observe that £(H) is the identity component of C,. We obtain an embedded chain &(H)Z(G° c C,Z(G°)" Cc Aut'(H) Cc Aut(H) of normal subgroups of finite index. Therefore €(H)Z(G°)! is of fi- nite index in Aut(H), and the two groups have the same identity component. Oo Let (H,H, 8, €) be a fixed endoscopic datum. One is interested in the L-homomorphisms of Wp into “G whose image is contained in &(H). (Recall that an L-homomorphism between two extensions of Wp is a homomorphism which commutes with the projection onto Wr.) One might like to be able to identify such objects with L- homomorphisms of Wr into the L-group “H of H. However, this is not always possible. The L-group is a semidirect product, HxWe relative to an L-action of Wr on H [20, 1.4]. (The action of Wr of course factors through the quotient I of Wr.) But the two extensions Hand “H of Wr by Hf need not be isomorphic. In other words, thereUNIPOTENT AUTOMORPHIC REPRESENTATIONS 11 might not be an L-embedding of “H into “G which co-incides with the image of €. Fortunately the problem is not serious. In the case that G = G®, the question can be resolved by taking a z-extension of G, as has been explained in [20, (4.4)]. In the general case, Shelstad has pointed out that it is necessary to work directly with extensions of the endoscopic groups H. Suppose, then, that (2.2) 1+ Ao HO 1 is a central extension of quasi-split groups over F. We shall review the question of whether there exists an L-embedding &:H 3 ley which extends the canonical embedding H © Hy of dual groups. Consider first the kernel Kp of the projection Wr > I'r, a con- nected group. It would be no trouble to construct an embedding for the preimage H! of Kr in 1, For it follows easily from (2.1) that &(H’) equals the subgroup €(H) x Kr of “G°. In other words, there is a splitting 9: Kr > H’ such that (2.3) hO(k)A7? = O(wpkw;*) , heH, ke Kp, where wp is the image of h in Wr. Now by assumption, the map of Wr into Out(H) defined by 1 is the same as the L-action h — w(h), he, weWr, used to define “H. It follows that @ can be extended to a section from Wr to H such that O(w)hO(w)-! = w(h) , heH, weWr. Keep in mind that it is only the restriction of @ to Kp which is a ho- momorphism. However, 6 is uniquely determined up to multiplication by elements in the center Z(H) of H. Therefore O(w1)O(w2) = b(w1, w2)O(w1, we) 5 w1,w2 Wr, where }(wy, wa) is a 2-cocycle from Wr to Z(H). By (2.3), b(wi, we) depends only on the images of w; and w2 in T'p. We shall write 312 JAMES ARTHUR for the image of b in H?(Wp, Z(H;)), relative to the embedding of Z(H) into Z(H;). Then @ is the inflation of a class in H?(I'p, Z(£1)) which is independent of 6. Suppose that is trivial. That is, B(wy, we) = 2(wy)w;(2(we))2(wiwe)! , wi, v2 € We, for a function z : We > Z(H) which is uniquely determined up to a l-cocycle. Every element in 1 can be represented uniquely in the form hO(w) , he, weWr, and the map (2.4) :(h0(w)) = hz(w) x w is then an L-embedding of H into /H,. Conversely, if an embedding &, exists, the function 2(w) in (2.4) will split the class . Assume that the embedding €, exists. Suppose also that the central subgroup Z, of H; is connected. Then we can form the L-group “2, = Z, x Wp, and there is a canonical projection “Hy > “Z,. We also have an exact sequence 1 — 2) — 2h) 4% 1 of complex abelian groups. Let 21(w) be the projection of 2(w) onto 2. Then 2 is a L-cocycle from Wr to Z;. In fact, if we agree not to distinguish between a cocycle and its corresponding cohomology class, 21 is just the preimage of the class b € H?(W, Z(H)) determined by the long exact sequence . 2 H'(We,Z(fh)) + H (Wr, 21) = H?(Wp,Z(H)) + H?(We, 2h) > It is uniquely determined modulo the image of H'(Wp,Z(i)) in H'(Wp,Z,). The map ax(w) = 2(w) xw, weWr, is an L-homomorphism of Wr to “Zi.UNIPOTENT AUTOMORPHIC REPRESENTATIONS 13 Suppose that L~ + Wr is some extension of Wp. Suppose also that % : Le + "Gis an L-homomorphism whose image is contained in &(H). That is, = oy, for some L-homomorphism qq : Lp > H. Then vy = € 0 Wy is an L-homomorphism of Lp into 4H. Set vault) = 7(t)A(w) , teLr, where w, is the image of t in Wp and 7(t) belongs to H. Then &i(vu() = o(t)z(wr) x we , tele. It follows that the composition of y; with the projection “H, > 4Z, equals a (or rather, the pullback of a; to Lr.) Conversely, any L- homomorphism 1 : Le — “Hy whose projection to ¥Z; equals ay is easily seen to be of the form & 0 #7. We can summarize these remarks in a commutative diagram Ly ay wy bn a ZA — te & HH 4 lg In conclusion, we want to associate pairs (Hy, £1) to endoscopic data, where Hj is a central extension (2.2) and & is an L-embedding (2.4). We shall call such a pair a splitting for the endoscopic datum. We shall say that (Hy,€,) is a distinguished splitting if, in addition, the map H,(A) > H(A) between adéle groups is surjective, and the central subgroup Z, is an induced torus. That is, Z; is a product of tori of the form Resg~(Gm). In particular, Z; is connected, as we assumed in the discussion above. Any endoscopic datum has a distinguished split- ting. For example, the cocycle b(w1,w2) € Z(H) that we described above often splits. In this case, we can simply take (H,€,) = (H, Id). In general, we can always take H; to be a 2-extension of H [11, §1], the existence of which is established in [17, pp. 721-722]. The first condition follows from [11, Lemma 1.1(3)], while the second is part of the definition of a z-extension. It is also part of the definition that the derived group of Hj is simply connected. This in turn implies14 JAMES ARTHUR that Z(H) is a complex torus. It follows from [17, Lemma 4] that the class 8 € H?(Wp, Z(H;)) is trivial. The embedding £1 therefore exists, and (Hy,€:) becomes a distinguished splitting. In general, if (H,€:) is any distinguished splitting, one needs to know that the canonical map ker! (F,Z(H)) —> ker! (F,Z(1)) is an isomorphism. (As before, ker! (F,Z(H)) denotes the kernel of the map HUF, Z(A)) S Qe (F,2(A).) This follows from the proof of [12, Lemma 4.3.2(a)]. We will also use the injectivity of the map H(F,2Z(H)) — H'(F,2(ih)) » which is a consequence of the long exact sequence of cohomology, and the fact the group H°(F, Z:) = mo(Z?) is trivial. §3. THE DISCRETE PART OF THE TRACE FORMULA We are going to study a piece of the trace formula. It consists of those distributions on the spectral side of the trace formula which are discrete with respect to the natural measure on the relevant auto- morphic representations. This part of the formula contains the actual trace on the discrete spectrum. It is thus the payload, the part which will eventually be used to compare automorphic representations on different groups. Of course, there are serious problems relating to the other terms in the trace formula which will have to be overcome first. Our intention in this paper is simply to see what can be learned once these other problems have been solved. Let A be the adéle ring of F. We should first identify our space of test functions on G(A), the set of A-valued points in G. Consider the diagonalizable group Z(G) = Cent(G,G°). We shall fix a closed sub- group X of the group Z(G, A) of adéle points such that X 9 Z(G, F) is closed, and such that X Z(G, F)\Z(G,A) is compact. Let x be a character on X which is trivial on XN Z(G, F). Then Cx (GA), x) will denote the space of smooth functions f on G(A), of compact support modulo X, such that S(zx) = x(z)"F(2) , z€x, 2€G(A).UNIPOTENT AUTOMORPHIC REPRESENTATIONS 15 Let t be an arbitrary but fixed nonnegative real number. The cor- responding discrete part of the trace formula is the distribution Taiset(f) » fe C2(G(A), x), on C2°(G(A), x) which is given by the expression 1) SY olay {M} wEWS (ayy )reg [det(w — tag *tr(M(w,0)pp(0, f)) - (See [2, §4], [4, §II.9].) We shall describe very briefly the terms in this expression. The outer sum is over the finite set of G°(F)-orbits of Levi components M of F-rational parabolic subgroups P of G°. The inner sum is over the regular elements Way res = {w €WS(ay): det(w—1)ag # OF in the Weyl set Way) = Norm(Ay,G)/M of (G,M). As in earlier papers, we regard the Weyl elements as operators on the real vector space ay = Hom(X(M)r,R) which leave invariant the kernel a§, of the projection of ayy onto ag. For each M there is canonical isomorphism from Amo = Ama(R)° , Mg = Resrjq(M), onto ay. If A§,,. denotes the preimage of af in Ay,co, we can extend y uniquely to a character x4¢ on Xy = AGj,..X which is trivial on AG. Let Lise¢(M(F)\M(A), x57) be the subspace of L?(M(F)\M(A), x37) which decomposes under M(A) as a direct sum of irreducible representions whose Archimedean infinitesimal charac- ter has norm ¢. Then pps(0): f > ppa(0,f) = [ F(x) pp,(0, x)dx X\GA)16 JAMES ARTHUR stands for the corresponding representation induced from P(A) to the group G(A)+ generated by G(A). It acts on a Hilbert space Hp, of Xf -equivariant functions on G(A)*. Finally, M(w,0): Hp, — Hey, we WS(an)regs is the global intertwining operator which comes from the theory of Eisenstein series. For a given conductor, Igisct(f) is a finite linear combination of irreducible characters on G(A)*. There are some minor discrepancies between (3.1) and the origi- nal definition [2, (4.3)]. In (3.1) we have summed over the orbits {M} instead of all Levi components which contain a given minimal one. This is why |W(ayy)|-! appears instead of the normalizing constant [W!'||Wé|-! from [2]. The operator pp,(0, f) here comes from a representation of G(A)* induced from a subgroup of the con- nected component G°(A). It is a direct sum of |to(G*)| copies of the corresponding operator from [2], which comes essentially from the induced representation of G®(A). Hence the constant |ro(G*)|~" in (3.1). The difference between taking a y-equivariant function on G(A), as we have done here, and a function defined on the subset G(A)! of G(A), as in [2], is purely formal. In [2], there was also the additional assumption that f was K-finite, but this was only for dealing with other terms in the trace formula. The program for comparing trace formulas on different groups, as it is presently conceived, falls into the general framework of stabilizing the trace formula. The basic references for this problem are [18], [12], and [13]. The problem was solved completely for G = SL(2) in [15]. A general solution would include: a transfer map from functions for G to functions for endoscopic data, a stable distribution analogous to Iise,t for any quasi-split group, and an identity relating Taise,t to the corresponding stable distributions for endoscopic data. We shall discuss the transfer first, and then describe the expected properties of the other objects in the form of a hypothesis. Suppose that (H,H,s,€) is an elliptic endoscopic datum for G. ‘Assume also that we have fixed a distinguished splitting (H1,1) for the endoscopic datum. As we recall from §1, 1 determines an L- homomorphism a; : Wp > 'Z;. Let Gi ZA(F)\Z(A) — Cc be the character associated to a1 by the Langlands correspondence for tori. Now, in the special case that G = G°, the results [20]UNIPOTENT AUTOMORPHIC REP} TATIONS 17 of Langlands and Shelstad imply the existence of a canonical map f — f" from functions f € C°(G(A) to functions f"(yz,) on suitable stable conjugacy classes in H(A), with the property that F(a) = (a) Fn) s 21 € Z(A). (See also [12], [13] and [3].) The map must be constructed as a tensor product of the local maps fy + fi", fy € C2(G(F,)), which are defined explicitly in [20]. Langlands and Shelstad expect that fis the set of stable orbital integrals on H(A) of a function g in C(H,(A),G1). We shall assume that this is so. In fact, we shall assume that the transfer map fs", LeEce (GA), has been defined, and has this property, for general G. We should actually modify the transfer mapping so that its domain is the space C2°(G(A, x) considered earlier. Lemma 4.4A of [20] suggests how the functions f(t) = flex), 2 € 2(G,A), rE GA), § E CR(G(A)), should behave under the transfer map. In general, there will be a norm mapping z — 2’ from Z(G, F)\Z(G,A) into Z(H, F)\Z(H, A). We also have the exact sequence 1— 2 — 2M) — Z(H) 1. We can then expect a formula (3.2) (f2) Om) = Galt (arm), where (; is an extension to Z(H:,F)\Z(H1,A) of the character on 2Z,(F)\Z,(A), and 2; is any point in Z(H), A) whose image in Z(H, A) equals 2’, Recall that y is a character on the closed subgroup X of Z(G,A). We shall assume that x) = x’), zeX, where y’ is a character on the image X’ of X in Z(H,A). To define the transfer mapping for functions in C2°(G(A), x), we simply multiply18 JAMES ARTHUR each side of (3.2) by x(z), and integrate over z in X 0 Z(G, F)\X. Let X; be the preimage of X’ in Z(H;,A), and set xl) = G(a)x'(2) 5 for any point 2; € X; with image 2’ in X’. Then x; is a character on X;, and the triple (H;, X1, x1) satisfies the conditions we imposed on (G,X, x). In this context our assumption is that for any function f € C(G(A), x) there is a function g € C2 (H1(A), x1) whose stable orbital integrals are given by f'. The function g is of course not uniquely determined by f. However, if ST is any stable distribution on C2°(Hi(A), x1), SI(g) will be uniquely determined by f. We shall therefore write sig) = SI(9). ‘The ultimate goal is to give an expansion of Igigc,t as a linear com- bination of stable distributions on the equivalence classes of elliptic endoscopic data {H} for G. The coefficients will be certain constants u(G,H), which in the case G = G° were introduced by Langlands [18]. (Following the usual convention of metonymy, we shall often write H in place of a full endoscopic datum (H,H,s,€).) Kottwitz has established a simple formula for these constants [12, Theorem 8.3.1], again when G = G®. Let n(G) = 1(6%)r(@)" be the relative Tamagawa number of G° [12, §5]. (r(G°) denotes the ordinary Tamagawa number of G®, and G®. is the simply connected cover of the derived group of G®. Thus according to Weil’s conjecture, which has been established by Kottwitz [14] for groups without Es factors, 71(G®) simply equals 7(G°).) Kottwitz’ formula is then u(G,H) = 71(G°)n(H)""|n0(Aut(H))| - In the general case, the constants have not yet been defined. We shall have to get by with a makeshift definition that reduces to Kottwitz’ formula when G = G°. If we are given an equivalence class {H} of elliptic endoscopic data, we shall usually assume implicitly that H is a representative of the class such that € is the identity. That is, 1 is an embedded subgroup of “G°. Then Z(H)" is a subgroup of G° whose identity componentUNIPOTENT AUTOMORPHIC REPRESENTATIONS. 19 = (Ago) of G® also has Ag as its identity component, so that kg Z(H)! is a subgroup of finite index in Kg. For general G we shall simply define (3.3) (G, H) = 71(G°)n(H)7} | (Aut(H)) |" |ko/no A Z(H)" |". The fourth factor in the product on the right, which of course equals 1 when G = G®, is suggested by the calculations in §7. We can now state the hypothesis. Part of it applies to any (G, X, x) as above, and part applies to triples (G1, X1,x1) with the restriction that G; is a connected quasi-split group over F. equals Ag. The subgroup KG Hyporuesis 3.1. For any (G;,X1,x1) there is a stable distribution TG. on C2 (Gi (A), x1) with the property that for any (G, X, x), the distribution (3-4) Bascal f) = TG HVS (F™) equals Igisea(f). Here f stands for any function in C2(G(A),x) and His summed over the equivalence classes of elliptic endoscopic data for G. Oo Remarks. 1. It is understood that we have fixed a distinguished splitting (H1,€:) for each H. The distribution Si{t}, ,(f") should then depend only on H and not on the splitting. 2. The stable distributions SI{?, , are uniquely determined by the condition that Euisc,(f) equals Igisc,t(f). For suppose that they have been defined inductively for any group whose semisimple part has dimension less than that of G;. Setting G = G1, one simply defines STRAP) = Meise f)— YO AG, W)ST (I) - HAG, In this case, the hypothesis becomes the assertion that the right hand side is a stable distribution in f. This of course is highly nontrivial. It is likely to be resolved only by proving a similar assertion for all the other terms in the trace formula. There is a discussion of this question in the paper [19]. We shall need a slightly different formula for (G,H) in §7. For H as above, set ZA = Z(H) 2(6°)/Z(G°) = ZA) /Z(A nzG) . Since H represents an elliptic endoscopic datum, Z(H)! is a finite (abelian) group.20 JAMES ARTHUR LEMMA 3.2. The constant «(G,H) equals (3.5) |ker! (F, Z(6°))|7*|to(xa)|* |ker! (F, Z(A))|[Z(A)" |"! Aut(H)/HZ(G°)"|7*. ProoF: The main point is the formula n(G°) = |70(Z(G°)") Iker" (F, 2(G°))|* of Sansuc and Kottwitz for the relative Tamagawa number [12, (5.1.1)]. From this, it will be a routine matter to derive the ex- pression (3.5) from (3.3). For Lemma 2.1 tells us that |ro(Aut(H))|"! = |Aut(H)/HZ(G°)"-"HZ(G°)" Agel - Keeping in mind that Ag is the identity component of Z(G)", we deduce that |1Z(G°)" /H Age |" = 126°) /2(G)F NH Age) I = |ro(Z(G°)")/-1|(Z(G°)? 9 (A Age)) /Aeol + Moreover, 12(G°)P n (HAge)) Ago = | Z(G)"/H Agel = (ZB) 12(G)"/Z(H) Ne = |ro(Z(A)" 9. 2(G)") [bro Z(H)! Oa) = |ro( ZB) IZ" Irena)” ke/Z(H)" 1 Kc - The lemma follows from the formula above for 7: (G°) and its analogue for 71(H). §4. THE CONJECTURAL MULTIPLICITY FORMULA Our goal is to provide some motivation for the conjectures on non- tempered automorphic representations stated in [1] and [3]. The main global ingredient of the conjectures is a multiplicity formula forUNIPOTENT AUTOMORPHIC REPRESENTATIONS 21 automorphic representations in the discrete spectrum. It is a general- ization of similar formula for tempered automorphic representations which was implicit in the examples of [15] and was stated explicitly in [12]. We shall recall the various objects from [3, §8] needed to state the formula. The automorphic representations which occur in the spectral de- composition should be attached to maps (4.1) wb: Lp x SL(2,€) + 46°. such that the projection onto G° of the image Lr is bounded. Here Lr is hypothetical Langlands group, which we shall assume is an extension of the Weyl group Wr by a compact connected group. The maps themselves are subject to certain conditions. For example, should be globally relevant, in the sense that its image must not lie in a parabolic subgroup of “G° unless the corresponding parabolic subgroup of G® is defined over F. Another condition is designed to insure that y parametrizes representations of G°(A) which lift to G(A)*. Let Sy = S4(G) be the set of elements s € G such that each point su(t')s dt)! , t €Lpx SL(2,€), belongs to Z(G°), and such that the class of the 1-cocycle t —+ su(t)sty(t)" , téeLr, lies in the subgroup ker! (Lp, Z(G°)) of H' (Lp, Z(G)). The condi- tion on y is that Sy be nonempty. Recall also that two parameters , and wz are equivalent if there is an element g € G® such that (4.2) -do(t,u) = govilt,u)ga(t), —(t,u) E Lp x SL(2,€), where a(t) is a 1-cocycle of Lp in Z(G°) whose class in H! (Lp, Z(G°)) lies in ker’ (Lp, Z(G°)). Let Y(G) denote the set of equivalence classes of maps (4.1) which satisfy the required conditions [3, §8]. Let Yo(G) denote the subset of (equivalence classes of) maps 7 € U(G) such that the set 5y = 5y(G) = Sy(G)/Z(G°)22 JAMES ARTHUR is finite. In [4] we called these maps elliptic. They should parametrize automorphic representations which occur in the discrete spectrum. It will be convenient to define two other subsets of &(G). Let us say that y is weakly elliptic if the group 5y(G°) (obtained by replacing G with the identity component G®) has finite center. We shall say that w is discrete if it satisfies the weaker condition that the group St = s}/z(@°) generated by Sy has finite center. (Keep in mind that St, St, S4(G), Sy(G°), etc., are complex reductive Lie groups which are generally not connected.) Let ¥5(G) and Waise(G) denote the set of (equivalence classes of) maps € V(G) which are weakly elliptic and discrete, respectively. Then we have embeddings Wo(G) C WG) C Waise(G) C WG). _Let x be a fixed character on a subgroup X of Z(G,A) which satis- he conditions of §3. We may as well assume that X is contained in 2°(G,A), the adéle group of the identity component of Z(G). There is a canonical map from 4G° onto the L-group 4Z°(G) of Z°(G). The composition of any parameter € Y(G) with the map gives a parameter in @(Z°(G)), and therefore a dual character Gy: 2°(G,F)\Z(G,A) — Cr. We shall write 1(G, x), Vo(G, x), etc., for the set of parameters y in W(G), Vo(G), etc., such that the character Cy, coincides with y on X. Suppose that y € U(G, x). As in [3, §8], we can form the finite set = SG) = $y/9.2(6%). It is a coset of a Sy(G) = Sy(G°)/S}.2(G) in the finite group = Sy(Gt) = Sy(Gt)/S4Z(G") . Now the local conjectures in [3, §6] assert that there is a set IIy of representations attached to 7. The elements in II, should in factUNIPOTENT AUTOMORPHIC REPRESENTATIONS 23 belong to Tunit(G(A), x), the set of equivalence classes of irreducible unitary representations of G(A)* whose restrictions to G°(A) remain irreducible, and whose central character on X coincides with y. There should also be a canonical pairing , zest, relly, such that the functions « +< &,m > are characters of nonzero finite dimensional representations of Sf}. Finally, the conjectures assert the existence of stable distributions (4.3) fi fh), v1 € W(Gi,x), on C#(Gi(A),x1), for each (Gy,X,, x1) with Gy connected and quasi-split. Let us recall how the distributions (4.3) are supposed to behave with respect to endoscopic data. Suppose that s is a scmisimple clement in Sy. Take H to be the connected centralizer in G° of any point in s, and set H = Hb(Lp x SL(2,€)) . There is obviously an injection H — 1 and a surjection H = Wr. We are assuming that the kernel of the map Le + Wr is connected, and it follows that y maps both the kernel and $L(2,€) into HZ. Therefore 1+ AH Wp 1 is a short exact sequence. We can identify H, equipped with the canonical L-action of Wy induced by H, with the dual ofa well defined quasi-split group H over F. If € is the inclusion of H into +G°, then (H. 1, 5.) is an endoscopic datum for G. It has the property that equals € 0H for some L-homomorphism Uy of Le x SL(2,€) into H. Now, let (H1,1) be any distinguished splitting for the endoscopic datum. We can construct the character x1 on a closed subgroup X; of Z(H,,A) as in §2, and from our remarks in §2, we see that the parameter di = Govn belongs to (Hi, x1). According to our assumptions on the transfer map f — f¥1, the distribution f— fh), fe CR(G(A),x),24 JAMES ARTHUR makes sense. It should satisfy the formula (4.4) fn) = YO < 5057 > fol), r€{Ily} where J is the image of s in Sy, sy is the element *@(¢ 4) in $y(G®), and a(n) = tr( J(x)n(x)dz) - X\G(A) As in [3], {Ily} denotes the set of orbits in IIy under mo(G*)*, the dual of the finite component group, which acts in the obvious way on Tynit(G(A), x). Recall that the element sy was introduced in [3, $4] to describe the signs which occurred on the right hand side of (4-4). The objects we have just described, namely the packets IIy, the pairings < 2,7 >, the stable distributions (4.3), and the formula (4.4), are all consequences of the local conjectures [3, Conjectures 6.1 and 6.2]. The adélic versions described here are simply restricted tensor products of the local versions in [3]. We shall assume their existence in what follows. We should also recall the sign character ey: S$ — {£1} which occurs in the conjectural multiplicity formula. Set Lip = Lp x SL(2,€), and consider the representation Ty(s,t) = Ad(s¥(t’)) , seSt, tel, of St x L'p on the Lie algebra g of G. Let w= Pr = Pema) cs csUNIPOTENT AUTOMORPHIC REPRESENTATIONS 25, be the decomposition of ry in which Ag, fe and vg are irreducible (finite dimensional) representations of S$, Lp and SL(2,C) respec- tively. The global L-function L(s, 1.) will be defined as a product of local L-functions. We shall assume it has analytic continuation and satisfies the functional equation L(s, pe) = €(8, He) L(1 — 8, fix) , where ¢(s, 1g) is a finite product of local root numbers. It follows from the functional equation that if jz is equivalent to its contragredient ji, then e(}, x) = £1. Let us write g, for the direct sum of those irreducible constituents 7, such that (i) x © Fir, (ii) (4, me) = -1, and (iii) dim vp is even. The sign character is then given by (4.5) ey(z) = eG(z) = II det (Ac(s)) ; rest, where the product is taken over those k such that 7, is contained in g,,, and s is any element in Sy which projects onto x. In other words, (4.5’) eG(x) = det(s, Endy, (6,,)) - We could actually have replaced the first condition in the definition of g,, by the stronger assertion (i") 7 © 7. Indeed 1% is always equal to its contragredient, and det (Ae(s)) = detdg(s)7? . Therefore, the contribution to (4.5) of the distinct pairs (x, 7) equals 1. It should also be noted that: the condition (iii) above is not really necessary. For suppose that 7, sal (i/) and (ii), but that dim(1) is odd. Then v% corresponds to the principal unipotent in an odd orthogonal group. Since px is self-contragredient, its image must be contained in either the orthogonal or the symplectic group. We shall assume the generalization of the theorem of Frohlich and Queyrot [6] which, in view of the sign e(3, 1x) = —1, implies that jg is actually symplectic. Finally, since the representation 7, is self contragredient and preserves the Killing form, it must be orthogonal. For this to be so, the third representation in the tensor product must actually be symplectic. Therefore det \4(s) = 1, and 7x contributes nothing26 JAMES ARTHUR to (4.5). This explains the apparent discrepancy between the present definition (4.5) and the earlier one [3, (8.4)]. If ¢ is any vector in the Hilbert space L?(G°(F)\G°(A), x71), set (RQy)¢)(2) = 9(& ey) 5 2 € G(F)\G(A), for any points y € G(A)* and € € G*+(F) such that £-"y belongs to G°(A). This gives an extension of the regular representation to G(A)+. For any representation 7 € Monit (G(A), x), let mo(n) be the multiplicity with which m occurs as a discrete summand of R. Now, suppose that 7 belongs to a packet Ty, p € W(G, x). Then we have the nonnegative integer (4.6) my(m) = [StI SO eve) , rest given explicitly in terms of the pairing. The multiplicity formula amounts to the global component of our conjecture, and will be stated formally as a hypothesis. Hyporuesis 4.1. For any representation * € TWunit(G(A),x), we have the multiplicity formula (4.2) mo(n) = Syl) - og WEVo(G,x) Before discussing the conjectures, we shall collect a few simple ob- servations for our later use. Let be a fixed map in ¥(G). (We shall sometimes not distinguish between a map and its equivalence class.) Let Cy denote the centralizer in G® of the image of y. Then CyZ(G°) is a subgroup of Sy(G°). The quotient Cy = CyZ(G°)/Z(G") is a subgroup of 5,,(G°). Now, the image of the cocycle t 4 so(t)sy(t-?) , té Lr, s € Sy(G), in H'(Lp,Z(G°)) gives a map from $4(G°) into ker! (Lr, Z(6°)) whose kernel is easily seen to equal CyZ(G°). We therefore obtain a continuous injection WR 5y(G°)/CoZ(G) ¥ 5y(G)/Ey > ker" (Lr, Z(G").UNIPOTENT AUTOMORPHIC REPRESENTATIONS 27 According to Lemma 11.2.2 of [12], or rather its extension to the hypothetical group Lr, ker'(Ly,Z(G°)) is isomorphic to ker! (F,Z(G°)). In particular, ker'(Ly, Z(G®)) is a finite discrete group. Therefore, the connected component 5%, of Sy maps to the identity element in ker! (Lr, Z(G°)). We obtain an identity (4.8) Sy = CE of connected components. In particular, if we set Cy = CyZ(G°)/C42(G) = Cy/Ch, we can write the injection above as (4.9) Sy(G)/Cy > ker! (Lp, Z(G) - Suppose that s is a semisimple element in Sy. According to our conventions, Sy, denotes the centralizer of s in 5%, and 5%, , is the connected component of 1 in Sy,,. We can also take the centralizer Cy,» of s in C4, and its identity component C%, ,. In §7 we shall use the identities Sy,. = Cy,, and 5%, = CY,,. These of course follow immediately from (4.8). We shall also have occasion to consider some slightly different centralizers. Keeping in mind that ¢ is a coset in G/Z(G°), we write Sy,. for the centralizer in $9, of any clement in the coset s. Then 5y,02(G°)/Z(G°) is a subgroup of Sy,,, which by Lemma 2.2 is of finite index. In particular, we have an equality (4.10) 59, = $9,,2(G°)/Z(G°) of identity components. Similarly, if Cy,; denotes the centralizer in C9 of any clement in the coset s, we have (4.11) 53, = Ch, = Ch,2(G°)/Z(G°) . §5. THE EXPANSION OF Iuise,t(f) We have now stated two global hypotheses. As we have already noted, Hypothesis 3.1 should be a consequence of a stable trace for- mula. Once this is established, one could try to combine the formulas28 JAMES ARTHUR (3.1) and (3.4) to deduce something approaching the multiplicity for- mula in Hypothesis 4.1. Our aims in this paper are more modest. We shall simply show that the two hypotheses are compatible. We are actually going to establish that Hypothesis 4.1, together with the local assumptions of §3, §4 and [3, §7], implies Hypothesis 3.1. More precisely, we shall show that the formula for Iuisc(f) obtained by combining Hypothesis 4.1 with (3.1) equals the formula for Eaisc,«(f) provided by the definition (3.4). In the process we shall gain some insight into the role of the sign characters €¢. In this section we shall derive a formula for Igisc,t from Hypothesis 4.1. By combining (4.7) with (3.1) we will obtain an expansion for Taiset(f) » fe C2(G(A),x), as a linear combination of irreducible characters. In doing this we will need to apply a local conjecture from [3, §7] for the values of normalized intertwining operators. ‘According to (3.1), Taisc,t(f) equals the sum over {M} and over w € WS(ags)reg, of the product of |ro(G*)I-* [WS (ang)I det(w — 1a with (5.1) tr(M(w,0)pp.(0,f)) - Our first task is to expand (5.1) into a linear combination of irre- ducible characters. For any M, and w € W¢(ay)), we can form the component My = M.-w. It satisfies the same conditions as G. Now, recall that pp,.(0) is the representation of G(A)* obtained by parabolic induction from the action of M(A) on (5.2) Liixe «(M(E)\M(A), Xi This representation of M(A) has a canonical extension to the group M,(A)* generated by the coset M,,(A) = M(A)w. In particular, the space (5.2) can be decomposed into a direct sum of subspaces corresponding to irreducible representations 4» of M.,(A)+. There is a similar decomposition Hee = PHP(ow)UNIPOTENT AUTOMORPHIC REPRESENTATIONS 29 of the induced space into subspaces which are invariant under the operator (5.3) M(w,0)pp.2(0, f) - If the restriction of oy to M(A) is reducible, one sees easily that the trace of the operator (5.3) on Hp(oy) vanishes. Therefore, in computing the full trace (5.1), we need only consider representations Jw which belong to the space we have denoted by Tunit(Mu(A), xar)- According to Hypothesis 4.1 (applied to M,, rather than G), the multiplicity with which a representation 0, € Tunit(Mu(A), x37) 0¢- curs in (5.2) equals Ma (Fw) 5 Peo Vo Mua sx mest) where my, (dw) is the nonnegative integer defined by (4.6). We have written Wo(M,, xa1,t) to denote the set of parameters in Vo(Mw, am) whose Archimedean infinitesimal character has absolute value t. Any pair Py and oy, with my, (7w) # 0, determines a subspace of (5.2), and also a subspace of the induced space Hp. The restriction of (5.3) to this latter subspace can be expressed in terms of the operators studied in [3, §7]. It equals an expression 4) My, (Fw)"(hw) (Re (Fws bw) Zp (9, £)) s whose constituents we shall describe in a moment. The trace (5.1) becomes the sum over Jy € Yo(Mw,xar,t) and oy € Hy, of the trace of the expression (5.4). Given M and P, it is convenient to fix a dual parabolic subgroup “Pp = P »Wr in “G° with Levi component 4M = M » Wr. The choice of P and “ P determines an embedding of the L-group 4M into 4G. It also allows us to identify W(a,,) with the dual Weyl set W(ay) = Norm(Ay,,G)/M . Returning to (5.4), we note that Zp(o) stands for the induced repre- sentation of G(A)* obtained from the restriction o of oy to M(A). The operator RP(Gwsw) = @) RPO, wr)30 JAMES ARTHUR is a tensor product of local normalized intertwining operators defined in [3, (7.4)]. When this operator is evaluated at a smooth vector in Hp,, almost all the terms in the product reduce to 1. Finally, the scalar (yy) in (5.4) is obtained from an infinite product of local normalizing functions of the form [3, (7.2)]. It equals Jim (LO,Prw °Pye.r)6(0s PP 2 P40.) Ls Br © b.a)?) 5 where $y,.,. is the twist of the global parameter buy t — tw (.( wet) ’ teLrp, by the vector ) in auc = X(M)r@C = X.(Ay)@C, and fpw is the contragredient of the adjoint representation of “M on wo Apw/wfipw A fp. Here fap stands for the Lie algebra of the unipotent radical of “P. Applying the anticipated functional equation L(0, PP,w 0 Pur) = €(0, PPw 2 Pyu,AL(L, PP,w 0 Pred) » we write (5.5) rw) = Tim (L(1, pra © ¥0,r)E(L, Brim © Oven) ') + (See [16, Appendix 2].) Having described the terms in (5.4), we go back to the expression we have obtained for (5.1). Recall [3, §7] that Row tu) = (Mw) Rr(ows bw) » for any character ¢ in m(Mt)* = Hom(Mi/M8,C*) .UNIPOTENT AUTOMORPHIC REPRESENTATIONS. 31 This allows us to write (5.1) as the sum over Yy € Wo(Mw, Xa.) and over the orbits {0} € {Ty} of mo(M+)* in Hy_,, of the expression mp, (Fw)r Yw)tt(Re(Cw; Pw)ZP(a, f)) « where mg (Fw) = > my. (Cow)E(Mw) « Cemo(ME)* Applying Fourier inversion on mo(M}) to the formula (4.6) (with G replaced by My), while taking into account the property (i) of the local Conjecture 6.1 in [3], we obtain ms,(0w) = Sel? Yo evalu) . UeSy,, Therefore (5.1) equals UY elt vw {ow} ueSy,, Eva (U)r(w) < Uy ow > tr(Rp(ow, ¥»)TP(o, f)) « Suppose that ,, belongs to Yo(M,,, x,t). Let 4 denote the com- position of %» with our embedding 4M Cc 'G®°. We claim that 7 is well defined (as an equivalence class of parameters) in (G, y,t). Recalling (§4) the definition of equivalent parameters, we note that it is enough to show that the map (5.6) ker! (F,Z(G°)) — ker! (F, (1) is an isomorphism. By the obvious transitivity property, we can in fact assume that M is minimal, and hence a torus. Then Z(A1)/Z(G°) is a maximal torus in an adjoint group, on which the Galois action is dual to a direct sum of permutation representations. The bijectivity of (5.6) then follows from the exact sequence 70 ((Z(M)/Z(G°))") + H*(F,Z(G°)) 7 WYP,Z(M)) + HF, 2(I1)/2(6)), and its analogues for the completions of F. (See the proof of Lemma 4.3.2(a) of [12].) This proves the claim.32 JAMES ARTHUR Thus, y,, maps to an element y in Y(G, x, t), to which we can asso- ciate the objects Sy = Sy(G), Hy = My(G) and ey = eG for G. The next step is to apply a conjectural formula [3, §7] for the trace of the normalized intertwining operators in terms of the pairing on Sy x Ty. As it is stated in [3], the formula applies to the local intertwining operators and pairings, but the product over all valuations gives a formula for the global objects. In fact, certain constants in the local formula (namely, c(ax,%w), Aw(¥r) and c(7y,nG), in the notation of [3, §7]) have the property that their products over all valuations equal 1. The global formula is therefore simpler. If y4 denotes the param- eter yy, but regarded as an element in U(M) rather than ¥(Mw), then the orbits {oy} above will be in bijective correspondence with the representations 7 € Ily,, which extend to M,,(A)*. It follows from Conjecture 7.1 of [3] (and also the two remarks made after the conjecture), that DY < uw > tr(Rp Ow, Yw)Er(a, f)) equals SY fol), relly where zy stands for the image in Sy of the point u € Sy... We have now obtained an expansion Sel S> epelwlr(bu) D> < eum > falz), bw EWo( Musxar st) uESyy relly for the trace (5.1). We shall substitute this into our formula for Taisc(f). Observe that Do <2ut> folt) = |ro(G*)] DO < tut > fal). relly re{Ily} Therefore, Iaise(f) equals the triple sum over {M}, w € WS(ays)reg and tw € Vo(Mu, xa,t) of the product of IWS (ayy) Mdet(w - Vag |" with (57) WSyel? SD evelwr(du) SO < tut > alt) - UESiw re{ My}UNIPOTENT AUTOMORPHIC REPRESENTATIONS. 33 We propose to interchange the sum over the parameters with the sums over M and w. The outer sum will then have to be over all parameters ¢ € Y(G, x,t). For any y there will be an M, unique up to conjugacy, such that is the composition of a parameter jy € Wo(M) with the embedding 4M C 4G°. The condition that yy also belong to Wo(M,,), for a given w € W(agy)reg, is that the set S\y,(Mw) be nonempty. There is another way to state this. Recall that we have identified W(aj,) with the dual Weyl set W@(ay,). Then Sy,,(Mw) is nonempty if and only if w belongs to the subset Wy = W4(G) of elements in W¢(ay,) which, modulo the isomorphic groups (5.6), centralize the image of y. It will be convenient for us to regard this subset Wy, as the full Weyl set associated to Sy = Sy/Z(G°). It acts on the maximal torus = Ay,Z(6°)/2(G) of the connected component 59, = $9.2(G°)/Z(G°) For any w € Wy, we shall write det(w — 1) for the determinant of (w — 1), acting on the Lie algebra of Ty. One sees easily that Idet(w—1)] = [det(w—1)gge] = [det(w ~ 1)ag lidet(w ~ 1)gco |. Now it is well known that |det(w—1),¢0| equals the order of the kernel 9g of w, acting on the dual torus (2G) /(ZG@)") . (See [23, II.1.7].) The action of w on this torus is of course inde- pendent of w, and the kernel is just the finite group of components in ka = Z(G) n (2(G%)"). Therefore [det(w —1)ag |? = |det(w—1)[""|o(Ka)I"* In particular, w belongs to WG(agy)reg if and only if it lies in the set Woreg = {we Wy: det(w—1) # 0}34 JAMES ARTHUR of regular elements in Wy. When this is so, the associated parameter in Wo(M,) in fact belongs to o(My,xm,t). We shall denote it by tw, as above. ‘Actually, #w is not uniquely determined by Y and w. We must decide how many parameters in Yo(Mw, xa, ) lie in the equivalence class of }. Keeping in mind the isomorphism (5.6), we see that two parameters #y map to the same # if and only if they are conjugate by an element in W&(a4,). Moreover, two such conjugates are equivalent in Vo(Mu,xmt) if and only if they differ by an element in Wy(G°). The number of th, associated to y is therefore Way IW(G)I-* = Wael - Thus, our interchange of summation expresses Ijise,t(f) as the sum over p € U(G, y,t) and w € Wy,reg of the product of [0(%6c)" [Wy det (w — 1) with (5.7). Suppose that Y € U(G). As in the case of a local parameter, we can define the finite set Ny = Ny(G) = Norm(Ty,5y)/Ty Norm(A gy, 5v)/Ag, Z(G) « Let Si, be the subgroup of elements in Ny(G°) which act trivially on Ty. This group acts freely by translation on Ny, and the set of orbits can be identified canonically with Wy. One sees easily from the isomorphism (5.6) that Sy, = Suy-w = Show, weWy, for M as above. We also have the Weyl group W2 = Norm(Ty,5})/Ts Norm(A 47, 59)/A Z(G") i of the connected component $9. This too acts freely on Ny, and the set of orbits can be identified canonically with Sy. We obtain aUNIPOTENT AUTOMORPHIC REPRESENTATIONS 35 commutative diagram 1 1 1 | we = we 1a Sham OW 1 1— Sh > Sy > Ry — 1 Lo] 1 1 as in the local case [3, (7-1)]. The dotted arrows stand for splittings of short exact sequences determined by a fixed Borel subgroup of 5%, containing Ty. Similarly, one obtains a commutative diagram of groups if one replaces Ny, Sy, Wy and Ry by the respective finite groups Nf, S}, WY and Rf they generate. We shall write u — ry and u —+ w, for the projections of Nf onto St and W}. Notice that if ¢ is any element in Sy, and N(x) is the corresponding orbit of W} in Ny, the second projection maps M(x) bijectively onto a subset W(z) of Wy. We shall set W(2)reg = W(t) 1 Wyreg and N(e)reg = {UE N (a): wu € W(a)cog} « We apply these observations to our formula for Iise.t(f).. According to the horizontal exact sequence for Ny, in the diagram, the double sum over w € Wy reg and u € Sy, = Siw can be combined into a simple sum over the regular elements in Ny. We shall write (6.8) eY(u) = eva(w) for any point u € Ny whose projection onto Wy equals w. We also set (5.9) ry(w) = rb) -36 JAMES ARTHUR Then eM and ry extend to well defined characters on NJ and Wi respectively. The simple sum in its turn can be decomposed by the corresponding vertical exact sequence into a double sum over x € Sy and u € N(st)reg: Observe that [WyllSu.l = WollSol = Wel ISellWol = |SellW@)- It follows that Jaice(f) equals the sum over p in U(G,y,t) of the product of Iro(ea)I Sel“ with wert ye 2650 EN een eM (u)ry(wu)idet(wy — I" > fol) - re{My} Any clement w € Wf operates on Ty. It preserves the set Dy of roots of (89,T;). We shall simply write ¢(w) for the usual sign attached to this permuation, namely the number (-1) raised to the power e (Bh) 9 WE), where Ef is the set of positive roots in Dy relative to some order. PROPOSITION 5.1. We have ryt) = e(wueg(@uley (uy for any element u€ Nj. This proposition is the motivation for the introduction of the char- acters cy into the multiplicity formula of Hypothesis 4.1. We shall prove it in the next section. In the meantime, we can combine it with our formula for Iuise,t(f)-UNIPOTENT AUTOMORPHIC REPRESENTATIONS 37 PROPOSITION 5.2. The distribution Igisc,(f) equals the product of |to(Kq)|7} with (5.10) YS Wel YO F@ile) < 2.4 > felx), VEV(Gxs1) wE(Ty} rESy where (5.11) é(z) = [Wal SD e(w) |det(w - 17? . Oo WEW(2) reg, §6. THE SIGN CHARACTERS €y, AND Ty In this section we shall pause to study the characters ey and ry. Our goal is to prove Proposition 5.1. Recall that «y = eG is the one-dimensional character (4.5) on St = Sy(G+) which comes into the conjectural multiplicity formula. The function ry is the one di- mensional character ((5.5), (5.9)) on WT = Wy(Gt) defined by the global normalizing factors. We have seen that St and Wi are both quotients of the finite group Nf. We can therefore identify cy and ry, with characters on Nf. Proposition 5.1 can be regarded as a formula for the quotient of these two characters. We shall begin by expressing ry, in terms of the orders of certain L-functions at s = 1. Let S,y denote the set of roots of (G°, A yy). For each @ € Sqy there is a representation pa of “M on the root space &,. Having already fixed the dual parabolic subgroups P and “Pp = PxWp, we shall write Sp C Dy for the set of roots of (P, Aj). Fix an element w € Wy, and set Spy = {@€ Dp: wae (-Ep)}. Then there is a decomposition pew = @ pa ES re for the representation of “M which occurs in (5.5). Notice that the Killing form provides an isomorphism between p_g and the contra- gredient J. The formula (5.5) becomes (6.1) ry(w) = lim TT] L(1-2(4), pay) L(1+A(A), parody) s of rw38 JAMES ARTHUR since L(1,pa0 yr) = L(1+X(4), pao ou) - We are going to show that ry(w) equals the character (6.2) cy (L(s.050%4)) 6ES he We claim that for every root & € Dpjw, there is also a root G € Sp» such that Pao = paoy. To this end, observe that par % puacad(w) ow pua ot - The first of th » isomorphisms is given by the intertwining map Ad(w): 5 — 8y4> and the second follows from the fact that the image of w € Wt under the adjoint representation commutes with the image of Ly x SL(2, C). Now, consider the orbit O.(4) = {wia: j EZ} of & under the cyclic group generated by w. The representations {ogov: BE Ou(A)} are all equivalent, and are also equivalent to the contragredients {pgov: —B€Ow(a)} - But after a moment’s thought, we see that the intersections of O,(4) with Sp,» and (—Sp,») contain an equal number of roots. The claim follows. In particular, the terms in the product in (6.1) can be grouped in such a way that pa appears in the numerator as well as the denom- inator. This leads directly to the formula (6.2) for ry(w). Recall that (pao dy)(t) = pa (« (« Cc wet) . te Lp.UNIPOTENT AUTOMORPHIC REPRESENTATIONS 39 Now there is a decomposition (raov) = @ (uj Ow), esa) where each 1; is an irreducible unitary representation of Lp and vj is an irreducible representation of $L(2,C). Therefore, (6.2) can be written as a product (6.3) Il 0c 1)tde= (L(s,u5 825) GED pw JEG) where L(s, 4; x v;) stands for the L-function of the representation “2 0 ta asco; ("4 wet) : teLp, of Ly. From the discussion above we see that the contragredient acts as an involution 4; @ vj — jij @ V; on the constituents of pp.w». It is, moreover, an easy consequence of the unitarity of j1; L(s, nj @ vj) = L(5,71j @D;) , so that ordsa1(L(s, 4; ®vj)) = ordsa1(L(s, fi; ® 7) « In particular, the contribution to (6.3) of a distinct pair of contragre- dient constituents cancels. The product (6.3) need only be taken over those constituents with Hj @vj Ajp@r;. Since any finite dimensional representation of SL(2,€) is self contra- gredient, the condition is just pz; © jij. The question then is to determine the sign (6.4) (-1)erte=s Hsn@u)40 JAMES ARTHUR for any irreducible representation .@ v of Lr x SL(2,C) such that is unitary, and ji & p. Set m = deg(y) and n = deg(v). Then v maps the matrix It} 0 a)> teLr, ("es cir to the diagonal matrix ding( |t]F°-P, [eFO, «5 JOY) in GL(n,€). Therefore 1 1 . Ls,0@v) = T[L(s+ 50-2 +0); ¥) - ia We must therefore describe the order of zero or pole of L(s, 1) at any real half-integer. Hypothesis 4.1 includes the global Langlands correspondence for GL(m), which asserts that L(s,n) = L(s,") for some unitary, cuspidal automorphic representation 1 of GL(m, A). (See [3, §2].) Then L(s, #1) can have a real pole only if is the trivial one dimensional representation, in which case there is a simple pole at s =O and s = 17, Corollary 13.8]. Results of Jacquet and Shalika [8, Theorem (1.3)], 9, Theorem 5.3] imply further that the only possible zero of L(s, 2) at a real half integer is at s = 3, the center of the critical strip. The poles of L(s, 42) will contribute to (6.4) if n is odd. However, if is trivial and n is of odd dimension greater than 1, the poles at 0 and 1 will both contribute, and their effect will cancel. The zeros of L(s,)2) will contribute to (6.4) if n is even. From the functional equation L(s,p) = e(s,n)L(1 - 8,4) » we see that L(s, jt) has a zero at s = } of even or odd order, according to whether <($,#:) equals +1 or We have establishedUNIPOTENT AUTOMORPHIC REPRESENTATIONS 41 LEMMA 6.1. If n = deg(v) is even, the sign (6.4) equals (4,1). If n is odd, (6.4) equals 1 unless 1 @ v is the trivial representation of Lr x SL(2,€), in which case (6.4) equals (—1). a If we substitute the formula of Lemma 6.1 into the product (6.3), we obtain a new expression for ry(w). To describe this in a convenient way, we shall define a character ¢(/™ which is closely related to the original characters ¢G and ef. Let rh denote the Lie algebra of f, and let Adg/j denote the adjoint representation of “M on g/1h. The group Nf = Norm(Ty, S$) = Norm(Ay, S$)/Z(G°) also acts by the adjoint action on g/rn, and it commutes with the composite representation Adgyy, 0 of Lip = Lr x SL(2,€). Now, we have a decomposition Adgmod = QB QB wer) GED yy JEI(G) into irreducible representations of L’p. Let us write (§/1i)y, for the direct sum of those irreducible constituents 1; @ vj such that (i) ji; = Hj, (ii) e(4, uj) = —1, and (iii) deg(v;) is even. Then (g/1i),, is an invariant subspace of both Lp» and Nj. Define (6.5) eG! (u) = det(ii, Endy, ((8/18)y)) , wens, where w is any element in NJ whose projection onto Nf = Nf/Ty equals u. Observe that iv = DB b> where 29 & Bay ~ Ba The subgroup Sj, of Af leaves invariant each of the subspaces &, ’ of (&/1h)y. Since the actions of S}, on , ,, and 8, ,, are contra- gredient, eG!” quotient is trivial on Sj, and descends to a character on the Nt/sy = Wt.42 JAMES ARTHUR Of course the main reason for defining |0(Sy,5)|7* VEW(G,x,t) s€0rb(S2,Sy,1in) Sus 159, 0 ZA) ST (f%). By assumption, STS) = o( Fad) Fv). Part of our local assumption in §4 is that f¥"(4,) depends only on the image x = 5 of s in the set Sy = Sy/S}.Z(G°) = 5y/59, = m0(Sy).50 JAMES ARTHUR More precisely, formula (4.4) asserts that f(y) = > < 5ya,7 > fa(m), we{ly} where < -,- > is the global pairing on Sy x ly. We can therefore write Euise,(f) as the product of |ro(«c)|~* with a) YY wrt VC YL wv re{Ilo} r€mo(Sy) s€Ord(5%, 2010) Iro(Sy,s)I"'7(w,8)) < Sy, % > fal), where Bin = TO Syfin and 7(,8) = [Sul 159, ZB) oth v1) ‘The similarity with the formula in Proposition 5.2 appears promising. We must try to define the constants o( Hj, ¥1) so that the two formulas for Iaise(f) always match. Suppose that G is an arbitrary connected quasi- F, and that y; € ©(G;). Then we shall set lit group over (7.12) o(Gi,%) = |Sy ~ ten (5p, a(S)» es where ¢y, = €@? is the sign character (4.5) and o(5},,) is a constant, to be determined, which depends only on the isomorphism class of the complex, connected reductive group 58, = ($4, /Z(Gi)). We also ask that this latter constant have the property that (7.18) ($1) = o(S1/Z) |Z", for any complex connected group $1, and any subgroup Z; of the center of S;. In particular, o(S;) is going to have to vanish if S, has infinite center. This implies that o(G1, #1) = 0 unless 1 € V4(G1), as we would expect.UNIPOTENT AUTOMORPHIC REPRESENTATIONS 51 We of course want to set G; = Hi. Then Sh, = Ch, = Ch, 2(Ai)/Z(Fh) = (Cyn H)Z(H)/Z(H) = C},.2(H)/Z(H), by the formula (7.6). Since C8. Z(H) = C3 ANZA) Z(G), we obtain Ch.2(H)/Z(H) & Ch, Z(H)’ 2(G)/Z( A)" Z(G) © (C},.2(G)9/2(G)) Z(H)" Z(H)" = C4,,2(H)"/Z(H), from (4.11). But Cf, = S},,, so that 5S = 59 Z(H) /Z(H)P & 59/59. Z(H). It follows from the property (7.13) that o(S},) = 0(54,s)159,. Z(H)". Therefore, (His) = Sol e¥ (54, )153,69 2A)! |o(8,.)s so that 7(, 8) = eff (5y, )o(S},»)- Lemma 7.1. For H; and x € Sy, as in (7.11), we have eft (5u,) = eG (Sur). Proor: As in §4, let Te = Dk @ He @ ve) k52 JAMES ARTHUR be the decomposition of the representation 14: St x Lp x SL(2,€) — GLB) into irreducible constituents. If I denotes the set of indices k in the direct sum, let I’ denote the subset of k such that (i) zx & fix, (ii) (4, wk) = —1, and (iii) dim(vg) is even. Then re Sy, eG (x) = TJ det(rx(s)), ker’ where s is any element in 5y which projects onto x. Notice that the clement sy lies in both Sf and SL(2,C). If k belongs to J’, we obtain Ae(sy) = ve(Sy) = 1, since dim 1 is even. It follows that £G(5y) = TI det(Ae(sy)) = TT. ker’ ker’ Now H, is a central extension of the endoscopic group H attached to s. The Lie algebra of H equals the centralizer of Ad(s) in g, and the Lie algebra of #, can be identified with the direct sum of this algebra and a ceatral ideal. For each k, let Xj be the space of s-fixed vectors for Xx. This of course is just the intersection of the underlying space of x with the Lie algebra of H. Recalling the relation between w and yy, and applying the formula for eG(3y) to Hy, we obtain ey.) = TT eninee. kel’ Finally, we observe that the number eG(x) = TT det(rx(s)) kel’ equals the product of all the eigenvalues, counting multiplicities, of the operators Pals): ke T'}.UNIPOTENT AUTOMORPHIC REPRESENTATIONS, 53 Now the contragradient \, —+ Ax defines an involution on the rep- resentations A, with k € J’. In particular, if € is an eigenvalue, not equal to +1, then €~! is also an eigenvalue with the same multiplicity. Therefore, ele) = (-r, where m(—1) is the total multiplicity of the eigenvalue (-1). By the same token, Se (dim(Ag) — dim(Ag)) = m(-1) ker’ is an even integer. Consequently, (-1)"Y = eG (2). We obtain eG (Sur) = ep (weg (Sy) = eg (z)eg (By)! = ef? Bun), as required. a The lemma allows us to write T(,8) = €G (5yx)o(S%, ,). Substituting this into (7.11), and setting (7.14) e(z)= Sto 5y,s)1720(59,0), s€0rb(5,,2Fin) we see that Euisc,e(f) equals the product of |79(«g)|~! with 3 a [Spl > eG (5yx)e(x) < 5y2,7 > fan). VEW(G.x,t) TELM y} r€mo(Sy) The point sy € Sy(G®) belongs to the center of So(G*). Conse- quently, for any point s in the component x, the group 5, equals 5S} .ys- It follows that e(x) equals e(3y2). Substituting this into the formula above, and changing variables in the sum over x € 7(Sy), we see that Euise¢(f) equals the product of |7o(kq)|7! with (715) YD Sel SO eG (a)ela) < 2,4 > fala). VEW(Gxst) TE (Ty) zEmo(Sy)54 JAMES ARTHUR, We have now reached the stage in §7 at which we concluded §5. Taking the two sections together, we see that Hypotheses 3.1 and 4.1 yield two parallel expansions for Iaisc,e(f) and Evise,(f) into ir- reducible characters. Our goal is to show that these two expansions are in fact the same. The expansions are given by (5.10) and (7.15). They differ only in the coefficients i(x) and (x), which are defined for any component z € 779(5y) by (5.11) and (7.14). We must then show that the coefficients are equal. Recall that e(x) depends on a constant (5), which is to be defined for any complex, connected reductive group S; and which satisfies (7.13). We must show that o(;) can be defined for each S; in such a way that (2) and e(zr) are equal for any x. This is a property of Weyl groups which we shall establish in the next section. §8. A COMBINATORIAL FORMULA FOR WEYL GROUPS Suppose that Sis a union of connected components in an arbitrary complex, reductive algebraic group. Then S* is the reductive group generated by S, and S° is the connected component of 1 in St. Recall also that we are writing S, for the centralizer in $° of any element s € S. This group is of course not always connected. As a slight generalization of (5.11), we set (8.1) i(S) = |W) SP e(w)|det(w -— 1)", weWereg where wW° = W(S°) = Norm(T, S°)/T is the Weyl group of S° relative to a fixed maximal torus T, and Weeg = W(S)reg is the set of elements w in the Weyl set W = W(S) = Norm(T,S)/T such that det(w — 1) # 0. The determinant can be taken on the real vector space ar = Hom(X(T),R). As in §5, e(w) = +1 is the parity of the number of positive roots of ($°,T) which are mapped by w to negative roots. As in §7, we shall write Orb(S°, 2) instead of 5/S° for the set of orbits under conjugation by S° on an invariant subset © of S. This will prevent any confusion of orbits with cosets. The main example is when © equals the subset Sin = {8 ES: |Z(S3)] < co},UNIPOTENT AUTOMORPHIC REPRESENTATIONS 55 in which case the set Orb(S®, Sgn) is finite. Our object is to prove THEOREM 8.1. There are unique constants o(S;), defined for each connected and semisimple complex group S;, such that for any § the number (8.2) eS) = > |ro( 5)" (89) s€Orb(S°, Sein) equals i($). The constants have the further property that (8.3) ($1) = o($1/Z)|Z,|* for any central subgroup Z; of S,. Remarks. 1, It is obviously enough to prove the theorem when S is just one connected component. We shall assume this in what follows. 2. Let us agree to write o(S;) = 0 if 5; is any complex, connected algebraic group which is not semisimple. In particular, this constant vanishes if $; is a reductive group with infinite center. The equation (8.2) can then be written eS) = SO Irro(S,)|-t0(82) . s€Orb(S9,5) 3. Theorem 8.1 is what remains to be proved of the comparison of Teise,t(f) and Eaise,a(f) that we began in §5 and §7. It is interesting to observe that Theorem 8.1 is actually a miniature replica of the origi- nal problem. It is a formal analogue for Weyl groups of the problem of comparing Idisee(f) and Baisc,(f), and indeed of many of the com- parison problems, both local and global, that arise from endoscopy. I do not know whether it is part of a larger theory of endoscopy for Weyl groups, or whether results of this nature are already implicit in the representation theory of Weyl groups and finite Chevalley groups. We shall begin the proof of Theorem 8.1 by taking note of the uniqueness of the constants o(S;). For a given semisimple Sj, as- sume inductively that o(S;) has been defined for any S/ of dimension smaller than $;. Then o(S}) is determined by the formula, 9($1)|Z(Si)] = a(S) - > Iro(S1,s)I-*0(S?.,) s€0rb(S?,S;—2(51))56 JAMES ARTHUR which follows from the required equality of e($;) with i(S,). In other words, the special case of (8.2) that S = S° = S, provides a definition of the constant 0(S,). Having defined the constants o($}) we shall next establish the prop- erty (8.3). The argument is similar to part of the discussion of §7. Sup- pose that S is an arbitrary component, and that Z is a finite subgroup of Z(S°) which is invariant under conjugation by S. Then 5 = S/Z is a connected component of the reductive group 5+ = $+/Z, of which the identity component $° equals $°/Z. LEMMA 8.2. (i) i(S) = i(5). e(S) = e(S). (iii) If $° = S, then o(S) = 0(5)|Z|-!. PRooF: The property (i) follows easily from the definition (8.1). We shall establish the other two properties together. To this end we shall assume inductively that (ii) holds for any connected group of dimension smaller than S. If the group Z(S) = Cent(S,Z(5°)) is infinite, the quantities e(5), e(S), (5°) and o(S°) all vanish, and there is nothing to prove. We can therefore assume that Z(S) is finite. This implies that the group Z(S*)/ S is also finite. Let 5 be a coset in Sin = Siin/Z which does not lie in Z(S+) S. Then S: Cent(s, S°) is a proper subgroup of 5°. Since 58 = $92/Z = S9/S8NZ for any element s in the coset 5, our induction assumption implies that o($8) = o(S)1S2 92] . Let $3 be the normalizer in S° of the coset 3. The set of orbits in Orb(S°, S) which meet 5 can be identified with Orb(S;,3), a set of cardinality [Z||S3/S3|"* -UNIPOTENT AUTOMORPHIC REPRESENTATIONS 57 Observe that DX lr0(S.)-*0(52) s€Orb(5y,3) = [2|I8s/$.|-118./$9|-10(89)|89 9 Z|? = I($s/Z)/(S8Z/Z)|-40( 59) = |r0(53)|"*0(52) - Summing over alll such 5, we obtain e(S) - o($°)|Z(St) NS] = e(8) — o(5%)|Z(5+) NS} . If Z(5*)/S is empty, it follows immediately that e(S) equals e(5). Suppose that Z(S*) MS is not empty. Then S acts on the group S° by inner automorphisms, and we have e(S) = e(S°) = i($°) = i(S) from the definitions. (The equality of e(S°) and i(S°) was part of the definition of o($°).) Similarly e(5) = i($). The property (i) then implies that e(5) equals e(S) in this case as well. This is the required property (ii). Suppose that 5 = 5°. Then Z(S+)N S equals Z(S), a group which of course is not empty. The property (iii) follows from the fact that |Z(S)| = |Z(5)||ZI. The property (iii) of the lemma is the required condition (8.3) We still have the main part of the proof of the theorem, which is to show that e($) equals i(S). This of course is a problem only if $ is not equal to $°. As a warm-up, let us verify the equality of e($) and i(S) in the special case that S° = T is a torus. Then W consists of one element w, the adjoint operation of S on T. We can assume that this element is regular. Recall (23, II.1.7] that [det(w -1)] = IT"), where T™ denotes the kernel of w in T. Since e(w) = 1, we obtain (8) = |r’).58 JAMES ARTHUR On the other hand, T’ = Cent(s,T°) = Ss, for any element s € S. The regularity of w means that s belongs to Sn, and that S° = {1}. Therefore o($°) equals 1. But the T-orbit of s equals the product of s with {t~!w(t) : t € T}, a subtorus of T. This subtorus has the same dimension as T, and must therefore equal T. In other words, the orbit of s equals S, so there is only one summand on the right hand side of (8.2). We obtain e(S) = |ro(S.)* = IT*I* = HS), as required. Now suppose that $ is arbitrary, We shall use Lemma 8.2 to effect a simplification. First, observe that i(S) and e($) depend only on S° and the sct of automorphisms of $° induced from conjugation by S. We may therefore assume that S* is the semidirect product of 5° with 19(St). Now, let 5%. be the simply-connected covering of the derived group of 5°, and let S®.,, = Z($°)? be the connected component of the center of $°. Then = 8° x 3°. cent is a finite covering group of $°. In particular, S° equals $°/Z, where Z is the finite central subgroup of 5°. It is then readily verified that S = §/Z, where § = Sy. X Scent is a component which normalizes Z and such that the identity components ($)°, (Sse)® and (Seent)° equal the respective groups 5°, $9, and $®,,, above. Applying Lemma 8.2 and the calculaion above for tori, we obtain e(S) — i(S) = e(Ssc x Scent) — #(Sse % Scent) = €(Ssc)e(Seent) — t(Ssc)é( Scent) (e(Sse) — #(Ssc))i(Scent) - (We have also used the fact, easily verified from the defi e and i are multiplicative on products.) It is therefore enough to show that e(S) equals i($) in the special case that S° is semisimple and simply connected. We shall assume this from now on. If s is any semisimple element in S, the group itions, that S. = Cent(s,$°)UNIPOTENT AUTOMORPHIC REPRESENTATIONS 59 is then connected, by [24, Theorem 8.1]. In this case, it is part of our definition that e(S,) equals i(S,). If ¢ is a semisimple element in S°, the connectedness of S; implies that the set St = Cent(t,S) is either connected or empty. We can assume inductively that if dim(S*) < dim(S), then e(5*) equals i($*). LEMMA 8.3. The required equality of e(S) and i(S) is equivalent to the formula (8.4) YS (= Yo vst). s€Orb(S°,S) tEOrb(S°,S°) Proor: If s € S and t € S° are elements that commute, we write Sea = Cent({s,t}, 5°) . It is obvious that To((Ss)t) = mo(Ss4) = to((St)s) - The left hand side of (8.4) then equals i(S.) s€0"0159,5) = d es) s€Orb(S®,S) = > YE dr0( 5,0)! 06524) s€Orb(S°,S) t€Orb(Sy,5s) = >» [o(Sst)|-10( $2.) {(s,)€ESx S%:st=ts}/S° = > SY r0( 85,4) 0($24) t€Orb(S°,S°) s€Orb(S,,St) = yest). tEOrb(S°,S°) This last expression would just be the right hand side of (8.4) if e(S*) were replaced by i(S‘), But if t does not belong to Z(S), dim(S*) is60 JAMES ARTHUR smaller than dim($), and e($*) equals i(5*) by our induction assump- tion. It t belongs to Z(S), St is just $ itself. Therefore, the equality of e(S) and i($) is indeed equivalent to the identity (8.4). Qo It remains for us to establish the formula (8.4), in which S is a component such that 5° is semisimple and simply connected. We shall deal with each side separately. According to [24, Theorem 7.5], any semisimple element in $ normalizes some pair (Ti, B1) of groups, where 7; is a maximal torus in S° and B, is a Borel subgroup of S° which contains T,. Let B be a fixed Borel subgroup of S° which contains our fixed maximal torus T. Then any semisimple orbit of 3° in S contains an element which normalizes T and B. The normalizer of T and B in S can be written Twa, where we is a fixed semisimple element in S which preserves some splitting of B. Let 5’ and T’ denote the centralizers of wg in S° and T. Since S° is simply connected, Ss is a connected reductive group which contains 7’ as a maximal torus. In particular, we can form the usual sign character e’ on the Weyl group W' of (S',7"). LEMMA 8.4. The left hand side of (8.4) equals the number (8.5) Awe!) = [Wirt SO ew). wewy, PRooF: Let N’ denote the normalizer of T’ in S’. Then W = N/T = TN'/T. We claim that (8.6) Norm(T’,S°) = TN’. To see this, we shall consider the (open) chambers in the real vector spaces ap = Hom(X(T’),R) © ar = Hom(X(T),R) determined by the roots of $’ and S°. The fact that wa preserves a splitting in (B,T) implies that the simple roots of (BN S’,T’) are just the orbits under powers of ad(ws) of the simple roots of (B.T). The corresponding positive chambers are therefore related by aj, = ap Mat. By an argument of symmetry, any chamber in ayy becomesUNIPOTENT AUTOMORPHIC REPRESENTATIONS 61 the intersection of a with a uniquely determined chamber in ar. Suppose that n is an element in Norm(T", $°). Then n also normalizes T, since T is the centralizer of T’ in $°. The chamber Ad(n)(at) contains an open subset of a7, and therefore a chamber Ad(n’)(af,) , n'EN’, in ay. The map Ad(n’)~!Ad(n) will then send the chambers at, and at to themselves. This justifies the claim (8.6). We have agreed that any semisimple orbit of $° in S intersects Twp. Suppose that two elements s; and s in Twg are S°-conjugate. Since T’ is a maximal torus of both S, and S,,, s, and s are conjugate by an element in the group (8.6). From this it follows that there is a canonical bijection from Orb(TN’, Tw) to the semisimple elements in Orb(S®, $). It is of course only semisimple orbits which are relevant ‘0 (8.4). We can therefore write the left hand side of (8.4) as XS us.) s€Orb(S°,5) = ~~ «s.) s€Orb(TN’,Twr) = > IW(S)I-P SD eg(w)Idet(w - 1)? , s€Orb(TN’ Twp) wEW (Saree where ¢, stands for the sign character on the Weyl group W(S,) of Se. If s belongs to Twa, S, need not be a subgroup of S’. However, the elements in W(S,) normalize 7’, and are induced from the group (8.6). Therefore W(S.) = Cent(s,N’)/T! © TCent(s,N')/T . In particular, W(S,) is a subgroup of W’. Thus, the simple reflections in W(S,) are also reflections in W’, and since €, and ¢’ both take the value (—1) on any such reflection, we see that ¢, equals the restriction of e’ to W(S,). We can substitute this into the expression above. Our characterization of W(S,) also suggests that we should change the sum over Orb(TN’, Tw) to a sum over the smaller set Tz = Orb(T,Twr) = {t"!watwy: t€T}\ Tue.62 JAMES ARTHUR The expression for the left hand side of (8.4) becomes WT SS SE e(w)|det(w — 1/77 s€Tp WEW(Sa)reg The group W’ operates on Tg. It is easy to check that W(Ss)eeg = {WE Weg + (8) = 3} - The last expression can therefore be written jwyt > e'(w)|det(w — 1)|7? . {(s,w) €Tp x W}.g:0(8)=s} Now T’ is a finite covering of the torus {t"lwytw5! : te TH\T , and |det(w—1)] equals the number of fixed points of w in either torus. In particular, this number equals the order of the fixed point set Ty of w in Tp. We can therefore write our expression as wht SOS ew) ITB wEWeg SETH = wit SO ew) wew), = A(W',e'). o LemMa 8.5. The right hand side of (8.4) equals the number (8.7) A(We) = [WI > e(w) . wEWreg PROOF: Since any semisimple conjugacy class in S° meets T, we have a bijection from Orb(W°,T) to the set of semisimple elements in Orb(S°, $°). The right hand side of (8.4) can then be written i(S*) t€Orb(S°,S*) = SY wert YE tw)ldet(w-1)1-, t€Orb(W?,T) wEW(S*)regUNIPOTENT AUTOMORPHIC REPRESENTATIONS. 63 where ¢' stands for the sign character on the Wey] set w(s'). We need only consider elements t € T' such that S$‘ is not empty. For any such t, T is a maximal torus in the connected group (S‘)° = Si, and W(S*) is a subset of W = W(S). We claim that ¢! is the restric- tion of e to W(S*). The group W(S;) is generated by reflections which lie in W(S°). Since this group acts simply transitively on W(S'), it suffices to check that € and e' coincide on one element in W(S'). Let s be a semisimple element in S$‘. Then there is a conjugate s1 = gsg', ges, of s which lies in Twg. We can in fact choose g so that t; = gtg~! lies in the maximal torus T’ of $,,. It then follows that t; equals w(t) for some w; € W°, and that t is fixed by the element wy wew, in W(S). In other words, wy!wpw; belongs to W(S*). Since this element normalizes the Borel subgroups w,!Bw; and wy! Bw; 1S; of S° and $,, we have e(wy'waw;) = e'(wrtwaw,) = 1. This establishes the claim. Since W(S;) is the centralizer of t in W®, the right hand side of (8.4) becomes IW S™ SF e(w)|det(w — 1)|-* . 1€T wEW(S* reg set W = W(S) operates on T, and W(S')reg = {we Wreg: w(t) =t}. The last expression can then be written [we|-t > e(w)|det(w — 1)|-! {(t,w)€T X Wreg:w(t)=t} = WE SS ewer weWeeg teT* = [Wt SS ew) seWeeg = A(W,e), since |det(w — 1)| equals the order of the fixed point set T® of w in T. The lemma is proved. o64 JAMES ARTHUR Lemma 8.6. A(W’,c’) = A(W,e). Proor: The numbers A(W’,e’) and A(W,e) depend only on the Weyl sct W. They are independent of the isogeny class of the un- derlying component S. We shall assume inductively that the required formula holds if $ is replaced by a component of strictly smaller di- mension. We have the fixed Borel subgroups B and B’ of $° and S', so we can speak of standard parabolic subgroups. Suppose that A is a standard torus in T’. In other words, A is the split component of a parabolic subgroup P’ of $’ which contains B’. Let M’ be the Levi component of P’ which contains T’. Then A equals Ay = Z(M')®, the connected component of the center of M’. Write Wi = W(M'/A) for the Weyl group of M’/A, acting on T’/A. We can also take the centralizer M of A in S. Then M°® is the Levi component of a standard parabolic subgroup of S°. Write Wa = W(M/A), for the Weyl set of the component M, acting on T/A. The element we obviously embeds into Wa, and W', is just the centralizer of wp in W9 = W(M°/A). If A is nontrivial, our induction hypothesis tells us that (8.8) A(Wi,e%4) = A(Wa,ea) , where ey and e4 are the sign characters on Wi, and Wa. Suppose that w is an arbitrary element in W’. The identity com- ponent of the fixed point set (T’)” is a torus in 7’, and equals a W’'-translate w7'(A), wreW', of a standard torus A in T’. The clement w,ww;’ then lies in W4 reg: It is also clear that el(w) = ey(wiwwy?) .UNIPO NT AUTOMORPHIC REPRESENTATIONS, 65 Now the pair (A,w;) is not uniquely determined by w. The number of such pairs actually equals n(A)IWal where n(A) is the number of chambers in Hom(X(A),R) cut out by the hyperplanes orthogonal to the roots of the corresponding standard parabolic subgroup. The elements in W’ can be enumerated up to this ambiguity, however, as conjugates wrlwaw, wa € Whregs wi EW. We obtain Wit So ew) wew" = Let) walt SS ea(wa) waeW cog = Sona awe’) « A If W’ is not equal to {1}, the sign character e’ is nontrivial, and the left hand side of the equation equals 0. Applying (8.8) to the right hand side, we conclude that the expression (8.9) AW") + D> (A) AW, 4) AA{1} vanishes if W’ # {1}. Now suppose that w is an arbitrary clement in W. The identity component (T")® of the set of fixed points of w in T is a torus which commutes with any representative in $ of the Weyl element w. Copy- ing an argument from the proof of Lemma 8.5, we see that (rey = wr"(A), where w; belongs to W° and A is a torus in T’, In fact, we can assume that the centralizer M° of A in 5° is the Levi component of a standard parabolic subgroup of $°, This implies that A is standard torus in 7”. The clement w,ww;? then lies in Wyre, and e(w) = e4(wiww;) .66 JAMES ARTHUR For a given w, how many such pairs (A,w1) are there? We can cer- tainly replace w; by a product w'wMw) , w' €W', wa €W(M?), in which w! maps A to another standard torus in T’, However, this is the only possible ambiguity, so the number of pairs equals n(A)|W(M°)| = n(A)|Wal . We obtain [Wel S> e(w) wew = Vea ywart S23 eawa) A waeWajreg = Y(t a(Wa,ea) « a If W contains more than just the one element wa, the left hand side of the equation equals 0. Therefore, the expression (8.10) A(W,e)+ So n(A)“'A(Wa, ea) A¥(} vanishes if W # {wp}. The simple reflections in W' correspond to the orbits of simple roots of (B,T) under powers of ad(wg). It follows that W’ = {1} if and only if W = {wg}. In this case, both (8.9) and (8.10) are trivially equal to 1. We can therefore conclude that the expressions (8.9) and (8.10) are equal. The equality of A(W’,’) and A(W,¢) then follows. oO We have reached the end of the lemmas that make up the proof of Theorem 8.1. We obtain the general inequality of i(S) with e($) immediately by combining Lemmas 8.3, 8.4, 8.5 and 8.6. The proof of Theorem 8.1 is now complete. oUNIPOTENT AUTOMORPHIC REPRESENTATIONS 67 §9. CONCLUDING REMARKS Theorem 8.1 tells us that the coefficients i(r) and e(z) in (5.10) and (7.15) are always equal. It follows that there is a term by term iden- tification of the expansions for Iuisc,(f) and Euisc,(f). We conclude that Hypothesis 3.1 is a consequence of Hypothesis 4.1 (together with the local assumptions of §3, §4 and [3, §7]). This was the task we originally set for ourselves. We have in fact shown that the contributions to Izisct(f) and Evise,t(f) of each parameter y € U(G, x,t) are equal. Now there are some parameters for which the representation theoretic hypotheses are known. Consider the special case that G is a connected quasi- split group. Suppose that y is the image of a parameter Yo € W(Mo) for a minimal Levi subgroup Mg of G. Since Mg is a maximal torus in this case, Yp is trivial on SL(2,C), and is the parameter of a uni- tary character on Mo(F)\Mo(A). We can take Ty to be the set. of irreducible constituents of the corresponding induced representation of G(A). The parameter yp factors through the quotient Wr of Lr, so there is no problem with the hypothetical Langlands group. In particular, Sy, equals the centralizer in G of the image of Wp, and the quotient Sy, is just the R-group Ry. The pairing on Sy x Hy is then determined by the global normalized intertwining operators. In fact, Conjecture 7.1 of [3], which we assumed in §5, is already known in this case thanks to Keys and Shahidi (10, Theorem 5.1]. If Hi is associated to a point in a component x € Sy, we could just define the distribution f + f"(,) by fh) = YO falx). relly Then with these interpretations, the notions that went into the discus- sion in §5-§8 are all understood. The reader who dislikes arguments based on unproven conjectures can regard the earlier discussion as pertaining only to the parameters just described. It establishes that the contribution of these parameters to (9.1) Euiset(f) — Taise(f) vanishes. This paper has concerned the conjectures in [3] on unipotent (and more general) automorphic representations. The long term goal is to68 JAMES ARTHUR prove them, at least in part, with the help of endoscopy and the trace formula. A first step towards the creation of a logical structure for the argument is to verify the compatibility of the notions involved and to analyze the reasons for it. This has been our emphasis, and we continue with some informal comments on the proof envisaged. In general, Hypothesis 3.1 asserts the vanishing of the distributions (9.1). As we mentioned earlier, one should first try to deduce this from the trace formula. One would then use (9.1) to establish some ver- sion of the multiplicity formula (4.7). The formula could be assumed inductively for any proper Levi-subgroup. This would permit the ap- plication of the arguments in §5~-§8 to any parameter y € U(G,x,1) which is not the image of an elliptic parameter for an elliptic endo- scopic datum. The contribution to (9.1) of all such parameters could then be shown to vanish. The only remaining contribution to (9.1) would come frora parameters 7 such that Sy,, is finite for some cl- ement s in Sy = S,/Z(G°). It is from this that we would hope to deduce some form of (4.7), again using arguments of Sections 5, 6 and 7. The sign characters eG would be forced on us at this stage, essentially because of Proposition 5.1. Of course, it would not be feasible to apply the arguments of §5-§8 in precisely the way they were presented here. The correspondence from maps Wy — /G to automorphic representations is much deeper than multiplicity formulas such as (4.7), and in any case, we would certainly not want to assume the existence of the Langlands group Lr. We would instead have to replace the parameters y by the families o = {oy : 0 ¢ S} of conjugacy classes in LG attached to automorphic representations. (Sce [3, §1, §8].) For many G we can expect a bi- jection from ¥(G) onto the set 5(G) of such families. In these cases, the idea would be to define the centralizer Sy in terms of o. This could probably be done by considering the set of endoscopic groups H for which g lies in the image of the map Z(H) + L(G). It is of course necessary to determine Sy in order to state the multiplicity for- mula (4.7). By definition, a parameter y has a Jordan decomposition (hss, Punip), Where Was: Lp —+ 'G and Punip : SL(2,€) — Sy bas We would describe the Jordan decomposition in terms of o by first determining the family 4, attached to 55, and then describing theUNIPOTENT AUTOMORPHIC REPRESENTATIONS 69 group S,_ in terms of oss. In the case of a general G, some un- derstanding of the fibres of the map &(G) — (G) will probably be needed. We have not said much about the local side of the conjectures. This des the definition of the stable distributions f; + fO1(d), the nstruction of the packets IIy and the pairing < -,- >, and the proof of the local character identity (4.4). Once the stable distributions have been defined, the packets and the pairing are determined by (4.4) (together with the maps f > f). The essential part of the local conjecture is then the assertion that for a given y, certain linear combinations of the distributions f— sh), £eCce2(G(A),x), are actually characters, as opposed to more general invariant distri- butions. Ideally, it would be best. to deduce this locally. However, the global Hypothesis 3.1 itself carries some local information. For it ultimately implies some version of (4.7), and any such multiplicity formula tells us that certain distributions are in fact characters. I do not know how far this can be pushed. It is perhaps best to wait until Hypothesis 3.1 has actually been established. The case in which Hypothesis 3.1 will lead to the most complete results is the example of outer twisting of GL(n). The hypotheses of §4 (interpreted without reliance on the parameters € W(G)) are now known for GL(n). Moeglin and Waldspurger [21] have recently characterized the residual discrete spectrum for GL(n) in terms of the cuspidal spectrum, and it is clear how to interpret this in terms of the Jordan decomposition [3, §2]. On the other hand, the twisted endo- scopic groups for GL(n) include all of the quasi-split classical groups of type B, C and D (up to isogeny). One should try to deduce the con- jectural properties of the spectra of these classical groups from what is known for GL(n). We will conclude with a very brief discussion of this example. Set G° =GL(n). If ce then On(9) = In'('97')In 5 gE,70 JAMES ARTHUR is an outer automorphism of G° which leaves invariant the standard Borel subgroup. Set G= Gx. If 6, denotes the same outer automorphism of G° = GL(n,C), then G = @ xd, . It is easy to describe the elliptic endoscopic data for G. For each integer r, with 1 = <5,my >= 1, FESy. Tt follows from (4.4) that. f*(br) = falty) fece(GA)). A similar formula holds for the corresponding stable distributions on the local groups G(F,). However, this formula may not determine (9.3) completely. The problem is that the anticipated injection {ptr fECR(G(A))} Oo (hr: fre CE (H(A))} obtained by transfer of twisted orbital integrals, could be a strict inclusion. This difficulty is tied up with the question of how many loca] parameters ve = Que vy € V(H/Fy),UNIPOTENT AUTOMORPHIC REPRESENTATIONS 73 lift to y. If H, is symplectic or odd orthogonal, the only such param- eter will be y, itself. However, there can be a number of y, in the even orthogonal case, and the formula then determines only a sum of distributions (9.3). Once the distribution (9.3) has been defined (for H, and its en- doscopic groups), the packet Ty, and the pairing on Sy, x Ty, will be uniquely determined. Leaving aside the question of whether the required local properties of these objects can be deduced from Hy- pothesis 3.1, let us simply assume that the local assumptions of §3, §4 and [3, §7] hold for H,. The next problem is to determine the stable distribution (9.4) SI (fr) = oH te) fl (br) - (See the notation of §7.) The distribution f/(y),) is a local object which we are assuming is known, so it is the global constant o(H,, Uy) which must be found. According to Hypothesis 3.1, we should take the contribution of % to (9.1), and set it equal to 0. I have not thought through the details, but it should just be a question of running backwards over a couple of the more trivial arguments of §5 and §7. The result will be a special case (9.5) (Hes br) = Suef e Fu.) of the general formula (7.12) we determined was compatible with Hy- pothesis 4.1. Observe that the sign character €/f"(5y,.) appears. It originates, through [3, Conjecture 7.1] and Proposition 5.1, from the normalizing factors for (nontempered) intertwining operators for GL(n). Having determined the stable distributions (9.4) (for H, and its endoscopic groups), we can apply Hypothesis 3.1 to H,. The contri- bution of a, to Baise,e(fr) can be calculated as an easy special case of the arguments in §7, or it can simply be read off from the formula (7.15), (applied to H,, instead of G). It equals DY (Sel t DO etre) < 2m > )ix(a(f,)) - nélly, 2ESy, On the other hand, the parameter , € Vo(H,) is elliptic. Its contri- bution to Iaiset(fr) equals DF mo(n)te(a(f-)) - relly,74 JAMES ARTHUR. Identifying the coefficients in these two linear combinations of irre- ducible characters, we obtain the multiplicity formula mo(m) = |Spl2 S> eft(2) < 2,0 > TESy, Notice that the only contribution to mo(x) should come from the parameter ¢,. ‘This suggests that the map Y(H,) + 5(H,) is b tive, at least if H;, is not an even orthogonal group, the case we left ambiguous. This discussion has been very sketchy. We have simply tried to indicate that since the spectrum of GL(n) can be understood in terms of a Jordan decomposition, the same should be true for the spectrum of its endoscopic groups. The arguments of §5-§8 will be essential for this, in that they allow for the elimination of the irrelevant parameters from the study of (9.1). REFERENCES [1] J. Arthur, On some problems suggested by the trace formula, in Lie Group ‘Representations II, Lecture Notes in Math., vol. 1041, Springer-Verlag, 1983, pp. 1-49. [2] J. Arthur, The invariant trace formula II. Global theory, J. Amer. Math. Soc. 1 (1988), 501-554. [3] J. Arthur, Unipotent automorphic representations: Conjectures, to appear in ‘Astérisque. [4] J. Arthur and L. Clozel, Simple Algebras, Base Change and the Advanced Theory of the Trace Formula, to appear in Annals of Math. Studies, Princeton University Press. [5] L. Clozel, J.-P. Labesse and R.P, Langlands, Morning seminar on the trace formula, Lecture Notes, Institute for Advanced Study, Princeton, N.J., 1984. [6] A. Frohlich and J. Queyrot, On the functional equations of the Artin L- Junction for characters of real representations, Invent. Math. 20 (1973), 125— 138. [7] R. Godement and H. Jacquet, Zeta Functions of Simple Algebras, Lecture Notes in Math., vol. 260, Springer-Verlag, 1972. [8] H. Jacquet and J. Shalika, A non-vanishing theorem for zeta functions of GLn, Invent. Math. 38 (1976), 1-16. [9] H. Jacquet and J. Shalika, On Buler products and the classification of auto- ‘morphic representations I, Amer. J. Math. 108 (1981), 449-558. [10] D. Keys and F. Shahidi, Artin L-functions and normalization of intertwining operators, Ann. Scient, Ec. Norm. Sup., 4° série, t.21 (1988), 67-89. [11] R. Kottwitz, Rational conjugacy classes in reductive groupe, Duke Math. J. 49 (1982), 785-806.UNIPOTENT AUTOMORPHIC REPRESENTATIONS, 75 [12] R. Kottwitz, Stable trace formula: cuspidal tempered terms, Duke Math. J. 51 (1984), 611-650. [18] R. Kottwitz, Stable trace formula: elliptic singular terms, Math. Ann. 275 (1986), 365-399. [14] R. Kottwitz, Tamagawa numbers, Ann. of Math. 127 (1988), 629-646. [15] J.-P, Labesse and R. Langlands, L-indistinguishability for SL(2), Canad. J. Math. 81 (1979), 726-785. [16] R. Langlands, On the Functional Equations Satisfied by Eisenstein Series, Lecture Notes in Math., vol. 544, Springer-Verlag, 1976. {17] R. Langlands, Stable conjugacy: definitions and lemmas, Canad. J. Math. $1 (1979), 700-725. [18] R. Langlands, Les Debuts d’une Formule des Traces Stable, Publ. Math. Univ. Paris VII, Vol. 13, Paris (1983). [19] R. Langlands, Fisenstein serice, the trace formula, and the modern theory of automorphic forms, to appear in Number Theory, Trace Formulas and Discrete Groups: Symposium in Honour of Atle Selberg, Academic Press. [20] R. Langlands and D. Shelstad, On the definition of transfer factors, Math. Ann. 278 (1987), 219-271. [21] C. Moeglin and J.-L. Waldspurger, Le spectre residuel de GL(n), preprint. [22] J. Rogawski, Auomorphic Representations of Unitary Groups in Three Vari- ables, to appear in Annals of Math. Studies, Princeton Univ. Press. [23] T. Springer and R. Steinberg, Conjugacy classes, in Seminar on Algebraic Groups and Related Finite Groups, Lecture Notes in Math., vol. 131, 1970, pp. 167-266. [24] R. Steinberg, Endomorphisms of Linear Algebraic Groups, Mem. Amer. Math. Soc. 80, 1968. Mathematics Department, University of Toronto, Toronto M35 1A1, Ontario, Canada.Motifs et Formes Automorphes : Applications du Principe de Fonctorialité LAURENT CLOZEL Sommaire 0. Introduction... . . . 78 1. Catégories de représentations automorphes . 81 1. Principe de fonctorialité; opérations tannakiennes . 81 2. Catégories de représentations automorphes et admissibles 2.2... 0... . 87 3. Formallisme et problémes tannakiens . 93 4, Catégories abéliennes de représentations isobares 96 2. Certains cas connus du principe de fonctorialité 99 1. Induction et restriction astomorphes a) 2. Carré symétrique . 100 3. Fonctorialité associ 100 4, Groupes unitaires. . 2... 100 3. Questions de rationalité 101 1. Le corps de rationalité d’une représentation | automorphe .. 2.2... 2 ee 101 2. Représentations non ramifiées : paramétistion | rationnelle . 2... eee 105 3. V’action de Aut(C) sur le type & Vinfini des représentations algébrique 106 4. Conjectures, et vérifications ; 107 5. Cas des représentations réguliéres . . 11 4. Motifs et représentations algébriques . 130 1. Motifs... ee . 131 2. Représentations algébriques rationnelles . 131 3. Motifs et représentations algébriques 136 4. Conséquences du formalisme motivique 143 5. Conjecture de Hasse-Weil et théorie réciproque pour les représentations algébriques 147 5. Le cas des représentations autoduales 149 1. Enoncé du théoréme 149 2. Remarques. .. . . .. 149 3, Esquisse de démonstration . . 151 Sloan Fellow. Auiomorphic Fos, Shimura Copyright ©1890 by Academie ress, Ine Vertis and Furcons ‘Allsigh of reproduction in wy fo eseved ISBN 0-12-176651-978 LAURENT CLOZEL 0. INTRODUCTION Ne t’attarde pas é Vorniére des résultats. Seule les traces font réver. —René Char Les conjecture générales de Langlands aménent A subodorer une relation étroite entre la théorie des formes automorphes et la théorie arithmétique des variétés algébriques - plus précisément, entre repré- sentations automorphes et motifs au sens de Grothendieck. Dans son article & Corvallis, Langlands avait esquissé l'idée d’une théorie “tan- nakienne” des formes automorphes: si l’on suppose que cette théorie existe, on est amené & conjecturer I’existence d’un “groupe de Galois automorphe”, qui devrait étre lié au “groupe de Galois motivique” postulé par Grothendieck. La théorie moderne des formes automorphes, cependant, a conduit - depuis Vintroduction par Maass des formes qui portent son nom - A Vétude de nombreux objets qui, & la différence des formes modulaires classiques, n’ont qu’une relation ténue avec des objets arithmétiques. Crest déja vrai des “Gréssencharakterer” de Hecke: dans ce cas, Weil avait été amené a définir les “Gréssencharakterer de type Ao” - ou caractéres de Hecke algébriques - qui sont ceux qui apparaissent, par exemple, dans la théorie de la multiplication complexe. Le propos de cet article est de définir, dans le cas non abélien, la catégorie des représentations automorphes qui devraient de méme étre liées aux “motifs” - analogue non abélien des caractéres de Hecke. Une fois acquise la “classification de Langlands” pour les groupes récls, la définition d’une telle catégorie s'impose assez naturellement (Borel avait proposée dans [10b]). On s’apergoit assez vite, cepen- dant, que la définition “évidente” ne remplit pas deux des condi- tions minimales requises: @tre accessible au formalisme tannakien, et vérifier les propriétés de rationalité qui sont satisfaites par les motifs, et par les caractéres de Hecke algébriques. Lorigine du probléme est que la classification de Langlands, basée sur induction unitaire, n’est pas “algébrique”: elle ne respecte pas la rationalité des représentations. En prenant pour guides le formalisme tannakien et les propriétés de rationalité, nous proposons ici une définition de la catégorie des représentations automorphes GL(n, Ar), A désignant les adéles dun corps de nombres F. Nous montrons alors, 4 l'aide de l’extension A GL(n) de la théorie d’Eichler-Shimura, que ces représentations -MOTIFS ET FORMES AUTOMORPHES. 79 au moins quand leur type a l’infini est convenablement régulier - ont les propriétés de rationalité attendues: si = @ ym, est une repré- sentation automorphe algébrique, supposée cuspidale, de GL(n, Ar), sa partie finie ms = ®vfinieTy est définie sur un corps de nombres EC C, et toutes les conjuguées de rs par Gal(Q/Q) sont encore parties finies de représentations cuspidales. De plus, si l'on admet le “principe de fonctorialité” de Langlands, la catégorie des représen- tations algébriques devrait étre une catégorie tannakienne. Nous précisons ensuite, de fagon assez détaillée, les relations conjec- turales entre représentations algébriques et motifs. Par construction, les représentations algébriques sont munies de structures qui devraient étre reliées aux structures diverses dont sont munis les motifs: théorie de Hodge; représentations des groupes de Galois dans la cohomologie Ladique. Les propriétés de pureté des motifs se traduisent par la “Conjecture de Ramanujan”. Dans quelques cas, que nous passons en revue, on dispose dé} d'une interprétation “motivique” des représentations automorphes: il s’agit essentiellement de formes automorphes de type arithmétique sur GL(2), ou des caractéres de Hecke. Dans le dernier paragraphe, nous montrons comment associer, sous certaines hypothéses locales, des systémes compatibles de représentations l-adiques & des représen- tations automorphes autoduales de GL(n) sur un corps totalement réel. En particulier, nous démontrons la conjecture de Ramanujan ( A presque toutes les places) pour de telles représentations. Le lecteur de penchants classiques pourra s’étonner de nous voir étudier les formes automorphes sur GL(n) plutét que sur les groupes réductifs liés aux domaines hermitiens symétriques: les formes modu- laires holomorphes sur ceux-ci admettent souvent une interprétation “motivique” évidente. L’importance des groupes linéaires tient au fait que, d’aprés le principe de fonctorialité, ils “recoivent” toutes les représentations automorphes des groupes classiques. Ainsi, toutes les formes automorphes sur ceux-ci qui ont une interprétation motivique doivent étre liées 4 des représentations algébriques. La réciproque n'est pas vraie: en fait, seuls des motifs satisfaisant des conditions trés fortes d’autodualité apparaissent, semble-t-il, dans la cohomolo- gie d’intersection des variétés de Shimura (cf. §5.2.4). Par ailleurs, le fait de considérer GL(n) permet d’utiliser - au moins conjecturale- ment - le formalisme tannakien. Voici maintenant une description plus détaillée des différents para- graphes. Le paragraphe 1 décrit le formalisme tannakien pour GL(n).80 LAURENT CLOZEL Nous avons repris les constructions de Langlands [38 f] en les explici tant: noter que le résultat fondamental qui est a la base du formalisme (Thm. 1.1), et qui était supposé par Langlands, a été démontré par Jacquet et Shalisa. Tl donne un sens, en particulier, & la notion de représentation isobare. La catégorie abélienne des représentations iso- bares est décrite on 1.4.3. Les représentations algébriques sont définies en 1.2.3. On a décrit en détail comment elles se traduisaient dans les cas “classiques” (formes modulaires de Hilbert). Le paragraphe 2, utilitaire, décrit certains cas du principe de fonc- torialité - établis pour la plupart pendant les dix derniéres années - qui nous seront utiles. Le paragraphe 3 étudie le corps de rationalité des représen- tations automorphes de GL(n). On y démontre tout d’abord que celles-ci sont défnies sur leur corps de rationalité (Prop. 3.1). On ex- plique ensuite comment transformer la paramétrisation de Langlands des représentations non ramifiées pour la rendre rationnelle. On for- mule ensuite les conjectures fondamentales (Conjectures 3.7, 3.77, 3.8) sur la rationalité des représentations algébriques (en particulier, la Conjecture 3.8 est l’analogue de la conjecture de Weil, démontrée par Waldschmidt, et caractérisant par leurs valeurs les caractéres de Hecke algébriques). On vérifie que ces conjectures sont compatibles avec notre définition des représentations algébriques ( §3.4.3) ainsi qu’au changement de base ( §3.4.4). On démontre ensuite la Conjecture 3.7 pour les représentations réguligres (Thin. 3.13). La démonstration repose sur la théorie @Eichler-Shimura, reformulée dans le cadre adélique par Harder. Le point essentiel consiste 4 montrer que la cohomologie parabolique, considérée commes sous-espace de la cohomologie & support compact des groupes arithmétiques, est stable par automorphismes de C. On le déduit des propriétés d’autodualité de la cohomologie parabolique: le résultat crucial (Prop. 3.18) a été démontré aussi, de facon différente, par Harder [30 a]. Il est ici déduit d’un lemme simple mais fonda- mental concernant image de la cohomogie & support compact dans la cohomogie L? (Lemme 3.17). Une conséquence importante de la démonstration est que la cohomologic parabolique (dans le cas des coefficients constants) est définie sur Q: c'est le Théoréme 3.19, qui étend & GL(n) un résultat classique d’Eichler et Shimura. Le paragraphe 4 est consacré au “dictionnaire” reliant motifs et re- présentations algébriques. On y définit les représentations algébriques rationnelles ( §4.2), analogue des motifs “absolus” (sans coefficients).MOTIFS ET FORMES AUTOMORPHES 81 Aprés avoir introduit le groupe de Galois (conjectural) associé aux re- présentations algébriques, et discuté ses relations avec le “groupe de Galois automorphe” de Langlands, on donne la conjecture fondamen- tale concernant la relation entre motifs et représentations algébriques (4.3.2, Conjecture 4.5). Elle est précisée en 4.3.3. On donne quelques verifications concernant la rationalité et le com- portement fonctoriel en 4.3.4, 4.3.4. En particulier, on démontre le fondamental Lemme 4.9, concernant la pureté des représentations algébriques a l’infini; il est utilisé dans le paragraphe 3 et pour définir a priori la notion de poids pour les représentations algébriques. Dans le paragraphe 4.5, on formule un probléme concernant la “théoric réciproque” pour les représentations algébriques. Enfin, le paragraphe 5 décrit comment l’on peut associer, sinon des motifs, au moins des systémes compatibles de représentations I- adiques, aux représentations automorphes de GL(n,Ar) (F totale- ment réel) qui Les démonstrations re- posent sur des résultats récents de Kottwit: Pendant la rédaction de cet article, j’ai bénéficié de discussions avec J.-L. Waldspurger sur les questions de rationalité du paragraphe 3. Jai aussi eu accés & un manuscrit de Harder [30 a] concernant la coho- mologie des groupes arithmétiques. Lors de la conférence, et ensuite, j’ai profité des explications de D. Blasius et D. Ramakrishnan sur les propriétés de la catégorie des motifs. Si le paragraphe 4 ne contient pas d’erreur grossiére, cela leur est da. J. Tits m’a décrit la coho- mologie galoisienne des groupes unitaires. Enfin, R. Kottwitz m’a, a plusieurs reprises, expliqué les propriétés miraculeuses des groupes décrits au paragraphe 5 et la stabilisation de la formule des traces tordue. Je lui suis tout particuliérement reconnais communiqué avant rédaction les résultats essentiels de [36]. Qu’ils soient tous ici remercié: sont régulitres et autoduale: nt de m’avoir 1. CATEGORIES DE REPRESENTATIONS AUTOMORPHES Dans ce chapitre on décrit certaines catégories de représentations automorphes des groupes linéaires qui devraient étre passibles d’un formalisme tannakien. 1.1. Principe de fonctorialité; opérations tannakiennes. Dans tout ce qui suit, on s’intéresse aux représentations automorphes de GL(n, Ay) ot F est un corps de nombres et A = Ar. Rappelons [12, §4.6] qu’une représentation irréductible de G(A) = GL(L,A)82 LAURENT CLOZEL est automorphe si elle est isomorphe un sous-quotient irréduc- tible de la représentation de G(A) sur l'espace des formes automor- phes (pour les problémes que cela pose du point de vue tannakien, cf.§1.3). On notera Aut (n) la classe des représentations automor- phes de GL(n,Ar), Aut =I], Aut (n). Rappelons que les représentations cuspidales sont celles qui in- terviennent dans l’espace des formes paraboliques Aeusp(G(F)\G(A)) [12]. La théorie des séries d’Eisenstein montre (Langlands [38 e]) que toute représentation automorphe est isomorphe a un sous-quotient @une induite GA) (1.1) p= Inde) aya? ® 1) oti ’on désigne par Ind l'induction unitaire a partir d’un parabolique P=MN dc G, ct o une représentation cuspidale de M(A). Dans tout ce qui suit, on considérera uniquement, sauf mention explicite, les paraboliques standard de GL(n) formés de matrices tri- angulaires supérieures par blocs. Si P est de type (m1,...,%r) ot ny ter++n,y =n, o peut s’écrire sous la forme 01 ® ++: @ or oil 0; est. une représentation cuspidale de GL(n;,A). Le résultat suivant est dit a Jacquet et Shalika. THEOREME 1.1 [33 a,b]. Soient n = ny +--+, = my to tm, deux partitions de n, S un ensemble fini de places de F, contenant les places infinies. Pour tout i = 1,...,7, soit o; une représentation de Gp,(A) sur un sous-espace irréductible V; de Acusp(Gn.(F)\Gns(A))s de méme, soit (t;,W;) une représentation de Gm,(A) (j = 1,---+8)- On suppose les 0; et les 7; nonramifiées en dehors de S. Supposons que pour v ¢ S, les matrices diagonales tor tre tone tre of les te,,»; (resp. trs,») désignent les matrices de Hecke, sont égales modulo 6. Alors r = 8, et il existe une permutation a € G, telle que mai = nj et les deux sous-modules Vi et Wai de Acusp(Gni(F)\Gn:(A)) coincident. L’analogue local de ce résultat est la classification de Langlands. Supposons pour un instant F local. Soit P:n = mi +: +n, uneMOTIFS ET FORMES AUTOMORPHES 83, partition de n, et 0; des représentations de carré intégrable mpodulo le centre de GL(nj, F). Soit x; le caractére central de o; sur FX. Ona xi =| |**, 51 € Re A permutation prés, on peut supposer $1 > s2 > +++ > sp. La représentation (1.2) p= Ind p}(01 @ ++ Bor @1) = p(1y-.-40r) a alors un unique quotient irréductible, qu’on appelle son quotient de Langlands; il apparait avec multiplicité 1 dans p et donc dans cha- cune des représentations p(o5.1,---,0s.r) ob s € G,. On V’appelle encore sous-quotient de Langlands de chacune de ces représen- tations (cf.[14, 50]). On notera que la définition habituelle du quo- tient de Langlands fait appel 4 induction a partir de représentations tempérées. Pour traduire ceci en notre définition, il faut savoir que pour GL(n) les induites paraboliques & partir de représentations de carré intégrable unitaires sont irréductibles: [5, 4 ; 51 a D’aprés Langlands, on sait que toute représentation irréductible de GL(n,F) est isomorphe au sous quotient de Langlands d’une p(o1,--+,r) et que les oj sont alors bien déterminées & permutation prés. Pour étendre la définition du quotient de Langlands aux com- posantes locales de (A.1), il faut introduire les représentations génériques. Soit 1a + No=qn=[- Bn-1 0 1 ¥ un caractére additif non nul de F, 0(n) = Y(a1 +++ + tp—1) le caractére non dégénéré de No(F) associé. Rappelons qu’une repré- sentation (7,V) de GL(n,F) (F local) est générique s’il y a un vecteur distribution dans son dual se transformant sous No(F) selon 9. On sait que les composantes locales d’une représentation cuspidale sont génériques [40, 49]. D’aprés des résultats de Kostant et Vogan dans le cas réel, de Bernstein et Zelevinsky dans le cas p-adique, on sait que toute représentation irréductible générique_s’écrit. comme Vinduite totale p(o1,...,0+) (1.2) pour des représentations de carré intégrable 01,...,0%-84 LAURENT CLOZEL Considérons alors les composantes locales de la représentation p de (1.1), ot o = 01 ®-+-@ ar est un produit tensoriel de représentations cuspidales de GL(n, Af). La composante a la place v s’écrit (1.3) po = Ind FE} (o1,9 @ ++ B Fre @1)- Puisque les 0,5 sont génériques, elles s’écrivent a leur tour comme des induites, a la (1.3), & partir de représentations de carré intégrable: (1.4) fig = Indo (7h, @ + @ rf, @ 1) (induite totale). On en déduit, par induction par éapes, que Py est Vinduite totale & partir du produit tensoriel des 7; et contient donc, avec multiplicité 1, le sous-quotient de Langlands correspondant. Definition 1.2 (Langlands). Soit 7 = @m, une représentation au- tomorphe irréductible de G(A). On sait que ™ est un sous-quotient @une représentation (1.1) p = @pv. On dit que m est isobare si 7, est, & toute place, le sous-quotient de Langlands de py. Remarquons tout d’abord que, si py est non-ramifiée, son quotient de Langlands est la représentation sphérique associée [31]. En par- ticulier la représentation isobare associée 4 p a bien la propriété de continuité exigée des représentations automorphes. Si o1,...,0, sont des représentations cuspidales de Gn,(Ar), 07 notera o, B---Ho, Lunique sous-quotient isobare de la représentation (1.1). Par construction, il ne dépend pas, a isomorphisme prés, de l’ordre des o;. Par définition, toute représentation isobare est ainsi obtenue; enfin, d’aprés le Théoréme 1.1, deux représentations 7, B--- Ba, et a} E+. Fa! sont isomorphes si et seulement si les a; sont obtenues par permutation des 0}. Exemples: Dans (1.1), prenons P égal au sous-groupe de Borel, et Meme Nica aaa) p=Ind ( ot | | désigne la norme d’idéle, La représentation isobare associée est ntation triviale.MOTIFS ET FORMES AUTOMORPHES. 85 Plus généralement, soit P homogene de type (m,...,m) (n = am) et w une représentation cuspidale de GL(m,A). La représentation = telle que, A toute place v, my est unique sous-quotient irréductible de (15) Ind $e (wo, wv | | 5-++,@v| (7) est isobare et intervient dans le spectre discret [51 b]. Moeglin et Waldspurger [39] ont montré récemment que ces représentations for- ment tout le spectre discret. On notera unique quotient irréductible de (1.5) J(wp,a) et le produit global J(w, a). Rappelons maintenant comment le principe de fonctorialité s’illustre dans le cas de GL(n): ¢ Sommes : Soit n =n; +n2+---+N,, 7; des représentations auto- morphes de GL(n;,A). On peut considérer la représentation induite p=Indpayt @--@m Ol. Elle a pour matrices de Hecke, aux places non-ramifiées, tay tow = t Autrement dit, les matrices de Hecke sont dé terminées par Vhomo- morphisme de L-groupes L 1X x PGO 4 FGe envoyant GL(n1) x +++ x GL(n,) vers GL(n) par blocs. Si les représentations 7; sont isobares, on peut considérer unique sous-quotient isobare de p (il existe d’aprés les mémes arguments isés pour m; cuspidales). On le note n= Bs Ba. * Produits. Soit n = ab, m une représentation automorphe de GL(a,A), m2 une représentation automorphe de GL(b,A).86 LAURENT CLOZEL Conjecture 1.3. II existe une représentation automorphe de GL(ab, A) telle que, pour v en dehors d’un ensemble fini S de places: (1.6) tro = ny @ trav (le produit tensoriel de deux matrices diagonales définit une matrice diagonale, qui ne dépend que de 1; et t2 modulo G,). Cette conjecture mérite d’étre appelée la conjecture principale pour GL(n). Notons qu’il suffirait de la démontrer pour des représentations cuspidales. On peut la raffiner pour des représentations isobares: soient 1, mp isobares. Si la conjecture est vraie pour 7 et 72, il existe une unique représentation isobare de GL(n, Ar) vérifiant (1.6). Definition 1.4. Si la représentation isobare de GL(n, Ar) vérifiant (1.6) existe, on l'appelle produit tensoriel extérieur de ™ et m2; on la note 7 B72. Du cété du L-groupe, cette opération correspond bien sir & Vhomomorphisme £G® x EGP bad donné par le produit tensoriel: GL(a) x GL(b) + GL(n). e Puissances. Plus généralement, soit r une représentation ra- tionnelle de GL(n), de degré N. SiG = GL(n) et H = GL(N), on considére r comme un homomorphisme “G° > "H°. On Vétend aux L-groupes, trivialement sur les groupes de Galois. D’aprés le principe de fonctorialité on en déduit, de méme une application conjecturale, envoyant les représentations isobares 7, de GL(n, A) vers celles, 1, de GL(N, A), et déterminée par le fait que tan =P (trae) comme classe de conjugaison dans GL(N, C). Des cas particuliers importants sont les puissances symétriques 5%, extérieures A‘ de la représentation standard de GL(n). Plus généralement, on peut considérer des représentations ra- tionnelles de G = GL(n) x +++ x GL(n,). On retrouve ainsi le cas des produits. « Changement de base. Soit E/F une extension de corps de nom- bres. On considére GL(n)/E comme un F-groupe par restriction des scalaires. Son L-groupe est alors [10a,§5] LH =GL(n,C) x --- x GL(n,€) x oF;pe MOTIFS ET FORMES AUTOMORPHES 87 ily a une copie de GL(n, €) pour chaque plongement EB = F sur F, et ar opere par permutation. On en déduit deux opérations fonctorielles: (a) Restriction. Posant G = GL(n)/F, on a un plongement diagonal “G = +H qui envoie “G° diagonalement dans “H°. Dualement, on en déduit une opération de restriction pry envoyant les représen- tations isobares de G(Af) vers celles de G(Az). (b) Induction. Soit d = [E : F], Gin = GL(dn)/F. On aun plongement "HA 'Gan envoyant les matrices de “H®, diagonalement par blocs, vers des ma- trices de GL(dn,C). On Pétend & “H° en faisant opérer les éléments de gy par des matrices de permutation. On en déduit, dualement, une opération (conjecturale) d’induction automorphe cf envoyant les représentations isobares de GL(n,Ag) vers celles de GL(dn, Ap). Pour la description, dans le cas oi E/F est cyclique d’ordre premier, de pryr et th en termes de matrices de Hecke, voir [18 d]. Le changement de base commute avec les opérations “géométriques” et & (les représentations étant considérées & isomorphisme pri Dans le chapitre 2, on passera en revue certains cas oit le principe de fonctorialité a été établi. Il sera commode de décrire tout d’abord certaines catégories de représentations. 1.2. Catégories de représentations automorphes et admissi- bles. Il s’agit maintenant de définir certaines catégories de représen- tations automorphes, qui puissent étre ibles au formalisme tan- nakien (cf. §1.3). On se contente ici de définir les objets. Il faut que les ensembles de représentations automorphes considérés soient sta- bles par les opérations tannakiennes & et &, ainsi que par l’adjonction qui & 7 associe sa contragrédiente #. Pour que celles-ci soient bien définies, on se limite dés 'abord aux représentations isobares. 1.2.1. La catégorie Isob. Ses objets sont les représentations iso- bares. La somme directe 7’ y est bien définie; le produit tensoriel Vest (A isomorphisme prés) modulo la Conjecture principale. On notera Isob(n) l'ensemble des objets de degré n (représentation de GL(n,A)); de méme pour d’autres catégories. 1.2.2. Les catégories Temp et IC..88 LAURENT CLOZEL Les catégories Temp sont les représentations automorphes irréducti- bles de GL(n,A)- pour les différents n- dont les composantes A toutes les places sont tempérées. LEMME 1.5. Sia € Temp(n), est isobare. En tant que représentation automorphe, 7 est en effet un sous- quotient de Ind$(c1 @---@a-@1) pour des représentations cuspidales a; de GL(nj,A). A presque toutes les places, 7, est non-ramifiée, Les représentations (g;)» le sont donc aussi, et 7» est, par induction par étages, le sous-quotient de Langlands de Indgt}(Xiy-+-)Xn) ot B est le sous-groupe de Borel, et les x; sont non-ramifiés. Comme 7, est tempérée, Punicité dans la ¢ ion de Langlands implique que les x; sont unitaires ct Vinduite irréductible. On en déduit que les 0; doivent étre unitaires (A presque toutes les places, et done partout); done 7 & IndG(a1 ® «++ @ oy @ 1) est isobare. Rappelons la: CONJECTURE 1.6 (“CONJECTURE DE RAMANUJAN”). Si a est une représcntation cuspidale unitaire de GL(n, Ap), les composantes lo- cales de m sont tempérées. Tl est clair que la somme directe de deux représentations de Temp est encore tempérée. Le Lemme 1.5 montre en fait que si 7 est une re- présentation automorphe dont presque toutes les composantes locales sont tempérées, elle est somme (pour §B) de représentations cuspidales unitaires. Si Pon admet la Conjecture 1.6, et bien stir la Conjecture 1.3, on voit que Temp est préservée par le produit tensoricl. Soit IC la catégorie de représentations unitairement induites de cuspidales unitaires. La conjecture de Ramanujan impliquerait que IC est incluse dans Temp. Réciproquement, la démonstration du Lemme 1.5 montre que Temp est incluse dans IC. I est souvent utile (cf. [2 ch. II) de considérer directement la catégorie IC, qui se décrit simplement sans attendre la solution de la conjecture de Ramanujan. Elle est formée de représentations isobares, et est stable par . La conjecture de Ramanujan a une conséquence intéressante pour les représentations de Isob: LEMME (QUOTIENT DE LANGLANDS GLOBAL). Supposons la Con- jecture 1.6. Alors, sit € Isob(n), x est l'unique quotient irréductible d’une représentation. IndyiQy(o1 @ ++ 9, @1)MOTIFS ET FORMES AUTOMORPHES 89 ot les 0; sont des représentations cuspidales de G,,(A), P est le parabolique de type (n1,...,n+), 0; & 09 | [** of 0° est unitaire et 81 ER, 8) > 92 2-+ Dd sy. Cela résulte immédiatement des formules (1.3) et (1.4) puisque les sont alors tempérés. 1.2.3. La catégorie Aly.. La théorie des formes automorphes sur GL(1) - i.e., des caractéres du groupe des classes d’idéles - montre que les catégories introduites jusqwici sont trop vastes pour étre liées & des objets arithmétiques. On veut maintenant définir la catégorie de représentations de GL(n) qui joue un réle analogue & celui des Gréssencharakterer de type Ao, ou caractéres de Hecke algébriques, de Weil. Comme on va le voir (cf. aussi §3), cela améne A modifier la définition de B. Rappelons que d’aprés Langlands [38 d, 51] il y a une bijection canonique entre représentations admissibles irréductibles de GL(n, F) (F =R ou C), modulo équivalence, et représentations complexes de degré n, & image semi-simple, & équivalence prés, du groupe de Weil Wr. On a une suite exacte 13 We > We ge 1 et We est canoniquement isomorphe & €* [53]. Il est parfois com- mode de considérer €* comme l'ensemble des points réels de § = Rese ~(Gm). On notera r les représentations du groupe de Weil récl ou com- plexe. Insistons sur le fait que la paramétrisation que nous utilisons est celle de Langlands. En particulier, elle envoie les représentations tempérées de GL(n, F) sur les représentations A image bornée de Wr. Soit x = @ym, une représentation irréductible de GL(n, A). Soit v une place infinie de F, ct fixons un isomorphisme de F, avec R ou €. La classification de Langlands associe & v une représentation 7, de Wr,. Par restriction dans le cas réel, on en déduit une représentation ~ encore notée ry- de C*; dans le cas complexe, ry est la représentation de €% définie par le choix d’isomorphisme. Enfin, soit | |g la valeur absolue complexe sur C*. Definition 1.8. Alg(n) est ensemble des représentations isobares de GL(n,A), 7 = @rp, telles que, pour toute place infinie v, ry soit90 LAURENT CLOZEL égale & y1 @ ++ @ Xn oid yj est un caractére de S(R) = CX tel que don xil [eZ soit un caractére algébrique. Autrement dit, yi(2) = 2027" (2)"t*F", avec pi, qi € Z. Remarque. Soit v unc place infinie de F. Soit H le tore maximal (diagonal) de G. Si v est complexe, soit Ho = Reseye (Lie H); si v est réelle, ho = LieH. Soit = 0 @C. Ona h = €" x €* (cas complexe), & €" (cas réel), le groupe de Weyl complexe W opérant comme 6, x 6, (resp. Gy). D’aprés Harish-Chandra, le centre 3 de U(g), ot go = Reseya LieG ou LieG suivant le cas, et g = go @ C, s‘identifie 4 5(h)"”. Le caractire infinitésimal de 7, s’identifie alors 4 1’élément (p,q) € €2" (cas complexe), ou A (Pi. P2rsPor4is+++1Ph) € C” Si ry est de Ja forme (2122, 27 2P1,,..,(2z)Pirtt,...,(z2)P&) avec pi # pj (cas réel). _Laumon 1'a fait remarquer que - modulo le centre - les re- présentations de Alg sont donc associées & des formes automorphes satisfaisant des équations différentielles associées & un caractére de 3 qui est entier pour la structure entiére sur déduite d’une base de Chevalley. Exemple. GL(2) sur un corps récl. Il sera utile de décrire ici comment la définition de Aly se traduit dans les cas classiques. Commengons par le cas de GL(2) sur Q. Soit f une forme parabolique de poids k € Z sur le demi-plan de Poincaré pour le groupe To(n), associée & un caractire w de (Z/NZ)*: 1 (S42) = warns), 7 TaN), ott j(7,2) = cz+d. Le relevement de poids k [26] lui associe une forme parabolique yy sur GL(2,Q)\GL(2,A), dont le caractére central est le caractére d’Artin w de A* associé & YW. On suppose que f est fonction propre des opérateurs de Hecke pour p|N, et est holomorphe si k > 0, antiholomorphe pour k < 0, et fonction propre du laplacien hyperbolique: e & _ re A=, As-y¥ (S+#)MOTIFS ET FORMES AUTOMORPHES 91 pour k = 0; enfin on peut supposer que f est une newform. Plus généralement, on peut tordre le relevement de poids k par un caractére |det g|* du déterminant, remplacant ainsi w par w| |?*. ¢ Cas holomorphe & > 2. Si f est holomorphe, ys engendre alors une représentation 7 = 7 ® 7 de GL(2,Ag) telle que 7. est associée, par la correspondance de Langlands, & un couple de car- actéres (2?2%, 272”) (p # q) de C%. Elle est algébrique si p = po + 4, 4 = + } avec po, qo € Z; son caractére central sur RX est «> 2”, avec m = p+q = po+4o+1 = po—qo+1 (mod 2); le poids de f est lp — qi + 1 = [po — go] +1. Done: Les relévements 1 de f qui sont algébriques sont ceux de caractére central 2 — 2(sgnz) avec m= k (mod 2). « Le cas antiholomorphe (k < —2) se traite de méme. Formes de poids 0 ou 1. Si f est de poids 0 ou £1, la représen- tation m associée est telle que To. = € Ags, ot €; = |x|" (sgn x)* sont des caractires de RX. Elle est algébrique si s; € 3 + Z. Comme elle est unitaire mpodulo le centre, on a ; — s2 € iR U J—1,+1[. On doit donc avoir 5 = 82 € +Z. Sie, =e, f est une forme de poids 0, de valeur propre du laplacien \ = —(s — 4)(s + 4) o& s = 5; — 59. Ona done A = }. Enfin, le caractére central de 7. sur RX, x — |a|"**, est de la forme x” avec m € 1+ 2Z: Si f est une forme de Maass, la représentation 7 associée ne peut étre algébrique que si \= +. Si f est de poids 0, -1, ou 1, les relévements algébriques sont les représentations n de caractére central x — 2"(sgnz)* avec m=1 (mod 2). Supposons maintenant F totalement réel de degrér ct soit f une forme parabolique de poids hy, ...,kr) (ki € Z). On dispose également. du relévement de poids k = (k1,...,k,) de f; il engendre une repré- sentation cuspidale irréductible s de GL(2,Ar) sous les hypothéses précédentes sur f. On peut le tordre par des caractéres de la forme |det g|* (rappelons par ailleurs que les caractéres de Hecke algébriques de AX sont de la forme w(x) |x|", m € Z, w un caractére d’Artin). Cela étant, siz. = ® _ 1; est algébrique, (7; doit étre associée A vy; infinie PEW, 245278) avec pi = Piot4, Gi = dio tbs Pio s Gio € Z. Comme de plus 7, & torsion prés par un caractére, est unitaire , on voit de plus Que Pi, + gio = w est indépendant de i; comme |k;| = [pio — gio| +1 Si Pio # Gio, 0,1,-1 sinon, on a le résultat suivant. Posons €; = kj si |ki| > 2, €; = 1 si [ki] <1. | La représentation 7 associée & f ne peut étre algébrique92 LAURENT CLOZEL que si ¢; = ¢; (mod 2) pour tous i,j. Si cette condition est vérifiée, les relevements algébriques 7 de f sont ceux de car- actére central (2) |x|", w étant un caractére d’Artin et m= ¢; (mod 2). Pour Vinstant, Alg n’est pas stable par la somme. On définit une nouvelle opération 6 a l’aide d’un twist ala Tate. Siw € Isob(n), soit | |* la représentation 7(g) |det gl’, ot | | désigne la norme d'idéle. Definition 1.9. Soit 7 € Isob(n;), 2 € Isob(n). On définit 7 somay | (2xtea=! mins = (m1 | Bm [F) PP) Par construction, Popération EB conserve la catégorie Alg = [] Alg(n) puisque, aux places infinies, { se traduit par la somme de représen- tations du groupe de Weil. Modulo la Conjecture 1.3, on a de méme: Definition 1.10. On pose T 1 nyng— ring = (mi | [2 Bre | |) LO. Elle conserve aussi la catégorie Alg. De plus les propriétés de B et g impliquées par le formalisme tannakien (§1.3) sont conservées: par ex- emple, (1E2ir»)EBis est isomorphe (en fait canoniquement isomorphe, T T T T par induction par étages) A 7 E5(w2EBms), Ei et, BY satisfont (conjec- turalement et a isomorphisme pres) les compatibilités habituelles... T Par contraste avec la somme de Langlands, Popération &8 ne com- mute pas an passage & la contragrédiente. Definition 1.11. Si € Aut (n), son dual de Tate est la représen- tation #| |"7!. rn lest clair que x — # est une involution. On vérifie que (716372) = T ai BBity sim, 72 € Isob(n1), Isob(n2). La décomposition des représentations isobares prend une forme nouvelle dans Alg:MOTIFS ET FORMES AUTOMORPHES 93 LEMME 1.12. Si € Alg(n), ou peut écrire T T n= mB... Bz,, oit t; € Alg®(n;) (ensemble des représentations cuspidales de Alg(n;)); les (7;,n;) sont uniques 4 permutations prés. Ceci résulte directement des définitions. 1.2.4. Les catégories Tan et Art. La catégorie Tan est formée des représentations de Aly qui, par restriction automorphe ( §1.1) & une extension finie B/F, deviennent des sommes de caractéres de Hecke (algébriques) de GL(1, Ag): T T T pret © xi By2B... xn, xi € Alg(1, E). La catégorie Art est formée des représentations de Alg telles qu’on ait, aux places infinies, avec les notations 1.2.3, ry = | |e? @--- © T T Ces catégories sont stables par et, conjecturalement, par %. 1.2.5. Sous-catégories self-duales. Dans chacune des catégories considérées, on peut considérer ’en- semble des représentations stables par la dualité (w+ # pour [sob ou Temp, ++ i pour les sous-catégories de Alg). On obtient alors des T sous-catégories (stables par FB ou HH et, conjecturalement, par &) de représentations. On verra (§5) que pour celles-ci la correspondance entre motifs et représentations automorphes parait plus accessible. 1.3. Formalisme, et problémes, tannakiens. 1.3.1.- On suppose le lecteur quelque peu familier avec le formalisme des catégories tannakiennes di a Grothendieck et Saavedra Rivano (cf. (45, 21]). Rappelons simplement qu’une catégorie tannakienne neutre est une catégorie abélienne k -linéaire (pour un corps k ) C munie dun foncteur @ : C > C, (X,Y) 4 X @Y, de certaines données (contraintes d’associativité et de commutativité) exprimant les pro- priétés habituelles du produit tensoriel, et d'un foncteur k-linéaire94 LAURENT CLOZEL (exact, fiddle, compatible avec @ ) C + Vece, oft Vecg est la catégoric tensorielle des espaces vectoriels sur k. Dans [38 f], Langlands exprime Pespoir qu’on puisse munir l’en- semble Isob des représentations isobares d’une structure de catégoric tannakienne. Au nivean des objets, & isomorphisme prés, la somme directe devrait étre donnée par lopération (x,7’) + 7 Hi’ et le produit tensoriel par (n,7") + Bx! (rappelons que Vexistence de -elui-ci est conjecturale!). Le corps & devrait étre ici le corps des complexes. Comme le remarque Langlands “the attempt {to define Tannakian categories of automorphic representations] may be vain but the prize to be won is so great that one cannot refuse to hazard it”. En effet, méme en supposant démontrée la Conjecture principale (§1.1), les problémes a résoudre semblent formidables. Le premier probléme a résoudre consiste & munir la catégorie Isob - ou les autres catégories du §1.2 - d’une structure de catégorie abélienne C-linéaire. Dans le §1.4, on propose une définition de Homysos(*,7') pour 7 yr’ dans un sous-ensemble convenable de Isob. On verra que, méme pour définir Homy,o5(7,7"), on ne peut pas travailler avec les représentations isomorphes & des sous-représen- tations (isobares) de A(G(F)\G(A)). On est contraint A se limiter & un sous-ensemble, contenant & G(A)-isomorphisme pres toutes les re- présentations de Isob (on vérifie alors qu’on obtient bien une catégorie abélienne). Le probléme consistant 4 définir un sous-ensemble convenable de Isob (ou des autres catégories de 1.2) devient encore plus crucial si Von considére le produit tensoriel. La définition des catégories tan- nakiennes impose qu’on puisse exhiber, pour , 7’ € Isob, leur pro- duit tensoriel xB’, un objet bien défini de Isob. Méme en supposant la Conjecture 1.3, il parait peu plausible que Yon puisse définir na- turellement un objet Bz! si x, 7’ sont seulement isomorphes & des représentations apparaissant dans A(G(F)\G(A)). En revanche, on peut espérer qwun théoréme de multiplicité 1 con- venable (cf. Th. 1.1) implique par exemple, l’existence d’une unique représentation (isomorphe 4) 7 & 7’ et apparaissant comme sous- module de A(G(F)\G(A)). On voit que cela améne & chercher des modeles des représentations automorphes, i-c., des représentants bien déterminées des classes d’isomorphisme de représentations auto- morphes (isobares).MOTIFS ET FORMES AUTOMORPHES 95, 1.3.2. Modéles de Whittaker. Soit ~ un caractére non trivial de F\Ar, @ le caractére de 1a * No(A) = : 1 défini par O(n) = Y(a +--+ 2p-1). Sit = Ory est générique, 7 a un unique représentant W(7) dans l’espace W(0) des fonctions sur G(A) se transformant a gauche suivant 6 [40, 49]. Ceci s’applique en particulier aux représentations de Temp (ou IC). Notons l’analogue local: on suppose F local non-archimédien; y est alors un caractére additif non-trivial de F, 6 le caractére associé de No(F). Par définition, une repré entation générique m de G(F) aun représentant -alors unique- dans espace Wo des fonctions sur G(F) telles que f(ng) = 6(n) f(g) (n € NF) (28, 33 a] (cf. §1.1). 1.3.3. Modéles automorphes. Soit € Aut?(n). D’aprés le Théor’me 1.1, il existe un unique sous- module de Acusp(G(F)\G(A)) isomorphe & x. On se demander s'il existe un résultat analogue pour les représentations de Isob ou Temp. Le rédacteur ne le sait pas, mais on peut remarquer les faits suivants: (1) La théorie des séries d’Eisenstein [38 a] montre que les repré- sentations de IC (de Temp, modulo la conjecture de Ramanujan) in- terviennent avec multiplicité 1 dans la décomposition hilbertienne de L(G(F)\G(A)). (2) Pour les représentations (isobares) de Speh introduites dans le §1.1, on a un théoréme, di a Moeglin et Waldspurger. Fixons un tére central x : F*\A* — C*. Soit L3.,(G(F)\G(A), x) la par- te de Pespace des fonctions sur G(F)\G(A), se transformant suivant y sous Z(A) & AX, et de carré intégrable modulo le centre. TuforbME 1.13. (cf. [39]). Sit & J(w,a) est un module de Speh (notations 1.1) de caractére central x, m intervient dans L?(G(F)\G(A), x) avec multiplicité 1. 1.3.4, Modéles principaux. Soit 7 une représentation isobare de GL(n, A). On peut Vécrire sous la forme 0; B---Ec,, ot les 0; sont des représentations cuspidales de GL(n1,A) (§1.1). Puisque toute représentation gj a un unique modéle96 LAURENT CLOZEL dans Acuep(G(F)\G(A)), on obtient - pour tout choix de V’ordre des g; - un unique modéle de 7, l’unique sous quotient de Indgiq}(a1 @ + @ 4 @ 1) de type m. Iy a antant de “modéles” (sic) que de permutations possibles des a;. Ces modéles seront appelés modéles principaux. 1.3.5. ‘Terminons ce paragraphe en rappelant que les représentations admissibles irréductibles de GL(n, F), ot F est local devraient étre paramétrées par les représentations complexes de degré n du groupe do Weil-Deligne W Dr (53, 38 f]. En particulier, les opérations et BH devraient étre définics (au moins A isomorphisme prés) pour les représentations locales. Dans le cas archimédien, oi la classification est connue, cela suggere le probléme suivant: PROBLEME 1.14 (F = R ov C). Définir une sous-catégorie convenable de la catégorie des repr sentations admissibles irréductibles de GL(n,F) et la munir d’une structure de catégorie tannakienne (pour une description d'une petite sous-catégorie dont on peut faire au moins une catégoric abélienne, cf. 1.4.2). 1.4, Catégories abéliennes de représentations isobares. 14.0. Dans ce paragraphe, on construit une (petite) catégorie abélienne de représentations isobares, oi l'on décrit les espaces Hom qui devraient étre ceux de la catégorie tannakienne. Elle contient un objet isomorphe & toute représentation isobare; elle se décrit simple- mont a Vaide des modéles principaux. Il y a une construction locale analogue. Cotte catégorie a le défaut que application (ou niveau des ob- jets) ow la restriction global - local ne peuvent y étre défini le dernier paragraphe, on indique briévement la construction d'une autre catégorie, équivalente, oft ces applications sont définies. Tl est clair que ces constructions sont insuffisantes au regard des espoirs du §1.3: en particulier, on n'a pas montré ici comment rendre Yopération 6, ou méme la restriction global-local, fonctorielles. 1.4.1. On appelle Isob = Isob\(F) la catégorie dont les objets sont les modéles principanx (1.3.4). Un objet 7 € Isob, s‘identifie 4 In donnée dune partition n =n; +--+ +ny et de r repré sentations cus- pidales 0; de Gp,(A). On appellera ici représentation cuspidaleMOTIFS ET FORMES AUTOMORPHES 97 une sous-représentation de Acusp(Gn,(F)\Gn,(A)): elle est unique- ment définie par sa classe d’isomorphie. On notera Homg(z, x’) l'espace des homomorphismes de a vers 1! dans Isob; (soit §4 pour la notation). Definition 1.15. (1) Sic, r sont cuspidales, on pose Home(o,7) =0 si ofr Homg(o,0) = Homaiay(o,0) = € (2) On pose Homg(o, B+ Bo,,7 B+ B75) = 15 Homg(oi, 7). Dans la partie (1), noter que Homgiay(o, 0) est canoniquement iso- morphe & C! Composition. Les morphismes se comme des matrices. composent de la maniére évidente, On définit Vobjet nul comme lunique représentation cuspidale de GL(0,Ar). Soit 7 = 0, B---Bo,, 7’ = 7 B--- Br, deux éléments de Isoby, avec oj, 7; cuspidales. Comme 7 7’ est, par définition, égal a o, &---Bo, 87, B--- B75, on vérifie aisément que c’est a la fois une somme directe et un produit direct de x et 7’. La notation H est donc justifiée. On vérifie aisément l’existence de noyaux et conoyaux (il est. com- mode de remarquer que tout 7 € Isob, s’écrit de fagon unique comme somme directe dans Isob; de représentations isotypiques, i.e. de la forme o 1... fo; on est ramené a l’existence de noyaux et conoyaux dans la catégorie des espaces vectoriels munis d’une base). On voit de méme que la catégorie ainsi définie est abélienne. On remarquera que, méme en supposant la Conjecture 1.3, le pro- duit tensoriel n’est. pas défini dans Isob,. Dans une catégorie munie de produit tensoriels, on définit le dual d’un objet & l'aide du Hom in- terne Hom(X,Y) ((21 , p. 110}). Dans le cas de Isoby, on peut définir le dual de = 0, --- Bo, comme l'objet 6; H---H4,; notons-le 7. On vérifie aisément qu’on obtient ainsi une involution contravariante de Isob,. 1.4.2. Si F est un corps local , l’analogue local de Isob, se définit ainsi. Tout d’abord, il n’y a pas a imposer des restrictions aux re- présentations locales: on considére toutes les représentations irréduc- tibles admissibles. Une telle représentation de G,(F) est isomorphe98 LAURENT CLOZEL & unique sous-quotient de Langlands d’une induite Ind(P(o1 @-- Gr @1) (formule 1.2) cit les oj sont de carré intégrable. On note ce sous- quotient 0 B-+- Boy. Pour avoir analogue des modéles principaux (1.3.4), il faut fixer des modiles des représentations de carré intégrable. On le fait & aide du modéle de Whittaker local (cf. 1.3.2, [28]). On a alors un modéle unique pour chaque représentation de carré intégrable; on définit alors la catégorie abslienne [sob (F), comme en 1.4.1, & l'aide des modéles principaux (remarquons qu’on aurait aussi pu utiliser les modéles de Whittaker 1.3.2 au lieu des modéles automorphes 1.3.3 des représen- tations cuspidales pour définir Isob,(F) pour F global). Le lecteur se convaincra qu’il n'y a pas d’application naturelle Isob,(F) — Isoh,(Fy) si v est une place du corps global F. 1.4.3. Indiquons briévement la construction d’une nouvelle catégorie Tsoby, équivalente & Isob,, et ot le produit tensoriel est bien défini modulo la Conjecture principale. Ses objets sont les classes d’équiva- lence de représentations isobares: on peut les identifier aux données o; B+. Ba; de Isob;, prises modulo permutation. Chaque objet de Isoby peut étre identifié & une somme formelle oo Sonea (somme finie), o& o décrit [] Aut °(n). On définit net Homg(@n90,@moa) = Go Hom(C"*,C™”). On vérifie aisément qu’on a défini une catégorie abélienne, avec la composition évidente des homomorphismes. Tl y a une application évidente Isob, 2 Isobp; utilisant le fait, remarqué en 14.1, que Homg(m, 7’) pour x, 7" € Isob, se décompose suivant les composantes isotypiques, on voit que c’est un foncteur, et que Homg(@z, &r’) = Home(x,7") pour 7,1! € Isoby, Les deux catégories sont donc équivalentes (ef. e.g. [16, Prop. 1.19]). ‘On construit de maniére analogue la catégorie Isobe(F) pour F local, & partir de 1.4.2. On remarque alors que: (1) Modulo la Conjecture principale, i est bien défini comme ap- plication (Jsobs)® + [s0by (F global). (5) Si v est une place d'un corps global FF, il y a une application de restriction Isobo(F) — Isobo(F.).-MOTIFS ET FORMES AUTOMORPHES 99 1.4.4. Les constructions des paragraphes 1.4.1 et 1.4.3 peuvent étre transposées dans la catégorie Alg, en remplacant bien sir la somme directe par Vopération &. En particulier, on munit ainsi des sous- catégories convenables Alg, ou Alga de Alg d’une structure de caté- gorie abélienne. Par restriction, on a donc des structures de catégories abéliennes sur toutes les catégories de 1.2. 2. CERTAINS CAS CONNUS DU PRINCIPE DE FONCTORIALITE, Dans ce paragraphe on passe en revue certains cas ott le principe de fonctorialité a été établi. On s’intéresse surtout aux groupes liés A GL(n) et aux exemples décrits en 1.1. Rappelons tout d’abord l’énoncé du principe de fonctorialité. Soit H un groupe réductif connexe sur F , “H = “H® x Wr la forme de Weil de son L-groupe. Sis = 7, est une représentation automorphe de H(A), non ramifiée en-dehors d’un ensemble fini de places $ (contenant les places infinies), on lui associe, pour v ¢ S , sa matrice de Hecke : c’est une classe de conjugaison (sous “H® ) dans “H , s’envoyant par la projection "H — Wr vers un élément de Frobenius pour Fy . Soit alors Hi, Hz deux groupes, y : “H; — Hz un homomor- phisme de L-groupes [10 a]. Si 7; est une représentation automorphe de H,(Ap), le principe de fonctorialité affirme V’existence d’une re- présentation 72 , automorphe, de H(A) telle que, en presque toute place, tz,,» = y(tz,,v) En Pabsence de “théorémes de multiplicité 1” (cf. Thm. 1.1), elle n’est pas n irement unique. Rappelons quelques illustrations du principe qui ont été démontrées pour GL(n) et certains groupes classiques. 2.1. Induction et restriction automorphes. Les homomorphismes de L-groupes sont décrits en 1.1. Dans le cas une extension cyclique , le théor’me suivant a été démontré par Arthur et lauteur: THEOREME 2.1. Soit F,/F une extension cyclique de degré d de corps de nombres. Alors: (i) Si mp est une représentation induite de cuspidale (catégorie ICde 2) de GL(n, Ar,), la représentation vf, tp, de GL(nd,Ap) existe, et est uniquement déterminée; elle appartient 4 IC.100 LAURENT CLOZEL (ii) Si mp est une représentation de GL(n,Ar) appartenant & IC, la représentation ppyr,tF de GL(n,Ar,) existe; elle est uniquement déterminée et appartient a IC. Dans le cas ot n = 1, la partie (i) est due & Kazhdan et Flicker [34, 25 al: cest la fonctorialité “de Hecke” envoyant Gréssencharakterer sur F; vers représentations de GL(d, Ar). Pour une description plus détaillée des résultats, et de image de pr/p; Alans le cas cyclique, on renvoie le lecteur & (18 d, « 2.2. Carré symétrique. Tl s'agit de la fonctorialité entre GL(2) et GL(3) déduite de Vhomomorphisrae de L-groupes GL(2,€) + GL(3,€) donné par le carré symétrique. Elle a été démontrée par Gelbart et, Jacquet (27). Flicker, a la suite de travaux antéricurs de Jacquet et Langlands, a an- noncé une démonstration différente, A l'aide de la formule des trace: qui permettrait d’en caractériser Pimage. 2.3. Fonctorialité associée au carré extérieur. Soit GSp(4) le groupe des similitudes symplectiques sur un es- pace de dimension 4. C'est un groupe déployé, dont le L-groupe est GSp(4,€) x Wr . La fonctorialité *évidente” associée au morphisme GSp(4,€) — GL(4,€) devrait avoir pour image les représentations 7 de GL(4,C) telle que la fonction L, L(s,7, Ao), associée an carré extéricur, a un pale. Ceci a été démontré par Jacquet et Shalika (voir Pexposé de Jacquet dans ce volume, et [6 a] pour un énoncé), 2.4. Groupes unitaires. Soit F,/F une extension quadratique de corps de nombr G un groupe unitaire sur F', associé A une forme F,/F -hermitienne sur un espace de dimension n sur F,. Son L -groupe est GL(n,€) x Wr, Wp opérant via Gal(F;/F), dont Pélément non trivial 7 agit sur GL(n,€) par Vautomorphisme extérieur d'ordre 2. On a G Xr F, & GL(n)/Fi; le changement de base devrait relier représentations automorphes de GL(n,Ap,) fixées par 'automorphisme g 7 (497!) et représentations automorphes de GL(Ar)- En général, le probleme est compliqué par les phénoménes de L ndiscernabilité: il est résolu, pour n = 3, par Rogawski [43 b]. Rapoport, Zink et Kottwita ont découvert que la L -indiscernabilité était mise en échec pour certains groupes unitaires venant d’algebres a division sur F, munies d’une involution de seconde espéce (ce cas a aussi étudié, sous certaines hypothéses locales et en utilisant un